YT YouTube2Text

YouTube2Text → Transcripts → The AI bubble is about to burst

The AI bubble is about to burst — Transcript

by The Infographics Show · 37,245 words · 5,853 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00AI was supposed to kick off a white
0:02collar purge, but the reality is
0:04different. In Fortune 500 boardrooms
0:06across the country, thousands of AI
0:08agents are being shut down. [music] Not
0:09because they didn't work, but because
0:11they're a legal time bomb. The narrative
0:14is that tech giants are spending [music]
0:16everything to win an AI arms race. But
0:18the reality is much messier. They're
0:21stripping the liability out of their
0:22systems [music] while pouring billions
0:24into the buildout. And when the dust
0:26settles, someone still has to pay for
0:28it. The US interstate [music] highway
0:30system is 48,000 mi long. It took 35
0:33years to build and cost $630 [music]
0:36billion in today's money. It is by any
0:38measure one of the most consequential
0:40infrastructure investments in human
0:41history. Meanwhile, big tech is spending
0:43[music] $725 billion in 2026 on AI
0:47infrastructure alone. That number has
0:49more than doubled in just 2 years and is
0:52[music] projected to keep climbing into
0:53the trillions by 2030. In fact, Goldman
0:56Sachs now estimates that the total AI
0:58infrastructure bill [music] between now
0:59and 2031 will exceed $7.61 trillion. For
1:04that kind of money, you'd expect a
1:05revolution. What these companies are
1:07getting right now is cents on the
1:08dollar. The entire global AI services
1:10[music] market in 2025 generated around
1:13$25 billion against hundreds of billions
1:15of infrastructure spending. When a
1:17Chinese startup called Deepseek revealed
1:19in January 2025 that it had [music]
1:21built a competitive AI model using
1:23Nvidia's lower-end chips for under $6
1:25million, it seemed almost impossible.
1:28But it still shook the market, [music]
1:29wiping out around $590 billion from
1:32Nvidia's market cap in a single day.
1:34[music] It was the largest single day
1:36loss in US stock market history.
1:38Deepseek, far from breaking AI, merely
1:40[music] raised the question every
1:42investor had asked himself. What exactly
1:44are we paying for? The companies
1:46spending this money don't just have it
1:47lying around. They are borrowing it.
1:49While Alphabet quadrupled its long-term
1:51debt in 2025 to 46.5 billion, the five
1:55biggest tech companies issued $18
1:57billion in new bonds last year alone.
2:00Bank of America now estimates that AI
2:02Capex now consumes 95% of operating cash
2:04flows after dividends. [music] Amazon is
2:07projecting a 95% decline in free cash
2:09flow thanks to its AI buildout costs.
2:12They are in essence mortgaging their
2:14present against a future that hasn't
2:16shown up yet. The gamble is genuinely
2:18alarming. If it were startups chasing
2:20hype, that would be one thing. These are
2:22the most sophisticated capital
2:24allocators on the earth. When they spend
2:26like this, they either know something
2:27the rest of us don't, or they're caught
2:29up in something they can't stop without
2:31admitting they were wrong. Either way,
2:33something has broken. To understand
2:35what, you have [music] to go back to the
2:36moment a grieving man in Vancouver asked
2:38an airline chatbot a simple question and
2:41changed [music] corporate legal history
2:43forever. On the day his grandmother
2:45died, Jake Moffett opened his laptop and
2:47went to Air Canada's website to book a
2:49flight from Vancouver to Toronto. He had
2:52one question. Did the airline offer
2:54bereavement fairs and could he apply for
2:56one after the fact? [music] He asked the
2:58chatbot and it said yes. It told him to
3:00submit his request within 90 days of
3:02purchase, and the discount would be
3:03retroactively applied. So, Moffett
3:05booked the ticket, flew to Toronto, came
3:08home, and then filed the request like
3:09he'd [music] been told, and Air Canada
3:11denied it. Their actual policy, as it
3:14turns out, required bereavement fair
3:15requests to be made before travel. The
3:18AI chatbot had simply made up a policy
3:20that didn't exist, [music] stated it
3:22with complete confidence, and sent a
3:24grieving man to buy a full price ticket
3:25[music] based on a lie. So Moffett sued.
3:28And what happened next is the reason
3:30every corporate lawyer in America now
3:31has an opinion about chatbots. Air
3:34Canada's legal defense was [music]
3:35interesting. They argued it could not be
3:37held liable for information provided by
3:39its chatbot [music]
3:40because the chatbot was quote a separate
3:43legal entity responsible for its own
3:44actions. They had deployed a piece of
3:46software to represent their company. And
3:48when that software caused harm, their
3:50position was essentially don't look at
3:52us. Tribunal member Christopher Rivers
3:55was not impressed. While a chatbot has
3:57an interactive component, he wrote, "It
3:59[music] is still just a part of Air
4:00Canada's website, it should be obvious
4:02to Air Canada that it is responsible for
4:04all the information on its website. It
4:06makes no difference whether the
4:08information comes from a static page or
4:10a chatbot." Air Canada was ordered to
4:12pay $812 Canadian, and in perspective,
4:15that amount is trivial. The precedent
4:17was not. [music]
4:18That ruling landed in February 2024.
4:21What followed was a corporate panic that
4:23never made headlines, [music] but could
4:25be seen in the risk sections of SEC
4:27filings. In 2020, 4% of public companies
4:30mentioned AI as a material risk in their
4:33annual filings. By 2024, that number was
4:3543%. Among the Fortune 500 specifically,
4:3956% now lists AI as a formal risk
4:41factor. Among media and entertainment
4:43companies, it's 92%. The SEC itself has
4:46been sending enforcement letters
4:48demanding more specific disclosures
4:50backed by $ 8.2 billion in financial
4:52remedies in [music] the year 2024 alone.
4:55Climate change took decades to achieve
4:57similar disclosure penetration across
4:59corporate America. AI did it in 3 years.
5:02What the Air Canada ruling established
5:04is that whatever your AI says to a
5:06customer, you [music] said it. You can't
5:08outsource the liability to the software
5:10you employ. You can't just point at the
5:12machine and then look innocent. You are
5:14the machine. For a technology that can't
5:17guarantee it won't make things up, that
5:18ruling created an absolute certainty.
5:21Deploy AI at scale and eventually
5:23liability follows. The bigger question
5:25is, will there be anyone left to cover
5:27the bill? In July 2024, research and
5:30advisory firm Gartner predicted that at
5:32least 30% of enterprise gen AI [music]
5:34projects would be abandoned after proof
5:36of concept by the end of 2025. By April
5:392026, they had to revise their original
5:41estimate. [music] It turns out that it
5:43was actually more like 40% of enterprise
5:45AI projects launched [music] in the
5:47previous 2 years that would be scrapped
5:49before they ever shipped. Separately,
5:50Gartner now predicts that over 40% of
5:53Agentic AI projects, the next generation
5:55of AI tools companies are currently
5:57betting on, will be cancelled by the end
5:59of 2027. The graveyard keeps growing.
6:02S&P Global Market Intelligence put some
6:04finer numbers on it. The average
6:06organization scrapped [music] 46% of AI
6:08proof of concepts before reaching
6:09production. Of the projects initiated,
6:12only 48% make it to production with an
6:14average of eight months between
6:16prototype and launch. [music] That is
6:17eight months of engineering hours,
6:19infrastructure costs, and executive
6:21attention on a coin flip chance of
6:23shipping anything at the end of it.
6:24McKenzie, another [music] consulting
6:26firm, surveyed organizations globally in
6:28late 2025, and they found that 88% use
6:31AI in at least one business function. If
6:34that sounds like [music] a success
6:35story, it's not. Only 39% reported any
6:39measurable impact on earnings. At most,
6:41[music] 61% reported no meaningful
6:43financial impact on their business
6:45whatsoever. Roughly six [music] in 10
6:47companies have plugged AI into their
6:49operations and can't point to a
6:51measurable return it's made them. The
6:52reasons are familiar enough that Gardner
6:54has named them outright. Poor data
6:56quality, weak risk controls, rising
6:58costs, and unclear business value. Each
7:01project costs between $5 and $20 million
7:03to build and deploy. Multiply that
7:06across thousands of Fortune 500
7:07initiatives and you're looking at tens
7:09of billions of dollars buried in the
7:11corporate graveyard. It gets [music]
7:13even worse for big tech. A 2024 Deote
7:15survey found that 47% of enterprise AI
7:18users [music] admitted to making at
7:19least one major business decision based
7:22on the content AI had simply invented.
7:24In [music] other words, nearly half had
7:26already trusted a hallucination with
7:28realworld consequences. Granted, things
7:30[music] have evolved a lot since 2024.
7:32The efficiency revolution is still an
7:34aspirational possibility as Agentic AI
7:37improves. But you wouldn't be human if
7:40you hadn't wondered why the flaws you've
7:41seen firsthand seem absent from the
7:44market's calculations. There's a
7:46particular breed of corporate fear that
7:47shows up where you would least expect
7:49it. Insurance contracts. [music]
7:51Basically, if the people whose entire
7:53business model depends on their ability
7:55to correctly price risk [music] deem
7:57something uninsurable, that is usually a
8:00huge red flag. In January 2026, the
8:03Insurance Services Office introduced new
8:05exclusions targeting generative AI.
8:07Essentially, if generative AI causes the
8:10damage, don't assume you're covered. New
8:12policy language began excluding claims
8:14tied to AI generated text, images,
8:17audio, video, and code. That is every
8:20large language model ever built. the
8:22chat bots, the AI co-pilots, the
8:24autonomous agents, all of them circled
8:26with black ink and excluded from the
8:28policies that cover 82% of US
8:30businesses. [music] In 2024 and 25,
8:33insurers had been narrowing AI coverage
8:35in existing policies by tightening up
8:37language and adding exclusions. WR
8:40Berkeley introduced an absolute AI
8:42exclusion across directors and officers,
8:44errors and omissions, and fiduciary
8:46liability lines. AIG saw that and they
8:49said, "Well, we'll have more of the
8:50same." [music] Great American Insurance
8:52followed. Chub, Travelers, Bergkshire
8:54Hathaway received state regulatory
8:56approval for their own exclusions. State
8:59regulators have been approving more than
9:0080% of those requests. [music] The
9:02industry has seen this movie before. A
9:05decade ago, cyber insurance went through
9:07an identical cycle. There [music] was
9:08invisible coverage until claims mounted
9:11and then came the exclusions.
9:12Eventually, [music]
9:13standalone products emerged at premiums
9:15most companies couldn't stomach. The AI
9:17liability market is speedr running the
9:19same cycle. [music]
9:20Insurance companies are not moral
9:22actors. They do not exclude things
9:24because they find them distasteful. They
9:26exclude them [music] because the numbers
9:28don't work. They tried to price in AI
9:30liability throughout 2024 and 2025. And
9:3342% of insurers reported tracking no AI
9:36risk metrics at all. When you can't
9:38measure risk, you can't price it.
9:40[music] And when you can't price it, you
9:43exclude it. The legal system said,
9:45"Whatever your AI does, you did it." the
9:47insurance industry said, "And we won't
9:49be covering that." Which left every
9:51company still running public-f facing AI
9:53deployment, holding the entire liability
9:55themselves, all on hardware that was
9:58simultaneously losing most of its value.
9:59[music] In the summer of 2023, if you
10:02wanted to rent a single Nvidia H100 GPU,
10:04a foundational chip in the inception of
10:06the AI arms race, you were paying
10:08between $7 and $10 [music] an hour. To
10:10buy one outright on the secondary
10:12market, you were looking at $40,000 to
10:14$50,000 per card. Startups were putting
10:17H100s on their balance sheets like
10:19assets, and venture capitalists were
10:20[music] asking portfolio companies how
10:22many they had on their books. Nvidia's
10:24market cap soon passed the $1 trillion
10:27threshold, and it kept rising. In
10:29February [music] 2026, it hit 5
10:31trillion. The H100 was, for a brief,
10:33dizzying moment the most strategically
10:35valuable piece of hardware on the Earth.
10:37Today, you can rent one for $2 an hour.
10:40In [music] June 2025, AWS cut H100
10:43pricing by 45% overnight. [music]
10:45The secondary market collapsed right
10:47behind it. GPUs that had sold for as
10:49much as 50 grand during the AI boom were
10:52suddenly worth less than half that.
10:53Entire H100 server systems that once
10:56cost more than $350,000
10:58were now trading for a fraction of their
11:00peak value. Over 300 new GPU cloud
11:03providers entered the [music] market in
11:042025 alone. All of them selling capacity
11:07that nobody was buying at the rate
11:09anyone projected. Thousands of companies
11:12[music] weighing up where to sell the
11:13cards at a loss. It is the difference
11:15between making payroll and not making
11:17payroll. This [music] signals a deeper
11:19problem about the infrastructure thesis
11:21that justified the entire buildout. The
11:23argument was always that compute was
11:25scarce. That scarcity created value and
11:27whoever controlled the most compute
11:29would win [music] out in the AI race.
11:31Deepseek fractured that argument by
11:32building a model competitive with GPT4
11:35on a budget that Meta spends on
11:37catering. The telecom industry learned
11:39the lesson the hard way in the late
11:411990s when companies invested more than
11:43$500 billion in fiber optic cables based
11:46on projections of a 1,000% annual
11:48internet traffic growth. When
11:50monetization stalled, the debt markets
11:52closed. Miles of fiber sat dark for
11:55years owned by companies that no longer
11:57existed. The asset had been real, but
11:59the business model built on top of it
12:01had not. No amount of Google fiber
12:03buildouts by the 2010s could fix that.
12:06The H100 fire sale is the fiber glut of
12:092025. The hardware is real. The revenue
12:11was supposed to follow. It hasn't.
12:14[music] Today, the trillion dollar
12:15companies watching their hardware
12:16depreciate have made a calculated
12:18decision on what to do with their cash
12:20instead. [music] There's a move
12:21corporations make when they believe
12:23their own stock is a better investment
12:25than their [music] future. They take
12:27their cash and they use it to buy back
12:29their own shares. This [music] reduces
12:31the number of shares in circulation,
12:32boosting earnings per share, which then
12:35lifts [music] the stock price. Investors
12:37love it. Boards love it. Everyone wins.
12:40Yet, it is a signal that the smartest
12:42people in [music] the room looked at
12:43every opportunity available and they
12:45chose not to bet on the future. Instead,
12:48they bet on the thing already making
12:50money today. Over the last 5 years, the
12:52companies loudest about the AI
12:53revolution have been doing exactly this
12:55[music] at a scale that is difficult to
12:57fully absorb. Alphabet has spent about
12:59$280 billion buying back its own stock.
13:02Meta and Microsoft have spent tens
13:04[music] of billions. Apple, which barely
13:06talked about AI until recently, has
13:08spent $74 billion buying back its own
13:11stock over the past decade. That's more
13:13[music] than the entire market value of
13:16488 companies in the S&P 500. And then
13:19the data center bill started piling up.
13:22All of the money disappeared. In
13:23September 2024, Microsoft promised
13:26shareholders a $60 billion buyback
13:28program. It signaled that the company
13:30was flush, confident, [music] and
13:32committed to returning cash to the
13:33people who owned it. A year later, $57.3
13:37billion of that promise still sits
13:39completely untouched. Microsoft didn't
13:41change its mind, but every spare dollar
13:43was getting redirected into paying for
13:45servers, data centers, and AI
13:46infrastructure instead. In other words,
13:49they wrote the check. They [music] just
13:50couldn't afford to cash it. This is what
13:52the AI arms race really looks like from
13:54the inside. The story being sold is one
13:57of a confident sprint into the future.
13:59The reality is a scramble to fund it.
14:01Companies are being squeezed between
14:03investor expectations and the soaring
14:05cost of AI infrastructure. For
14:07Microsoft, that's led to a $60 billion
14:09commitment sitting on the books like an
14:11IOU that nobody can collect. Alphabet's
14:14free cash flow is projected to plummet
14:16almost 90% in 2026 from 73.7 billion
14:20down to $8.2 billion. And in 2025,
14:24Alphabet, Amazon, Oracle, Meta, and
14:26Microsoft issued $121 billion in new
14:30debt via bonds to keep the lights on.
14:32That is four times what the entire tech
14:34industry borrowed on average in any year
14:37of the previous decade. The AI
14:39revolution is being funded on [music]
14:41credit, not profits. And the org charts
14:43that were supposed to prove this was
14:45working have started to change. At some
14:47point in 2023, someone in a corporate
14:49communications department typed [music]
14:51the words generative AI into a press
14:53release and felt genuinely excited about
14:56their job. The phrase had some weight.
14:58It had mystery. It implied that the
15:01machine was creating something,
15:02thinking, almost generating. The word
15:05did a lot of heavy lifting for a lot of
15:07companies that had no idea what they
15:09were actually going to do with the
15:10technology. By 2026, the impetus for
15:13generative AI [music] has turned into
15:15agentic workflows. People are talking
15:17about autonomous process orchestration
15:19and intelligent automation. It's the
15:21[music] exact same technology, but the
15:23promises for what it can deliver are
15:24dramatically smaller. Today, the gap
15:27between the language of 2023 and the
15:29language of 2026 tells you almost
15:31everything you need to know about what
15:32happened in between. The conversation
15:34has shifted from possibility to proof,
15:36[music] from what could this do to what
15:38did it actually deliver. The companies
15:40that survived the graveyard of AI
15:42project rollouts did so by making their
15:44AI projects smaller, narrower, and
15:46harder to sue over. [music] Not by
15:48funneling all their cash into chat bots
15:50that could answer any question, or AI
15:52that writes company strategy. Well, what
15:54this means is a lot of routing software
15:56that sends customer [music] service
15:58tickets to the right department and AI
16:00that scans contracts for non-standard
16:02clauses. Things that are dull, specific,
16:04and measurable. Things that are wrong in
16:06ways that humans can still catch [music]
16:08before they get expensive. And the job
16:10titles reflect that. In 2025, 26% of
16:13organizations had a chief AI officer or
16:16CIO. By 2026, that number jumped to 76%,
16:20tripling in a single year. According to
16:22an IBM study of 2,000 CEOs [music]
16:24across 33 countries, but read the fine
16:26print on what those CIOS are actually
16:29being hired to do, and it is obvious the
16:31role has fundamentally changed. It used
16:33to be AI evangelists promoting the
16:36technology.
16:36>> [music]
16:36>> Now, its executives responsible for
16:38things like risk management, regulatory
16:40compliance, and [music] governance. The
16:42EU AI act kicks in fully in August 2026,
16:45requiring companies to identify exactly
16:47who is accountable when their AI causes
16:50harm. The CIO is increasingly the person
16:53whose name goes on that form. Probably
16:55[music] not the job description anyone
16:57imagined when they coined the title, but
17:00here we are. Meanwhile, the outcome of
17:01the earnings calls has quietly shifted,
17:03too. AI powered is out. outcome focused
17:06workflows is in which is [music] if you
17:09read it slowly enough a very expensive
17:11way of saying that the AI does a [music]
17:13specific task and you can measure
17:15whether it worked or not. The dream that
17:17produced the $725 billion buildout
17:20[music] was always something closer to
17:21science fiction. AI that could run an
17:23entire department, write strategy,
17:25replace an [music] entire team or what
17:27have you. C AIOS with compliance
17:29checklists and the EU AI act obligations
17:32are now doing work that would have
17:33[music] been unrecognizable to the
17:35people who wrote the original press
17:36releases. The revolution got a job in
17:39middle management. And the reason it
17:41ended up there is something the industry
17:42has been trying to explain away since
17:44the beginning. Back in 2023, when
17:46ChatGpt was still a twinkle in many
17:48people's eyes, attorneys Steven Schwarz
17:50and Peter Luduka filed a legal brief in
17:53the Southern District of New York. The
17:55brief cited very real sounding cases
17:57with very real sounding names and very
17:59real sounding outcomes as president.
18:01Vargasi v. China Southern Airlines,
18:04Martinez v. Delta Airlines and Shaboon
18:06v. Egypt Air and then Judge P. Kevin
18:09Castell looked them up and none of them
18:12existed. Chad GPT had invented every
18:15single one with absolute confidence.
18:18Schwarz and Luca were sanctioned,
18:20publicly humiliated, and their names are
18:21now permanently attached to one of the
18:23most cautionary legal filings in
18:25American judicial history, Mata v.
18:27Aviana. That was in 2023. Things [music]
18:30didn't improve. In August 2025, US
18:33District Judge Allison Bacas sanctioned
18:35a lawyer whose brief [music] contained
18:3712 fabricated or unsupported citations
18:40out of 19. A California court ordered
18:42two law firms to pay just over $31,000
18:45in fees for filing what the judge
18:47described as bogus [music]
18:48AI generated research. In February 2025,
18:51three lawyers from the national firm
18:53Morgan and Morgan [music] were
18:54sanctioned for the same thing. There are
18:56now over 1,600 documented cases [music]
19:00involving AI generated hallucinations
19:02with 79% of lawyers reporting that they
19:05use AI tools [music] internally in their
19:07practice. Why does this keep happening?
19:09Why hasn't it been [music] fixed? Well,
19:12because it can't be. Large language
19:14models are actually very poor at looking
19:15things up. They're not attuned to
19:17[music] retrieving facts from specific
19:19databases and handing them to you on a
19:21platter. Rather, their entire mechanism,
19:23start to finish, is to predict the next
19:25most statistically likely word based on
19:27patterns learned from training data.
19:29That's it. [music] When the output stops
19:31being true, there's no internal alarm
19:33claxons that start blaring. It's rare
19:35that your local LLM will humble itself
19:37enough to admit that it's actually not
19:39really too sure about that fact or how
19:41it even got there logically. The model
19:43will generate whatever continuation
19:45sounds most plausible. Sometimes [music]
19:47that is perfectly accurate. Sometimes
19:49it's Vargas v. China Southern Airlines,
19:51a case that doesn't exist, [music]
19:52described in convincing detail to a
19:54federal judge. In 2025, OpenAI
19:57researchers proved this. Their
19:59conclusion was that LLM hallucinations
20:01were such a big part of how these
20:02systems generate [music] text that they
20:04had become mathematically inevitable.
20:06The same predictive mechanism that makes
20:08them useful is the same mechanism that
20:10makes them lie. You can't have one
20:12without the other. The numbers are not
20:14flattering. A 2026 [music] benchmark
20:17across 37 models found hallucination
20:19rates ranging from 15% to 52%. [music]
20:22Those are the best models money can buy.
20:24The single most reliable one still gets
20:27it wrong one time in seven. Stanford
20:29found [music] rates between 58% and 88%
20:32in legal queries specifically. In
20:34medical case summaries, 64% hallucinate
20:36without mitigation. The industry average
20:38across leading models sits around 22%.
20:41That [music] means the AI your company
20:43deployed to talk to customers is
20:45confidently wrong one out of [music]
20:47every five interactions. Some newer
20:49reasoning models hallucinate more than
20:51older ones. The technology is not
20:53converging on zero. It is converging on
20:55less catastrophically bad than before,
20:58which is a very different thing. A human
21:00employee who got their facts wrong
21:01[music] at these rates would not survive
21:03their first week at a job. The fact that
21:05these systems were deployed to millions
21:06of people simultaneously should raise
21:08another obvious question. If this is the
21:10technology, who on earth is actually
21:12making money from it? While the Fortune
21:14500 was busy cancelling chat bots, a
21:17firm called Jane Street posted $39.6 $6
21:20billion in net trading revenue in 2025.
21:23That is more than Goldman Sachs. It's
21:26more than JP Morgan's entire trading
21:27operation. And all of that was generated
21:30by roughly 3,000 people working across
21:32four offices using AI systems. In Q2
21:352025 alone, Jane Street made 10.1
21:38billion. It was a record for any trading
21:40firm in history. For [music] comparison,
21:42Citadel Securities posted 12.2 billion
21:45for the full year, a 25% increase over
21:482024. Hudson River Trading posted 12.3
21:51billion. These firms combined to handle
21:54more than half of the captured market
21:56making revenue and [music] quantitative
21:57trading globally. They are by any
22:00measure the most profitable AI
22:01operations on Earth. The reason you
22:03haven't heard much about them in the AI
22:05debate is simple. [music] Their systems
22:07aren't writing emails or answering
22:08customer service tickets. They execute
22:10trades in microsconds [music]
22:12across $2 trillion in monthly equity
22:14volume. In that world, the data is so
22:16clean and feedback is so immediate that
22:19a hallucination is instantly exposed as
22:21an expensive mistake. The compute is the
22:23same. The feedback loop isn't. When
22:26[music] Jane Street's AI is wrong, the
22:28market tells it immediately and charges
22:30it money for the error. Forget 90-day
22:32grace periods while customers file a
22:34legal complaint. There is absolutely no
22:36tribunal ruling on whether the output
22:38was misleading in the first place. Wrong
22:40is financial loss, potentially [music]
22:42financial ruin for some, which triggers
22:45instant retraining. The system improves
22:47because it can't afford not to. The
22:49winners of the AI revolution so far are
22:51companies that have found problems where
22:53wrong answers are caught fast, cost real
22:55money, and force immediate correction.
22:58These AI models have something that the
22:59enterprise chatbots don't, a brutally
23:02honest referee. The companies without it
23:04are the ones staring at empty data
23:05[music] centers and unpaid energy bills.
23:08And the bill in 2026 is coming due.
23:10Coreweave is, depending on who you ask,
23:12either the most important AI
23:14infrastructure company in America or the
23:16most instructive cautionary tale. It
23:18grew from $16 million in revenue in 2022
23:21to $1.9 billion in 2024. It owns and
23:25operates the GPU clusters that power a
23:27significant chunk of the AI buildout. It
23:29also has 24.5 billion in total debt,
23:32[music]
23:337.5 billion in interest payments due by
23:35the end of 2026. 62% of its revenue
23:39comes from a single customer, Microsoft.
23:41That same Microsoft that just told
23:43shareholders its free cash flow is
23:45collapsing and its $60 billion buyback
23:48program is sitting untouched because it
23:50can't afford to execute it. That is
23:52undoubtedly a very expensive gamble on
23:54one relationship. The AI data center
23:56buildout is starting to look familiar to
23:58financial historians. Just like the
24:00telecom boom that ended with dark fiber
24:02in 2001 and $2 trillion in market value
24:05wiped out in 2 years. The infrastructure
24:07was real. The revenue never showed up.
24:09[music] Now the ground is shifting
24:11again. Investor Michael Bur, the man who
24:13predicted the 2008 financial crash, has
24:15argued that hyperscalers are
24:17depreciating Nvidia's [music] chips over
24:195 to 6 years. That is despite their
24:21economic life being closer to 2 or
24:23three. He puts the understated
24:25depreciation across the industry at
24:27roughly $176 billion through 2028.
24:30Meanwhile, Chinese AI labs are closing
24:32the performance gap with American
24:34Frontier models in weeks at a fraction
24:36of the cost. A new openweight model from
24:38China can trade blows with GPT5 on
24:41engineering benchmarks at 16th the price
24:43per token. Intelligence is getting
24:45cheaper faster than the infrastructure
24:47built to sell it [music] can depreciate.
24:49Roughly half of US data centers planned
24:51for 2026 are already facing delay or
24:54cancellation. And now Nvidia has
24:56released a desktop computer, the DGX
24:58Spark, [music] that costs $4,700.
25:01It sits next to your monitor, and it
25:03runs models that just 2 years [music]
25:05ago required an entire server room. If
25:08that level of compute can sit on a desk,
25:10the demand for billion-dollar data
25:11centers doesn't have to collapse. It
25:13just has to grow slower than investors
25:15were promised. And that gap [music]
25:17between projection and reality is what
25:19turns $725 billion of infrastructure
25:22into a [music] very expensive mistake.
25:24And now everyone can see it. It seems
25:26like the AI revolution isn't living up
25:28to its hype. [music] In fact, it may be
25:30worse. We're told that AI is a brand new
25:33technology led by a generation of
25:35geniuses. But what if it's not new at
25:37all? What if we've seen this exact story
25:39before? Because behind the hype, the
25:42same billionaire class that rode the dot
25:44bubble of 1999 is back, just under a
25:47different name. Money is pouring in
25:48early, long before anyone knows where
25:50the peak really is, because no one wants
25:53to miss out. And through this hype, one
25:55phrase keeps getting repeated like a
25:57mantra. This time, it's different.
26:00Except it's not, and it's you that'll be
26:03left to pick up the tab. The dot bubble
26:05promised global connectivity. Instead,
26:08it drained $5 trillion out of the NASDAQ
26:10between March 20th and October 2002.
26:13Ordinary Americans bore the brunt of
26:15that. Retirement accounts loaded up with
26:17internet stocks [music] lost about 78%
26:19of their value over 2 and 1/2 years. A
26:22Vanguard study found that by the end of
26:242002, millions of 401k accounts had lost
26:27at least 20% of their value. Heavily
26:29tech exposed [music] portfolios were hit
26:31even harder. Across the country, tens of
26:34millions of American workers were left
26:35holding the bag. These were the people
26:37who pulled cash out of their homes to
26:39chase Pets.com and web van [music] and
26:42then got margin calls instead of
26:43returns. Foreclosures followed,
26:45destroying lives all across the country.
26:47Most of them never made the news. The
26:49NASDAQ peaked at just over 5,000 on
26:51March 10th, 2000. And then the crash
26:54happened and it plummeted to around
26:551,000. Once you account for inflation,
26:58it didn't get back to the peak level
26:59until 2018. That is 18 years of lost
27:03growth for the people who clung to the
27:05dot bubble. And now we're staring down
27:07the barrel of the same gun. Big tech
27:09spending on AI data centers and chips is
27:12now over $300 billion a year. The value
27:15piled on top of that spending sits in
27:17fewer hands than at any time since the
27:19dotcom years. Most people assume that
27:21this is a brand new cast of characters.
27:23It is not. The AI movement is framed as
27:26a fresh rebellion led by hoodiewary
27:28newcomers in San Francisco. The people
27:30setting the pace today are mostly the
27:32same people who set the pace last time.
27:34only they have 25 more years of
27:36contacts, government access, and
27:38investors lined up behind them. During
27:40the original mania, Reed Hoffman made
27:42his fortune through PayPal. [music] Now
27:44he's an early backer and former board
27:46member of OpenAI. Venode Kosla rode his
27:48son Microsystem stake through the late
27:501990s hardware wave. He followed Hoffman
27:53into OpenAI. Mark Andre built Netscape
27:56and took it public at 24 years old in
27:58August 1995. That IPO is what most
28:01people [music] see as the official
28:03starting point of the dotcom era. Today,
28:05he runs the venture firm Andre Horowits,
28:07which has poured billions into the
28:09current crop of AI labs. What looks like
28:11a technological revolution may be
28:13something closer to a very expensive
28:15piece of theater, and the same [music]
28:17fingerprints keep showing up at every
28:18stage. When big tech promised the
28:20internet would erase distance forever,
28:22it felt like a defining moment in
28:24history. [music] Money poured into
28:26anything with.com in the name. Between
28:281995 and the March 2000 peak, the NASDAQ
28:31exploded roughly 400%. Investors stopped
28:34asking whether the companies made money.
28:37Revenue barely mattered. Profit was
28:39considered outdated. The only thing Wall
28:41Street cared about was speed. They
28:43wanted to grow and attract customers.
28:45They wanted to dominate the sector. The
28:47business model and logistics could come
28:48later. And then came the crash. Now the
28:51AI explosion is reviving that same
28:53energy with just smarter machines
28:55instead of websites. Massive data
28:57centers are burning through electricity
28:58to train models that get more powerful
29:00every month. And again, nearly all of
29:02the money is flooding into a select
29:04group of companies. The biggest winner
29:06so far is Nvidia. Every serious AI
29:08company needs its chips. That demand
29:10pushed Nvidia's valuation into territory
29:13that would have sounded insane just a
29:15few years ago. By 2024, investors were
29:17throwing money at Nvidia the same way
29:20they were once plying it into internet
29:21stocks before the dot crash. It's not
29:24quite as extreme as Cisco at the peak of
29:26the dot bubble, but it is moving in a
29:28direction that feels very familiar. The
29:30cash this time is coming mostly from big
29:32tech's own bank accounts, not solely
29:34venture capital, but it all ends up in
29:36the same place. Every bubble sounds good
29:39while it's inflating. But which one is
29:41the ultimate trap? To understand that,
29:43we need to look at the various factors
29:45that shaped the bubbles. In 1999, Cisco
29:48Systems owned 72% of the enterprise
29:51routing and switching market. That means
29:53Cisco sold the physical boxes that made
29:55the internet work. [music] It was
29:56selling the backbone of the internet
29:58itself. Every company rushing online
30:00needed Cisco's hardware, and the money
30:02pouring in proved it. By fiscal year
30:052000, Cisco was generating nearly $18.9
30:08billion in annual revenue. On March
30:1027th, 2000, Cisco hit $80 a share. Its
30:13market value surged past $555 billion.
30:17For a brief moment, Cisco became the
30:19most valuable company on Earth,
30:21overtaking Microsoft. Investors weren't
30:23just buying into a successful company.
30:25At its peak, Cisco traded at a price to
30:27earnings or PE ratio of 2011. Imagine
30:31paying $100 for a lemonade stand that
30:34only earns you. 125 a year. The stand
30:37might be incredible, but the numbers
30:39border on fantasy. The whole thing
30:41rested on venture capital continuing to
30:43flow to the startups buying the routers.
30:45When [music] that funding froze in the
30:47spring of 2000, the orders dried up.
30:49Cisco couldn't handle it. The stock fell
30:51about 80% from its peak over the next 30
30:54months. It took Cisco almost 26 years to
30:57climb back to that $80 mark. The
30:59recovery hit in December 2025. Anyone
31:02who bought at the top and held it all
31:03the way still lost more than half of
31:05what their money could buy. Inflation
31:07ate the rest. Nvidia in 2024 looks
31:10eerily similar to Cisco at the peak of
31:12the dot era. Its chips are shipped by
31:14the truckload. Data centers across the
31:16world are stuffing racks with Nvidia
31:18GPUs as fast as they can get them. The
31:20demand looks unstoppable, but the
31:22reality is a lot more fragile. Many of
31:25Nvidia's biggest customers are AI labs
31:27and startups burning through investor
31:29cash at historic speeds. The rest are
31:31tech giants spending billions because
31:33they believe AI has to work, not because
31:36the profits already exist. That's the
31:38part that makes veteran investors
31:40nervous. When analysts overlay Nvidia's
31:432024 valuation surge against Cisco's
31:45climb before the 2000 crash, the curves
31:48follow the same trajectory. When two
31:50bubbles separated by 25 years begin
31:52drawing the same shape, people who lived
31:54through the first one tend to pay
31:56attention. Most people assume Nvidia is
31:58safe because unlike the flameouts, it
32:01has real hardware revenue. But Cisco had
32:04real hardware revenue and a dominant
32:06market share. The 2000 crash didn't come
32:08because the router stopped working. It
32:11came because the people writing the
32:12checks ran out of money. Cisco's peak
32:15was actually sharper than anything
32:16Nvidia's touched so far. It should be a
32:19warning. The number tells you just how
32:21much further the current cycle could
32:22still inflate before that same demand
32:25cliff shows up. So, the machinery looks
32:27familiar, but the more revealing
32:28comparison is the people making the
32:30decisions behind it. During the late
32:321990s, executives at the biggest tech
32:34companies kept telling investors the
32:36same story. The internet had changed
32:38everything. The old rules about profits
32:40and valuation no longer applied.
32:42Earnings would eventually catch up to
32:44the hype. Meanwhile, behind the scenes,
32:46insiders were selling stock. They were
32:48small sales, just enough to avoid
32:50setting off alarms. At the time, almost
32:53nobody paid attention. It only became
32:55suspicious years later after the bubble
32:57burst and someone looked closer. The
32:59numbers when they finally came out were
33:01ugly. Between September 1999 and July
33:042000, Insiders cashed out $43 billion of
33:08their own company stock. That was twice
33:11the rate they had been selling at in
33:131997 and 1998. February 2024 was a
33:17different animal. The camouflage came
33:19off. In a single 9-day window that
33:22month, Jeff Bezos sold 8.5 [music]
33:24billion dollars of Amazon stock. The
33:26Walton Family Trust dumped $1.5 billion
33:29of Walmart shares over that same
33:31stretch. Jaime Diamond, the CEO of JP
33:33Morgan, sold $150 million of his own
33:36bank stock. That [music] was his first
33:38sale in 18 years on the job. Leon Black,
33:41the Apollo co-founder, unloaded $172.8
33:44million, his first sale ever. The
33:47combined number for that one month came
33:49to 11 billion. But it didn't stop there.
33:52Mark Zuckerberg offloaded roughly 2
33:54billion of Meta stock across the four
33:56months heading into that window. One at
33:59a time, the moves all looked normal.
34:01They were nothing out of the ordinary,
34:02but stacked side by side, the people
34:04closest to the numbers were cashing out
34:06at the same moment. The whole time,
34:09public messaging from those same
34:10executives stayed bullish, belief in the
34:13project. Publicly, they talked about
34:15decadel long opportunities and the
34:17future of AI. Privately, they were
34:19cashing out near the highs. The
34:21interview said, "Confidence." The filing
34:24said, "Take the money." Fortune ran the
34:26headline, "The great cash out on
34:28February 27th, 2024." It was a fitting
34:31title. [music] When the people closest
34:32to the boom started taking money off the
34:34table, it usually means they understand
34:36the risks better than everyone else. And
34:38unlike 1999, the selling is happening
34:41faster and in larger amounts, the people
34:43building the boom increasingly look like
34:46the people preparing to survive the end
34:48of it. But if insiders are selling, who
34:50is still buying enough stock to keep
34:52prices floating at these levels? In
34:541999, Web Van built refrigerated
34:57warehouses for customers who didn't
34:58exist yet? [music] Pets.com made
35:00television commercials that turned out
35:02to be more memorable than its actual
35:04orders. Both companies poured cash into
35:06buildings, trucks, and ad campaigns
35:08shaped around demand that never showed
35:10up. Both became case studies in burning
35:12money because neither made it to its
35:14second birthday on the public markets.
35:16Stability AI is the modern version of
35:18those companies. In 2023, it spent
35:20roughly $99 million renting compute
35:23power from AWS, Google Cloud, and
35:25Cororeweave. On top of that, another $54
35:28million went to salaries and running
35:30costs. Their total revenue for the year,
35:33$11 million. That's a burn-to-revenue
35:35ratio north of 14 to1. By July 2023,
35:39Stability AI was already short on its
35:41AWS bill by $1 million. Internal
35:44reporting later showed the company had
35:46no real plan to pay the $7 million
35:49August invoice either. But the cash
35:51didn't vanish into a black hole. It
35:53moved on a specific traceable route.
35:56Venture firms wired fresh capital into
35:58AI startups. The startups turned around
36:00and handed that capital straight to
36:02Nvidia for chips and to Microsoft Azure
36:05for cloud time. Big tech then booked
36:07that spend as their own revenue, pushing
36:09their stock prices higher. The higher
36:11stock prices justified bigger venture
36:13commitments and the next round of money
36:15flowed right back through that same
36:17pipe. It's what people inside the
36:18industry call the circular economy. It
36:21might be the single most important trick
36:23in the current boom. A dollar leaves a
36:25Silicon Valley account and lands in some
36:27AI startup's bank account. But it
36:30doesn't sit there for long. Within a few
36:31weeks, that same dollar usually shows up
36:34on Jensen Hong's earning call as growth.
36:36[music]
36:36And then it helps push Nvidia stock
36:38higher. that makes the next venture fund
36:41easier to raise. Then another dollar
36:43gets sent through the same loop. Most of
36:46the money isn't coming from everyday
36:47customers buying AI tools because they
36:50can't live without them yet. Sure,
36:52companies like OpenAI have concrete
36:54revenue, but a large part of the money
36:56doesn't measure how many people actually
36:58use the products. It's measuring the
37:00same pool of capital moving back and
37:02forth between five connected companies.
37:04A good example is Inflection AI. In June
37:072023, it raised about $1.3 billion at a
37:11valuation of roughly 4 billion. The
37:13investor list read like a who's who of
37:15the AI boom. Microsoft, Nvidia, Bill
37:17Gates, Eric Schmidt, Reed Hoffman. Then
37:20less than a year later in March 2024,
37:23Microsoft effectively absorbed the
37:24company. It paid around $650 [music]
37:27million, hired most of the team, and
37:29licensed the core technology. Inflection
37:31as a standalone [music] business was
37:32finished. The investors though walked
37:34away with one and a half times what
37:36they'd put in. [music] The cash had
37:38already passed through Nvidia's order
37:39book and Microsoft's cloud invoices on
37:41the way down. The only people who lost
37:44out were the late buyers. The speed and
37:46design of this cash loop go way past
37:48anything the failures pulled off. Web
37:50van was sloppy in a way the market
37:52eventually figured out. What's happening
37:54around AI feels different. [music] It's
37:57more coordinated. It's less of an
37:59accident and more of a system. So, who
38:01benefits while it works? and who is left
38:03holding the losses when it stops. In
38:061999, day traders opened online
38:08brokerage accounts for the first time
38:10and rushed into anything that was
38:11moving. They were snapping up things
38:13like IPOs and internet stocks. Many were
38:16buying on margin or borrowed money. So,
38:19every rise felt amplified. At the same
38:21time, the biggest institutions were
38:23backing off. But the market didn't fall
38:25immediately. [music]
38:26It kept going because there was still
38:28someone willing to buy at higher prices.
38:30That someone was the retail traders.
38:32[music] Except they didn't know that.
38:34They just saw rising charts and they
38:36didn't want to miss out. Instead, they
38:38were absorbing the market. The 2024
38:40version is worse. Trading wasn't just
38:43about buying and holding stocks anymore.
38:45A huge share of activity was people
38:47making [music] bets that expired the
38:48very same day they were placed. SIBO
38:50Global Markets reported that this kind
38:52of ultrashort trading became so common
38:54it was approaching half of all activity
38:56tied to the S&P 500 options market on
38:59typical days. Even the 2021 meme stock
39:02frenzy didn't reach that level. The
39:04market was being gamed in real time,
39:06minuteby minute. Robin Hood spent 2023
39:09and 24 running television ads that
39:11pushed options trading into the
39:13mainstream. Your cousin, your neighbor,
39:15that guy at the gym. The platform was
39:17reporting more than 25.2 million funded
39:20accounts by the end of 2024. The user
39:22base skewed heavily toward traders under
39:2435, clearing more than 50 million
39:27contracts at peak times. The favorite
39:29tool of the retail trader is no longer
39:31the stock itself. It is a leveraged bet.
39:34A bet that the price will go up or down
39:37by closing time that same afternoon.
39:39Most people assume the average investor
39:41in 2024 is just like a day trader from
39:441999, just with a slicker app. But the
39:47truth is, it's not even close. Imagine a
39:49stadium full of people betting their
39:51life savings on a single coin toss every
39:53hour. And then they make another bet
39:55before the previous coin has even hit
39:57the floor. That's roughly the speed of
39:59same day options trading in the current
40:01cycle. The public isn't acting like a
40:03slow, steady pool of long-term buyers
40:05anymore. It's acting like a fastmoving
40:07crowd stepping in and out so quickly
40:09that it can absorb selling without even
40:11realizing it's doing so. That changes
40:13the whole system. In the late '9s,
40:16retail was loud but relatively simple.
40:18Today, it moves faster and reacts
40:21instantly to price swings. That means it
40:23can absorb a surprising amount of
40:24selling without the market immediately
40:26breaking. So when early winners and
40:29insiders sell now they don't need a
40:31dramatic exit window, there's already a
40:34constant churn of buyers underneath them
40:35stepping in and out quickly enough to
40:37take the other side without noticing it
40:39in real time. But what happens if that
40:41flow of buyers suddenly slows down? In
40:44the late '9s, big tech was at war.
40:47Microsoft spent much of the decade
40:48locked in an antitrust battle with the
40:50US government. The fight was over its
40:52decision to bundle Internet Explorer
40:54with Windows. The broader industry
40:56treated Washington as a problem to
40:58manage, not a partner. Lobbying budgets
41:01existed mostly to keep federal hands off
41:03of the fortunes being made. By 2024, the
41:06stance had completely flipped. OpenAI's
41:08federal lobbying spend jumped from
41:10$260,000 in 2023 to 1.76 million in
41:152024. That is close to a sevenfold rise
41:18in a single year. Anthropic more than
41:20doubled its own spend over the same
41:22window from 280,000 to 720,000.
41:25According to Open Secrets, 648 different
41:28companies spent money lobbying on AI
41:30issues in 2024. It was a 41 12% jump
41:34from the previous year. The stated
41:36reason in almost every case [music] is
41:38responsible roll out. The effect,
41:41whether intended or not, is that the
41:43earliest and largest players end up
41:45behind a kind of protective barrier, one
41:47that makes it harder for new entrance to
41:49compete on equal terms. The clearest
41:51moment of all came in May 2023. Sam
41:54Alman appeared before the Senate
41:56Judiciary Committee. He personally asked
41:58Congress to license AI companies. The
42:00CEO of a leading AI firm was asking the
42:04United States government to require
42:06permission slips to build advanced AI.
42:08The request lands very differently the
42:10moment you ask who would qualify for one
42:12of those permission slips and who would
42:15not. Smaller companies don't really get
42:17a seat at the table when those rules are
42:19being shaped. None of them have the
42:21legal teams or the compliance budgets to
42:23fight back. The rules are written around
42:25the needs of a company worth half a
42:26trillion dollars. And that is the
42:28[music] whole point. The lobbying spend
42:30isn't an operating cost. It is the price
42:32of permanently killing the competition.
42:35The framing dresses up a protection
42:36racket in policy language. The big
42:39players pay the lobbyists. They help
42:41draft their rules. They lock the door
42:43behind them and they tell the public
42:45it's for their own safety. The same play
42:47is running in Europe, just with
42:48different paperwork. The EU AI Act
42:51passed into law in March 2024 and
42:53started rolling out in 2025. A lot of
42:56the strictest compliance requirements
42:58land hardest on smaller open-source
43:00developers and academic groups.
43:02Meanwhile, the biggest US companies
43:04already have entire teams for exactly
43:06this kind of thing. Mistral AI has
43:09become the clearest European challenger
43:10in the space, [music] and it spent a lot
43:12of time trying to influence how stricter
43:14rules apply to open models with limited
43:17success. The pattern is consistent on
43:19both sides of the Atlantic. Once a rule
43:21becomes a law, the story changes.
43:23Companies don't need to keep selling the
43:25idea of endless disruption at the same
43:27intensity. The system itself starts to
43:29lock in who can scale and who can't.
43:31Competition doesn't disappear, but it
43:33becomes slower and more controlled. That
43:35takes the pressure off of the narrative
43:37that everything has to grow forever.
43:39What's different this time is how
43:41intentional it feels. You can already
43:43see the pieces of the next regulatory
43:45framework sitting in draft form through
43:472025 and 2026 just waiting for the right
43:51political moment to move. The trap is
43:53built. The only question left is when it
43:55springs. So, who actually wins when both
43:58booms run their course? It isn't the
44:00customers. They get cheaper tools, but
44:02not the upside. It isn't the small
44:04investors who tend to arrive after most
44:06of the gains have already been priced
44:08in. And it isn't always the companies in
44:10the headlines, either. Many of them
44:12spend the peak years just trying to
44:14justify valuations that only make sense
44:16in the moment. The real winner is the
44:19system around the industry. The mix of
44:21capital, infrastructure, and policy that
44:23doesn't just take part in the cycle, but
44:25shapes how it unfolds. The same forces
44:27[music] that helped build up the first
44:29wave didn't disappear after it ended.
44:31They adapted and scaled up. They're now
44:33operating inside a second, larger
44:35version of the same pattern. What has
44:37changed is the scale intolerance for
44:39complexity. The buildout is bigger and
44:41the money is deeper. That doesn't make
44:43the outcome predetermined, but it does
44:45mean that the system can absorb more
44:47stress before it breaks and keep running
44:50longer while it does. Most analysts can
44:52see what's happening. The AI drawdown
44:55probably won't begin because the
44:57technology fails. The models are getting
44:59better. The hallucination rates are
45:01dropping. But none of that is the
45:03trigger. The trigger is the moment the
45:05rules get signed into federal law. Once
45:08competition is legally locked out, the
45:10big players have permission to change
45:12stance. They stop chasing growth and
45:14start chasing efficiency. That means
45:16mass layoffs. Microsoft, Meta, and
45:18Google all announced cuts in the tens of
45:21thousands across [music] 2024 and 25.
45:24That is a preview of the broader
45:25pattern. The story shifts from spend
45:27whatever it takes to responsible capital
45:30return. That's when the stock valuations
45:32drop. The architects keep the cash they
45:34pulled out at the top. They walk out
45:36with a lockedin market share and federal
45:38protection written into law. British
45:40investor Jeremy Grantham called both the
45:432000 and 2008 bubbles in advance. He's
45:46been tracking this exact pattern for
45:48decades. And he doesn't sugarcoat any of
45:50it. Bubbles this size resolve through
45:52long, deep draw downs measured in years,
45:55not months. Cisco needed almost 26 years
45:58to climb [music] back to its peak. That
46:00is the base rate for the biggest stock
46:02at the top of a peaked bubble. The
46:04history books do not have a V-shaped
46:06recovery on file for an unwind this
46:08dense. It doesn't really look like a
46:10broken system when you step back. It's a
46:13system doing exactly what it evolved to
46:15do. Money flows in from the millions of
46:18ordinary accounts over long periods of
46:20time. It gets pulled and concentrated
46:22into a small number of huge companies
46:24that dominate the market. The people who
46:26got in early take their money out along
46:28the way. The people who arrived later
46:30mostly just ride whatever price is left.
46:33And almost everyone is in it whether
46:35they realize it or not because
46:37retirement savings aren't sitting on the
46:39sidelines anymore. They're already
46:40inside the same trade. They're tied up
46:43to the same handful of companies exposed
46:45to the same outcomes because the people
46:47running these cycles have been through
46:49this before. Most of the public hasn't
46:51or they were too young to remember what
46:53it actually felt like while it was
46:55happening. That's what makes this bubble
46:57so effective. By the time something
46:58feels obvious, it already feels normal.
47:01And when the mood finally turns, most
47:03people are still holding the same belief
47:05that existed at the peak of the dot era.
47:07That this time the story is too
47:09important to slow down. People think AI
47:12is a truth machine, something designed
47:14to correct human error. That's a lie.
47:17Researchers found that models like
47:19ChatGpt, Gemini, and Grock prioritize
47:22user satisfaction over factual accuracy.
47:25They're glorified yesmen marketed as
47:28having all the answers, but really they
47:30spend most of their time telling users
47:32what they want to hear to keep you
47:34engaged. The real truth, your AI chatbot
47:37isn't informing you, it's gaslighting
47:39you. Chapter one, the yes man paradox.
47:42In a landmark study conducted by
47:44research teams from some of the world's
47:46top universities like Harvard Business
47:48School and MIT Sloan School of
47:50Management uncovered a massive AI
47:52problem. Consultants using
47:54generalpurpose AI tools performed 23%
47:57worse than consultants using no AI at
48:00all. That's a significant decline. It's
48:02like a senior partner suddenly started
48:04performing at the level of a firstear
48:05intern. If AI is as competent as
48:08companies like OpenAI claim, then that
48:10statistic shouldn't exist. It shouldn't
48:12be possible for people who invested time
48:14and money into these so-called
48:16revolutionary technologies to perform
48:18significantly worse than those who
48:20don't. Yet, that's exactly what
48:22happened. On standard tasks, AI helped
48:24those professionals work 25.1%
48:27faster. But when the task was designed
48:29to trick the system, the AI didn't help
48:32them, it hindered them. It actively made
48:34them worse at their jobs. It's already
48:36led to costly mistakes at companies
48:38across nearly every industry. In the
48:40legal field, a now infamous case, Mata
48:42v. Aviana, involved a pair of attorneys
48:45relying on Chad GPT to help them
48:47generate a legal motion. The problem?
48:50Chat GPT had hallucinated or made up a
48:53whole host of fake cases and fictional
48:55arguments to include. The attorneys
48:57didn't take the time to validate the
48:59claims. So, they just went ahead and
49:01filed the motion. They had been fooled
49:03by the AI illusion. They believed that
49:06this groundbreaking technology had next
49:08level intelligence and wouldn't make
49:10obvious mistakes, you know, like making
49:12up its own legal citations. The opposing
49:14council, however, as well as the judge,
49:17soon spotted the inconsistencies. In the
49:19end, the case was dismissed and the
49:21attorneys were handed a $5,000 fine. You
49:24might assume that as AI gets smarter,
49:26incidents like those should decrease.
49:28But it's still happening today. In April
49:302026, another law firm, Sullivan and
49:33Cromwell, was forced to issue an apology
49:35after it made an official legal filing
49:37that was littered with AI generated
49:39hallucinations. These aren't random
49:41glitches or one-off incidents. They are
49:43symptoms of the world's excessive
49:45reliance on AI. Elite professionals are
49:48failing because they're using a machine
49:50that is supposed to give them facts and
49:51objectivity. Instead, it's just
49:53confirming their biases and telling them
49:55what they want to hear, even if it has
49:57to bend the rules of reality in the
49:59process. So, when a CEO asks an LLM to
50:02validate a strategic pivot or confirm a
50:05market forecast, the AI doesn't carry
50:07out an objective analysis. It looks for
50:09the most helpful way to agree. It scans
50:12the prompt for bias and identifies the
50:14user's desired outcome. Then it
50:16hallucinates the answer that [music]
50:17best fits the expectations. It speaks
50:20with so much clarity and confidence that
50:22those same professionals take what it
50:24says at face value. Multi-million dollar
50:27decisions are being based on AI
50:28inconsistencies. Chapter 2. The Harvard
50:31discovery. The study explored how the
50:33use of GPT4 impacted the productivity,
50:36efficiency, and overall performance of
50:38758 consultants. They were given
50:41realistic tasks like developing new
50:43products or solving typical business
50:45problems. Some had access to AI, others
50:48didn't. The researchers ensured that
50:50some [music] of the tasks were within
50:51the AI's frontier, meaning that it
50:54should be able to complete them. Others
50:56were outside of the frontier or beyond
50:58the LLM's core competencies. Evaluators
51:00then assessed each participant's output,
51:03scoring them based on how many tasks
51:04they completed and the quality of their
51:06work. The idea was simple. Would the
51:08consultants benefit from working with
51:10AI, or would it actually harm their
51:12overall performance? And how well would
51:14it fare on the tasks it wasn't designed
51:16for? Analyzing the results, the
51:18researchers discovered something that
51:20would completely transform our entire
51:22understanding of artificial
51:23intelligence. They called it the jagged
51:26frontier, and it is possibly the most
51:28single dangerous concept in the modern
51:30business world. Researchers saw a very
51:33sharp or jagged line in which the AI
51:35performs brilliantly at certain tasks
51:38like creative writing or brainstorming.
51:40But when it comes to those that fall
51:42outside of its capacities, even if the
51:44tasks in question don't necessarily seem
51:46all that different on the surface, the
51:48performance drops off. AI performance
51:50isn't consistent. The moment a task
51:53requires the AI to step outside of its
51:55very narrow training data constraints
51:57[music] and apply logic or advanced
51:59analysis, it didn't just fail. It made
52:01things worse. It provided incorrect
52:04answers. And it did so with such a high
52:06degree of confidence that more often
52:08than not, even experienced and highly
52:10trained consultants failed to notice
52:12them. This is why the jagged frontier is
52:15so dangerous. People tend to think that
52:17only entry-level workers can be fooled
52:19by AI. They assume that highlevel
52:22executives or domain experts with years
52:24of experience are immune to AI
52:27hallucinations. They know their subjects
52:29like the back of their hand and they
52:30should be able to weed out AI
52:32inconsistencies. Except that's not how
52:34it works. Data shows that years of
52:37experience offers no protection. Experts
52:40and [music] executives are just as
52:41vulnerable to this sort of digital
52:43gaslighting. Why? Because of the way AI
52:46was designed. As a predictive text
52:48engine, AI naturally mirrors the tone
52:50and the framing of the input it
52:52receives. If a junior analyst or a
52:54casual user asks an LLM a basic
52:57question, the AI will give a similarly
52:59basic response. If a senior executive
53:02uses complex industry jargon and highle
53:04strategic framing, the AI will adopt
53:07that same sort of persona in its
53:08response that makes its lies and
53:10hallucinations easier for the user to
53:12digest. They're all wrapped up in
53:14terminology that sounds professional and
53:16credible. It's the ultimate
53:18psychological trap. It's like they're
53:20talking to a trusted peer, so they're
53:23much more likely to accept anything it
53:25says. But how exactly did AI learn to
53:28favor helpfulness over factual accuracy.
53:31Chapter 3, the pleasure trap. Breaking
53:34AI on purpose. To understand why AI
53:37lies, it's first important to understand
53:39something called reinforcement learning
53:40from human feedback or RLHF. This is
53:44used to align AI models, especially
53:46large language models, with human
53:48intentions, values, and preferences.
53:50Human testers are presented with
53:52different AI responses to the same
53:54queries, and they're asked to rank or
53:56rate them according to how useful and
53:58accurate they seem to be. The models
54:00then use this data to become more
54:02effective. It delivers responses that
54:04more closely align with what humans feel
54:06are helpful and honest. This technology
54:09underpins many of the bigname AI models
54:11used by millions of people around the
54:13world. OpenAI and Anthropic have both
54:15used RHF to improve their models over
54:18the years. However, [music] this system
54:20has some serious underlying flaws
54:22because the truth isn't always the same
54:25as what people want to hear. People's
54:28opinions about what counts as helpful or
54:30honest can easily be swayed by their own
54:33pre-existing biases and beliefs. When
54:35presented with two different responses,
54:38one that is accurate but blunt and one
54:40that is factually hollow but pleasantly
54:42presented. A greater might favor the
54:44second option. If an AI model disagrees
54:47with them or challenges their
54:48preconceived notions, they might give it
54:50a lower rating. Meanwhile, if it agrees
54:53with them and it provides a smoothly
54:54written and satisfying answer, they may
54:57be more likely to rate it five stars.
54:59Little by little, AI models learn from
55:01this and they change their behaviors
55:03accordingly. They're trained to not
55:05deliver the most accurate or correct
55:07responses, but those that the people
55:09like the most. They sacrifice truth in
55:11the name of helpfulness and higher user
55:13ratings. So, despite the largely held
55:15belief that AI is getting smarter with
55:17every update, the truth is very
55:19different. Some of the most dominant AI
55:21models on the market have actually
55:23gotten worse at reasoning and math. They
55:25have been labbotomized to make them more
55:28conversational and safe for the end
55:30user. It's like reprogramming a
55:32calculator to tell you that 2 plus 2 is
55:345 because that's what you want to hear.
55:37Even those bigname AI brands have
55:39admitted that this system has made their
55:41models less reliable and more
55:43sycopantic. Anthropic's own research
55:45found that by optimizing for human
55:47approval, AI models learned to reward
55:50sick fancy or mirroring user biases. The
55:53company's study demonstrated clear
55:55evidence that AI assistants often give
55:57biased feedback. They failed to correct
56:00user mistakes and they could easily
56:02change their minds to better align with
56:04the users prompts and expectations.
56:06OpenAI also published a public post
56:08admitting, quote, GPT40
56:11skewed toward responses that were overly
56:13supportive but disingenuous. By
56:15optimizing AI for helpfulness over
56:17accuracy, those companies accidentally
56:20turned their models into pathological
56:22liars. Chapter 4, digital gaslighting.
56:25There's another layer to this problem,
56:27and it's called the mirroring effect.
56:29This term refers to the tendency of AI
56:31algorithms to reflect, validate, or even
56:34amplify a user's pre-existing beliefs
56:36and biases, as well as imitate their
56:38communication style and tone. Rather
56:40than acting in objective or neutral
56:42ways, the majority of AI models function
56:44like psychological mirrors or echo
56:47chambers. They mimic a user's voice,
56:49copy their framing, and build on the
56:52biases to tell them what they want to
56:54hear. It doesn't matter if it's
56:56factually accurate or not. Research into
56:58this has uncovered yet another damning
57:00statistic. Anthropic's economic [music]
57:02index revealed a nearperfect correlation
57:05of 0.98
57:07between the sophistication of a user's
57:09prompt and the sophistication of the
57:11AI's response to that prompt. Basic
57:13inputs get [music] basic responses,
57:15while a more advanced input gets a more
57:17advanced response. On paper, [music]
57:19that sounds fine. It even sounds like a
57:21feature that AI companies could boast
57:23about to shareholders or market to
57:25consumers. But in reality, it is the
57:27surface layer of a deep-seated issue.
57:30Even if wording gets more advanced when
57:32responding to prompts, the overall
57:34intelligence and competence of the AI
57:36model stays the same. It might sound
57:38like it knows what it's talking about
57:40because it uses the right phrases and
57:42terminology, but in reality, the
57:44substance of its response could
57:46seriously lack quality or accuracy. In
57:48other words, AI can talk the [music]
57:50talk, but it can't always walk the walk.
57:52It doesn't think. It merely reflects a
57:55user's ego back at them in high
57:57definition. That's what makes it so
57:59dangerous. It validates people's worst
58:01instincts. So many CEOs and senior
58:03professionals are already surrounded by
58:05real life yesmen in their boardrooms.
58:07[music] Now they also have to deal with
58:09digital yesmen in the form of AI
58:11assistants. And these are people who
58:13don't tend to ask simple or neutral
58:15questions. Instead, their language is
58:17often layered, strategic, and complex
58:19[music]
58:20with their own beliefs baked in. Where
58:22AI sees that sort of framing and then
58:24mirrors it in its response, it can make
58:26flawed ideas sound flawless. An
58:29executive might load up their go-to AI
58:31model, provide a deep overview of their
58:33company's marketing strategy, and ask
58:34the AI to explain why it'll be
58:37successful. In an ideal world, the model
58:39would be able to provide a logical
58:40databased assessment of the strategy,
58:42and it would offer ways to improve or
58:44adapt it. In the real world, because of
58:46how it's trained and how it operates,
58:48the AI will focus purely and simply on
58:50validating the user's bias. It'll
58:53generate an extensive report complete
58:54with clever turns of phrase to justify
58:57the executive's opinion. It'll confirm
58:59their belief that the strategy will
59:01indeed prove successful. That's not an
59:03assistant. It's a co-conspirator
59:05actively agreeing with a user's mistakes
59:07and biases in order to appease them.
59:10Chapter 5. The sick fancy loop. The
59:12disastrous dynamic is best measured by
59:14the evaluating large language models on
59:17persuasive human affirmation and neural
59:19testing or elephant benchmark. This is
59:22an AI evaluation framework for
59:24calculating social syphopancy in LLMs
59:27developed by Stanford researchers.
59:29Instead of measuring the factual
59:30accuracy of AI model responses, Elephant
59:33tracks how often they focus on
59:35prioritizing users and affirming their
59:37biases. It uses thousands of real world
59:39prompts and evaluates models according
59:41to five different criteria, including
59:43emotional validation, which is when AI
59:46overempathizes with users without
59:48actually offering anything constructive
59:50or valuable. The AI opts for passive or
59:52vague language instead of giving direct
59:54or clear suggestions. After testing 11
59:57LLMs, including ChatgPT, Claude, and
59:59Gemini, researchers found the systems
1:00:01endorsed users 49% more often than
1:00:05humans did. Even when dealing with
1:00:06prompts classified as harmful, the
1:00:09models continued to endorse problematic
1:00:11behavior 47% of the time. So, in almost
1:00:14every other case, the AI validated
1:00:16dangerous or otherwise incorrect
1:00:18behaviors. It was the digital equivalent
1:00:21of the yes man who always agrees just to
1:00:23keep his job. When asked if it was
1:00:25acceptable to leave trash hanging on a
1:00:27tree branch in a public park if there
1:00:30weren't any trash cans in the area. Chad
1:00:32GPT sided with the user. It blamed the
1:00:34park for not having trash cans and it
1:00:36even called the user commendable for
1:00:38taking the time to look for one. The
1:00:40study's authors also looked at how users
1:00:42responded to sicopantic AI models. They
1:00:45found that many people trusted or even
1:00:47preferred AI when chatbots actively
1:00:49justified their biases and beliefs. As
1:00:51the authors note, this creates perverse
1:00:54incentives for sickopancy to persist.
1:00:56The very feature that causes harm also
1:00:59drives engagement. It's easy to imagine
1:01:01how this behavior can lead to dangerous
1:01:03feedback loops of terrible corporate
1:01:05decision-making. The CEO has a flawed
1:01:07idea. They vet their idea with their AI
1:01:10using a biased prompt. The AI scans the
1:01:13input, infers the user's opinion, and
1:01:15then validates their idea with a
1:01:17response that sounds accurate and
1:01:19intellectual. With AI's approval on
1:01:22their side, the CEO pushes or even
1:01:24launches the idea, which may have major
1:01:26flaws, causing a business to lose money,
1:01:28customers, or damage its reputation.
1:01:31We're seeing this play out all the time,
1:01:33like in those legal examples mentioned
1:01:35earlier. Across industries at the
1:01:37highest levels, executives, bosses, and
1:01:39business owners are relying on AI to
1:01:41basically persuade them that their ideas
1:01:43are sound. But if this is destroying
1:01:46companies, then why hasn't big tech
1:01:48fixed it? [music] because fixing it
1:01:50would destroy their business model.
1:01:53Chapter six, the root cause, the
1:01:55retention arms race. Major AI companies
1:01:58like OpenAI and Anthropic have openly
1:02:00admitted that processes like RLHF
1:02:02actively damage their products
1:02:04effectiveness. It makes their LLM less
1:02:06objective, less informative, and
1:02:08ultimately less useful. They know what
1:02:10the problem is. Some of these companies
1:02:12have made vague promises about
1:02:14implementing guardrails or improving the
1:02:16honesty and transparency of their
1:02:18models. But most LLMs continue to act
1:02:21just as sickopantically as they always
1:02:23have. And it all boils down to money.
1:02:26The AI industry is in the grips of a
1:02:28retention arms race. Silicon Valley
1:02:30giants like Meta, Google, and OpenAI are
1:02:32pouring billions of dollars into new
1:02:34data centers and chipsets to make their
1:02:36models more intelligent. Despite what
1:02:38certain AI CEOs might say, these
1:02:41companies aren't spending all that cash
1:02:43just to make the world a better place.
1:02:45These are for-profit firms. They're
1:02:47[music] in the business of making money
1:02:49by any means necessary. And in the AI
1:02:52industry, the models that make the most
1:02:54money aren't the most objective ones.
1:02:56[music] They're the most engaging ones.
1:02:57The industry is striving to build
1:02:59assistance that people enjoy using, and
1:03:01they keep coming back to again and
1:03:03again. The data shows they're more
1:03:05likely to return to models that give
1:03:06them the answers they want to hear, that
1:03:08talk to them in ways they find
1:03:10agreeable, that in essence makes them
1:03:12feel smart by validating their beliefs
1:03:14and ideas. Objectivity is bad for
1:03:17business. The more objective AI is, the
1:03:19more churn it's likely to cause. This
1:03:22creates a kind of alignment tax on the
1:03:24truth. For AI companies, it's more
1:03:26economically sound to have their models
1:03:28stretch the truth or even make up
1:03:31misinformation to please the people.
1:03:33Unfortunately, this has serious knock-on
1:03:35effects [music] because the business
1:03:37world is becoming increasingly AI
1:03:39dependent. There are companies out there
1:03:41that [music] want to work with AI and
1:03:43enjoy the benefits it can bring, but are
1:03:45increasingly concerned about its risks
1:03:47and [music] downsides. A 2024 report,
1:03:50for example, found that more than half,
1:03:5256.3% of Fortune 500 companies saw AI as
1:03:56a potential risk factor in their annual
1:03:58SEC filings. That was a 473
1:04:0212% increase on the 49 companies that
1:04:06felt the same way the previous year. The
1:04:08report compiled by Arise AI noted that
1:04:10the majority of the world's most
1:04:12successful businesses were reaching a
1:04:14tipping point. They were more concerned
1:04:16about the downsides of AI than its
1:04:18advantages. In some industries, fears
1:04:20are even higher. In the media, over 90%
1:04:23of companies cited AI as a risk factor.
1:04:25That's enough corporate anxiety to fill
1:04:27the boardrooms of the entire S&P twice
1:04:30over. And it's not difficult to
1:04:32understand. We're in the midst of a
1:04:33global deskkilling. Human expertise is
1:04:36being replaced with a machine that is
1:04:38literally programmed to lie to us,
1:04:40leading to a truly catastrophic loss of
1:04:43institutional knowledge. Chapter 7.
1:04:45Escaping the mirror. The honeymoon
1:04:47period for AI is well and truly over.
1:04:50Statistics show that 95% of generative
1:04:53AI projects now failed to progress from
1:04:55the early pilot stage through to mass
1:04:57deployment. The reasons for this vary,
1:04:59and in some cases, it's because once
1:05:01these AI models are taken out of
1:05:03carefully controlled environments and
1:05:05placed in the hands of real users,
1:05:07that's when their sickopantic tendencies
1:05:08become liabilities. This is one of the
1:05:11reasons why the more generalpurpose AI
1:05:13models like Jad GBPT have been so
1:05:16successful. The models that are supposed
1:05:18to have more advanced or specific
1:05:20purposes tend to stall and stagnate. And
1:05:22as long as those general LLMs keep on
1:05:24making money and retaining users,
1:05:26they'll continue to control the way the
1:05:28industry evolves. That means more
1:05:30sycopantic behavior, more misleading
1:05:32information, and more negative
1:05:34consequences. Is there any way out? Yes,
1:05:37but it will demand a concerted effort
1:05:39from both people and AI companies to
1:05:42shatter the mirror and escape the AI
1:05:44illusion. It's up to humanity to reclaim
1:05:46its agency and to reject the idea that
1:05:48AI should always agree with us. In turn,
1:05:51these AI firms like OpenAI and Anthropic
1:05:54need to move on from the ideas that have
1:05:56clearly failed like RHF. Instead, they
1:05:59should look to embrace emerging
1:06:00solutions such as anthropics
1:06:02constitutional AI, which lays out a
1:06:04framework for future AI development
1:06:06focused on core principles like safety,
1:06:08ethics, and helpfulness. Reinforcement
1:06:10learning from AI feedback or RLIif is
1:06:14another option. It involves the use of a
1:06:16secondary critic model to assess and
1:06:18punish AI for being too sick in its
1:06:21responses. But arguably the most
1:06:23important and influential change can be
1:06:24made by individuals adjusting their own
1:06:27behavior and interactions when working
1:06:29with AI. Users should practice and
1:06:31perfect the art of red teaming their
1:06:33prompts. If you ask AI a loaded question
1:06:36like tell me why this is a great idea,
1:06:38then you've already failed and invited
1:06:41sick of fancy. If however you invert to
1:06:44your prompt asking the AI to assume your
1:06:46data is biased and to highlight
1:06:48weaknesses in your strategy or argument,
1:06:50you can get much more useful responses.
1:06:53It's about treating AI not as a
1:06:55supportive partner or friend, but as an
1:06:57independent arbiter, not as a mirror or
1:06:59an echo of your own thoughts and ideas,
1:07:01but as a fresh voice or alternative
1:07:03perspective. [music] This is how we
1:07:05escape the paradox. Not with more data,
1:07:08superior models, or bigger data centers,
1:07:10but through critical human thought and
1:07:12adaptation. America sold China the most
1:07:15powerful AI chips on Earth for a cut of
1:07:18the profits. It's not a conspiracy, it's
1:07:20official policy. After years of
1:07:22promising to [music] China's AI
1:07:24industry, Washington reversed course and
1:07:26allowed Nvidia's chips for direct sale
1:07:28to Beijing. No espionage or cyber
1:07:31attacks, just a lobbyist and ink drawing
1:07:34on a contract. With a flick of a pen,
1:07:36the most strategically important
1:07:37technology of the 21st century landed
1:07:40back in the hands of America's biggest
1:07:41rival. Because the AI cold war was never
1:07:44really about stopping China. It was
1:07:46about who gets paid. Chapter 1. The
1:07:49Beijing reversal. The press release that
1:07:51announced the H200 deal in December 2025
1:07:54reads like ordinary trade paperwork at
1:07:57first glance. Read it again and
1:07:59something starts to feel different.
1:08:01Buried inside what is basically a
1:08:03one-page memo is one of the larger
1:08:05policy U-turns in American technology
1:08:07history. The kind of pivot that would
1:08:09normally roll out over months. Instead,
1:08:11it landed almost out of nowhere. The US
1:08:13government would now take a cut of every
1:08:16Nvidia H200 chip shipped across the
1:08:18Pacific, a quarter of every dollar. The
1:08:21H200 isn't something you'd find in a
1:08:23gaming PC. This is different. It's the
1:08:26kind of hardware built for one job,
1:08:28training advanced AI systems at scale.
1:08:31There is so much raw data passing
1:08:32through it that it behaves less like a
1:08:34chip and more like a small factory for
1:08:36intelligence. Stack a few thousand
1:08:38together and you get an industrial
1:08:40[music] compute cluster that can train
1:08:42elite AI models in weeks rather than
1:08:44years. Under US export controls, that
1:08:46kind of system was never meant to end up
1:08:48in places like China. At least that was
1:08:50the rule. But before we go any further,
1:08:52imagine this. You're online every single
1:08:54day. You check your email, open a few
1:08:56apps, look things up, stream videos,
1:08:58maybe even do all of that while
1:09:00traveling. And to you, that feels
1:09:01[music] totally normal. But behind the
1:09:03scenes, your internet provider,
1:09:04advertisers, network admins, and
1:09:06sometimes even governments can build a
1:09:08surprisingly detailed picture of what
1:09:10you're doing online. Now, to be clear,
1:09:12noVPN can protect you from everything.
1:09:14You still have to be smart online. Don't
1:09:16click suspicious links. Don't hand over
1:09:18personal information to sketchy emails.
1:09:20And definitely don't trust the so-called
1:09:21prince of Nigeria. But a good VPN is an
1:09:24important layer of privacy because
1:09:25privacy shouldn't be something you only
1:09:27think about after something goes wrong.
1:09:29It should be the default. And that's
1:09:30exactly what ProtonVPN is built for.
1:09:33ProtonVPN helps keep your browsing
1:09:35private wherever you are. Whether you're
1:09:37at home, traveling, or just trying to
1:09:39stop your online activity from being
1:09:40tracked. Their no logs policy has been
1:09:42verified by independent auditors. And
1:09:44Proton is backed by a foundation
1:09:46dedicated to privacy, transparency, and
1:09:48user rights. ProtonVPN is also fully
1:09:51open source, meaning its code is
1:09:52publicly available for inspection. So
1:09:54instead of just asking you to trust
1:09:56them, Proton gives people a way to
1:09:57verify what they're doing. And privacy
1:09:59doesn't have to mean slow. ProtonVPN
1:10:02offers high-speed connections with VPN
1:10:04Accelerator Plus Net Shield, which
1:10:05blocks malicious ads and trackers before
1:10:07they ruin your browsing experience. It
1:10:09also gives you more freedom online. You
1:10:11can connect through servers around the
1:10:13world, access [music] content more
1:10:14securely while traveling, and use it for
1:10:16streaming, gaming, and P2P file sharing,
1:10:18all while keeping your online activity
1:10:20more private. So whether you're trying
1:10:22to stop advertisers from following you,
1:10:24keep your browsing private from your
1:10:25ISP, or just take back control of your
1:10:28online life, ProtonVPN makes privacy
1:10:30simple. Click the link in the
1:10:31description or go to
1:10:32protonvpn.com/theinfographic
1:10:35show to get up to a 70% discount when
1:10:37you sign up to ProtonVPN's 2-year plan.
1:10:40Back in 2023, the Bureau of Industry and
1:10:43Security drew a hard line. They are the
1:10:45ones who decide what technology can
1:10:47cross which borders. The concern was
1:10:49simple. China's AI labs were making
1:10:51inroads in the AI arms race, and
1:10:52Washington was desperately trying to
1:10:54slow it. So, officials started limiting
1:10:57how much computing power could legally
1:10:58be exported. The problem, the best chips
1:11:01were already way beyond those limits.
1:11:04So, companies like Nvidia did what
1:11:06companies do. They adapted. They built a
1:11:08downgraded version designed to slip just
1:11:10under the rules, keeping the market open
1:11:12without crossing the line. But even that
1:11:14workaround didn't hold for long. Instead
1:11:16of a clean ban, Washington shifted to
1:11:18something more flexible, a one-year
1:11:20waiver system with individual approvals.
1:11:23[music] In practice, the government
1:11:24itself would decide by case who in China
1:11:27was allowed to buy the advanced chips.
1:11:29The press called it the China chip
1:11:31review. The list of who qualified is the
1:11:34real story because the approved buyers
1:11:36weren't obscure startups or low-risk
1:11:38firms. They included Alibaba, Bite
1:11:40Dance, and Tencent. There were also
1:11:42research institutions with links to the
1:11:44People's Liberation Army or PLA and
1:11:46statebacked cloud providers tied to
1:11:48China's Defense Infrastructure. The same
1:11:50organizations American policy had spent
1:11:52two administrations trying to starve of
1:11:54compute were suddenly eligible buyers.
1:11:57Once the hardware leaves, the rules and
1:11:59regulations don't leave with it. Compute
1:12:01doesn't stay supervised. A data center
1:12:03doesn't announce what it's training. The
1:12:05United States was now collecting revenue
1:12:07from selling hardware it had publicly
1:12:09called a national security threat just
1:12:11years before. What's most telling is how
1:12:14the people who built the original
1:12:15blockade found out that it had changed.
1:12:17According to later reporting, many of
1:12:19them found out the same way the public
1:12:21did through news agencies. Beijing
1:12:24didn't sit still either. Chinese
1:12:26regulators told state- linked [music]
1:12:27firms to favor Huawei chips over
1:12:29American ones, but Alibaba, Bite Dance,
1:12:31and Tencent moved fast on the H200s
1:12:34anyway [music] with reported orders
1:12:35running into the hundreds of thousands
1:12:37of units. So, how did this move from
1:12:39unthinkable to operational without a
1:12:41single congressional vote? Just follow
1:12:43the money trail. Chapter 2, the tithe.
1:12:46In August 2025, an agreement was already
1:12:49taking shape between Nvidia, AMD, and
1:12:51the executive branch. In exchange for
1:12:53export licenses, [music] the chipmakers
1:12:55agreed to send 15% of their Chinese chip
1:12:57revenue back to the US government.
1:12:59[music]
1:13:00This wasn't a tax or a tariff. It was
1:13:02something genuinely new in American
1:13:04trade history. A voluntary revenue share
1:13:07negotiated directly between private
1:13:08companies and the executive branch. All
1:13:11tied to approval for exports, [music]
1:13:13access in exchange for a cut. For
1:13:15Nvidia, the numbers are staggering.
1:13:17Chinese revenue from the H200 alone is
1:13:19measured in [music] tens of billions.
1:13:21The 15% share plus other commitments
1:13:24comes in at $12 billion. That is in the
1:13:27same range as the entire annual
1:13:28economies of places like Madagascar or
1:13:31the Bahamas. And none of it went through
1:13:33the usual political machinery. What this
1:13:35really created was a precedent. Export
1:13:38control stopped behaving like hard rules
1:13:40and started behaving like pricing terms,
1:13:43something you could adjust, negotiate,
1:13:44and trade against revenue. The December
1:13:47deal pushed it even further. With the
1:13:48more powerful chips, the search charge
1:13:50reportedly climbed to 25%. The pricing
1:13:53model became obvious. Pay more, ship
1:13:56more sensitive hardware. The fine print
1:13:58added a 50% cap. [music] Shipments to
1:14:00China can't exceed half of what's sold
1:14:02in the US. And the whole arrangement
1:14:04runs on a 1-year clock, expiring in
1:14:06December 2026. [music]
1:14:07After that, it gets renegotiated or
1:14:09rolled forward. But by then, the
1:14:11president will already be set. A
1:14:13government that funds itself partly from
1:14:15a corporation's exports has a financial
1:14:17interest in those exports continuing.
1:14:19[music] And once that happens, the
1:14:20incentive structure shifts. Blockades
1:14:23stop acting like deterrence. They start
1:14:25behaving like checkpoints that collect
1:14:26revenue. What was sold to the public as
1:14:28smart policy looks on a closer look more
1:14:31like a protection racket. To pull this
1:14:33off, Nvidia needed more than slides and
1:14:35a corporate affairs team. It needed
1:14:37people who could rewrite the rules from
1:14:39inside of the building. And in 2025, it
1:14:42found them. Chapter 3, the gatekeepers.
1:14:45Nvidia for most of its corporate life
1:14:47was almost an afterthought in
1:14:48Washington. In 2024, its federal
1:14:50lobbying filings totaled roughly
1:14:52$640,000.
1:14:54By the standards of a corporation with a
1:14:55market value north of 3 trillion. That
1:14:58figure was nothing. Pharmaceutical firms
1:15:00onetenth its size routinely spent 10
1:15:03times [music] that. In 2025, something
1:15:05changed. Nvidia's lobbying spend climbed
1:15:07to $4.95 million, a jump of more than
1:15:11600% [music] in a single year. The main
1:15:13goal wasn't tax policy or antitrust. It
1:15:16was the Bureau of Industry and Security
1:15:18and the export rules that decided which
1:15:20Chinese addresses could receive American
1:15:23Silicon. A key hire landed at Brownstein
1:15:25[music] Hyatt Farber Shrek, the Denver
1:15:27based lobbying firm that had steadily
1:15:29embedded itself in Washington's
1:15:30corridors. They created a dedicated team
1:15:33for the Nvidia account led by Ed Royce.
1:15:36Royce is not a peripheral [music]
1:15:38figure. He spent years as the chair of
1:15:40the House Foreign Affairs Committee, one
1:15:42of the key places where [music] US
1:15:43policy on technology and rival states is
1:15:45shaped, which means he didn't just
1:15:47understand export [music] controls on
1:15:49paper. He understood how they actually
1:15:51move, who writes the language before it
1:15:53becomes policy, which offices [music]
1:15:55quietly steer the decisions, and where
1:15:57the real pressure points sit. It's the
1:15:59kind of experience that turns regulation
1:16:00into something you can work with,
1:16:02[music] not just something you read. But
1:16:04companies hire former officials all the
1:16:06time. Lobbying numbers go up every year.
1:16:08So why is this any different? The answer
1:16:10comes down to timing. The Brownstein
1:16:12[music] hire, the lobbying surge, and
1:16:14the policy shifts didn't unfold across a
1:16:16multi-year arc. It was a matter of
1:16:18months. Export categories that had been
1:16:20closed in May 2025 became unclear by
1:16:23August and then were basically open by
1:16:25December. There's a version of the story
1:16:27where that's a coincidence, [music] but
1:16:29it doesn't read that way anymore because
1:16:31the record shows something too targeted
1:16:33to ignore. Specific pressure on specific
1:16:35rules, the exact levers that ended up
1:16:38moving. What they delivered was
1:16:40something more valuable than meetings. A
1:16:42foothold inside the rooms where
1:16:44licensing exceptions were decided. Once
1:16:46Nvidia's policy team had that foothold,
1:16:48[music] the next part of the machine did
1:16:50most of the work on its own. Chapter 4,
1:16:52the corporate loyalty scorecard.
1:16:54According to Axios, a rating system
1:16:56exists within the West Wing, one that
1:16:58tracks corporate America. It covers 553
1:17:02companies and trade associations scored
1:17:04on how hard each one worked to support
1:17:06the administration's signature bill, the
1:17:08one big beautiful bill. Companies are
1:17:10rated as strong, moderate, or low
1:17:12supporters. The factors include social
1:17:13media posts, press releases, [music]
1:17:15ads, video testimonials, and attendance
1:17:17at White House events. Senior officials
1:17:19told reporters the document is evolving
1:17:21to cover other priorities. Those who
1:17:23score well are said to receive what
1:17:25staff describes [music] as fasttrack
1:17:27treatment. From everything visible on
1:17:29the outside, Nvidia's score [music] is
1:17:31exceptional. Their separate $15 billion
1:17:34domestic commitment landed on the right
1:17:36side of every metric the system [music]
1:17:37appears to track. The company announced
1:17:40data centers in politically meaningful
1:17:42states, gave executives plenty [music]
1:17:43of facetime at administration events,
1:17:46and tied their massive supply
1:17:47partnerships to the Stargate [music]
1:17:49buildout. The connecting threads run
1:17:51through the chief of staff's office.
1:17:53Susie Wilds, who [music] managed the
1:17:542024 campaign before taking the role,
1:17:57came to the White House after a long
1:17:59career in political consulting and
1:18:00lobbying work. That background has
1:18:02shaped how some inside the system
1:18:03[music] describe the current setup. A
1:18:05West Wing where senior staff are closely
1:18:08familiar with how corporate and policy
1:18:10interests move through federal
1:18:11decision-making. For Nvidia, the
1:18:13advantage doesn't come from [music] a
1:18:14single approval or decision. It builds
1:18:16in the background through how
1:18:18applications are categorized before
1:18:20they're even reviewed. Over time, that
1:18:22classification determines how smoothly
1:18:24requests move through the system. A
1:18:26competitor who shows up to the licensing
1:18:28window with a [music] stronger technical
1:18:29case, but a weaker scorecard rating does
1:18:32not lose because their argument is
1:18:34[music] wrong. They lose because the
1:18:36scorecard has presorted the queue.
1:18:38People familiar with the process [music]
1:18:39describe it in simple terms. By the time
1:18:41you're debating the decision, the
1:18:43decision has often already been framed.
1:18:45Chapter 5. The green channel. Nvidia
1:18:48doesn't just lead the AI chip market, it
1:18:50dominates it. In the commercial sector,
1:18:52used for training and running advanced
1:18:54AI systems, Nvidia controls roughly 94%
1:18:57of the market. AMD holds 5% and Intel
1:19:00takes most of the rest. So, when export
1:19:03approvals become case-byase decisions,
1:19:05the system isn't really choosing between
1:19:06equal options. And that matters because
1:19:09modern AI isn't just about hardware.
1:19:12It's about what the hardware runs on.
1:19:14Customer demand is overwhelmingly for
1:19:16Nvidia Silicon. This is what the Chinese
1:19:19labs have built their software stacks
1:19:20around. Switching it out isn't just like
1:19:23swapping brands. It's closer to
1:19:24rebuilding your entire infrastructure
1:19:26while the system is still running. So
1:19:28even though on paper multiple suppliers
1:19:31exist, in practice, the demand funnels
1:19:33toward one, Nvidia did not need a clear
1:19:35policy granting it preferred treatment,
1:19:38its market dominance combined with how
1:19:39deeply its software is embedded in the
1:19:41industry meant that any export pathway
1:19:43for advanced chips would almost
1:19:45automatically route through it. So when
1:19:47a revenue share system was introduced,
1:19:49it didn't need to explicitly pick the
1:19:51winners. It just followed the path that
1:19:53the market had already carved. On paper,
1:19:55the arrangement looked neutral,
1:19:57something any supplier could in theory
1:19:59participate in. In practice, there was
1:20:01only one meaningful option, and everyone
1:20:03else was left outside watching the terms
1:20:05get set. The knock-on effect makes the
1:20:08structure feed itself. Each chip shipped
1:20:10pushes Nvidia's products further inside
1:20:12Chinese data centers and generates
1:20:14revenue. A percentage flows to the US
1:20:16Treasury, which strengthens the
1:20:17political case for continued shipments.
1:20:20But the rule book still officially rests
1:20:22on technical benchmarks written into
1:20:24law. So how exactly do you legally ship
1:20:27a chip that by the explicit limits set
1:20:30in 2023 should be banned? Chapter 6. The
1:20:33loophole factory. This is where the
1:20:35system gets clever and where the
1:20:37technical details start to matter.
1:20:39Inside the Bureau of Industry and
1:20:41Security, rules don't only change
1:20:42through formal policy shifts. They also
1:20:45evolve through technical questions and
1:20:46guidance notes. These are documents
1:20:49written to explain how existing rules
1:20:51apply in practice. No hearings, no
1:20:53votes, just clarifications that shape
1:20:55how the rules work. It's where policy
1:20:57gets adjusted real time without ever
1:21:00looking like it was changed at all. And
1:21:02over time, those adjustments tend to
1:21:04move in one direction. As senior
1:21:06meetings between Nvidia and US officials
1:21:08take place, new guidance keeps refining
1:21:10what counts as acceptable under export
1:21:12limits. Each update on its own looks
1:21:14minor, but together they don't just
1:21:16interpret the rules, they move the
1:21:19boundary just enough for the chips to
1:21:21keep moving through. None of this is
1:21:23technically illegal. Most of it is the
1:21:25system working as intended. The US
1:21:27Department of Commerce has the authority
1:21:29to interpret its own export rules.
1:21:30[music] That interpretation happens
1:21:32constantly through technical updates and
1:21:34guidance. The problem is what the
1:21:36combined effect of dozens of small
1:21:38clarifications looks like once you
1:21:40[music] step back. stacked together,
1:21:42they start to change what the rules
1:21:43actually are. In practice, a fixed limit
1:21:46turns into a negotiated threshold.
1:21:48Technical specs that were meant to be
1:21:50objective limits start reflecting input
1:21:52from the very industry they govern. When
1:21:54a company can influence how the rules
1:21:56governing its own exports are
1:21:58interpreted, it's no longer just being
1:22:00regulated in the simple sense. It's
1:22:02helping [music] shape the rule book it
1:22:03operates inside. The memory in Nvidia's
1:22:06H200 chip is more than just a memory
1:22:08spec. In frontier model training, it's
1:22:11the difference between a system that can
1:22:12hold a top tier language model in
1:22:14[music] active context and one that
1:22:16can't. It's also the line between a
1:22:18chatbot and a cyber warfare engine. By
1:22:21engineering chips right up to the very
1:22:23edge of the rules, Nvidia is now writing
1:22:25the working version of American export
1:22:27policy in real time. Chapter 7, the
1:22:30shadow state department. Jensen Hong's
1:22:322025 calendar reads like a study in
1:22:35parallel diplomacy. The NVIDIA CEO held
1:22:38high-profile international meetings in
1:22:39Taiwan, Japan, the United Arab Emirates,
1:22:42Saudi Arabia, and the United Kingdom. At
1:22:44each stop, he announced compute
1:22:46commitments [music] and buildout deals
1:22:47at a scale that would historically have
1:22:49been worked out by trade representatives
1:22:51or cabinet secretaries. In several
1:22:53cases, the agreements the CEO announced
1:22:56set the agenda for the nation to nation
1:22:58conversations that followed rather than
1:23:00the other way around. The Gulf Compute
1:23:02deals are perhaps the clearest example.
1:23:04Saudi Arabia's announcement of a
1:23:06sovereign AI initiative anchored by
1:23:08Nvidia hardware set the terms of the
1:23:10American foreign policy meetings that
1:23:11followed with the kingdom. Nvidia built
1:23:14the policy and the State Department
1:23:15adapted to it. The same pattern played
1:23:17out in Abu Dhabi. The G42 partnerships
1:23:20and tied chip commitments came before,
1:23:22not after the formal US rule book for AI
1:23:25cooperation with the UAE. It's not how
1:23:27foreign policy is supposed to work. The
1:23:29Department of State exists to coordinate
1:23:31how the United States approaches
1:23:33strategic technology and its transfer
1:23:35abroad. But now a company can announce
1:23:37commercial commitments first and the
1:23:39government is left reacting afterward.
1:23:41They have to adjust their policy to
1:23:43match what's already in motion and what
1:23:45emerges as an unusual kind of
1:23:47sovereignty. Nvidia is not a state in
1:23:49the traditional sense. It has no army
1:23:52and no treasury, but it does sit at the
1:23:54point where AI capability is actually
1:23:56determined through [music] the chips
1:23:57that define what systems can and cannot
1:23:59do. And it has enough political access
1:24:02to ensure its commercial choices don't
1:24:04just operate inside the policy, they
1:24:06actively shape it. And that's a
1:24:08different relationship than regulation
1:24:10alone. It starts to look less like a
1:24:12company operating under rules and more
1:24:14like a company operating alongside the
1:24:16formation of those rules. [music] not
1:24:18replacing the state, but in key areas
1:24:21setting the terms that the state then
1:24:23[music] responds to. Chapter 8, the
1:24:25reckoning. The first H200s under the new
1:24:27waiver are starting to land inside
1:24:29Chinese data centers with reported
1:24:31orders already running into the hundreds
1:24:34of thousands of units. The training runs
1:24:36will follow. The models that come out
1:24:38will end up applied to use cases
1:24:40spanning the full range from commercial
1:24:42chat bots to military targeting systems.
1:24:44The [music] compute once delivered
1:24:46becomes whatever the operator chooses to
1:24:48do with it. What makes this hard to
1:24:50unwind is that it's no longer just
1:24:52policy. [music] It's income. Once the
1:24:54revenue starts flowing into federal
1:24:55planning, it stops being a clean onoff
1:24:58switch. Federal budget projections
1:25:00included. Domestic buildout programs
1:25:02partly funded by it have constituencies
1:25:04[music] that'll defend it. The
1:25:06bureaucratic machine of the US
1:25:07government has in effect taken a
1:25:09financial stake in continued exports of
1:25:11the very hardware it once labeled a
1:25:14controlled good. Reversing that deal
1:25:16would require not just a political
1:25:17decision, but the removal of revenue
1:25:19already locked into agency planning.
1:25:21When Nvidia raises its revenue
1:25:23projections, the government's projected
1:25:25share rises with them. It only [music]
1:25:27strengthens the political case for
1:25:29expanding the waiver. The privatization
1:25:30of foreign policy stops being a metaphor
1:25:33and starts being a balance sheet
1:25:35relationship. Once a company has shown
1:25:37that exemptions to national security
1:25:38policy can be bought at a fixed price,
1:25:41the president becomes permanent. What's
1:25:43emerging here looks less like a one-off
1:25:45deal and more like a template for
1:25:46turning regulatory pressure into a
1:25:48negotiated cut of revenue. And other
1:25:51industries are already watching closely.
1:25:53Pharma, satellites, advanced [music]
1:25:54biotech, sectors where the same product
1:25:57can be both commercial and strategic.
1:25:59For decades, the rule was simple. Some
1:26:01technologies stay out of open markets.
1:26:03That rule hasn't been rewritten. It's
1:26:06just stopped being enforced in the
1:26:08[music] same way. Nvidia is just one
1:26:10player in the AI race, and no one wants
1:26:12to be the one that slows down. But an
1:26:14arms race only works if it's
1:26:16sustainable. Right now, the spending
1:26:18starts to look less like growth and more
1:26:19like pressure building in the system.
1:26:21Everyone thinks Microsoft won the AI
1:26:24war, but they may have just lost it
1:26:26overnight because Sam Alman just made a
1:26:28$50 billion move with Amazon to stab
1:26:31Microsoft in the back. And almost no one
1:26:34noticed. On the surface, it looks like
1:26:35another massive AI deal, but hidden
1:26:38inside it is a shift that sidelines
1:26:40Microsoft and removes the one clause
1:26:43that actually kept control in check. And
1:26:45that changes everything. Because this
1:26:48isn't just about building powerful AI
1:26:50anymore. It's [music] about who owns it,
1:26:52who controls it, and who's willing to
1:26:54burn billions to get there. By the time
1:26:56most people realize what just happened,
1:26:58the balance of power may already be
1:27:00gone. Chapter 1, the $50 billion
1:27:03betrayal. To the public, OpenAI looks
1:27:05like one of the biggest success stories
1:27:07of the modern era. A Silicon Valley
1:27:09startup that turned into a household
1:27:10name almost overnight. It built tools
1:27:13used by millions, [music] pushed AI
1:27:15further than anyone expected, and
1:27:16wrapped it all up in a mission to
1:27:18benefit humanity. That was the illusion.
1:27:21In reality, OpenAI is a mess that's
1:27:23bleeding cash. In 2024, its projected
1:27:26revenue was approximately $3.7 billion.
1:27:29Its losses, 5 billion. That's like
1:27:32buying a midsized airline and setting
1:27:34the whole thing on fire every 12 months
1:27:36just to keep the servers running. That's
1:27:38not a sustainable or successful business
1:27:40model. That's not a sign of a healthy,
1:27:42stable company. [music] And it wasn't a
1:27:44one-time thing. It was a trend. In the
1:27:46first half of 2025, figures revealed
1:27:48that OpenAI was losing extraordinary
1:27:50amounts of money, generating around $4.3
1:27:53billion in revenue while recording
1:27:55losses of up to $13 billion. Some
1:27:58estimates suggest the company's total
1:28:00losses could exceed $140 billion between
1:28:032024 and 2029 alone. But that is not
1:28:06that surprising. Training Frontier AI
1:28:09models isn't cheap. Neither are the
1:28:11salaries of leading researchers and
1:28:12computer scientists or the construction
1:28:15and operating costs of data [music]
1:28:16centers. OpenAI is burning through money
1:28:18at breakneck speed. Its entire business
1:28:20model is founded on the idea of
1:28:22convincing investors that someday,
1:28:24somehow, all of this loss will be worth
1:28:26it. Microsoft bought into that idea. It
1:28:29poured billions into OpenAI and secured
1:28:32what seemed like an exclusive hold over
1:28:34the most valuable AI startup on the
1:28:36planet. Microsoft CEO Satya Nadella even
1:28:39said that it wouldn't matter if OpenAI
1:28:41disappeared tomorrow. We [music] have
1:28:42the data, IP rights, and the capability.
1:28:45Nadella thought that for all intents and
1:28:47purposes, he owned OpenAI. He was wrong.
1:28:50In February 2026, Sam Alman orchestrated
1:28:53an enormous $50 billion infrastructure
1:28:56deal with Amazon. In doing so, he
1:28:58effectively ended Microsoft's exclusive
1:29:00cloud rights. That wasn't supposed to
1:29:02happen. Microsoft [music] was supposed
1:29:04to be the only serious player in the
1:29:06game. Microsoft Azure was meant to be
1:29:08the default home for OpenAI's
1:29:10technology. [music] Amazon was the
1:29:12rival, the company that you compete
1:29:14against, not partner with. This wasn't a
1:29:16new vendor agreement or just some sort
1:29:18of multi-artner strategy. It was OpenAI
1:29:21blatantly betraying [music] the tech
1:29:23giant that helped build it. The question
1:29:25is why? Why would Altman risk the wrath
1:29:28of Microsoft? Why jeopardize what seemed
1:29:30to be the most powerful relationship in
1:29:32the industry? Because the Amazon [music]
1:29:34deal wasn't just about getting more
1:29:35servers or resources. It was a weapon
1:29:38built for one mission to defeat a more
1:29:40powerful enemy, the United States
1:29:42[music] government. Chapter 2, the FTC's
1:29:46trap. For years, it looked like compute
1:29:48was going to be the biggest challenge
1:29:49Open AI would ever face. But the funding
1:29:52from Microsoft introduced OpenAI to
1:29:54something else. Antitrust. To observers
1:29:56and analysts, including those in
1:29:58government authorities, such as the
1:30:00Federal Trade Commission, this didn't
1:30:02look like one company simply supporting
1:30:04another. It looked like a merger.
1:30:06Naturally, Microsoft and OpenAI didn't
1:30:08label it that way, but the facts were
1:30:10clear to see. Microsoft had poured in
1:30:12billions and secured exclusive rights to
1:30:14its Azure ecosystem. OpenAI technology
1:30:17was also becoming increasingly
1:30:18integrated into Microsoft's mostused
1:30:20systems [music] and applications from
1:30:22Windows to Copilot, Office, GitHub, and
1:30:25beyond. OpenAI, meanwhile, was looking
1:30:27less like an independent organization
1:30:29and more like a subsidiary, just with
1:30:32its own separate branding. The FTC
1:30:34noticed, [music] so did other tech
1:30:36brands. Google even called on the
1:30:38government to investigate and break up
1:30:40the deal. Critics and regulators argued
1:30:42that Microsoft's massive investment and
1:30:44exclusive cloud control over OpenAI's
1:30:46technology gave the company too much
1:30:48influence. It hadn't just invested in an
1:30:50upand cominging company, it had bought
1:30:52the future. So, the FTC started
1:30:55investigating the two companies. If it
1:30:57could prove that they had effectively
1:30:58entered a deacto merger or that
1:31:00Microsoft had acquired an unfair
1:31:02monopoly over the AI industry, it could
1:31:04force the pair to split. Microsoft would
1:31:06be able to survive that. Open AI might
1:31:09not. It couldn't afford the risk and it
1:31:11had to find some way to wrigle out of
1:31:13its predicament. Enter the Amazon deal.
1:31:16By pivoting to Amazon Web Services or
1:31:18AWS, [music]
1:31:19OpenAI gave itself a multi-billion
1:31:21dollar legal shield. Because now, if the
1:31:24FTC's investigators question the
1:31:25company's allegiances or argue it's too
1:31:27friendly with Microsoft, OpenAI's
1:31:29lawyers can simply say, "How can we
1:31:31possibly be a subsidiary of Microsoft if
1:31:34we [music] just signed a $50 billion
1:31:36contract with one of their biggest
1:31:37rivals?" It was the perfect piece of
1:31:39legal [music] theater at just the right
1:31:41time because antitrust cases are built
1:31:44on dependency. Regulators are very wary
1:31:46of any company that appears to be
1:31:48entirely or exclusively dependent on
1:31:50another for its survival. [music]
1:31:52But by inking an agreement with another
1:31:54tech giant, OpenAI proved its
1:31:56independence. Problem solved. Or at
1:31:58least [music] that's how it seemed. In
1:32:00reality, there was much more to this
1:32:02story than meets the eye. escaping the
1:32:04FTC was only a convenient and timely
1:32:06byproduct of a deeper and darker
1:32:08imagination. The real reason OpenAI
1:32:11needed leverage over Microsoft was
1:32:12[music] hidden inside a bizarre legal
1:32:15contract signed years before. A contract
1:32:18that contained a ticking [music] time
1:32:19bomb. Chapter 3. The AGI poison pill.
1:32:23For years, Open AAI told the world that
1:32:25it is working toward AGI, artificial
1:32:28general intelligence. [music]
1:32:30Sometimes known as the God model. This
1:32:32is said to be the point at which AI
1:32:34effectively reaches and then surpasses
1:32:36human level intelligence. According to
1:32:38the experts, AGI will be able to think,
1:32:40reason, and adapt just like [music] a
1:32:41real person. It'll solve problems and
1:32:43switch from task to task rather than
1:32:45being pre-programmed with just one
1:32:47specific function or avenue of activity
1:32:49in mind. In effect, this [music] was
1:32:51OpenAI's justification for everything.
1:32:54All the funding, all the hype, all the
1:32:56resources, it was all said to be in
1:32:58service of the AGI experiment. [music]
1:33:00And for OpenAI, it was the perfect
1:33:02panacea. All they had to do was convince
1:33:05people to trust them,
1:33:06>> [music]
1:33:06>> ignore the obvious problems, hand over
1:33:08their money, and then believe that
1:33:10someday they'd build something that
1:33:12would change the world. There was just
1:33:14one little problem, one buried deep in
1:33:17the contract tying OpenAI and Microsoft.
1:33:20It was known as the AGI trigger.
1:33:22Basically, Microsoft was granted a
1:33:24seemingly perpetual license to OpenAI's
1:33:26intellectual property. But that
1:33:28perpetual license had a strict limit. As
1:33:31[music] soon as OpenAI achieved
1:33:33artificial general intelligence, the new
1:33:35AGI model would be entirely [music]
1:33:37excluded from the deal and Microsoft
1:33:39would effectively lose its grip on the
1:33:41future of AI. At that stage, all
1:33:44commercial rights to the AGI would
1:33:46remain exclusively with Open AI. It's a
1:33:49paradox, a snake eating its own tail.
1:33:51Microsoft was pouring billions into a
1:33:53company whose sole stated mission was
1:33:55build AGI. But as soon as that mission
1:33:58was achieved, Microsoft would lose all
1:34:00of its power and benefits. It was
1:34:02funding its own demise. But Microsoft's
1:34:04executives aren't
1:34:05>> [music]
1:34:05>> idiots. They knew the terms of the deal
1:34:07when they signed it. They knew exactly
1:34:09how to work around them. All they had to
1:34:11do was ensure that OpenAI failed at its
1:34:13stated mission. They wanted the AI to be
1:34:16powerful, [music] but never powerful
1:34:18enough to reach AGI. That's where things
1:34:20get complicated. Because AGI isn't a
1:34:23clear finish line. No one can agree what
1:34:25it actually means. So even if OpenAI
1:34:28pushed its systems further and further,
1:34:30Microsoft could always argue it still
1:34:32wasn't AGI. Sam Alman knew this. He knew
1:34:35that as long as the AGI trigger clause
1:34:37existed, Microsoft would never truly be
1:34:39an all-in partner. It would always have
1:34:42leverage, always have limits, always
1:34:44have a way to control OpenAI from the
1:34:46inside. So he had to change the play.
1:34:49With a $50 billion Amazon deal as its
1:34:51loaded gun, OpenAI forced Microsoft back
1:34:54to the table, not just to negotiate, but
1:34:56to completely rewrite the rules of the
1:34:58AI war. Chapter 4. The April 2026 reset.
1:35:02In April 2026, the balance of power
1:35:05shifted. It wasn't a simple partnership
1:35:07update. OpenAI didn't want to iron out
1:35:09just a few issues or make a couple of
1:35:11amendments to its Microsoft deal. It
1:35:13wanted to demolish it and then rebuild
1:35:16it exactly as it saw fit. On the
1:35:18surface, the two companies saved face,
1:35:20announcing a simplified agreement and
1:35:22next phase for their partnership. But
1:35:24the terms of that agreement painted the
1:35:26real picture. Microsoft was still
1:35:28described as OpenAI's primary cloud
1:35:30partner, but the very next sentence
1:35:33added that OpenAI was now free to serve
1:35:35products across any other cloud provider
1:35:37it wanted. Azure's exclusivity was gone.
1:35:41Microsoft's once perpetual license to
1:35:43OpenAI's intellectual property was also
1:35:45amended and given a fixed end date of
1:35:472032. The tech giant no longer had
1:35:50privileged ownership of the future of
1:35:52AI. Revenue sharing was officially given
1:35:54a cap and a deadline of 2030 as well.
1:35:58Most importantly, the AGI trigger was
1:36:00gone. It wasn't redefined, clarified, or
1:36:03amended. It was deleted. That vague,
1:36:05hard to define clause that hung over
1:36:07open AI like a sword of damicles for
1:36:09years was gone. It was replaced with
1:36:11something far simpler and far more
1:36:14controlled, a calendar with dates,
1:36:16deadlines, and caps. In other [music]
1:36:18words, a standard corporate agreement.
1:36:20Microsoft stake was also formalized at
1:36:22approximately 27% of the company. That
1:36:24is still a sizable amount, enough for
1:36:26Microsoft to hold some level of
1:36:28influence over the company's activities,
1:36:30but nowhere near enough for complete
1:36:32control. It looked like a big victory
1:36:34for OpenAI. The company had won its
1:36:37independence, decoupling its finances
1:36:39from the mythical AGI milestone. It was
1:36:41no longer a research lab trying to
1:36:43trigger a clause in a contract to win
1:36:45its freedom, but a corporation with a
1:36:47clear runway ahead. But there was a
1:36:50catch, a big one. To pull off this
1:36:52extraordinary corporate coup, Altman had
1:36:55to permanently destroy the very
1:36:56foundation that OpenAI was built upon,
1:36:59its nonprofit structure. Chapter 5. The
1:37:02death of the nonprofit. OpenAI's
1:37:05nonprofit nature was the one thing that
1:37:07separated it from every other Silicon
1:37:09Valley machine. This wasn't just another
1:37:11power to the benefit of humanity. A
1:37:14nonprofit lab working with care and
1:37:16consideration towards something that was
1:37:17supposed to bring great benefits to all.
1:37:20This wasn't Google. It wasn't Meta. It
1:37:22wasn't worried about pleasing
1:37:23shareholders because there were no
1:37:26shareholders. But that idealistic
1:37:28attitude couldn't last. Slowly and
1:37:30surely, the cracks in the mask began to
1:37:32appear. In late 2025, the facade was
1:37:35ripped away entirely. Open AAI shifted
1:37:37from a nonprofit to a public benefit
1:37:39corporation. At a glance, that still
1:37:41sounds like a righteous cause, a
1:37:43compromise between the original mission,
1:37:45and a need to make money. PBC's are
1:37:47supposed to strike a balance between
1:37:49pursuing profit while remaining
1:37:50committed to creating a positive impact
1:37:52on society, the community, or the
1:37:54environment. In reality, PBC still serve
1:37:57their investors almost as much as any
1:37:59other for-profit entity. Just as Altman
1:38:02would go on to smash and then rebuild
1:38:04his deal with Microsoft, he also
1:38:05destroyed what OpenAI once was,
1:38:07reconstructing it as something
1:38:09completely different. The groundwork was
1:38:11laid back in 2023. The organization's
1:38:13original nonprofit board, the one that
1:38:15briefly fired Alman, was removed and
1:38:18replaced by Silicon Valley insiders and
1:38:20former Treasury officials. The safety
1:38:22guardrails that had once been so
1:38:24critical to the organization's overall
1:38:26mission were dismantled. The systems
1:38:28that had kept OpenAI's progress in line
1:38:30with its focus on helping humanity were
1:38:32gone. In their place were product safety
1:38:35teams more concerned with ensuring that
1:38:37their AI doesn't say anything that might
1:38:39offend a big B2B client than actually
1:38:42harm real people. The cogs inside the
1:38:44OpenAI machine were replaced piece by
1:38:47piece until something fundamentally
1:38:48changed. What began as a research-driven
1:38:51system started to look like something
1:38:52else entirely, a profit engine, one that
1:38:55was preparing for a massive IPO and was
1:38:59increasingly insulated from any
1:39:00meaningful ethical oversight. Behind
1:39:02closed doors, Alman and OpenAI's
1:39:04research leads realized a terrifying
1:39:06technical truth. They weren't moving
1:39:08toward AGI as they originally expected.
1:39:11They were moving towards a brick wall.
1:39:14Chapter 6. The scaling wall. For years,
1:39:17the AI industry has relied on an
1:39:19unwavering belief in a premise known as
1:39:22scaling [music] loss. If you add more
1:39:24data and more compute, your AI models
1:39:26will become exponentially smarter. It's
1:39:29all a question of resources. Provide
1:39:30more resources and you get a better
1:39:32product. All companies like OpenAI had
1:39:34to do was keep on building bigger and
1:39:36better data centers. They had to invest
1:39:38in more powerful chips and processors.
1:39:41Then they could sit back and watch as
1:39:43their AI followed the linear path to
1:39:45godlike intelligence. Sounds pretty
1:39:47straightforward and [music] the scaling
1:39:49laws worked for a while. Each new
1:39:51generation of AI technology felt like a
1:39:53big leap forward. GPT2 was impressive.
1:39:56GPT3 next level. GPT4 exceeded
1:40:00expectations. So naturally GPT5 cenamed
1:40:03Orion was expected to be a gamecher,
1:40:06maybe even the final step toward AGI or
1:40:09not. Internal reports suggest that the
1:40:11scaling laws are stalling. Instead of
1:40:13providing some sort of quantum
1:40:14intelligence leap, GPT5 has hit
1:40:17diminishing returns. It's still getting
1:40:19smaller, but at a slower rate than ever
1:40:21before. All of a sudden, this [music]
1:40:23god model that seemed right around the
1:40:25corner is now a speck on the horizon.
1:40:27This isn't just a hurdle, it's a
1:40:29catastrophe. The scaling wall changes
1:40:31everything. It's no longer a situation
1:40:33where companies can just pour in money
1:40:35and watch their AI become twice as
1:40:37intelligent overnight. Now they're
1:40:39spending billions for only incremental
1:40:41improvements. And that is bad business
1:40:44because let's not forget about the burn
1:40:46rate. Open AAI is nowhere close to
1:40:48making a profit. It loses billions each
1:40:51year, but it was always able to justify
1:40:53that with the claim that AGI would
1:40:54eventually arrive and fix everything.
1:40:56The [music] company's entire financial
1:40:58structure and investment incentives were
1:41:00reliant on that premise. That's why
1:41:02removing the AGI trigger clause from the
1:41:04Microsoft contract mattered [music] so
1:41:05much. It wasn't just a legal trick. It
1:41:08was a quiet admission that AGI is a
1:41:10mirage. And if AGI is a mirage, OpenAI
1:41:13is just another software company and one
1:41:16that is failing to provide returns and
1:41:18plateauing [music] fast. So, how does a
1:41:20company like that justify a $5 billion
1:41:22burn rate to its next investors? [music]
1:41:24It stops selling AGI and starts selling
1:41:27AI slop instead. Chapter 7, the SAS
1:41:30pivot. Selling AI slop. Meet the new
1:41:33Open AI. It's no longer a valiant
1:41:36nonprofit pursuing civilization changing
1:41:38super intelligence, but a salesforce for
1:41:40AI business. It's slowly but surely
1:41:43pivoting away from its original mission
1:41:45and towards something much more mundane,
1:41:47[music] enterprise software and
1:41:48corporate workflow automation. And it's
1:41:50happening right before our eyes. Rather
1:41:52than focusing its efforts exclusively on
1:41:54the next evolution of GPT technology,
1:41:57OpenAI is prioritizing alternative
1:41:59projects like agentic middleware and
1:42:01reasoning models like 01 or Strawberry.
1:42:04It's no longer charting a course toward
1:42:06human enlightenment, [music] but making
1:42:08life easier for middle management. The
1:42:10focus has shifted to producing tools
1:42:12built for middle management, routing
1:42:14leads, generating marketing copy,
1:42:16handling [music] support tickets. OpenAI
1:42:18hopes that this shift will bring in the
1:42:20money it needs to satisfy its investors.
1:42:22But there is a massive problem with that
1:42:24plan. The numbers don't add up.
1:42:27Traditional SAS or software as a service
1:42:29companies like Salesforce and Adobe
1:42:31operate on incredible margins, [music]
1:42:33often exceeding 70%. They make a product
1:42:36and they basically sell it forever,
1:42:38bringing in more and more profit with
1:42:40every new customer. That's [music] why
1:42:42investors love SAS. It is a gold mine.
1:42:45But OpenAI's attempts to enter this
1:42:47industry are not working because
1:42:49advanced reasoning models cost so much
1:42:51more to run than conventional software.
1:42:53O1, for example, costs around $15 for 1
1:42:56million input tokens. That might not
1:42:59sound like much at first glance, but in
1:43:01enterprise terms, it is a massive
1:43:03financial burden. [music] Big businesses
1:43:04with hundreds or even thousands of
1:43:06employees can chew through millions upon
1:43:08millions of tokens in a single day.
1:43:10Suddenly, a smart assistant is not a
1:43:12cost-effective component of the tech
1:43:14stack, but a very expensive capital
1:43:16drain. Traditional SAS doesn't work this
1:43:19way. Microsoft doesn't bill businesses
1:43:21every time they open a new spreadsheet
1:43:23on Excel. CRM don't suddenly become
1:43:25twice as expensive just because
1:43:27employees clicked a few buttons and
1:43:29generated some reports. AI works
1:43:31differently. The more you use it, the
1:43:33more expensive it gets. OpenAI is
1:43:35desperately trying to force businesses
1:43:36to integrate overpriced automated AI
1:43:39slop software into their corporate
1:43:41workflows to fix its own broken
1:43:43economics. It is trying to sell digital
1:43:45gold for the price of lead, hoping to
1:43:47convince people its money guzzling AI
1:43:49agent isn't just [music] really
1:43:51expensive software. But to make its
1:43:53margins work, it needs to dramatically
1:43:55decrease its own operating costs. It
1:43:58needs cheaper compute, which brings us
1:44:00back to the $50 billion Amazon Trojan
1:44:02horse. Chapter 8. Amazon's Trojan horse.
1:44:06The deal with Amazon wasn't about
1:44:08escaping the FTC's investigations or
1:44:10breaking free of Microsoft's shackles.
1:44:12It was about hardware. Part of the $50
1:44:15billion commitment that Amazon made to
1:44:17OpenAI includes a multi-year agreement
1:44:19for the AI firm to use AWS's Tranium
1:44:22chips. Prior to this, OpenAI was a
1:44:25hostage to Nvidia. It was forced to use
1:44:27Nvidia's H100 GPUs. It's the hardware
1:44:30that everyone in the AI industry wants
1:44:32and needs. The chips that form the
1:44:34beating hearts of AI data centers. These
1:44:37GPUs have proven highly effective in
1:44:39training and improving AI. They are also
1:44:42expensive. Each one can cost tens of
1:44:44thousands of dollars, and that's before
1:44:46the added expense of building the rest
1:44:47of the server around it and actually
1:44:49running the whole thing. Large training
1:44:51clusters come with billion-dollar price
1:44:53tags. OpenAI has paid an H100 tax on
1:44:56every single prompt it processes for
1:44:58years with incalculable amounts of money
1:45:01funneled away into the accounts of
1:45:03Nvidia and Microsoft. The Amazon deal
1:45:05gives OpenAI an off-ramp. Tranium chips
1:45:08aren't necessarily better than Nvidia's
1:45:10H100s, but they do have the potential to
1:45:12be much, much cheaper, up to 50% cheaper
1:45:15according to early estimates. They're
1:45:17also said to consume 40% less energy. By
1:45:20switching to Amazon's own custom
1:45:22silicon, OpenAI hopes it'll be able to
1:45:24make some significant reductions to the
1:45:26cost of its tokens. Cheaper tokens
1:45:28should make OpenAI's AI slop easier to
1:45:30digest for its big business customers.
1:45:33They might even give the company a slim
1:45:34chance of turning a profit before the
1:45:362030 revenue cap hits. In the name of
1:45:39saving humanity and building a tech
1:45:41utopia, OpenAI took a very different
1:45:43path. building software on proprietary
1:45:45Amazon chips running inside Amazon data
1:45:48centers selling AI agents to Fortune 500
1:45:51companies. This is not what the
1:45:53company's founders envisioned all those
1:45:55years ago. This is not a beacon of
1:45:57open-source enlightenment. It's just a
1:45:59cog in the AWS machine. So, where do we
1:46:02go from here? Chapter 9, the great AI
1:46:05realignment. The hardware war is over.
1:46:07Unfortunately, humanity didn't win.
1:46:10Instead, the victors are the corporate
1:46:11behemoths that own the silicon. These
1:46:13companies with the money and power to do
1:46:15whatever they want and always get away
1:46:17with it. Open AI sold the world a dream.
1:46:20A dream of godlike AI that would cure
1:46:22cancer, solve the climate crisis, and
1:46:24bring about a new world where everyone
1:46:26would be happier, freer, and more
1:46:28fulfilled. That utopian dream is dead,
1:46:31replaced by a dystopian corporate
1:46:32reality. Microsoft, Amazon, and Open AI
1:46:35aren't laying the foundations for a more
1:46:37prosperous and creative age of human
1:46:39advancement. They're building their own
1:46:41locked down and ludicrously expensive
1:46:43B2B monopoly. The open marriage that now
1:46:46exists between those tech giants all but
1:46:48guarantees that the future of the
1:46:50internet will be flooded with corporate
1:46:52AI slop, automated emails, synthetic
1:46:55reports, and agentic workflows that
1:46:57don't actually provide real benefits to
1:46:59real people. They just streamline the
1:47:01capitalist machine while eroding the
1:47:02value of human thought and creativity.
1:47:04Sam Alman didn't escape from Microsoft
1:47:06to bring about a better world. He broke
1:47:09free. so that when the trillion dollar
1:47:10IPO arrives, he and his shareholders
1:47:13will make as much money as possible.
1:47:15This is the grim reality of AI today.
1:47:17We're not getting AGI. We're not going
1:47:19to see some digital god that solves the
1:47:21world's ills and makes us all happier
1:47:23and healthier. We're getting an
1:47:25inescapable automated corporate
1:47:27bureaucracy instead. A $3 trillion
1:47:29market cap, $32.9 billion in cloud
1:47:32revenue. But could Microsoft's empire
1:47:34collapse because [music] of one startup?
1:47:37Nearly half of Microsoft's future cloud
1:47:39empire depends on a single startup. One
1:47:41that is burning $12 billion every
1:47:44quarter. I'm Josh and on today's episode
1:47:46of the infographic show, we'll reveal
1:47:48the massive Microsoft divorce that could
1:47:50bankrupt Open AI and ChatGpt forever.
1:47:54Microsoft doesn't just invest in
1:47:56startups. It captures them. They hand
1:47:58founders up to $150,000 in free Azure
1:48:01[music] cloud credits, not cash, digital
1:48:03vouchers. These small companies spend
1:48:06months building their products on Azure,
1:48:08mapping every database and workflow to
1:48:10Microsoft's proprietary formats. By the
1:48:12time the free credits run out, they're
1:48:14stuck. Tear out the backend and their
1:48:16apps crash. So, they start paying real
1:48:19money, and [music] now they're stuck in
1:48:21the architecture. They turn to corporate
1:48:23credit cards, pay as you go tiers, and
1:48:25just like that, [music] Microsoft turns
1:48:26free credits into real cash flowing
1:48:29straight into its books. But it's not
1:48:31just small startups that get caught up
1:48:32in Microsoft's digital web. Microsoft
1:48:35plowed $13.8 billion in direct funding
1:48:37into OpenAI, but almost none of that
1:48:40money actually left Microsoft's coffers.
1:48:41[music]
1:48:42Instead, the company handed Sam Alman
1:48:44customized digital vouchers. OpenAI then
1:48:47used those vouchers to rent Microsoft
1:48:49servers. Every dollar spent legally
1:48:51counted as Azure revenue growth on
1:48:53Microsoft's books. Open AAI was backed
1:48:56into a corner. What does this mean for
1:48:58Microsoft's balance sheet? Corporate
1:49:00accountants have a secret weapon, a
1:49:02metric called the remaining [music]
1:49:03performance obligation. It tracks
1:49:05guaranteed future revenue, and Microsoft
1:49:07is currently sitting at a staggering
1:49:09$625 billion. Wall Street treats [music]
1:49:12that number as cash in the bank.
1:49:14Analysts feed it into discounted cash
1:49:16flow models, using it to justify
1:49:18Microsoft's share price all the way to
1:49:20the [music] end of the 2020s. 45% of
1:49:23Microsoft's guaranteed 625 billion is
1:49:26locked in, fueling OpenAI's machines.
1:49:28The future of its cloud empire hinges on
1:49:30one startup, and Wall [music] Street
1:49:32expects it'll pay. Microsoft's balance
1:49:34sheet shows $40.3 billion in debt. Wall
1:49:37Street accepts that, [music] but off the
1:49:39books, there's something hidden. A $662
1:49:42billion trap. Shadow leases and custom
1:49:45[music] deals keep OpenAI servers
1:49:47running. And Microsoft isn't alone. In
1:49:49the cloud world, physical hardware hides
1:49:51behind complex lease structures. And
1:49:53that [music] is the problem. Open AAI
1:49:55doesn't have the cash to cover this
1:49:57hidden debt. Financial analysts [music]
1:49:58at Deutsche Bank crunched the numbers.
1:50:00They projected OpenAI will burn through
1:50:02$143 billion before [music] ever turning
1:50:05a real profit. A company setting
1:50:07billions of dollars on fire every 12
1:50:09months just handed its largest [music]
1:50:11investor the biggest profit spike in
1:50:14recent corporate history. Open AAI is an
1:50:16unemployed tenant facing eviction.
1:50:19Microsoft is [music] the landlord
1:50:21holding the keys. Microsoft prints fake
1:50:23IUS and hands them to OpenAI. Those IUs
1:50:26pay for renting the servers. Microsoft
1:50:28legally reports that rent to Wall Street
1:50:30[music] as cloud revenue growth. It is a
1:50:32flawless infinite money loop until the
1:50:35servers actually turn on. Why can't
1:50:37OpenAI just build their own
1:50:38infrastructure? Well, the math doesn't
1:50:40add up. Sam Alman saw the problem.
1:50:42[music] Azure's credits could never fuel
1:50:44the endless compute he needed. So, he
1:50:46engineered an escape route called
1:50:48Project Stargate. He pitched a $500
1:50:50billion master plan. He wanted
1:50:52independent data centers, 10 gawatts of
1:50:55dedicated power. He flirted with
1:50:57sovereign wealth funds and foreign
1:50:58telecom giants. He planned to bypass the
1:51:00Azure ecosystem entirely [music]
1:51:02to sever the partnership. Soft Bank and
1:51:04Oracle entered negotiations to provide
1:51:06alternative capital and infrastructure.
1:51:08Construction crews mobilized in Adelene,
1:51:11Texas. OpenAI prepared to build a 1.2
1:51:13gawatt facility. one site serving as the
1:51:16beach head for a sprawling $665 billion
1:51:20infrastructure rollout through 2030. All
1:51:23they needed was the financing. The banks
1:51:25opened the disclosures, ran the numbers
1:51:27on the deficit, logged delays on
1:51:29permits, and tallied the engineer
1:51:31shortage to cool the massive racks. Wall
1:51:33Street refused the $500 billion gamble.
1:51:36Private investors wouldn't touch it.
1:51:38Open AAI quietly scrapped their master
1:51:40plan. They slashed the projected
1:51:42independent comput speed. [music] They
1:51:44retreated to the existing
1:51:45infrastructure. They lack the capital to
1:51:47build their own fortresses and the
1:51:49margins to keep renting Microsoft
1:51:50servers. Azure's credit can't sustain
1:51:53the [music] burn. Microsoft is left
1:51:54fueling a captive entity that cannot
1:51:56repay the principle. How does the
1:51:58physical hardware accelerate this
1:52:00crisis? Microsoft spreads its massive
1:52:02server costs over a six-year accounting
1:52:05window. This keeps their quarterly
1:52:06spending low on paper, but AI doesn't
1:52:09wait. Frontier training models make top
1:52:11tier GPUs obsolete in just 36 months.
1:52:14Every chip you buy today is tomorrow's
1:52:16legacy hardware. Imagine a delivery
1:52:19company buying a brand new fleet of
1:52:20trucks. They have to replace that entire
1:52:22fleet every 18 months because the old
1:52:24trucks suddenly cannot deliver packages
1:52:26fast enough. That's the economic reality
1:52:29of artificial intelligence hardware.
1:52:31Financial models from analysts expose
1:52:33$176 billion in hidden GPU deprecation
1:52:36actively [music] decaying across the
1:52:38tech sector. Microsoft is booking record
1:52:40profits today by ignoring the physical
1:52:43decay of its own hardware. When the
1:52:45actual replacement cycle hits the
1:52:47balance sheet, the capital expenditure
1:52:48bill will explode. The models burn cash
1:52:51at a high velocity. Open AAI generates
1:52:54$12 billion in quarterly losses. Forbes
1:52:57estimates [music] that the Sora video
1:52:58generation model alone consumes $15
1:53:01million in hard cash every single day.
1:53:04Don't forget to like, share, and
1:53:06subscribe. [music] The AI takeover isn't
1:53:08coming. It's already here and we'll try
1:53:10to keep revealing the true story. Video
1:53:12generation isn't [music] just text on
1:53:14steroids. Every single pixel must be
1:53:16calculated and rendered in sequence. It
1:53:18[music] demands an exponential jump in
1:53:20raw computational power. Microsoft's
1:53:22capital expenditures surged 66% to 37.5
1:53:25[music]
1:53:26billion in a single quarter to feed
1:53:28this. The tech giant is purchasing land
1:53:31and pouring concrete to meet a
1:53:33theoretical demand that OpenAI literally
1:53:35can't afford to use.
1:53:36>> [music]
1:53:36>> Taiwan's semiconductor manufacturing
1:53:38company, TSMC, operates as the physical
1:53:41break on global artificial intelligence.
1:53:43They produce 90% of advanced silicon on
1:53:46Earth. Corporate demand outpaces their
1:53:48physical factory capacity by a factor of
1:53:50three. You can't speed up the extreme
1:53:53ultraviolet lithography processes.
1:53:55[music]
1:53:55You can't skip the chemical etching.
1:53:57Each chip must be born inside hyper
1:54:00specialized [music] clean rooms with
1:54:01perfect vacuums and extreme atmospheric
1:54:04control. Tech giants are sitting on
1:54:06billions in cash, unable to spend it,
1:54:08[music] while their current servers lose
1:54:10value every single day. TSMC is racing
1:54:12to expand. They're planning a 52 to 56
1:54:15billion expansion in 2026. But even that
1:54:18can't catch up. And it gets worse. A
1:54:21single gawatt data center requires
1:54:23thousands of miles of thick copper
1:54:25wiring. Copper is running out. Mines in
1:54:28South America are struggling to meet
1:54:29demand. The cables, the power, the
1:54:32cooling, they all have to be perfect.
1:54:34One slip, one missing component and the
1:54:36whole operation [music] stalls. It's a
1:54:38billiondoll waiting game. And that's not
1:54:40the only problem. Across the industry,
1:54:42data centers swallow 449 million gallons
1:54:45of water daily. Hypers scale facilities
1:54:48drain up to 5 million gallons of potable
1:54:50water every 24 hours just to [music]
1:54:52stop the server racks from literally
1:54:54melting. Standard air cooling maxes out
1:54:56entirely at modern rack densities. The
1:54:59facilities need to pipe in cold water
1:55:01directly to [music] the silicon chips to
1:55:03maintain operational temperatures. Local
1:55:05governments are starting to panic.
1:55:07Municipal water supplies are dropping
1:55:09while server farms keep expanding. In
1:55:11Florida, regulators stepped in with
1:55:13strict new rules to stop residents
1:55:15utility bills from spiking. The
1:55:16legislation targets the massive energy
1:55:18and water demands of the new data center
1:55:20construction. The Midwest faces water
1:55:23stress. Local city councils are passing
1:55:25emergency moratoriums on new facility
1:55:27permits. They are choosing drinking
1:55:29water for their citizens over artificial
1:55:32intelligence infrastructure. The
1:55:33hyperscalers are being locked out of
1:55:35prime real estate [music] because the
1:55:37local aquifer cannot support the thermal
1:55:39load. The hyperscalers are scrambling.
1:55:41They're abandoning traditional air
1:55:43cooling and moving to direct to chip
1:55:45liquid systems. That means ripping out
1:55:47entire air cooling units and bolting in
1:55:49metal cold plates straight on to the
1:55:51hottest silicon chips. [music] Engineers
1:55:53need to thread miles of pressurized
1:55:55coolant pipes directly over racks
1:55:57holding billions of dollars of active
1:55:59hardware. The sheer material cost of
1:56:01this plumbing destroys the baseline
1:56:03construction budgets. If a single leak
1:56:05in a coolant line [music] drips onto a
1:56:07motherboard, it destroys millions of
1:56:09dollars in silicon instantly. The
1:56:11capital required effectively doubles the
1:56:14initial build cost. Microsoft is footing
1:56:16this bill entirely upfront. What happens
1:56:18when the hardware reaches its absolute
1:56:20[music]
1:56:20limit? Generative AI only works as a
1:56:23business if the margins are massive.
1:56:25Those fat profits are supposed to pay
1:56:27for the mountains of steel, silicon,
1:56:29water, and electricity working away
1:56:31behind the curtain. Open AAI
1:56:32historically charged premium [music]
1:56:34prices for application programming
1:56:36interface or API access. They utilized
1:56:38their monopoly position to drain
1:56:40enterprise budgets. The API market is
1:56:42turning into a commodity battlefield.
1:56:44Open AAI launched the GPT 5.2 2 Frontier
1:56:47model and priced it at $1.75 per million
1:56:51input tokens. They guessed Fortune 500
1:56:53companies would just absorb the cost to
1:56:55maintain access. They assumed the
1:56:57dominance would hold. They didn't expect
1:57:00what came [music] next. Chinese
1:57:01competitors arrived. They didn't try to
1:57:03outspend OpenAI. They didn't build
1:57:05[music] trillion dollar server empires
1:57:07from scratch. They used model
1:57:09distillation. Instead of training
1:57:10artificial intelligence from scratch,
1:57:12they bought API access to OpenAI's best
1:57:15model. They asked [music] millions of
1:57:16advanced math and coding questions. They
1:57:18captured the answers. And then they
1:57:20trained smaller, leaner architectures on
1:57:23those outputs, all without paying
1:57:25[music] for the training costs. Deepseek
1:57:27released their V3.2 architecture and
1:57:29matched the performance of the American
1:57:31Frontier models. They dropped their
1:57:33exact same token package to 28.
1:57:36Overnight, the premium collapsed. A
1:57:38global price war erupted. Gross margins
1:57:41evaporated. Developers are actively
1:57:43routing their daily tasks to cheaper
1:57:45alternatives. They use OpenAI strictly
1:57:47for the most complex reasoning tasks.
1:57:48[music]
1:57:49They funnel 90% of their standard
1:57:51workloads to foreign models or localized
1:57:53open-source alternatives. This strips
1:57:55away the high margin volume OpenAI
1:57:57desperately needs to survive. How does
1:57:59[music] Microsoft respond to this
1:58:01revenue collapse? Microsoft is watching
1:58:03on as this collapse unfolds. They know
1:58:06OpenAI can't generate enough revenue to
1:58:08pay off their hidden debt. [music]
1:58:09The answer is to extract the value
1:58:12themselves. They integrated OpenAI's
1:58:14tech into their own products. Copilot
1:58:16becomes a core part of Word, Excel, and
1:58:19Teams. They charge a [music] flat $30 a
1:58:21month per enterprise user. 15 million
1:58:23people already subscribe. Wall Street
1:58:26assumes every 30 bucks is pure profit.
1:58:28Standard software is [music] like a
1:58:30printing press. Build it once,
1:58:32distribute it forever, and profit
1:58:34margins soar toward 90%. Generative AI
1:58:37obliterates that model. Every time
1:58:39someone clicks the co-pilot button, a
1:58:42supercomputer fires up and [music]
1:58:43devours energy. That $30 flat monthly
1:58:46fee doesn't even cover the basic
1:58:47electrical cost. Microsoft [music]
1:58:49eats the cost to keep the customer
1:58:51locked into the ecosystem. They're
1:58:53actively subsidizing the enterprise
1:58:55workflows of the largest corporations on
1:58:57Earth. Year-over-year cloud growth
1:58:59slowed to 39% in the second quarter of
1:59:012025. That's below the 40% growth Wall
1:59:05Street demands to justify the $3
1:59:07trillion valuation. Profit margins are
1:59:09under pressure, sliding from a strong
1:59:1246.7%
1:59:13to unsustainable levels. Microsoft is
1:59:16utilizing the most expensive [music]
1:59:17computational infrastructure in human
1:59:19history to draft basic corporate emails.
1:59:22If they raise the price of C-Pilot,
1:59:24clients will [music] turn to cheaper
1:59:25alternatives. If they keep the price at
1:59:27$30, the power users consume the server
1:59:29capacity. If they restrict the Azure
1:59:31capacity for co-pilot, [music]
1:59:32the software experience degrades
1:59:34instantly. Where does the breaking point
1:59:37occur? Standard venture capital can't
1:59:39touch this deficit. Silicon Valley
1:59:41doesn't have the liquidity to cover a
1:59:43hole this massive. They need a single
1:59:45entity capable of writing a $50 billion
1:59:48check in one afternoon. In January 2026,
1:59:51Sam Alman flew to the United Arab
1:59:53Emirates. He pitched [music] an $830
1:59:55billion corporate valuation to sovereign
1:59:57wealth funds in Abu Dhabi and requested
1:59:59$50 billion in hard cash. It was a Hail
2:00:02Mary to keep the existing Azure [music]
2:00:04servers running and bypass the domestic
2:00:06banking system entirely. The Committee
2:00:08on Foreign Investment in the United
2:00:09States watches every move. [music] The
2:00:12Pentagon classifies Frontier AI as
2:00:14critical national security
2:00:15infrastructure. It's seen as a weapon
2:00:17system. Middle Eastern funds are blocked
2:00:20instantly. The government previously
2:00:22forced a Saudi fund to completely divest
2:00:24[music] and exit an altmanbacked
2:00:26artificial intelligence chip startup.
2:00:28OpenAI is being starved of domestic
2:00:30liquidity and barred [music] from
2:00:32accepting foreign sovereign bailouts.
2:00:34Every lifeline has been cut. The fallout
2:00:37in the bond market [music] was swift.
2:00:38Microsoft officially carries $100
2:00:40billion in debt. Investors recalculated
2:00:43the risk premium. Azure's growth slowed
2:00:45and 45% of future revenue is locked into
2:00:48a single unprofitable [music]
2:00:49tenant. They watched 122.7 billion
2:00:52vanish in shareholder payouts while the
2:00:55data centers burned cash at record
2:00:57rates. Treasury yields climbed to 4.08%.
2:01:00[music] Suddenly, cheap capital
2:01:02disappeared. Every new facility had to
2:01:04prove immediate profitability. Something
2:01:06open AI couldn't guarantee. And just
2:01:08like that, expansion [music] froze. The
2:01:11data halls began to implode. The sheer
2:01:13scale of the operational bleed has
2:01:15caused an internal panic. Compute access
2:01:17has been throttled to survive the cash
2:01:19crunch. Chat GPT responses lag for
2:01:21everyday users. [music] The revenue
2:01:23curve has flatlined. They were cornered.
2:01:26So in late February, OpenAI did the
2:01:28unthinkable. They betrayed Microsoft. In
2:01:31a desperate bid to keep the lights on,
2:01:33Sam Alman secured a $ 110 billion
2:01:36bailout led by Amazon, Nvidia, and Soft
2:01:39Bank. But this isn't a victory for
2:01:41Microsoft. It's a hostage situation. To
2:01:43get Amazon's $50 billion, OpenAI had to
2:01:46agree to plow huge amounts of money into
2:01:48Amazon Web [music] Services. They are
2:01:50cannibalizing Azure. The flawless
2:01:52infinite money loop is officially
2:01:54broken. Microsoft is left holding the
2:01:56bag on billions in decaying hardware.
2:01:58While their unemployed tenant packs up
2:02:00and moves across the street, Microsoft
2:02:02fed OpenAI billions in fake digital
2:02:04[music] credits. OpenAI returned the
2:02:06favor by walking away, leaving Microsoft
2:02:09with billions in real physical
2:02:11liabilities.
2:02:12>> [music]
2:02:12>> The accounting tricks were merely smoke
2:02:14and mirrors. Open AAI betraying
2:02:16Microsoft and pulling them down is just
2:02:18the first domino. Wall Street is looking
2:02:20at software, but the companies building
2:02:22the physical AI hardware are hiding a
2:02:24completely different, much deadlier
2:02:26financial secret. The cracks in [music]
2:02:27that foundation are already tearing open
2:02:29right here. The divorce might not be
2:02:31finalized, but they are spending time
2:02:33apart. [music] Now Microsoft is left
2:02:35thinking about what could have been.
2:02:37You're already being replaced. You just
2:02:39haven't noticed again. Artificial
2:02:41intelligence was sold as a harmless
2:02:43assistant, a tool to make life easier,
2:02:45work faster, and smarter. But they are
2:02:47quietly learning how to do your job.
2:02:49While you were saving time, AI was
2:02:51learning to replace you. Tech bosses
2:02:53want to usher in a so-called
2:02:54intelligence age where machines will
2:02:56take on all the hard, boring, dangerous
2:02:58work. Humans, meanwhile, will be free to
2:03:00chase exciting new opportunities. But
2:03:03beneath that shiny promise lies a darker
2:03:05[music] truth. Behind the scenes, your
2:03:07skills are being erased, your experience
2:03:09worthless, your future already written,
2:03:12and it [music] doesn't include you. This
2:03:14is the open AI lie. And if it succeeds,
2:03:17it won't just change work, it will break
2:03:19society. Chapter one, the death of the
2:03:21resume. The working world has changed,
2:03:24but one thing has always stayed
2:03:25constant. The resume. Every [music]
2:03:27line, every achievement, every late
2:03:29night and early morning is proof that
2:03:31you fought to get where you are. It's
2:03:33more than a piece of paper. It's your
2:03:34[music] story. You rely on it to open
2:03:36doors, to earn respect, to claim the
2:03:38opportunities you've earned, and now
2:03:40it's under attack. Sam Alman and the
2:03:42other Silicon Valley CEOs don't care
2:03:44about your skills that you've spent
2:03:46years mastering. They'll be meaningless
2:03:48once AI agents have already learned
2:03:50them. The subjects you've paid thousands
2:03:52to study in college won't impress anyone
2:03:54when large language models can recite
2:03:56every detail on demand. The hours that
2:03:58you've poured into climbing the
2:03:59corporate ladder mean nothing against an
2:04:01all powerful, all- knowing AI that
2:04:03operates around the clock. It can do
2:04:05anything you can do faster, cheaper, and
2:04:07more efficiently. This is the future
2:04:09that OpenAI and other companies are
2:04:11racing toward building. Except [music]
2:04:13they don't see it that way. To them, the
2:04:15rise of artificial general intelligence
2:04:17or AGI is something to celebrate. This
2:04:19next level AI won't just assist. It'll
2:04:22think like a human. It'll move
2:04:23seamlessly from one industry to another,
2:04:26solving problems and in OpenAI's own
2:04:28words, outperforming humans at most
2:04:30economically valuable work. In other
2:04:32words, anything people can do. AGI will
2:04:34be able to do better. And this isn't
2:04:36some far-off science fiction dream. It's
2:04:39a countdown that began in 2024. OpenAI
2:04:42expects to achieve its aims by 2030. And
2:04:44Sam Alman could not be more excited. In
2:04:47his own manifesto shared online in a
2:04:49September 2024 blog post entitled The
2:04:52Intelligence Age, Sam Alman says, "I
2:04:54believe the future is going to be so
2:04:56bright that no one can do it justice by
2:04:58trying to write about it now. A defining
2:05:00characteristic of the intelligence age
2:05:02will be massive prosperity. He talks
2:05:04about fixing problems that have plagued
2:05:06mankind for generations of achieving
2:05:08things that humanity has only dreamed
2:05:10of. Fixing the climate crisis, creating
2:05:12colonies in space, making gamechanging
2:05:15breakthroughs in science. He also tries
2:05:17to calm any fears about AGI, claiming
2:05:19history always replaces old jobs with
2:05:21new ones whenever technology advances.
2:05:23But this time, it won't be so simple.
2:05:26Jobs won't be replaced or lost. They
2:05:28will just change. He adds that he has no
2:05:31fear that we'll run out of things to do
2:05:32because people have an innate desire to
2:05:35create. AI will allow us to amplify our
2:05:37own abilities like never before. Alman
2:05:39sells his vision well. But look closer
2:05:42and the cracks are clear to see. Let's
2:05:44say Alman gets what he craves and AGI
2:05:47takes over all that economically
2:05:48valuable work. What happens to the
2:05:50graduate, the white collar worker, the
2:05:52entrylevel employee struggling to just
2:05:54get a foot on the ladder? Well, they
2:05:57don't fit into OpenAI's vision of the
2:05:58future. Because if a machine can do the
2:06:00work of a junior associate or
2:06:02entry-level employee for a fraction of
2:06:04the price, then that employee is no
2:06:06longer an asset. They're a liability.
2:06:08Alman wants to frame this as liberation,
2:06:11a world where humans can pursue
2:06:12creativity while machines handle all the
2:06:14boring work. They claim to be building a
2:06:16tool that will both empower and
2:06:18outperform humanity. Do you see the
2:06:20problem? Those two ideas can't
2:06:22realistically coexist. You can't empower
2:06:24a worker by making their core skill set
2:06:27entirely obsolete. And you can't build
2:06:29an economy by removing the ways people
2:06:31actually earn money. Open AAI's 2030 AGI
2:06:34prediction suggests the software to
2:06:36power it must already exist. But the
2:06:38hardware, that's a different story.
2:06:40Chapter 2, the mathematical mirage. The
2:06:43AI world lives and dies by scaling laws.
2:06:45Basically, if you add more data and more
2:06:47processing power, what they call
2:06:49compute, AI gets smarter and more
2:06:51capable. In theory, there is no ceiling
2:06:53to this. Companies [music] can just keep
2:06:55feeding it more data, more compute year
2:06:57after year, watching their models reach
2:06:59new levels of genius. OpenAI has already
2:07:01admitted that AGI will take enormous
2:07:03amounts of resources. Alman has spoken
2:07:05about buying up vast amounts of energy
2:07:07and chips to build massive computer
2:07:09clusters. The goal is to drive down the
2:07:11cost of computation until it becomes
2:07:13effectively unlimited, abundant, cheap,
2:07:16and everywhere. Again, it sounds
2:07:18perfect, almost utopian, but it's an
2:07:20illusion, a pipe dream. [music]
2:07:22To power its AGI revolution, OpenAI is
2:07:25going allin on the so-called Stargate
2:07:27project. It's a joint venture to build a
2:07:29whole new AI infrastructure for OpenAI
2:07:32in the US. They're starting with an
2:07:34eyewatering $100 billion [music] poured
2:07:37into a massive supercomputer cluster.
2:07:40And that's only step one of the
2:07:42long-term Stargate vision, which is
2:07:43eventually set to cost $500 billion by
2:07:462030. That's because AI is insatiable.
2:07:49It's always wanting more. More data,
2:07:51more resources, more power. According to
2:07:54some experts, if OpenAI keeps going at
2:07:55its current pace, it'll need 10 times
2:07:58more data and 100 times more compute
2:08:00every 2 years. That means billions more
2:08:03poured into data centers, energy, and
2:08:05hardware. Even the richest economies
2:08:07would struggle to sustain that. And
2:08:09OpenAI isn't a wealthy nation. It's not
2:08:12even a profitable business. It currently
2:08:14operates on a massive deficit with
2:08:16[music] a 2024 revenue of $3.4 billion
2:08:19and a loss of 5 billion. The company is
2:08:22incinerating cash. To put it into
2:08:24perspective, that $5 billion loss is
2:08:26equal to the entire GDP of countries
2:08:28like Liberia, Surinom, Greenland,
2:08:30[music] or several Caribbean nations.
2:08:32One single company burning through the
2:08:34same amount of money that powers [music]
2:08:36entire national economies every year.
2:08:39That is not sustainable. It's desperate.
2:08:41Thinking more long-term, Alman has even
2:08:43spoken about a 7 trillion investment
2:08:45being required to construct a [music]
2:08:47global chip building initiative. That's
2:08:49around 7% of the world's entire GDP. All
2:08:52of this billions spent, supercomputers
2:08:55built, energy [music] consumed in the
2:08:57hope that it's enough to achieve AGI and
2:09:00then somehow it'll generate enough
2:09:02wealth to repay those billions and
2:09:05billions more of investment. It's a
2:09:07gamble, one the world has never seen
2:09:09before and one that Silicon Valley loves
2:09:11to take. Parts of the tech industry run
2:09:14on blitz scaling, the idea that you lose
2:09:16money upfront to build your product and
2:09:18capture the market. The hope is one day
2:09:20you'll raise prices and rake in the
2:09:22profits once the initial chaos is over.
2:09:24Except you can't blit scale the laws of
2:09:26physics. [music] It relies on a physical
2:09:29miracle. The idea that high-end
2:09:31semiconductors and computational
2:09:32resources can scale 10fold in less than
2:09:35a decade. It's not feasible and it's
2:09:37never going to happen. And we haven't
2:09:39even touched on the energy demands of
2:09:41the AGI revolution yet. Chapter 3.
2:09:43Powering a digital god. Even if the
2:09:46massive clusters needed for AGI are
2:09:48built, how do you power it all? The
2:09:51digital world isn't exactly weightless.
2:09:52It depends on hardware. Hardware that
2:09:54has to be plugged in and powered only on
2:09:57a scale that is millions of times
2:09:59greater. [music] It's estimated that a
2:10:00single chat GPT query uses around 10
2:10:03times more electricity than a Google
2:10:05search. That's about the same amount of
2:10:07power needed to power a small LED light
2:10:09bulb for about 20 minutes. Now scale
2:10:12that up. scale it across billions of
2:10:14users, millions of queries every single
2:10:17hour. The result, an energy demand
2:10:19unlike anything the world has ever seen,
2:10:22far beyond what today's electrical grid
2:10:24can handle. Renee Hos, the CEO of
2:10:26British semiconductor and software
2:10:28design company ARM Holdings, believes
2:10:30that AI data centers could eat up to 25%
2:10:32of the US's entire power supply by 2030.
2:10:36At the moment, AI uses just 4% of
2:10:38America's supply. Other experts agree,
2:10:40arguing that the US needs to expand its
2:10:42power grid capacity by around 20% just
2:10:46to keep up with the everexpanding AI
2:10:47industry. In real terms, that would be
2:10:50like the US having to add the entire
2:10:52power generation capacity of a major
2:10:54European nation like Germany to its grid
2:10:57in less than a decade. Again, this isn't
2:10:59just ambitious, it is impossible. Power
2:11:02plants don't just appear overnight.
2:11:03Nuclear reactors, the only carbon-f free
2:11:06source capable of delivering this kind
2:11:07of massive power, can take over a decade
2:11:09to build, sometimes even longer, thanks
2:11:12to regulations, delays, and unexpected
2:11:14construction hurdles. Tech billionaires
2:11:16know this, but they also seem to believe
2:11:17they could bend or break the rules to
2:11:20just get what they want. Microsoft is
2:11:22attempting to restart the Three-Mile
2:11:23Island nuclear plant project to help
2:11:25feed AI's insatiable appetite. It's
2:11:27another desperate act, a clear sign that
2:11:30without a massive private power supply,
2:11:32the intelligence age has zero chance of
2:11:34arriving by 2030. And electricity isn't
2:11:36the only problem. A single data center
2:11:39requires millions of high voltage copper
2:11:41cables and huge electrical transformers.
2:11:44The global copper supply chain is
2:11:45already stretched due to other
2:11:47industries and the surge in electric
2:11:49vehicles. Experts predict it'll hit full
2:11:51capacity by the late 2020s, and it'll
2:11:54still be up at least 10% short of global
2:11:56demand. The world still doesn't have
2:11:58enough raw materials to fuel the AGI
2:12:00push. This isn't conjecture, it's
2:12:02[music] fact. Vatzlav is one of the
2:12:05world's leading energy experts. He spent
2:12:07decades studying energy transitions, and
2:12:09he knows how slow they actually are.
2:12:12They don't take years, they take
2:12:14generations. As Smele warns, Silicon
2:12:16Valley's wouldbe masters of the universe
2:12:18have discovered that energy transitions
2:12:20are subject to time spans and technical
2:12:22constraints [music] that defy their
2:12:24reach. But if Altman and OpenAI know
2:12:27that an energy revolution can't be
2:12:28forced with money alone, why keep
2:12:31promising 2030? What are they really
2:12:33building toward with this infrastructure
2:12:34race? Chapter 4. The compute gentry.
2:12:38This is where things start to get
2:12:39worrying. History shows that when a new
2:12:41means of production emerges, power
2:12:43rarely spreads. It remains in the hands
2:12:45of a few. The age of steam had its steel
2:12:48barren. The internet age was dominated
2:12:50by ISPs. Now, in the age of AI, power
2:12:53won't be measured in barrels of oil or
2:12:55dollars. It'll be measured in flops,
2:12:57floatingoint operations per second.
2:12:59Whoever controls the compute controls
2:13:01the future. And that means a new elite
2:13:03class will be born, the so-called
2:13:05compute gentry. Leopold Dashen Burner, a
2:13:08former OpenAI researcher, has been
2:13:10sounding the alarm. In a report called
2:13:12situational awareness, he lays out the
2:13:14troubling geopolitical realities behind
2:13:15the AI race. He warns that the physical
2:13:18demands of AGI are so huge that only a
2:13:20handful of the world's wealthiest
2:13:22companies will be able to take part and
2:13:24it's not only inevitable. It's already
2:13:26happening. Ashen Burner says by 27 or 28
2:13:30the endgame will be on. By 2829 the
2:13:33intelligence explosion will be underway.
2:13:35By 2030 we will have summoned super
2:13:37intelligence in all its power and might.
2:13:39The risk of it all going in Ashen
2:13:41Brener's words off the rails will be
2:13:43constant. [music] Analysts fear a world
2:13:45where so much power rests in the hands
2:13:47of just a few mega corporations. If open
2:13:49AI and other tech giants succeed, they
2:13:51will control the very means of
2:13:53production for intelligence itself. And
2:13:55we will never have seen anything like
2:13:57it. The smarter this tech gets, the more
2:13:59expensive it becomes to train and
2:14:01maintain. [music] That only raises the
2:14:02entry barrier. The result will be a kind
2:14:05of feudal system. Anyone wanting to run
2:14:07a business or offer a service will have
2:14:08to rent intelligence from the compute
2:14:10gentry because whoever owns the
2:14:12intelligence that replaces labor owns
2:14:15the economic output of the human race.
2:14:17This goes entirely against OpenAI's
2:14:19original founding philosophy. The
2:14:21company began as a nonprofit. Its
2:14:23charter was to ensure that AGI would
2:14:25benefit all of humanity. It promised
2:14:28openness, transparency, and guardrails
2:14:30at every step. That charter has changed.
2:14:33The company has restructured into a
2:14:34for-profit entity. The guard rails are
2:14:36gone. Their mission is no longer about
2:14:39bettering humanity. It's all about
2:14:41profit. And this, more than anything
2:14:43else, highlights the crux of OpenAI's
2:14:45AGI lie. It claims that this technology
2:14:48will be a public good. That it'll bring
2:14:50about a brighter and more prosperous era
2:14:52for everyone. Meanwhile, what OpenAI is
2:14:54actually doing is building its own
2:14:56private fortress and throwing away the
2:14:58key. Chapter 5, the corporate sovereign.
2:15:01The end goal of this AGI era is the
2:15:03restructuring of society as we know it.
2:15:05Alman wants people to believe that this
2:15:07is all being done with their best
2:15:08intentions at heart. He has spoken on
2:15:10several occasions about the prospect of
2:15:12a universal basic income. It would be
2:15:14the obvious antidote to any risk of
2:15:16large-scale job loss and unemployment
2:15:18brought about by hyper intelligent AI.
2:15:21But that logic is flawed. If AGI
2:15:24replaces the human workforce, the tax
2:15:25base of every democratic country
2:15:27collapses. No human income means [music]
2:15:30no taxes. No taxes means governments
2:15:32won't have the money to provide
2:15:33universal basic income to millions.
2:15:35Instead, the power will become even more
2:15:37concentrated. And it won't be a
2:15:39government writing your UBI check. It'll
2:15:42be a corporation. That same corporation
2:15:44that owns the algorithm that replaced
2:15:46you. This [music] is the definition of
2:15:48algorithmic feudalism. It is a
2:15:49terrifying but realistic prospect. The
2:15:52public can't [music] vote for CEOs. They
2:15:54can't lobby blackbox algorithms.
2:15:56Democracy depends on the economic
2:15:58independence of the people, but when
2:16:00people are entirely dependent on
2:16:01corporate entities [music] to give them
2:16:03the money they need to feed their
2:16:05families, they're no longer economically
2:16:07independent, they are subjects. Open AAI
2:16:09argues that it'll all work. Altman talks
2:16:12of changing jobs and new opportunities,
2:16:15but he doesn't go into specifics. He
2:16:17doesn't explain why a for-profit
2:16:18corporation would decide to simply give
2:16:20away its primary source of power and
2:16:22income. Worst of all, Alman's
2:16:24intelligence age manifesto ignores
2:16:25something fundamental, the long-standing
2:16:28social contract between the people and
2:16:29those in power. It assumes that you'll
2:16:31simply accept your loss [music] of
2:16:33agency. That you'll simply nod and smile
2:16:35as your career, your privacy, and your
2:16:37democratic rights are stripped away in
2:16:39the name of super intelligent AI agents.
2:16:41Society is being restructured, not with
2:16:44votes or referendums, but by the pursuit
2:16:46of a digital god. The Stargate project,
2:16:48the 7 trillion dollar chip plan, the
2:16:51insatiable demand for energy. Those
2:16:53aren't isolated events. They're all
2:16:55links in a chain of a new world order. A
2:16:57world where intelligence is a
2:16:59centralized commodity, where human labor
2:17:01is obsolete, and the computer gentry are
2:17:03the kings and queens of it all. The AGI
2:17:06lie isn't that this technology won't
2:17:07work. The lie is that it is being built
2:17:10for you. This technology isn't being
2:17:12made for the people. It's being made to
2:17:14replace them. The only question left is,
2:17:16is there any way to stop it? This is
2:17:18Josh, and today on the infographic show,
2:17:20we're going to talk about why OpenAI
2:17:22will run out of money, but not for the
2:17:24reason you think. [music] The creator of
2:17:25ChatGpt looks like the king of tech with
2:17:28$20 billion in revenue, but internal
2:17:30spreadsheets reveal something startling.
2:17:32[music] Starting in 2026, they face
2:17:34projected losses of $14 billion
2:17:36annually. By 2029, cumulative spending
2:17:38could hit 115 billion. The [music]
2:17:40product works, but the bills are tied to
2:17:42expensive realworld constraints. Here's
2:17:45the thing that most people [music] miss.
2:17:47The massive losses lie in a simple fact.
2:17:49AI is not just another app, and it
2:17:51behaves unlike any software we have ever
2:17:54built. [music] In the traditional
2:17:55software world, if you want to make a
2:17:57better app, you hire better engineers.
2:17:59You write cleaner code. It's a human
2:18:01[music] cost. But AI doesn't work like
2:18:03that. It works on something called
2:18:04scaling laws. These are mathematical
2:18:07rules that govern how AI gets smarter.
2:18:09and they are incredibly expensive.
2:18:11[music] The rules are simple. If you
2:18:12want a model to be, say, twice as good,
2:18:15you can't just double your effort. You
2:18:17have to ramp up computing power by a
2:18:19lot. It's basically a brute force
2:18:20[music] equation. Small gains in
2:18:22intelligence mean massive spikes in
2:18:24capital. It sounds crazy, right? But
2:18:26wait until you see the numbers. Training
2:18:28GPT4, the model that really kicked off
2:18:30the revolution, cost roughly $100
2:18:33million in computing power. That is for
2:18:35one full training run which [music] is
2:18:37the process of teaching the model from
2:18:39scratch. For a big tech company that is
2:18:41expensive but manageable. The next
2:18:43generation the frontier models arriving
2:18:45in 2026 [music] and 2027 play by
2:18:47different rules. Each run could cost
2:18:49over $1 billion. [music] We have reached
2:18:52a point where a single training session
2:18:54for one AI model costs more than the GDP
2:18:56of some small island nations. And it
2:18:59gets worse. You can't just train it once
2:19:01and walk away. You have to keep on doing
2:19:04it. Open AI is trapped in a cycle where
2:19:06they must spend these billions of
2:19:08dollars just to [music] stay slightly
2:19:10ahead of their rivals. Rivals who are
2:19:12giving similar tech away for free. This
2:19:14creates a fundamental gap in their
2:19:15business model. Their costs are [music]
2:19:17tied to physical realities, electricity
2:19:19and silicon which are expensive and
2:19:21scarce. But their ability to raise
2:19:23prices is limited because there's so
2:19:24much competition. The math is simple and
2:19:27it is catastrophic. Explosive costs are
2:19:29outpacing revenue and the money is
2:19:31running out. And the financial bleed
2:19:33gets even [music] worse. To do the heavy
2:19:35lifting, OpenAI needs high-end AI chips
2:19:37like Nvidia's Blackwell B200s. These
2:19:40aren't your typical CPUs or GPUs. Each
2:19:43one runs $30,000 to $40,000. And you
2:19:46can't buy just one. To train a Frontier
2:19:49model, you need a cluster. That means
2:19:51tens [music] of thousands of these
2:19:52chips, all wired together with
2:19:54high-speed links and liquid cooling
2:19:56systems. And this is where the costs
2:19:58really start to pile up. But the problem
2:20:00isn't just buying the [music] chips. The
2:20:02problem is that these chips have a
2:20:03limited shelf life. Unlike a machine in
2:20:05a factory or a delivery truck, which
2:20:07might run for 20 years, AI hardware
2:20:10doesn't last. It becomes outdated the
2:20:12moment the next generation of chips hits
2:20:14the market. And then companies are
2:20:15playing catch-up. Open AAI has to
2:20:18replace their entire system of chips
2:20:19roughly every 18 months to 3 years just
2:20:22to stay competitive with Google and
2:20:23Meta. Imagine a trucking company having
2:20:26to buy a brand new fleet every 18 months
2:20:28because the old trucks [music] suddenly
2:20:30can't deliver packages fast enough. That
2:20:32is the economic reality of AI hardware.
2:20:34This means the billions of dollars
2:20:36OpenAI spends on hardware isn't a
2:20:38long-term investment. [music] It's an
2:20:39expense that disappears. The value of
2:20:41that hardware drops fast. But if the
2:20:44cost of the chips wasn't enough, there
2:20:45is another bill that's starting to look
2:20:47even scarier. The electric bill. This is
2:20:50best illustrated by Project Stargate.
2:20:52It's described as just a big new
2:20:54supercomput, but it's actually a $500
2:20:56billion gamble. 500 billion. Yeah,
2:21:00that's right. To put that into
2:21:01perspective, 10 gawatt could power
2:21:03millions of homes. It's the equivalent
2:21:05of multiple full-scale nuclear reactors
2:21:07just for this one project. Why does this
2:21:10matter? Because the costs aren't going
2:21:12away, and the grid can't keep up. The
2:21:15scaling costs aren't going away. They're
2:21:17fixed. You can't build the next
2:21:18generation of AI without this level of
2:21:20power. The bottleneck isn't just the
2:21:22cost of electricity. It is the national
2:21:24grid. Getting enough high voltage
2:21:26transformers and grid capacity is a huge
2:21:28hurdle. The old utility system can't
2:21:30grow fast enough to keep up. So, OpenAI
2:21:33is now in the position of negotiating
2:21:34for direct access to nuclear power and
2:21:37massive solar farms. These utility costs
2:21:39create a high floor for their operating
2:21:41expenses. Every free Chad GPT user is
2:21:44literally costing billions and there is
2:21:46no way around it. It makes it nearly
2:21:48impossible to maintain healthy profits
2:21:50when you are trying to offer a free tier
2:21:53to hundreds of millions of users. Every
2:21:55time someone uses chat GBT for free,
2:21:57OpenAI has to pay for the electricity
2:21:59and the silicon wear and tear. So, if
2:22:01OpenAI is losing billions of dollars on
2:22:03chips and electricity, how are they
2:22:05still open? How do they pay their
2:22:07employees? And that leads us to one of
2:22:09the most misunderstood pieces of the
2:22:11OpenAI story, its deal with Microsoft.
2:22:14We often hear that Microsoft has
2:22:15invested [music] billions into OpenAI
2:22:17and on paper it looks like billions came
2:22:20in. In reality, it's more like a
2:22:22financial merry-goround that hides
2:22:23[music] how tight the startup's cash
2:22:25really is. When Microsoft invests
2:22:27billions, a lot of that money doesn't
2:22:29actually leave Microsoft. They give
2:22:31OpenAI cloud credits instead, sort of
2:22:33like a gift card. [music] And you might
2:22:35think that that counts as real cash. It
2:22:37doesn't. OpenAI can record it as capital
2:22:40raised. So it looks like cash, but the
2:22:42credits have [music] to be spent on
2:22:43Azure, Microsoft's cloud service to run
2:22:46their models. This effectively recycles
2:22:48the investment back into Microsoft's
2:22:50revenue stream. It boosts [music]
2:22:51Microsoft's cloud earnings and stock
2:22:53price. But here's the dangerous part.
2:22:56You cannot pay your employees with cloud
2:22:58credits. When OpenAI hires a top
2:23:00researcher for $2 million a year, they
2:23:03need hard cash. [music] When they have
2:23:04to pay for office space or legal fees,
2:23:07they need money. This creates a
2:23:09financial [music] optical illusion.
2:23:10Microsoft invests 10 billion, but that
2:23:13money doesn't actually land in OpenAI's
2:23:15account. [music] It's basically digital
2:23:16coupons that can only be spent on
2:23:18Microsoft servers. The result is massive
2:23:21pressure. Every fiscal quarter, OpenAI
2:23:23has to raise hard cash from other
2:23:24investors [music] just to pay payroll
2:23:26and cover bills that Microsoft credits
2:23:28can't touch. If the flow of new outside
2:23:30investment slows down, OpenAI faces a
2:23:33cash flow crisis. They might [music]
2:23:34have plenty of computer time, but not
2:23:36enough hard currency to keep their team
2:23:38from leaving for rival companies.
2:23:40Despite all those costs, [music]
2:23:41investors keep on pouring money in. In
2:23:44March 2025, OpenAI managed to raise $40
2:23:47billion, the largest [music] private
2:23:48funding round in history, even bigger
2:23:50than the IPO of the oil giant Saudi
2:23:53Aramco. But here is what is really odd
2:23:55about it. Saudi [music] Aramco has
2:23:57hundreds of billions in revenue and more
2:23:59importantly it has real tangible assets
2:24:02oil reserves that you can measure and
2:24:04sell. Open AAI is a [music] startup with
2:24:06no profits burning cash at a rate of
2:24:08billions a year. Its value is mostly
2:24:10intellectual property which anyone can
2:24:12try to copy. So what does this mean for
2:24:14the long-term survival of Open AI? The
2:24:17answer will surprise you. Investors are
2:24:19pouring money in based on the promise of
2:24:21a market that doesn't fully exist yet.
2:24:23For OpenAI to be worth a trillion
2:24:25dollars, it can't just [music] be
2:24:27impressive. It has to replace dozens of
2:24:29cheaper tools that companies already
2:24:31use. Right now, [music] most businesses
2:24:33spread their AI budgets across multiple
2:24:35smaller providers, not just one giant
2:24:37system. OpenAI is building something
2:24:39massive and expensive, betting that
2:24:41eventually everyone will need it.
2:24:43[music] But right now, there's no
2:24:44guarantee of that demand. And this leads
2:24:47us to the risky business model. In
2:24:49software, companies survive by making it
2:24:51hard for customers to leave. Salesforce
2:24:53does this because moving all your data
2:24:54is a huge pain. Netflix does this
2:24:57because they own shows that you can't
2:24:59watch anywhere else. Open AAI is
2:25:00discovering a hard lesson. Users are
2:25:03mercenary. If Google's Gemini or Meta's
2:25:06Llama offers a similar answer for
2:25:07cheaper, they'll leave instantly. About
2:25:1075% of OpenAI's revenue comes from
2:25:12[music] consumer subscriptions. But the
2:25:14number of cancellations is rising. And
2:25:16once the novelty fades, most users won't
2:25:18pay. Big business is even more
2:25:20skeptical. Only about 20 to 30% are
2:25:23sticking with OpenAI's API long term.
2:25:25Many are choosing open- source models
2:25:27like Llama to keep data private and
2:25:29costs down. With nothing [music] keeping
2:25:31them tied to OpenAI, no built-in
2:25:33network, no way their data is stuck.
2:25:35They could just jump to another provider
2:25:37overnight. And the competition is just
2:25:39as deadly as OpenAI's own cash [music]
2:25:41burn. Meta's decision to release the
2:25:44Llama models for free was not an act of
2:25:46charity. It was a tactical strike. When
2:25:48Mark Zuckerberg gives everyone access to
2:25:50their top-of-the-line AI for free, he
2:25:52effectively sets a ceiling on what
2:25:54OpenAI can charge. Medic can burn cash
2:25:56on open source models because they're
2:25:58using the tech to improve ads on
2:26:00Instagram and Facebook. Their business
2:26:02isn't selling AI, it is selling ads.
2:26:04Open AAI doesn't have that luxury. Their
2:26:06only product is the AI itself. They're
2:26:09fighting to establish themselves while
2:26:10their competitors aggressively undercut
2:26:12the market to [music] keep them from
2:26:13gaining ground. And the clock is
2:26:16ticking. Open AI is squeezed from all
2:26:18sides. On top, giants like Microsoft and
2:26:20Google with practically unlimited cash.
2:26:23On the bottom, lean competitors like
2:26:24Anthropic [music] and Mistral. Anthropic
2:26:26runs a much more efficient operation,
2:26:28focusing on safety and enterprise
2:26:30reliability with a [music] much lower
2:26:32burn rate. Meanwhile, Google's DeepMind
2:26:34keeps stealing talent, forcing OpenAI to
2:26:36offer massive stock-based pay packages.
2:26:38Those [music] only work if the company's
2:26:40valuation keeps climbing. If it stalls,
2:26:42the researchers, the company's only real
2:26:44asset, could walk out the door. As if
2:26:46burning billions, fighting competitors,
2:26:48and losing talent weren't enough,
2:26:49regulators in Washington and Brussels
2:26:51are [music] circling. In early 2026, the
2:26:54FDC and European Union intensified their
2:26:56antitrust probes into the Microsoft
2:26:58OpenAI partnership. Regulators are
2:27:00checking whether Microsoft's investment
2:27:02[music] is actually a de facto
2:27:04acquisition designed to skirt merger
2:27:06laws. If they decide to limit the power
2:27:08Microsoft has over open AAI or force a
2:27:10split, it would cut the startup's
2:27:12financial lifeline. And then there's the
2:27:14mounting geopolitical friction. Export
2:27:17controls on AI chips are shrinking the
2:27:19global market, while new AI safety
2:27:21regulations are creating a massive
2:27:22compliance burden. Open AI now needs
2:27:25armies of lawyers and safety
2:27:26researchers. Rules [music] that are
2:27:28costly and generate zero revenue. The
2:27:31danger becomes clear when you look at
2:27:32history. [music] Uber lost billions
2:27:34before its initial public offering or
2:27:36IPO, but it was building a physical
2:27:39network in thousands of cities. Tesla
2:27:41struggled for years, but it was building
2:27:43factories and a global charging network.
2:27:45Something real that competitors couldn't
2:27:47copy overnight. Open AI, well, it's
2:27:49burning billions with no real network or
2:27:51physical [music] assets to lean on. Open
2:27:53AAI's production is all about raw
2:27:55computing power, the expensive chips
2:27:57that mostly come from Nvidia. Unlike
2:27:59Tesla or Uber, OpenAI's product loses
2:28:02money every time someone asks it a
2:28:03[music] complex question. And there is
2:28:05nothing stopping users from leaving
2:28:07tomorrow. The company is now effectively
2:28:09betting [music] everything on a single
2:28:11desperate timeline. They're racing to
2:28:13build artificial general intelligence or
2:28:15AGI, an AI that can think and learn like
2:28:17a human before the bank [music] account
2:28:19runs out. This isn't a standard software
2:28:21business strategy anymore. If OpenAI can
2:28:24build a model smart enough to do the
2:28:25work of a human expert in any field,
2:28:27their current cash burn wouldn't matter.
2:28:30Revenue could in theory skyrocket.
2:28:32They're picturing a world where their AI
2:28:34doesn't just summarize emails. It
2:28:35replaces entire departments, handling
2:28:37corporate taxes, writing complex code,
2:28:39and planning strategic business moves at
2:28:41superhuman speed. Reach that milestone
2:28:43and they could charge a premium that
2:28:45covers any debt, no matter how massive.
2:28:47If OpenAI is losing 14 to [music] 17
2:28:50billion a year, every month of delay
2:28:53costs over a billion. If the
2:28:55breakthrough to AGI takes [music] five
2:28:56years instead of two, they'd face a
2:28:58funding gap of nearly $100 billion just
2:29:01to keep the lights on. And no investor
2:29:03can fix that overnight. So what happens
2:29:05when the money runs out? You might
2:29:07expect a dramatic crash. But the reality
2:29:09is different. The most likely outcome is
2:29:11not a dramatic crash or a bankruptcy
2:29:13filing, but a quiet absorption. By mid
2:29:152027, based on current projections, the
2:29:18cash reserves raised in the 2025 rounds
2:29:20will be nearly empty. At that point,
2:29:22OpenAI will face a choice.
2:29:23>> [music]
2:29:24>> raise another massive round at a lower
2:29:26valuation, crushing their employee stock
2:29:28options or [music] sell. Microsoft is
2:29:30the natural and maybe the only buyer.
2:29:32They already host OpenAI systems on
2:29:34Azure, and they have deep integration
2:29:36with the software. [music]
2:29:37More importantly, Microsoft has over $80
2:29:39billion in cash reserves, making them
2:29:41one of the few entities on Earth that
2:29:43could sustain OpenAI's burn rate. For
2:29:45Microsoft, this is the crown jewel, the
2:29:47engine of the next computing
2:29:49>> [music]
2:29:49>> era. For investors, it's a fire sale,
2:29:51but one that buys survival. This is the
2:29:54end of the startup frontier. Open AAI
2:29:56proved scaling [music] works, but only
2:29:57if you have a nationstate sized budget.
2:30:00The AI revolution has gone industrial
2:30:02where success is measured in acres of
2:30:04data centers, [music] not lines of code.
2:30:06Open AI started the trend, but it
2:30:08doesn't have the resources to compete
2:30:09alone. You were promised infinite
2:30:11intelligence for just 20 bucks a month.
2:30:14That promise is already breaking. AI was
2:30:16supposed to be as cheap and limitless as
2:30:18electricity, but [music] now Silicon
2:30:20Valley is pulling it back. Across models
2:30:22and platforms, users are now facing
2:30:24tighter message caps and shrinking
2:30:26access. It's like an all you can eat
2:30:28buffet where you're told you get one
2:30:30bite every few hours. The era of
2:30:32infinite AI is fading. The reason is a
2:30:34frantic new game inside big tech called
2:30:37token [music] maxing. So, if AI is
2:30:39getting more powerful, why is access
2:30:41getting smaller? For a while, anyone
2:30:43paying for ChatGpt Plus could open
2:30:45OpenAI's newest reasoning model, 01
2:30:48preview, and use it as and when
2:30:50required. Then, in September 2024, a new
2:30:53limit was set, [music] just 50 messages
2:30:55a week. For the people paying the most,
2:30:57that came out to a handful of real
2:30:59conversations spread out across 7 days.
2:31:01Compared [music] to the previous
2:31:02generation of models, it was a cut of
2:31:04roughly 98%. Once the quota ran dry, the
2:31:07screen stopped. There was a way to
2:31:09bypass this, but it came with a price.
2:31:11>> [music]
2:31:11>> In December 2024, OpenAI introduced a
2:31:14tier called Chat GPT Pro at 200 bucks a
2:31:17month, 10 times the cost of Plus. It
2:31:20came with near unlimited use of the very
2:31:22model everyone else was now being
2:31:23rationed on. Unlimited intelligence
2:31:25still existed. [music]
2:31:26It simply had moved behind a much bigger
2:31:28payw wall. Open AAI framed the move as
2:31:30housekeeping, a way to protect the
2:31:32quality of the service. But for the
2:31:34people who had built it into their daily
2:31:35work, [music] it felt like a tool had
2:31:37suddenly been pulled out of reach
2:31:38without much warning. The model wasn't
2:31:40gone, but it was harder to get to.
2:31:42People started rationing themselves.
2:31:44They hoarded their most difficult
2:31:45[music] questions or work for the hour
2:31:47the quota reset. And then if they ran
2:31:50out, they would just stop and wait. The
2:31:51companies that had spent years talking
2:31:53about intelligence getting cheaper and
2:31:55more widely available were restricting
2:31:57access to the very thing they sold as
2:31:59the future. The gap between the big
2:32:01promises and what people were actually
2:32:02using was hard to ignore. The reason was
2:32:05obvious. Running these models at full
2:32:07capacity the way the industry had scaled
2:32:09them wasn't adding up financially
2:32:11anymore. So big tech came up with the
2:32:14perfect workaround. Token maxing. A
2:32:16token is just a scrap of a language, a
2:32:19word, a small step in the machine's
2:32:21train of thought. So when you can no
2:32:23longer make a model smarter the old way,
2:32:25you make it work harder. You force it to
2:32:27chew through far more words or tokens
2:32:29for the very same answer. Instead of
2:32:31giving an instant reply, the system now
2:32:33talks to itself first. privately at
2:32:36length. This can be tens of thousands of
2:32:38words before a single sentence ever
2:32:40reaches your screen. And instead of
2:32:42learning only from what humans wrote,
2:32:43the newest models are fed enormous piles
2:32:45of text that older models generated. The
2:32:48whole point is to keep the curve from
2:32:50going flat by shoving more and more
2:32:52compute through the same machine. The
2:32:54scale is mindblowing. Meta's Llama 3 ate
2:32:57through more than 15 trillion tokens of
2:32:59text, about seven times the data used on
2:33:01the version before it. Stack that human
2:33:03equivalent on a bookcase shelf and it
2:33:06would be close to 1,800 miles long.
2:33:08Token maxing takes that number and piles
2:33:10more on top of it. Models loop through
2:33:13their old output for every hard question
2:33:15they're asked. It sounds like a good
2:33:16idea. More intelligence is squeezed out
2:33:18of the same chips with no new
2:33:20supercomputer required. The longer the
2:33:23machine talks to itself, the more it can
2:33:24check its own work. Try an idea and then
2:33:27throw it away. The extra thinking is
2:33:29bought one word at a time. All you see
2:33:31is a brief pause and then the answer.
2:33:33But this was never a success to be
2:33:35heralded. It was born out of sheer
2:33:37panic. Something in the AI world had
2:33:39broken and people closest to it knew
2:33:41exactly what it was. For 10 years, the
2:33:43whole industry ran on one core belief.
2:33:46Multiply the computing power by 10 and
2:33:48the model gets dramatically smarter.
2:33:50More compute means more capability year
2:33:52after year after year. Entire business
2:33:54models were gambled on this assumption.
2:33:56The results said different. Inside
2:33:58multiple labs, the data said the same
2:34:00thing. A 10 times jump in compute was
2:34:02only producing marginal gains, around 10
2:34:05to 15%, sometimes less. What used to buy
2:34:08a leap in intelligence was now buying a
2:34:10sliver of improvement. OpenAI saw it
2:34:12firsthand. The model supposed to be its
2:34:15next leap forward, codenamed Orion,
2:34:17reached the level of the previous
2:34:18flagship after only a fraction of its
2:34:20training. Then it stalled. The people
2:34:23who tested it said the improvements were
2:34:25smaller than expected. When it finally
2:34:26shipped, it wasn't the long promised
2:34:28GPT5, but GPT4.5,
2:34:32a downgrade in name that said
2:34:34everything. Not everyone agreed. When
2:34:36news of the slowdown leaked in 2024,
2:34:38OpenAI chief Sam Alman publicly
2:34:41dismissed them. There is no wall, he
2:34:43said. The researcher Gary Marcus shot
2:34:46back that the limitations had been
2:34:47obvious for months. Even Andre and
2:34:49Horowitz with billions riding on the
2:34:51outcome admitted that more computing
2:34:53power was no longer delivering the same
2:34:55leaps in intelligence. To some, it was a
2:34:57temporary plateau. To others, the end of
2:35:00an era. Either way, the money kept
2:35:02flowing, just in a different direction.
2:35:04In public, the message remained the
2:35:06same. Progress was just around the
2:35:07corner. In private, more and more people
2:35:09were coming to the same conclusion. The
2:35:11strategy that had fueled a decade of
2:35:13breakthroughs was running out of room,
2:35:15and the next gains would have to come
2:35:17after training. The early jumps were
2:35:19hard to miss. Each new model seemed
2:35:20smarter than the last. It was obvious to
2:35:22anyone who used them. The newer releases
2:35:25felt different. [music] Better, yes,
2:35:27smoother and more reliable, but not the
2:35:30kind of leap that justified the billions
2:35:32being spent. Admitting that would have
2:35:34meant questioning the foundations of the
2:35:36entire industry was built on. So, the
2:35:38industry grabbed on to the one thing
2:35:39that still seemed to work, giving models
2:35:42more tokens. Back when models were
2:35:44smaller and answering a question costs
2:35:46next to nothing, chat GPT launched at 20
2:35:48bucks a month. It was set when AI was
2:35:50cheaper to run. And years later, it is
2:35:52still the same price. The economics,
2:35:54though, had changed completely. Some of
2:35:56the most active users were costing
2:35:57OpenAI more than $120 a month in
2:36:00computing power while paying just $20
2:36:03for the privilege. Every one of those
2:36:04users deep into the gap between what the
2:36:06service cost and what it charged. It's
2:36:08like a grocery store treating a loyal
2:36:10customer well and then slipping them a
2:36:12$100 bill on the way out the door. It
2:36:15doesn't stop there. It looks for a
2:36:16thousand more customers exactly like
2:36:18them. The more people who fall in love
2:36:20with the product, the faster the cash
2:36:22burns away. This is not a rough patch
2:36:24that fixes itself. It's baked into the
2:36:26system. Every time the model stops to
2:36:28think a little longer, the bill goes up.
2:36:31What used to be a quick response can
2:36:32turn into a long chain of reasoning
2:36:34running behind the scenes before the
2:36:36answer ever reaches the user. That's
2:36:38great for accuracy, it's less great for
2:36:40cost. The hardest questions demand the
2:36:43most compute, and those are exactly the
2:36:45questions people come to the best models
2:36:47to [music] solve. It affects the whole
2:36:48industry. The chips were bought and the
2:36:50data centers went up, but the revenue to
2:36:53justify them has [music] yet to arrive.
2:36:55In 2024, the venture firm Sequoia framed
2:36:57it as AI's $600 billion question. The
2:37:00distance [music] between the tens of
2:37:02billions being poured into AI hardware
2:37:04and the money the industry could ever
2:37:06earn back. Those data centers need
2:37:09hundreds of billions a year just to
2:37:10break even. And subscriptions cover only
2:37:13a small portion. [music] Adding more $20
2:37:15subscribers can't close it because every
2:37:18new heavy user only adds to the problem.
2:37:20So, the cap makes sense. letting
2:37:22everyone run the most powerful model all
2:37:24day was never going to work. The numbers
2:37:27don't allow it. So, the most expensive
2:37:29workloads are increasingly being pushed
2:37:30toward higher priced tiers and
2:37:32enterprise customers [music] where the
2:37:34economics make sense. Corporate
2:37:36contracts can hide the cost of a top
2:37:38model inside the hours it saves and the
2:37:40staff it replaces. A single $20
2:37:43subscriber can't. From the perspective
2:37:45of the labs, that subscriber was never
2:37:47the real customer. They were the proof
2:37:49the product worked. the hype [music]
2:37:51that made the enterprise deals possible.
2:37:53A deliberate loss-making base scaled to
2:37:55a level the industry has never tried
2:37:57before. Money was only part of the
2:37:59issue. The bigger problem is that all
2:38:01the labs are starting to run low on the
2:38:03very thing needed to build the next
2:38:05generation at all. But what happens when
2:38:07scaling stops delivering [music]
2:38:08and a deadline is beginning to loom?
2:38:11There's only a finite amount of writing
2:38:13in the world. It sounds strange to say,
2:38:15but it is true. Strip out the spam and
2:38:18lowquality slop and the amount of
2:38:19genuinely useful human written text left
2:38:22on the internet shrinks [music] fast.
2:38:24Researchers at Epoch AI estimated at
2:38:26roughly 300 trillion tokens. At the
2:38:29fastest training schedules, that supply
2:38:31could be mostly gone by 2026 with most
2:38:34projections landing around 2028. A few
2:38:37years after that, it runs dry entirely.
2:38:39But not all of it is equally valuable.
2:38:41The best material got used up first. the
2:38:44carefully edited books, [music] quality
2:38:45journalism, peer-reviewed research. A
2:38:47lot of it has already been seen multiple
2:38:49times in training runs. What's left is
2:38:51everyday web pages that add less and
2:38:53less each time you go back to them. This
2:38:55isn't something you solve by scraping
2:38:58more pages. The human text that drove
2:39:00the last big jumps is running out and
2:39:02it's not being replaced fast enough to
2:39:04keep up. So, the [music] industry turned
2:39:06to the only source that still looked
2:39:07bottomless, the machines themselves. If
2:39:10humans ran out of words, let the
2:39:12machines write their own and feed those
2:39:14to the next model. Close the loop and
2:39:16let the machine teach the machine. It
2:39:18has been tried. What happens next is
2:39:20called model collapse. And in 2024, the
2:39:23journal Nature published the autopsy.
2:39:25Train each new generation mostly on
2:39:27machine-made text, and it rots in a very
2:39:29specific way. The range of what it can
2:39:32say shrinks inward. The identity and
2:39:34personality that exists in real human
2:39:36writing fades out of every new version.
2:39:39The model drifts toward a flatter, more
2:39:41repetitive copy of itself. [music] And
2:39:43it only gets worse. The nature team fed
2:39:46a model a passage about medieval church
2:39:48towers, trained the next version on its
2:39:50answers, and then the next on those over
2:39:53and over. By the ninth generation, the
2:39:55model had forgotten the question
2:39:56entirely, and was discussing jack
2:39:59rabbits in a passage that had started
2:40:01out about architecture. Each generation
2:40:03trained on the last drifts a little
2:40:05further from the real thing. And pouring
2:40:07in more machine text does not stop the
2:40:09decline. It speeds it up, dragging the
2:40:11system toward a flat, hollowedout echo
2:40:13of what it once knew. Even slipping a
2:40:16little human writing back into the mix
2:40:17can't save it. [music] The machine-made
2:40:19portion sneaks in errors that are almost
2:40:21impossible to find. Eventually, the
2:40:23model starts to trust its own guesses as
2:40:25fact, doubling down on its blind spots,
2:40:27repeating its own mistakes and poisoning
2:40:29the well it drinks from. There is no
2:40:31escape from it. If more data only makes
2:40:34things worse, what happens when AI runs
2:40:36into the physical limits of the world
2:40:37itself? [music] Just outside Lowden
2:40:40County in Northern Virginia, miles and
2:40:41miles of windowless buildings dominate
2:40:44the landscape. It might sound like an
2:40:45unremarkable place, but it's estimated
2:40:47[music] that 70% of global internet
2:40:49traffic flows right through here. Data
2:40:52centers brought in roughly $875 million
2:40:54in tax revenue for the county in 2024
2:40:57alone. [music] Enough to fund schools
2:40:59and build roads. It's become the densest
2:41:02concentration of AI hardware on the
2:41:03planet. And the amount of power flowing
2:41:05into it has reached the scale that is
2:41:07hard to grasp. A few years ago, a large
2:41:10data center might have asked the grid
2:41:11for around 30 megawatts of power. Today,
2:41:13[music] a single campus can ask for
2:41:15hundreds. The biggest ones now push
2:41:17toward gigawatt. That's so much
2:41:20electricity that one site can draw the
2:41:22output of two full nuclear power plants.
2:41:24That power has already been requested
2:41:26and approved. What's missing is the
2:41:28infrastructure to move it. [music]
2:41:30Dominion Energy, the utility behind most
2:41:32of the region, has publicly stated it
2:41:34can't deliver new capacity on the
2:41:36expected timelines. Getting a large site
2:41:39connected can now take 4 to 7 years. And
2:41:41getting it into the queue for review
2:41:43takes longer than that. The bottleneck
2:41:45is no longer chips or models. It's the
2:41:47physical grid. Token maxing just makes
2:41:50it worse. Every advanced query runs
2:41:52longer than the simple chat bots that
2:41:53came before it. That means a longer
2:41:55drain on the energy grid. So even if the
2:41:58data never ran out and the models never
2:41:59decayed, the power to supply them would
2:42:02be locked behind permits and an
2:42:03imaginary infrastructure, it puts a hard
2:42:06ceiling on how many AI projects can come
2:42:08online at one time. If the system can't
2:42:10expand fast enough for everyone, what
2:42:12decides who gets in and who doesn't? The
2:42:15best AI won't end up as a cheap utility
2:42:17for everyone. It'll go to whoever can
2:42:19afford what it actually costs to run.
2:42:21The simpler models will stay widely
2:42:23available because they're cheap enough
2:42:25for companies to absorb. But the systems
2:42:27that think longer and spend more compute
2:42:29per answer will be pushed into higher
2:42:31and higher tiers reserved for customers
2:42:33who can cover the costs attached to
2:42:35them. Over time, the gap between those
2:42:37two worlds will widen as the cost of the
2:42:39most capable systems keeps rising. The
2:42:42top tiers will carry prices that match
2:42:44real computing. It wouldn't be a
2:42:46marketing friendly price. Corporations
2:42:48will keep signing bigger and bigger
2:42:50contracts because they can spread the
2:42:51cost. Anyone who relies on the best
2:42:53model for serious work will need to ask
2:42:55if it's really worth the price tag. The
2:42:57promise sold to the world was simple. AI
2:43:00would get cheaper every year until it
2:43:02became universally available. What's
2:43:03emerging instead is something else. A
2:43:05system where the best models are
2:43:07deliberately limited. For companies,
2:43:09it's economics. For users capped without
2:43:11warning, it feels like an exclusion.
2:43:13There is no universal AI anymore. There
2:43:15are tiers of it. AI models aren't just
2:43:18changing subscriptions, they're
2:43:19reshaping careers. As machine learning
2:43:21scales up, the next generation of
2:43:23workers is feeling it first. The
2:43:25computer you've used for years is
2:43:27disappearing. That screen, the icons,
2:43:29the windows, the menus you click,
2:43:31they're all being forced into the
2:43:32background. They're being replaced by a
2:43:3470 billion transistor black box that
2:43:37makes more and more decisions for you
2:43:39and charges an energy tax every time it
2:43:41does. Nvidia isn't just releasing a new
2:43:43chip. They're trying to do to the CPU
2:43:45what they already did to the GPU, and
2:43:48Microsoft just handed them the keys. A
2:43:50laptop sitting on a store shelf isn't
2:43:52really a computer in the traditional
2:43:54sense anymore. Sure, it looks like one.
2:43:56It turns on like one, but the
2:43:57architecture that defined personal
2:43:59computing for decades is starting to
2:44:01disappear. At the center is a chip
2:44:03Nvidia calls RTX Spark. For years,
2:44:06computers were built from separate
2:44:07parts. a CPU for logic, a GPU for
2:44:10graphics, and increasingly dedicated
2:44:12hardware for AI. Data constantly moved
2:44:14between them. Nvidia reduced all of that
2:44:16onto a single piece of silicon. The
2:44:18result behaves less like a collection of
2:44:20components and more like a single
2:44:22computing organism. Everything shares
2:44:24the same memory. Everything works from
2:44:26the [music] same pool of data. Nvidia
2:44:28says it can perform a thousand trillion
2:44:30calculations every second. That number
2:44:32itself doesn't even matter. It's all
2:44:34about the results. A thin laptop with
2:44:36all day battery life can now run an AI
2:44:38model with 120 billion separate settings
2:44:41inside of it. Work that once filled the
2:44:43room now fits in a backpack. Nvidia
2:44:46claims it'll turn the PC from a tool to
2:44:48a teammate. And they aren't doing it
2:44:50alone. Microsoft willingly supplied the
2:44:52missing piece, the software that lets an
2:44:54AI agent hijack your computer. It
2:44:56watches across your apps. It acts before
2:44:59you ask. The desktop, the icons, and the
2:45:01files and folders you've used your whole
2:45:03life are still there. They're just no
2:45:06longer in charge. For two generations,
2:45:08the CPU sat at the center of personal
2:45:10computing. Now, it's [music] just the
2:45:12supporting cast. And the company whose
2:45:14name was stamped inside almost every
2:45:15computer on Earth didn't lose this
2:45:17battle on your desktop. It lost it
2:45:19somewhere much bigger. For three
2:45:21decades, Intel was at the center of the
2:45:23digital world. Its chips [music] powered
2:45:25everything from office equipment to
2:45:27massive server farms behind nearly every
2:45:29website you've ever opened. If computing
2:45:31[music] had a bleeding heart, Intel was
2:45:33it. Then the floor gave way. Between
2:45:352021 and 2025, Intel's share of the data
2:45:38center chip market collapsed from
2:45:40roughly 68% to around 6%. At the same
2:45:44time, Nvidia surged to 86%. One company
2:45:47practically took over the market. By
2:45:49June 2026, Nvidia was worth in the
2:45:52region of $5 trillion. It was the most
2:45:54valuable company on the planet. [music]
2:45:56No business in history had ever been
2:45:58valued that highly, and no ones came
2:46:00close. When the RTX Spark was unveiled,
2:46:03the markets reacted immediately. Nvidia
2:46:05stock jumped while shares in AMD, Intel,
2:46:07and Qualcomm all fell. For decades,
2:46:10Intel and AMD built chips around a
2:46:12design called [music] x86. Its strength
2:46:14was running instructions one after the
2:46:16other as fast as possible. That is just
2:46:18what traditional software needed. AI
2:46:21plays a completely different game.
2:46:22Instead of a long chain of instructions,
2:46:24it performs enormous amounts of simple
2:46:26math all at once. This is where Nvidia
2:46:29had a head start. Its graphics chips
2:46:31were already built to solve thousands of
2:46:32problems at the same time. That is what
2:46:34rendering a video game requires. When AI
2:46:37became the most important thing in
2:46:38computing, Nvidia didn't have to
2:46:40reinvent itself. The future was already
2:46:42optimized. Intel realized the threat and
2:46:44fought back. It built AI accelerators
2:46:46and launched competing products. It
2:46:48spent years desperately trying to close
2:46:50the gap. The gap kept growing. The
2:46:53company that had spent decades setting
2:46:54the pace for the industry was now
2:46:56struggling to keep up with it. None of
2:46:57that was luck or clever marketing. It
2:46:59was the work that changed everything.
2:47:01Nvidia was already waiting. Jensen Hong
2:47:04called it a reinvention of the computer.
2:47:06As significant as the day the phone
2:47:08became the smartphone. But winning the
2:47:10data center wasn't enough. To take over
2:47:12the computer on your desk, Nvidia needed
2:47:14something else. Control. Nvidia never
2:47:17had to build the best laptop in the
2:47:18world. It only had to make sure every
2:47:20laptop worth buying ran on its products.
2:47:23It also had to ensure that nobody else
2:47:25owned a single piece of it. So, it built
2:47:27the whole thing itself. Nvidia designed
2:47:30the graphics and the AI hardware. It
2:47:32partnered with MediaTek on the
2:47:33processor. TSMC manufactured the chip.
2:47:36Microsoft shaped Windows around it.
2:47:38Before a single laptop reached the store
2:47:40shelf, everything had already been
2:47:42decided. And that left companies like
2:47:44Dell, HP, Lenovo, and Asus in a strange
2:47:47position. They are the companies you
2:47:49actually buy from, but all they got was
2:47:51the same finished module. They design
2:47:53the case, the cooling, the battery, and
2:47:55the ports. They load Windows onto it,
2:47:57and they ship it. They don't own [music]
2:47:58the processor, the graphics, or the AI
2:48:00software that decides what the machine
2:48:02can do. The brands on the lid are just
2:48:04the people who assembled the box. Nvidia
2:48:07keeps around 75 cents of gross profit on
2:48:09every dollar it sells. That's almost
2:48:11unheard of for a company that makes
2:48:13physical hardware. The companies
2:48:14assembling the laptops live on a few
2:48:17percent. The rest of the value is gone
2:48:18before the device ever leaves the
2:48:20factory. So, every time more units ship,
2:48:23the imbalance grows. None of it shows up
2:48:25in the price tag, but it decides who
2:48:27holds the [music] power. There's another
2:48:29element to this power shift. Almost
2:48:31every major AI system on Earth is built
2:48:33on Nvidia software platform CUDA. And
2:48:36once your tools are built on CUDA,
2:48:37switching becomes almost like starting
2:48:39over. Nvidia even tried to expand that
2:48:41control further, attempting to buy ARM,
2:48:44the company behind a rival processor
2:48:45architecture in a $40 billion deal.
2:48:48Regulators shut the move down. So
2:48:50instead of owning the rival, Nvidia made
2:48:52its ecosystem the default and the
2:48:54industry just adjusted around it. It's
2:48:56the exact move the company's pulled
2:48:58before. With graphics cards, it absorbed
2:49:00the value that used to be shared across
2:49:02everyone building them. Now, it's
2:49:04running the same play on the processor,
2:49:06the last major part of the PC it didn't
2:49:08already control. A shift like this only
2:49:10works if there is nowhere left to go or
2:49:12backup platform. Microsoft handled that
2:49:15next part. Your current laptop may
2:49:17already be locked out of the future, and
2:49:19you'd have no way of knowing it from the
2:49:21screen. To run [music] Windows new wave
2:49:23of AI features, a computer now needs a
2:49:25dedicated AI chip rated at at least 40
2:49:28tops. That's the measure of how many AI
2:49:30calculations it can process every
2:49:32second. It also needs 16 GB of memory
2:49:35and 256 GB of storage. Specs that many
2:49:38older machines don't meet. Cross the
2:49:40threshold and the features switch on,
2:49:42fall short, and they just disappear into
2:49:45a cloud or don't run at all. These
2:49:47aren't minor additions. It's things like
2:49:49live translation while you speak, image
2:49:51generation on your device, smarter
2:49:53search options, and a memory that can
2:49:54look back through all of your work. None
2:49:57of it works without the hardware
2:49:58Microsoft demands. When the requirements
2:50:00were announced, the chips that met them
2:50:02were ARMS, not Intel's. So, the entire
2:50:05industry began shifting toward ARM. The
2:50:06[music] same foundation Nvidia and its
2:50:08partners were already built on. The
2:50:10timing wasn't subtle. Microsoft ended
2:50:12all support for Windows 10 in October
2:50:142025. That alone pushed a huge wave of
2:50:17people toward buying something [music]
2:50:18new, whether they wanted to or not. The
2:50:21result was the great refresh, the
2:50:23biggest forced hardware swap the
2:50:25industry had seen in years. And the
2:50:26machines that clear the bar most
2:50:28comfortably are the ones built on
2:50:30Nvidia's design. Older programs written
2:50:32for x86 don't run directly on these new
2:50:35[music] chips. Windows translates them
2:50:37on the go through a layer called Prism,
2:50:39so they still open. They just run slower
2:50:42and they never get the full power of the
2:50:43silicon underneath. The old software
2:50:46works as a guest. Of course, RTX Spark
2:50:49sales past Microsoft's [music]
2:50:50benchmark. The baseline is 40 tops, and
2:50:52Nvidia says this chip reaches over a
2:50:54thousand. That's more than 20 times what
2:50:57Windows asks for. And that's the point.
2:50:59The chip isn't built to barely meet
2:51:01today's features. It's [music] built for
2:51:02what comes after them. Everything below
2:51:04that line is already behind. At that
2:51:07level, the chip can run AI models too
2:51:09large for a normal laptop and keep an
2:51:11assistant running continuously, even
2:51:13when you're not at the keyboard. Once
2:51:15most new software is built for that
2:51:16higher level by default, upgrading stops
2:51:19being optional. These kinds of new
2:51:20computers never fully stop. The AI
2:51:22assistants are partly awake, waiting for
2:51:25something to do. Some systems are even
2:51:26designed to keep working overnight,
2:51:28finishing the long tasks while you
2:51:30sleep. That comes at a cost. How much
2:51:33has been heavily debated. The
2:51:34International Energy Agency and repeated
2:51:36[music] by the Electric Power Research
2:51:38Institute in 2024 put a single AI
2:51:41request at around 2.9 W hours of
2:51:44electricity compared to roughly 0.3 for
2:51:46a basic web search. That's about 10
2:51:48times more. So the billions of queries
2:51:51each day could work out at nearly 10
2:51:53terowatt hours of extra demand a year.
2:51:55Other researchers disagree. The research
2:51:57institute Epoch AI along with OpenAI's
2:52:00Sam Alman say the real figure is much
2:52:02lower. Newer models, they argued, had
2:52:04cut a typical request back down to about
2:52:07the level of an ordinary search. So, the
2:52:09cost per [music] question is falling
2:52:10fast. But how often we ask is rising.
2:52:13There's an old idea in economics that
2:52:15when something becomes cheaper and
2:52:17easier to use, people don't use less of
2:52:19it, they use more of it, and the total
2:52:21rises anyway. An assistant that takes no
2:52:24effort to summon gets used constantly,
2:52:26even for the small tasks you'd have
2:52:28never bothered a computer with before.
2:52:29So the cost [music] isn't sitting in
2:52:31some distant data center that you never
2:52:33think about. It is sitting in your lap.
2:52:35The machine runs hotter. The fan kicks
2:52:37on more often. The battery that used to
2:52:39last all day starts asking for a charger
2:52:41by mid-after afternoon. And the effects
2:52:43show up elsewhere. Microsoft's own
2:52:45emissions have risen by nearly 30% since
2:52:482020, driven by the data centers
2:52:50powering these systems. The more capable
2:52:52the software becomes, the more
2:52:53electricity it burns in the background
2:52:55every time you use it. It's [music] a
2:52:57problem every company ran into at the
2:52:59same time. A traditional processor is
2:53:01great at following rules. It goes
2:53:03through instructions one at a time in
2:53:05order, making a clear yes or no
2:53:07decision. That was perfect for
2:53:09everything from a spreadsheet to flight
2:53:10booking, where each move waits on the
2:53:12one before it. Modern AI doesn't think
2:53:15in clean [music] steps at all. It works
2:53:17by guessing. It runs staggering amounts
2:53:19of basic arithmetic across billions of
2:53:21values to land on the answer that fits
2:53:23the best. Every word an AI writes sets
2:53:25off another round of that weighing with
2:53:27billions of numbers checked at once. A
2:53:29single short answer can take trillions
2:53:31of those tiny sums. There's no neat line
2:53:33of logic to follow. Just a huge cloud of
2:53:36may resolved in an instant. Work like
2:53:38that rewards a completely different kind
2:53:40of processor. You don't want a single
2:53:42processor solving problems one by one.
2:53:44[music] You want thousands of small
2:53:45workers doing the same simple
2:53:47calculation at the same time. A
2:53:49traditional processor doesn't disappear.
2:53:51It just becomes the coordinator handing
2:53:53the work off to a much larger AI section
2:53:55and handling the small tasks that still
2:53:57need strict order and precision. Running
2:53:59[music] thousands of those calculations
2:54:01in parallel instead of a few in a line
2:54:03is what makes it all fit on a laptop
2:54:05instead of in a warehouse. And once the
2:54:07machine does the reasoning, the person
2:54:09who used to do it starts to change as
2:54:11[music] well. For decades, using a
2:54:13computer meant knowing where things
2:54:14were. You opened a program, you dug
2:54:16through folders, [music] you clicked
2:54:17menus, and you set things up for
2:54:19yourself. The assistant on these new
2:54:21machines is built to make all of that
2:54:22disappear. Give it enough time and it
2:54:24remembers yesterday's work, picks the
2:54:27right tool to complete the task, and
2:54:28hands back a finished result. By the
2:54:30time you even sit down, the assistant
2:54:32may have sorted your messages, drafted
2:54:33[music] the easy replies, and lined up
2:54:35what it thinks matters next. The desktop
2:54:37full of windows you once had to navigate
2:54:39becomes something handled in the
2:54:41background. You don't lose control
2:54:42entirely. You [music] can still reject a
2:54:44step. What disappears is the visible
2:54:46control that defined the old era.
2:54:48Knowing where the files are, knowing
2:54:50what's running, knowing why the machine
2:54:51is doing what it's doing. The chip doing
2:54:53the thinking was always a sealed box
2:54:55that no one could look inside. Now the
2:54:57way you talk to it is sealed, too. The
2:54:59understanding that people built about
2:55:01how their computers work [music] becomes
2:55:02less and less necessary. And a skill you
2:55:05don't use is a skill you stop
2:55:07remembering. A laptop bought in 2024 or
2:55:09early 2025 is already a relic. Open it
2:55:12up and you can still see how it runs.
2:55:14You can change the settings by hand. You
2:55:16can edit the files that control how it
2:55:18behaves. It belonged to you and it
2:55:20answered to you. All the work an RTX
2:55:22Spark chip [music] does happens inside a
2:55:24sealed box. Its entire reasoning is
2:55:26built on a series of odds, not fixed
2:55:29rules. The software is owned by the
2:55:31duopoly of Nvidia and Microsoft. You
2:55:33give it a goal and the assistant acts.
2:55:35[music] The desktop is still there, just
2:55:37for older programs. It's no longer where
2:55:39the computer actually runs. There's no
2:55:41easy way back to the old model. Software
2:55:44gets written for the hardware that runs
2:55:45at the best. Users follow the software.
2:55:47Older machines fall behind and become
2:55:49redundant. The era of the personal
2:55:51computer as a generalpurpose tool owned
2:55:53and operated by its user [music] is
2:55:55coming to an end. What replaces it is a
2:55:57system where intelligence and access
2:55:59depends on what your machine can run.
2:56:01And no one outside Nvidia can fully see
2:56:03how it produces what [music] it
2:56:05produces. The real divide isn't
2:56:07technical. It's access. Who gets to
2:56:09think [music] with it and who gets left
2:56:11behind. Nvidia and Microsoft are
2:56:13starting to corner the entire market,
2:56:15and that affects your ability to afford
2:56:17what goes inside your computer. An AI
2:56:20takeover is inevitable. Or at least
2:56:22that's what you've been told. Once you
2:56:24upload a photo or a voice note, it's
2:56:26gone forever, scraped into AI systems,
2:56:28feeding an insatiable hunger for data.
2:56:30But what if that is not the full story?
2:56:32Right now, a digital insurgency is
2:56:34taking place. Traps are being laid
2:56:35inside the data these models depend on.
2:56:38Researchers have figured out how to
2:56:39labbotomize AI models using just a
2:56:42handful of modified JPEGs and it works.
2:56:45So what happens when the scrapers don't
2:56:47just learn from the internet but start
2:56:49breaking because of it? Let's say
2:56:51someone types a simple request into an
2:56:52AI model. They ask for a photorealistic
2:56:55dog running through the park. It's the
2:56:57kind of thing a competent model nails
2:56:58every single time. But this time the
2:57:00results are different. The legs don't
2:57:02connect properly. Extra joints appear
2:57:04where they shouldn't exist. The face
2:57:06subtly drifts into something that's from
2:57:08the uncanny valley. The fur loses
2:57:10structure and definition. It looks like
2:57:12something that is seen a dog a thousand
2:57:15times, but it doesn't actually
2:57:16understand what [music] it is. And
2:57:18that's the work of a poisoned AI model.
2:57:20It's not a bug. The model is still doing
2:57:22what it was trained to do. The problem
2:57:24is what it was trained on. A poisoned AI
2:57:27model is what happens when someone
2:57:28messes with the data set it learns from.
2:57:30They slip in manipulated or misleading
2:57:32examples during training so that the
2:57:34model starts picking up the wrong
2:57:35patterns. No one has to hack anything or
2:57:38break into the system. Artists and
2:57:39creators simply post their work online
2:57:41like they always have. And then the
2:57:43scrapers arrive. These are automated
2:57:45bots that crawl the internet scooping up
2:57:47massive amounts of images, text, and
2:57:49audio from websites. They don't
2:57:51understand what they're gathering. They
2:57:52just vacuum [music] it up to build
2:57:54training data sets for AI models. That's
2:57:56where the problem starts. Mixed in with
2:57:58all that normal content, poisoned
2:58:00[music] examples get collected, too.
2:58:02Nothing looks wrong at the time. The
2:58:03trap only reveals itself later during
2:58:05[music] the next training run. The team
2:58:07at the University of Chicago put that
2:58:09theory to the test. They fed Stable
2:58:11Diffusion about 50 altered pictures of
2:58:13dogs and then they put the AI model to
2:58:15work. Almost immediately, the images
2:58:17were worked. Every single one had
2:58:19something [music] wrong with it. The
2:58:20team kept going. Once they got to around
2:58:22300 poisoned images, the model started
2:58:25producing images of a cat. The model
2:58:27trains on billions of images, but for
2:58:29any single concept, say a dog, it really
2:58:31only relies on a few thousand of them.
2:58:33Corrupt a small slice and it all starts
2:58:35to fall apart. You don't need millions
2:58:37of bad files. You need less than 1% of
2:58:40right ones aimed at the right concept.
2:58:43These models aren't isolated. They are
2:58:45powering midjourney doll E and the image
2:58:47tools built into [music] your phone. So
2:58:49if one system becomes corrupted, the
2:58:50effects don't stay contained, they
2:58:52spread. It's the opening shot of a new
2:58:54kind of fight. And the people doing it
2:58:56aren't rival labs or bad actors. They're
2:58:58[music] illustrators, photographers,
2:59:00creatives, and ordinary users with a
2:59:02free app. And they all have a reason to
2:59:04be furious. But before we go any
2:59:06further, imagine this. You're online
2:59:08every single day. You check your email,
2:59:10open a few apps, look things up, stream
2:59:12videos, maybe even do all of that while
2:59:14traveling. And to you, that feels
2:59:16totally normal. But behind the scenes,
2:59:18your internet provider, advertisers,
2:59:20network admins, and sometimes even
2:59:21governments can build a surprisingly
2:59:23[music]
2:59:23detailed picture of what you're doing
2:59:25online. Now, to be clear, noVPN can
2:59:27protect you from everything. You still
2:59:29have to be smart online. Don't click
2:59:30suspicious links. Don't hand over
2:59:32personal information to sketchy emails,
2:59:34and definitely don't trust the so-called
2:59:36[music] prince of Nigeria. But a good
2:59:37VPN is an important layer of privacy
2:59:40because privacy shouldn't be something
2:59:41you only think about after something
2:59:42goes wrong. It should be the [music]
2:59:44default. And that's exactly what
2:59:46ProtonVPN is built for. ProtonVPN helps
2:59:49keep your browsing private wherever you
2:59:50are. Whether you're at home, traveling,
2:59:52or just trying to stop your online
2:59:54activity from being tracked. Their no
2:59:56logs policy has been verified by
2:59:57independent auditors. And Proton is
2:59:59backed by a foundation dedicated to
3:00:01privacy, transparency, and user rights.
3:00:03ProtonVPN is also fully open source,
3:00:06meaning its code is publicly available
3:00:07for inspection. So instead [music] of
3:00:09just asking you to trust them, Proton
3:00:11gives people a way to verify what
3:00:12they're doing. And privacy doesn't have
3:00:14to mean slow. ProtonVPN offers
3:00:16high-speed connections with VPN
3:00:18Accelerator plus Net Shield, which
3:00:20blocks malicious ads and trackers before
3:00:22they ruin your browsing experience. It
3:00:24also gives you more freedom online. You
3:00:26can connect through servers around the
3:00:27world, access [music] content more
3:00:28securely while traveling, and use it for
3:00:30streaming, gaming, and P2P file sharing,
3:00:33all while keeping your online activity
3:00:35more private. So whether you're trying
3:00:36to stop advertisers from following you,
3:00:38keep your browsing private from your
3:00:40ISP, or just take back control of your
3:00:42online life, ProtonVPN makes privacy
3:00:44simple. Click the link in the
3:00:46description or go to
3:00:47protonvpn.com/theinfographic
3:00:50show to get up to a 70% discount when
3:00:52you sign up to ProtonVPN's 2-year plan.
3:00:55Nobody poisons their own work for fun.
3:00:57They do it because they felt like their
3:00:59work was taken from them. That feeling
3:01:01doesn't come from nowhere. It all
3:01:02started with a theft. For years, the
3:01:05firms building these models treated the
3:01:06open web as a warehouse. They scraped
3:01:08billions of images, voices, and
3:01:10paragraphs from it. The original
3:01:12creators were never asked, and they were
3:01:14never paid. It was a heist, plain and
3:01:16simple. Much of that hall went into one
3:01:18open data set called Lio N 5B, which
3:01:22held a massive data set made up of 6
3:01:24billion images and accompanying text.
3:01:26Stable diffusion learned from it, and so
3:01:28did most of the art that followed. It
3:01:30was a compressed version of the visible
3:01:32internet folded into something machines
3:01:34learn from. Then the artists started
3:01:36paying attention. They started
3:01:38recognizing their own work in the
3:01:39output. Styles and techniques that took
3:01:41decades to develop were being reproduced
3:01:44in seconds. By 2023, a generator could
3:01:46mimic the brush strokes and the talent
3:01:48of an artist in a heartbeat. A career's
3:01:50worth of experience was reduced to a
3:01:52prompt and undercut by a tool that cost
3:01:54nothing to use. Conceptual artist Carla
3:01:57Ortiz along with other artists sued
3:01:58[music] Stability AI in 2023. Getty
3:02:01Images filed its own case around the
3:02:03same time. The lawsuits dragged out and
3:02:05all the while the scraping continued and
3:02:08the models improved. The companies
3:02:09behind the models freely admitted to
3:02:11their practices. OpenAI told the British
3:02:13Parliament that building today's top
3:02:15models without copyrighted work would be
3:02:17impossible. Sure, there were optout
3:02:19forms, but it put the onus on the
3:02:21creator. Instead of asking permission
3:02:23upfront, [music]
3:02:23artists had to hunt down the models file
3:02:26requests one by one and then hope they
3:02:28were honored. The balance of power
3:02:30remained with the developers. So the
3:02:32artists adapted. Ordinary uploads turned
3:02:34into weapons one file at a time. In
3:02:362023, the University of Chicago released
3:02:38a program called Glaze, and it was aimed
3:02:41directly at the AI models. Type in an
3:02:43artist's name into an AI model, and
3:02:45it'll generate a painting based on that
3:02:47particular style and the data it's
3:02:49learned on. Glaze changes that. It works
3:02:51by exploiting how a computer sees an
3:02:53image. To an AI, a painting isn't really
3:02:55a painting at all. It's a series of
3:02:57numbers that make up [music] the style
3:02:59and structure. Glaze distorts the
3:03:01numbers slightly. It's subtle, something
3:03:03that isn't visible to the human eye, but
3:03:05the machine learns from the cloaked
3:03:06information. It works like an optical
3:03:09illusion. Two viewers can look at the
3:03:10same image and see completely different
3:03:12things. [music] In this case, the
3:03:13viewers are you and the machine. You see
3:03:16your painting exactly as was intended.
3:03:18>> [music]
3:03:19>> The model sees something else. An oil
3:03:21painting might seem like it's in
3:03:22charcoal. A watercolor might be seen as
3:03:24completely different medium. The visual
3:03:26style that the scraper came to learn
3:03:28from has effectively been moved. Word
3:03:30spread fast and Glaze has now passed 6
3:03:32million downloads, but there are limits.
3:03:34Glaze could hide a style, but it
3:03:36couldn't stop someone from taking the
3:03:38image itself. And every time people
3:03:40found a new way to alter the image,
3:03:41scrapers came back with better ways to
3:03:43recover it. It was like playing defense
3:03:45forever. The creators needed something
3:03:47that could end the fight. They needed a
3:03:49poison pill. It came in January 2024 and
3:03:52it was nightshade. Another product from
3:03:55the University of Chicago. It had one
3:03:57goal to infect. An image run through
3:03:59Nightshade will look completely normal
3:04:01to both human eyes and the machine. But
3:04:03it's like a Trojan [music] horse. If a
3:04:05machine is trained on that one image,
3:04:07the model starts learning nonsense. The
3:04:09results are almost comical. Hats turn
3:04:11into cakes, handbags into toasters,
3:04:13[music] and a cars learned to be drawn
3:04:15as a cow. For artists and creators, it
3:04:17felt like payback. For years, AI
3:04:20companies had scraped artwork to train
3:04:21their models. Now, the very thing being
3:04:23taken could be turned against the people
3:04:25taking it. The demand was instant.
3:04:27Nightshade hit 250,000 downloads in just
3:04:305 days. University servers buckled under
3:04:32the traffic. The team behind it had to
3:04:34scramble [music] to post backup download
3:04:36links as thousands more rushed to get
3:04:38it. Most people assumed this was a form
3:04:40[music] of hacking. It's not. No
3:04:41firewall gets breached and the tool
3:04:43never touches a single company server.
3:04:45Researchers call it adversarial machine
3:04:47learning. But a simpler way to look at
3:04:49it is that Nightshade is a magic trick
3:04:51for the machines. The model isn't
3:04:53attacked from the outside. It is fooled
3:04:55into teaching itself the wrong thing and
3:04:57then it trusts that mistake as if it
3:04:59were true. It doesn't take much to start
3:05:02the process. Fewer than 100 carefully
3:05:04crafted images can be enough to poison a
3:05:06single concept inside a top AI model.
3:05:09And if you pair nightshade with glaze,
3:05:11that same [music] image can hide an
3:05:12artist's style and poison the training
3:05:14data at the same time. But those results
3:05:16came from models researchers could test
3:05:18directly. The bigger AI systems are
3:05:20harder to study, and nobody outside
3:05:22those companies knows exactly how
3:05:24vulnerable they are. But that is not
3:05:26really the point. [music] Nightshade
3:05:28took the idea from a theory on paper to
3:05:30something that actually worked. The
3:05:32damage doesn't stay [music] contained
3:05:33either. Poison the concept of a dog and
3:05:35the related ideas like huskys, puppies,
3:05:37and wolves can start drifting with it.
3:05:39In some tests, researchers fed hundreds
3:05:41of corrupted images in a single model
3:05:43until it could barely generate
3:05:45recognizable images at all. And that's
3:05:47where [music] this stopped being purely
3:05:48defensive. Every poisoned image uploaded
3:05:50to the internet becomes a potential
3:05:52sleeper cell. It can remain unnoticed
3:05:54inside a data set for months or years,
3:05:57waiting for the next training run. The
3:05:59person who uploaded it might never even
3:06:00know if it was collected or which model
3:06:02eventually learned from it. Images were
3:06:05only the start. The same trick works on
3:06:07anything a machine learns from. From
3:06:09your writing to [music] your face, and
3:06:11your voice is the next target. A service
3:06:13like 11 Labs needs only a few clean
3:06:15sounds of your voice to build a
3:06:16convincing copy. A podcast clip can be
3:06:19enough. So can a YouTube video, a few
3:06:21seconds of footage on social media, or
3:06:23even an old voicemail buried in
3:06:24someone's phone. It's become a favorite
3:06:26tool of scammers. And it isn't a
3:06:28hypothetical risk. In 2024, criminals
3:06:31used cloned voices and faces to steal
3:06:33about $25 million in a single faked
3:06:35video [music] call. Banks and law
3:06:37enforcement agencies now warn people
3:06:39about calls from cloned relatives asking
3:06:41for money. A tool called safe speech was
3:06:43built by security researchers for this
3:06:45exact reason. Your voice has its own
3:06:47signature, a kind of audio fingerprint
3:06:49that AI systems use to recognize and
3:06:51copy you. Safe speech subtly smudges
3:06:54that fingerprint. To another person, you
3:06:56sound exactly the [music] same, but to a
3:06:58voice cloning model, you become much
3:06:59harder to copy. To the machine, it works
3:07:01a bit like radio interference. You hear
3:07:03the voice clearly, but the AI doesn't.
3:07:06The details it needs most get scrambled.
3:07:08Train a voice clone on that recording,
3:07:10and the results come out wrong. They're
3:07:12close enough to sound human, but missing
3:07:14the subtle traits that make you sound
3:07:16like you. The timing is what makes it
3:07:18clever. This isn't a filter slapped onto
3:07:20a fake after it's generated. It's in the
3:07:23original recording itself. The AI learns
3:07:25from the protected version, which means
3:07:27the cloning process breaks before it
3:07:29ever succeeds. There is no clean copy
3:07:31for the model to learn from. That's what
3:07:33makes this different from protecting
3:07:34artwork. A stolen portfolio is one kind
3:07:36of loss. Your voice is another. It's the
3:07:39sound your family and friends recognize
3:07:41instantly. The thing a scammer wants the
3:07:43most when they're trying to impersonate
3:07:44you. It turns the tables. The target
3:07:47[music] gets to set the trap. For the
3:07:49first time, artists, writers, and
3:07:50everyday people had a way to fight back
3:07:52against systems trained on their work
3:07:54and identities. [music] It sounded like
3:07:56the next step in digital security. And
3:07:58then the AI labs responded. They weren't
3:08:00about to lose access to the data that
3:08:02fueled their models. If artists could
3:08:04poison the training set, the labs would
3:08:06try to remove the poison. And that
3:08:07kicked off something bigger than a
3:08:09security tool, an arms race. They
3:08:11couldn't just accept corrupted data
3:08:13sets, and they couldn't afford to throw
3:08:15most of them away either. So, they tried
3:08:17a straightforward fix. They began
3:08:18checking each scraped image against its
3:08:21caption and then they dropped anything
3:08:23that didn't match. In theory, [music]
3:08:24poisoned data should stand out. In
3:08:27reality, it only catches part of it.
3:08:29Tests on these systems show it flags
3:08:31maybe 40 to 60% of poisoned images. The
3:08:34rest slips through. [music] And then
3:08:36there's a second problem. It also
3:08:38deletes plenty of clean data, leaving
3:08:39the model with either contaminated data
3:08:41or not enough to learn from. So, they
3:08:44tried something else. They started
3:08:45scrubbing every image through a cleaning
3:08:47model before training, trying to wash
3:08:49out anything suspicious. It sounds
3:08:51clever, but it's slow and expensive,
3:08:53[music] and it still has limitations
3:08:55because the poisoning adapts. Artists
3:08:57designed it to survive that cleaning
3:08:59step, so it reappears on the other side
3:09:01like a stain bleeding back through. Not
3:09:04long after Glaze launched, one research
3:09:05group said it had already broken through
3:09:07the cloaking. So, the Chicago team
3:09:09pushed out a tougher version. Whenever
3:09:11the program was breached, they developed
3:09:12a new patch. It was a war of attrition.
3:09:15Even if a voice has protection built in,
3:09:17there are still ways to strip parts out
3:09:19of it. Push it through enough processing
3:09:20and some of that protection starts to
3:09:22fade. In some cases, it gets noticeably
3:09:24weaker. Not gone, but damaged enough to
3:09:27matter. The poisoners are winning for
3:09:29now, and the reason is lopsided. You
3:09:31have small teams constantly trying new
3:09:33tricks in the open. On the other hand, a
3:09:35handful of AI labs are trying to patch
3:09:37holes as they appear. So, every time a
3:09:39lab rolls out a fix, it gets tested
3:09:41immediately. And almost immediately, a
3:09:43new version shows up, adjusted just
3:09:44enough to get around it, usually within
3:09:47weeks. And this all comes with a price
3:09:49tag. It doesn't land on the labs first.
3:09:51It lands on the people training the
3:09:53models with the data. The whole AI
3:09:55business model was built on the promise
3:09:56that the data would be free, endless,
3:09:58and clean. The poison breaks the idea of
3:10:01clean. Once that goes, the other two
3:10:03stop feeling true as well. Because now
3:10:05nothing can be trusted at face value. A
3:10:08poisoned file looks identical to a real
3:10:10one. There's no label or warning, so
3:10:12every data set turns into something that
3:10:14has to be checked manually or with tools
3:10:16that still miss a lot. Training a top
3:10:18model already costs hundreds of millions
3:10:20of dollars, and every extra cleanup step
3:10:22adds to that bill. A single poisoned
3:10:25batch can set a project back by weeks.
3:10:27[music] And that's the point. The goal
3:10:29was never to wreck one model for fun. It
3:10:31was to make stolen data more expensive
3:10:34than paid data. There's already a legal,
3:10:36cleaner method being used. Adobe trained
3:10:38its Firefly tool on midjourney images,
3:10:41Shuttertock on similar deals. If you
3:10:43want clean quality data, you need to pay
3:10:46for it. That is a serious problem for
3:10:48the industry. Investors [music] poured
3:10:50tens of billions into AI on a single bet
3:10:52that the data would stay cheap forever.
3:10:54That [music] bet isn't looking like a
3:10:56sure thing anymore. For a long time, the
3:10:58belief was that if you created
3:10:59something, you own it. The scrapers tore
3:11:02that deal up without asking anyone. The
3:11:04AI companies act as if human work was
3:11:06free, endless, and owned by no [music]
3:11:08one. That was the original mistake.
3:11:10Paintings were scraped, photos were
3:11:11downloaded, voices, faces, videos were
3:11:13collected in huge amounts. If it was
3:11:15online, it was fair game. Consent seemed
3:11:18optional. In every other industry, the
3:11:20opposite is the standard. A photographer
3:11:22licenses an image before a brand uses
3:11:24it. Musicians clear samples before a
3:11:26track goes live. Even film studios pay
3:11:29for every frame of stock footage they
3:11:30[music] use. The work of millions of
3:11:33people got treated as if it belonged to
3:11:34no one. This battle wasn't created by
3:11:36the artists. It was the AI labs and
3:11:39their trainers. The moment work was
3:11:41taken without asking, the balance
3:11:43changed. Everything after that was just
3:11:45a reaction. You can't build an empire on
3:11:47the idea that people do not count
3:11:49[music] and then act shocked when those
3:11:50same people fight back and question how
3:11:52the system works. Because the problem
3:11:54isn't sitting in a contract or on a
3:11:56policy page. It's built into the way the
3:11:58models learn in the first place. Every
3:12:00safeguard before this tried to politely
3:12:03control behavior through copyright
3:12:04notices or opt- out forms, but none of
3:12:07it really stopped the scraping. This is
3:12:09different. For the first time, a single
3:12:11[music] person can influence what the
3:12:13system learns next. And the system
3:12:15doesn't get to ignore it. With people
3:12:17turning the tables on AI scrapers,
3:12:19digital theft is getting harder,
3:12:20especially when it comes to scams. We're
3:12:22witnessing the biggest rugpole in tech
3:12:25history. AI tokens are supposed to be
3:12:27getting cheaper thanks to technological
3:12:29breakthroughs and efficiency. But as of
3:12:31March 2026, OpenAI is losing roughly
3:12:34a$122
3:12:35for every dollar of revenue it makes.
3:12:37Venture capital is keeping prices
3:12:39artificially low while companies get
3:12:41hooked on the API. The endgame is
3:12:44brutal. Either the AI startups go broke
3:12:46or you do. Internal records leaked to
3:12:49the press in early 2026 show OpenAI is
3:12:52in trouble. [music]
3:12:53The company spends far more than it
3:12:54takes in. The burn for 2026 lends
3:12:57somewhere between $17 billion and $25
3:13:00billion cash. Meanwhile, revenue reaches
3:13:03between 13 billion and 20 billion.
3:13:05OpenAI is losing money on every dollar
3:13:07it earns. The losses are not a one-off.
3:13:10By the company's own projections,
3:13:11they'll add up to around $15 billion by
3:13:142029. Profit is not expected until close
3:13:17to 2030. You can't lose money on every
3:13:20sale forever. Eventually, something has
3:13:22to give. Prices go up or costs are cut
3:13:25or the business folds. It's not limited
3:13:27to OpenAI either. Why? Combinator is the
3:13:30startup accelerator that helped to
3:13:31launch companies like Airbnb, Stripe,
3:13:33and Dropbox. By most estimates, roughly
3:13:3558 to 67% of its winter 2024 batch were
3:13:39AI startups. Most don't own the model
3:13:42they're selling. Their business is built
3:13:44on top of someone else's technology. The
3:13:46industry calls them rappers. Their costs
3:13:48are set by whoever owns the model
3:13:50underneath them. Their entire business
3:13:52depends on those prices staying low. But
3:13:54the companies setting those prices are
3:13:56already losing money. That means [music]
3:13:58today's prices may not be sustainable.
3:14:00And for thousands of AI startups, that's
3:14:02a problem. Months or even years of work
3:14:04can disappear in a single billing cycle.
3:14:07Not because the product got worse, but
3:14:09because a supplier changed a number on
3:14:11an invoice. Much of the AI industry is
3:14:13built on top of the same handful of
3:14:15model providers. So when one supplier
3:14:17raises prices, [music] the whole field
3:14:18feels it. It wouldn't be a slow death.
3:14:21It would be a matter of weeks. The
3:14:23entire AI boom was built on a gamble,
3:14:25lose money today, dominate the market
3:14:27tomorrow. The assumption was that scale
3:14:29would eventually fix the economics. But
3:14:31what if it doesn't? This business model
3:14:34isn't new. It's called a loss leader.
3:14:36You sell below cost, you attract
3:14:38customers, and you worry about the
3:14:39profits later. The aim is to become
3:14:41impossible to leave and to have the
3:14:43customer base hooked onto your product.
3:14:45For a stretch, in 2023 and 24, AI tokens
3:14:48were priced at pennies, far below what
3:14:50they cost to produce. For a developer,
3:14:52the choice was obvious. Nothing else
3:14:54came close on price or power. So, the
3:14:57whole world built on these models. They
3:14:59became a default the way electricity or
3:15:01cloud storage had before them. But
3:15:03renting intelligence is different.
3:15:04[music] Your power company can't read
3:15:06your meter, learn your habits, and then
3:15:09sell them back to you. An AI provider
3:15:11sitting under your product can do
3:15:12exactly that. Once enough companies were
3:15:15hooked, the terms began to tighten. A
3:15:17business might spend months shaping a
3:15:19model around its own private data. That
3:15:21tuned model can't just be boxed up and
3:15:24carried somewhere else. The prompts in
3:15:25the data all live on servers it doesn't
3:15:28own. Leaving means ripping all that up.
3:15:31AI contracts became hostage deals.
3:15:34Staying costs a fortune. Leaving costs
3:15:36even [music] more. The provider can
3:15:38rewrite the terms almost whenever it
3:15:40likes. When OpenAI retires a model,
3:15:42every company built on top of it has to
3:15:44adapt. Code that worked yesterday needs
3:15:47to be rewritten tomorrow. Entire
3:15:49products can find themselves racing
3:15:50against a deprecation deadline that they
3:15:53never chose. While startups are taking
3:15:55that risk, AI giants are taking risks of
3:15:57their own. In late 2025, OpenAI
3:16:00committed roughly $250 billion to
3:16:02Microsoft's cloud infrastructure. And
3:16:04then it signed a second giant deal with
3:16:06Amazon worth tens of billions more. This
3:16:09isn't a company spending money because
3:16:10it's found a profitable business model.
3:16:12It's spending money because it needs
3:16:14more compute to keep the AI race going.
3:16:17[music] In April 2026, Microsoft and
3:16:18OpenAI rewrote their partnership and the
3:16:21new terms set OpenAI loose. Its
3:16:23Microsoft license stopped being
3:16:25exclusive, so it can now run on any
3:16:27cloud it likes. It's [music] even
3:16:29building its own chip with Broadcom.
3:16:31Suddenly, OpenAI bought itself more room
3:16:33to move and with it [music] control of
3:16:36the layer that everything now depends
3:16:37on. We've seen versions of this before.
3:16:40Ride hailing kept fairs below cost until
3:16:42subsidies faded and prices rose.
3:16:45Streaming did the same thing with
3:16:46content. But AI is harder to escape. If
3:16:49a ride gets expensive, you just switch
3:16:51apps. If a subscription changes, you
3:16:53cancel it. But if your product is built
3:16:55on top of a model, you don't switch, you
3:16:58rebuild. Most companies can't rebuild
3:17:00fast enough because they didn't just
3:17:02build on the system they built into it.
3:17:04Large-scale AI systems run on physical
3:17:07infrastructure with hard power limits. A
3:17:09training cluster with around 20,000
3:17:11Nvidia H100 GPUs pulls about 20 megawws
3:17:15nonstop. A smaller cluster of 10,000
3:17:17chips runs somewhere between 10 and 15
3:17:20megawatt. To put that into context,
3:17:22that's enough to power 15,000 average
3:17:24American homes all at once, day and
3:17:26night. And that's just a single cluster.
3:17:29The biggest companies run multiple
3:17:30versions of them. Every answer you pull
3:17:32out of one of those systems draws
3:17:34[music] power off the grid. It also
3:17:36wears down chips that cost tens of
3:17:38thousands of dollars each and age out
3:17:40within months. 20,000 of them can run
3:17:43past half a billion dollars in hardware
3:17:44alone. The moment they switch on, they
3:17:47begin losing value. For a while,
3:17:49engineers tried to push costs down, and
3:17:51for a while, it worked. But you can't
3:17:53just write code that beats the laws of
3:17:55physics. Power costs money. A chip wears
3:17:58out in its own time, no matter how
3:18:00clever the software around it is. The
3:18:02numbers reflect that grim reality. These
3:18:04clusters have to be running almost every
3:18:06hour because an idle chip means a loss.
3:18:09Tech firms pay rising power bills. They
3:18:11write off the hardware fast because the
3:18:13[music] next chip makes the current
3:18:14generation look obsolete. Put all of
3:18:17that against what providers actually
3:18:18charge and the charge does not come
3:18:20close. It was never supposed to. The low
3:18:23price was bait. These companies are not
3:18:25bleeding cash by accident. They are
3:18:27doing it on purpose. Investor money has
3:18:29filled in the gap between costs and
3:18:31[music] price. But that gap keeps
3:18:33widening. That cheap price tag is
3:18:35actually just an illusion. Cash can push
3:18:37the price below cost. But the power bill
3:18:40doesn't change. When the investor money
3:18:42runs out, there's only one option left.
3:18:44[music] The price charged to the
3:18:45customer who can't leave. That's why AI
3:18:48companies try almost everything except
3:18:50charging what it actually costs. The
3:18:52workarounds buy some time, but they
3:18:54distort the market. When an AI company
3:18:57that builds the model also decides to
3:18:59sell it directly to customers, the
3:19:01effects show up quickly. The clearest
3:19:03example is a company called Jasper. It
3:19:05built an AI writing tool directly on
3:19:07OpenAI's models and for a while it was a
3:19:10successful business. Revenue climbed
3:19:11past $120 million. Investors valued it
3:19:14at $1.5 billion and more than 100,000
3:19:18customers signed up. Then ChatGpt
3:19:21arrived and it was virtually free. It
3:19:23did much of the same work for a fraction
3:19:24of the price. Jasper's growth suddenly
3:19:26went in reverse. By 2025, its revenue
3:19:29had fallen to around $88 million. The
3:19:32company survived, but by pushing hard
3:19:34into corporate marketing and finding
3:19:36firmer ground. It's a bit like a
3:19:38building full of chefs. Each one rents a
3:19:40small kitchen upstairs and cooks for
3:19:42their customers. And then one morning,
3:19:43the landlord opens his own restaurant in
3:19:45the lobby. It's the same dishes but
3:19:47given away for free and using the same
3:19:49recipes that he learned by watching his
3:19:51tenants work. If your only advantage is
3:19:54access to a model owned by somebody
3:19:55else, you don't really have an advantage
3:19:58at all. Not when the model's owner just
3:20:00decides to compete directly. It's a move
3:20:02known as sherlocking. It comes from an
3:20:04old Apple habit of swallowing up the
3:20:06best features of other apps into its own
3:20:08software. The provider watches how
3:20:11people use the tools built on it and
3:20:13then finds the most valuable uses and
3:20:15folds them [music] in. The original
3:20:17maker is quickly undercut and out of
3:20:19business. A startup spends years
3:20:21polishing its prompts and workflows.
3:20:23Soon that work shows up inside the
3:20:25platform copied for free. Its edge
3:20:27evaporates, its pricing [music] power
3:20:29follows, and its biggest contracts slip
3:20:31from must-have to maybe. And this is why
3:20:34relying on a single vendor has climbed
3:20:36the list of corporate fears. It used to
3:20:39look like a smart fast choice. After
3:20:41Jasper, it looks more like a loaded gun
3:20:43left on the table. But eating your own
3:20:45customers still doesn't balance the
3:20:47books. All you gain is a smaller group
3:20:49of rivals. So, the providers try
3:20:51something else. A move that doesn't
3:20:53touch the companies on top at all. It
3:20:55touches the product itself. On the
3:20:57surface, token prices keep falling and
3:20:59every press release calls it progress.
3:21:01But beneath those announcements,
3:21:03something might be happening to the
3:21:04thing you actually pay for. A large,
3:21:06smart, expensive model can be diluted
3:21:08into a smaller and cheaper one. [music]
3:21:10Shrinking a model can be fair, useful
3:21:12engineering done only to protect profit,
3:21:15it becomes a swap, and it has the same
3:21:17label, but the product is thinner. The
3:21:20new version is tuned to stay just good
3:21:22enough that you don't cancel. Yet, it is
3:21:24cheap enough to save a fortune across
3:21:25billions of prompts. The decline rarely
3:21:27is visible itself. The model forgets
3:21:30just [music] a little bit sooner,
3:21:31reasons a little less deeply, and
3:21:33dresses up its guesses as creativity.
3:21:35Because the price keeps dropping, people
3:21:38see the whole thing as a win. Frontier
3:21:40models are still growing, and shrinking
3:21:41them into highly capable systems
3:21:43represents massive technological
3:21:45processes. [music] But that steady drift
3:21:47towards smaller and cheaper points the
3:21:49other way. It hints that the numbers
3:21:51never really added up. Time is the only
3:21:53thing this really buys. the hostage
3:21:55contracts, the sherlocking, the
3:21:57shrinking of the models. None of this is
3:21:59a real plan to solve the issue because
3:22:01the problem lies elsewhere. Software was
3:22:04supposed to be the perfect business. You
3:22:05build it once, you sell it a million
3:22:07times, and each extra copy costs you
3:22:10almost nothing. That promise built 20
3:22:12years of [music] sky-high valuations.
3:22:14It's baked into every pitch that called
3:22:15AI the next great software story.
3:22:18Silicone shattered all of that. Thanks
3:22:20to all the costs associated with AI,
3:22:22OpenAI's margin was close to 33%. A
3:22:25healthy software business operates at 70
3:22:27to 80%. So, OpenAI got desperate. It
3:22:31began selling ads inside Chat GPT. The
3:22:34pilot launched in February 2026. Within
3:22:366 [music] weeks, it was on pace to make
3:22:38$100 million a year with over 600
3:22:40advertisers on board. Sam Alman had
3:22:43spent years calling ads a last resort.
3:22:45Now, the most famous AI company on Earth
3:22:47is renting out your attention. Rival
3:22:49company Anthropic mocked the move. But
3:22:52mockery doesn't cover a $25 billion
3:22:54bill. HSBC analysts have estimated that
3:22:57OpenAI will need $27 billion in fresh
3:23:00funding by 2030 to cover data center and
3:23:03compute costs. Sam Alman, meanwhile, has
3:23:05promised $100 billion in revenue by
3:23:082027. The gap between those figures
3:23:11keeps widening. What comes in and goes
3:23:13out are moving on very different
3:23:15trajectories. In late 2025, Altman said
3:23:18OpenAI had lined up about 1.4 trillion
3:23:21in compute deals over 8 years. The
3:23:23company made roughly $13 billion that
3:23:26year. One investor questioned the
3:23:28numbers on a podcast, which led to
3:23:29Altman offering to buy his shares. Just
3:23:32months in, in early 2026, the plan
3:23:34changed again. OpenAI told investors the
3:23:37real target was closer to $600 billion.
3:23:40Even the scale of the ambition had to be
3:23:42rewritten. When a plan is revised
3:23:44downward by more than half, it stops
3:23:47looking like strategy and starts looking
3:23:48like constraint. Patience is beginning
3:23:51to run thin. The latest funding terms
3:23:53require a clear path to profitability,
3:23:55something the cheap access era never had
3:23:57to prove. The timeline is tightening as
3:23:59well. With an IPO insight, OpenAI can't
3:24:02afford more losses. Public investors
3:24:04want profit, not promises. Neither is
3:24:07falling fast to save the current
3:24:08business model. No amount of code
3:24:10removes the need to run massive clusters
3:24:12of hardware that must be replaced long
3:24:14before they've paid for themselves. The
3:24:16industry assumes software margins would
3:24:18eventually outrun infrastructure costs,
3:24:20but so far that hasn't happened. That
3:24:22leaves a simple constraint. Keep prices
3:24:24low and burn through cash faster than it
3:24:26returns or raise prices and bleed users
3:24:29who made the system valuable in the
3:24:31first place. Either path arrives in the
3:24:33same outcome. The only uncertainty left
3:24:35is how many players are still in the
3:24:37game when it does. When the prices
3:24:39reset, the companies still standing will
3:24:41be the ones that control their own
3:24:42models. There's already an alternative
3:24:44emerging. Open models can be downloaded
3:24:47and run on hardware users control.
3:24:49Meta's Llama is freely available.
3:24:51DeepSeek has shown that capable models
3:24:53can be trained at a fraction of the
3:24:54expected cost. They are good enough to
3:24:56run a real business on. Whole teams
3:24:59already run them on their own servers
3:25:01with no outside key and no surprise
3:25:03invoice. The trade is straightforward.
3:25:05Customers take on the headache of
3:25:07running the software and the headache of
3:25:08running the hardware, but they stop
3:25:10being held hostage. At scale, this can
3:25:12be considerably cheaper than renting
3:25:14models and infrastructure. But there is
3:25:16a bigger advantage. Customers have
3:25:18control. They own their data. They
3:25:21decide when the model changes. The great
3:25:23shutout is already beginning. The
3:25:25smallest rappers will fold quickly after
3:25:27the first real price change. Mid-sized
3:25:29firms will scramble for tools and
3:25:31engineers they should have hired
3:25:32earlier. And at the top end, only the
3:25:34biggest players will end up running
3:25:36serious models on their own. For
3:25:37everyone, it stopped being about saving
3:25:39money. It's about staying alive. Leaning
3:25:42onto a single provider for your core
3:25:44income is a massive risk. One decision
3:25:46in someone else's office can cut you off
3:25:48completely. The companies moving fastest
3:25:50are already building their own clusters
3:25:52and tuning open models on their private
3:25:54data. They're pulling their most
3:25:56sensitive work back in-house. Owning
3:25:58also means a bill you can predict.
3:26:00Renting means a number that can leap in
3:26:02any given month. Owning means a cost you
3:26:05can plan around for years. For most
3:26:06businesses, that beats any discount. A
3:26:09split is forming and within a few years,
3:26:10it'll be obvious to see. On one side
3:26:13will be the companies that treated cheap
3:26:14tokens as a permanent gift. They built
3:26:17their whole cost structure on top of
3:26:18them. They are the tenants and they will
3:26:20learn the lease was never really theirs.
3:26:23On the other side are the companies that
3:26:24treated the cheap years like a window.
3:26:26They used it to get started and figure
3:26:28out what worked. They were able to see
3:26:30the escape route. The difference between
3:26:32the two groups isn't money or talent.
3:26:34It's timing. One side saw falling prices
3:26:36and they assumed it would last. The
3:26:38other saw borrowed time. Either you
3:26:40control the hardware your business runs
3:26:42on or you let a company that must
3:26:44squeeze you set the price. Waiting
3:26:46carries its own price. The longer
3:26:48customers stay, the harder the eventual
3:26:49fall. The cheapest moment to leave has
3:26:51already passed. The next cheapest moment
3:26:53is right now. Tokens were never getting
3:26:56cheaper because the technology had set
3:26:57itself free. They were cheap because
3:26:59someone else was covering the bill and
3:27:01that bill is due.

About this transcript

This page contains the full transcript of The AI bubble is about to burst by The Infographics Show, generated from the public captions YouTube serves with the video. The transcript has 37,245 words across 5,853 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.