YT YouTube2Text

YouTube2Text → Transcripts → My Brutally Honest Thoughts About the AI Bubble.

My Brutally Honest Thoughts About the AI Bubble. — Transcript

by Steve Eisman · 6,489 words · 1,009 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00AI has become the dominant theme in
0:02investing today. It's driving the stock
0:04market, influencing the bond market, and
0:06to a remarkable degree, driving the
0:09entire US economy. But here's what
0:12concerns me. This entire AI investment
0:15boom has become increasingly dependent
0:17on the success of just a few companies.
0:20And if these companies don't deliver on
0:22their enormous promises, the
0:24consequences could be incredibly severe.
0:27To put it simply, the US could easily
0:30enter a recession, and we could almost
0:32certainly see a large market correction.
0:35Hi, this is Steve Eisman, and welcome to
0:37another episode of The Real Eisman
0:39playbook. Now, don't get me wrong. AI is
0:42an extraordinary technology, and I have
0:44no doubt it will change much of the
0:46world. But the [laughter] questions
0:48investors need to be asking are how
0:50quickly, at what cost, and most
0:53importantly, who's actually going to
0:55make money from it? Because right now,
0:58hundreds of billions of dollars are
0:59being poured into AI infrastructure.
1:02Companies are taking on enormous amounts
1:04of debt, and investors are betting that
1:07all this spending will eventually
1:09generate extraordinary profits. So in
1:12this video, I want to examine exactly
1:14where we stand in the AI boom. We'll
1:17look at who's financing it all, the
1:19bull, bear, and doomsday scenarios, and
1:22most importantly, the five major risks
1:25that I believe could derail the entire
1:27investment story. Not since the railroad
1:30has a new technology caused so much
1:32CAPEX confusion, tension, and
1:34excitement. Since for many the whole AI
1:37story is rather overwhelming, I thought
1:40it would be a good idea to go through
1:41the major aspects of the story, and
1:43importantly, the major players. First
1:46things first, what exactly is an LLM?
1:50Well, LLM stands for large language
1:52model. Now, I'm oversimplifying here,
1:55but think of it as an incredibly
1:57sophisticated word prediction machine.
2:01You ask it a question and it predicts
2:03what word should come next to produce an
2:06answer. That's how ChatGPT and Claude
2:09work. That's different, the LLM, from
2:11AGI, or artificial general intelligence.
2:15AGI is essentially the holy grail where
2:18machines could theoretically match or
2:20exceed human intelligence across a broad
2:23range of tasks. In the wonderful words
2:26of Spinal Tap, it goes to 11.
2:29>> The numbers all go to 11.
2:32>> Now, we're not there yet, although the
2:35people raising billions of dollars to
2:37develop it would certainly like you to
2:39believe we're getting close. Then
2:41there's agentic AI, which I think is
2:44particularly interesting. Whereas a
2:46traditional LLM answers your question,
2:50an AI agent can actually go out and
2:53perform tasks on your behalf. My wife
2:55Valerie, for example, has used an AI
2:57agent to search for theater tickets,
2:59compare prices, and purchase the theater
3:02tickets she wanted with her permission.
3:05That's genuinely useful technology, and
3:07I think it has enormous potential to
3:09change how we live and work. But here's
3:11where the investment story gets
3:13interesting. All of this requires
3:16extraordinary amounts of computing
3:18power, and that computing power lives
3:20inside data centers. Data centers are
3:23enormous facilities filled with
3:25sophisticated chips, servers, cooling
3:28equipment, and electrical
3:29infrastructure. And by the way, these
3:32things are incredibly expensive. A new
3:36large AI data center capable of
3:39supporting a gigawatt of power demand
3:42could cost around, get this,
3:45$30 billion to build and equip. That's
3:48just one data center. Now, you've
3:50probably heard that data centers are
3:53popping up across the country, and
3:55indeed across the world. But who's
3:57actually paying for all of this?
3:59Primarily, not exclusively, but
4:02primarily, it's the hyperscalers. That's
4:05companies like Microsoft, Google,
4:06Amazon, and Oracle, which have spent
4:09years building the cloud computing
4:11infrastructure businesses rely on. now,
4:15AI is forcing them to build data centers
4:17on an entirely different scale. In 2026
4:21alone, hyperscalers, just the
4:23hyperscalers are expected to spend
4:25roughly $850
4:27billion
4:28on AI infrastructure and potentially
4:31even more next year. Astonishing sums.
4:34And that leads us to probably the two
4:36most important players in the current AI
4:38story. At the center of this entire AI
4:42ecosystem sit two private companies,
4:45OpenAI led by Sam Altman and Anthropic
4:48led by Dario Amodei. They're developing
4:51the AI models that are driving much of
4:54this demand for data centers. And
4:57possibly the kingpin of this whole
4:58shebang is Nvidia. Nvidia creates the
5:01most sophisticated chips in the world,
5:04and Jensen Huang is at the epicenter of
5:06the AI build out. So, you have the AI
5:09companies developing the models, the
5:11hyperscalers spending hundreds of
5:12billions building the infrastructure to
5:15support it, Nvidia manufacturing
5:17billions of dollars in chips and
5:19financing them as well, and all kinds of
5:21debt and equity investors including
5:23private credit banks, insurance
5:25companies, corporations who buy the debt
5:28issuances, and stockholders, all
5:30providing enormous amounts of capital to
5:32finance the whole thing, which brings us
5:36to literally the trillion dollar
5:37question. Who is actually going to make
5:39money from all of this? Because as we
5:41saw with the railroads and later the
5:44internet, a technology can completely
5:47transform the world while leaving an
5:49enormous trail of bankruptcies behind
5:52it. And right now, the market seems to
5:54have a few very different ideas about
5:56how this AI story will play out. So,
5:59before we get into the underlying risks,
6:02let's look at the four possible
6:03scenarios starting with the bull case.
6:06AI for the win. AI will change the world
6:09and Anthropic and OpenAI will continue
6:11to drive the stock market to new and
6:13greater heights with Nvidia setting
6:15endless records. Result, the market
6:17keeps going up. Now this has been the
6:19dominant narrative for about two years,
6:22so it's not crazy and it needs to be
6:24taken seriously because the largest tech
6:26companies in the world
6:28believe it. I give this narrative a 30%
6:31chance. The bear case. AI is overhyped.
6:34LLMs are useful, but adoption will be
6:37very slow and even if adoption is fairly
6:39rapid, the capital required to build the
6:42data centers and pay Nvidia and others
6:45for the chips and pay Anthropic and
6:47OpenAI their exorbitant prices for their
6:49programming won't happen. The Chinese
6:52open weight models will put pressure on
6:54pricing, a price war will ensue, moats
6:57vanish, and the returns on investment
6:59don't measure up. The result, the bear
7:02case leads to a full economic recession
7:05and a market crash as well as an
7:08unwinding of a lot of private credit and
7:10private equity which is chasing AI.
7:12Banks are most likely involved to a
7:14degree. The only good news is that after
7:17the crash, the hyperscalers like Amazon
7:19and Google will still have rock solid
7:20businesses and franchises and the stocks
7:23will likely be available then at a good
7:26price. I give this narrative also a 30%
7:29chance. Then we turn to the intermediate
7:32case. [music]
7:33Nvidia keeps financing the whole
7:34shebang. The hyperscalers keep going
7:37into debt and issuing more stock to
7:39finance the whole shebang. The data
7:41centers get built, but much more slowly
7:44and with many more hiccups than the
7:46build out is modeling. The Chinese keep
7:49making a pretty good product that does
7:50the job for much of computing needs
7:53which siphons off demand from Anthropic
7:55and OpenAI and this does lead to a price
7:58war. Or the construction of AI data
8:01centers is so great that it dwarfs
8:03demand. One analyst I follow, Brad
8:05Safflo of PAA Research, put out a report
8:09It was a good report where he estimated
8:11that US data center capacity will be
8:13double the demand by the end of 2027.
8:17That's Brad's estimate, but if he is
8:19right, it will mean that AI CAPEX
8:21spending will then slow in 2028. It's
8:24also positive that the political winds
8:27stay negative on allowing data centers,
8:30[music] NIMBY.
8:32And there is not enough capacity coming
8:34online to meet demand. Either way,
8:36[music]
8:37the future looks iffy. This narrative
8:39assumes that AI is a great technology,
8:41but there is simply way too much CAPEX
8:43chasing demand. Result, the market
8:46unwinds much of its AI gains and we have
8:49a bear market for a time. [music] I give
8:51this narrative a 40% chance. And then we
8:54turn to the doomsday scenario. This is
8:58my favorite narrative [music] because
8:59it's so ridiculous and yet seems so
9:02possible to so many people. Just because
9:04some old sci-fi novels predicted [music]
9:07part of the future doesn't mean that all
9:09sci-fi predictions come true. Machines
9:11taking over the earth like in Terminator
9:14is not in the foreseeable future. I know
9:17that and more importantly, the heads of
9:19Anthropic and OpenAI who are peddling
9:22this narrative know it too. So
9:24why are they spouting such nonsense to
9:26every politician who will listen?
9:28Obviously, if everyone really, really,
9:31really thought that the world was coming
9:33to an end, we wouldn't need to take
9:35these companies public or build data
9:37centers or even listen to this podcast.
9:40We'd go home kiss our loved ones and hug
9:42[music] the dog. They peddle this
9:44nonsense and surf the fear curve they've
9:46created because they want to scare the
9:49government into regulating the
9:50competition out of existence. I give
9:52this narrative about a 0% chance. Now
9:55that we've set the stage, [music] let's
9:57turn to the AI investment story, both
9:59the positives and the negatives. Over
10:02the past few years, AI has sent the
10:03stock market into a frenzy. In fact,
10:07from 2023 through 2025, the Magnificent
10:09Seven were responsible for roughly half
10:12of the S&P 500s total [music] return.
10:15Hyperscalers are pouring hundreds of
10:17billions of dollars into AI data
10:18centers. At times, investors have become
10:21euphoric. But before we get to AI
10:23fundamentals, let me first explain how
10:25AI dominates the market. Today, Infotech
10:29constitutes 40% of the S&P 500 with most
10:32of that related to AI in one way or
10:34another. Then add companies that are not
10:37technically in Infotech, but are techy,
10:40like Google and Amazon, and you reach
10:4150%. The utility sector, which is 3% of
10:44the S&P, is also AI related and that it
10:46is spending 250 billion this year for
10:49the power needs of AI. Then add
10:52companies in the industrial sector that
10:54help build AI data centers like Eaton
10:56and Rockwell. Then add industrial
10:59companies that supply the power like GE,
11:01Vernova, and Quanta. Then add financial
11:03companies that are financing AI like the
11:05large investment banks and the
11:07alternative asset managers. And add all
11:09that up and you are well north of 65% of
11:12the S&P. The only two sectors that are
11:14not involved with AI at all, as far as I
11:17can tell, are staples and healthcare,
11:19which add up only to 14% of the S&P. So
11:23what about bonds, you say? About 15% of
11:25existing corporate bonds are AI related,
11:27but over 50% of this year's issuance is
11:31AI related. So bonds don't provide the
11:34same degree of diversification that
11:36investors are looking for. Again, I want
11:39to emphasize that the entire market is
11:41not only directly related or influenced
11:43by AI, but the direction of the market
11:46hinges on the future health and success
11:48of AI. Before we get to the risks of AI,
11:50let's first examine the bull case, and
11:52it's a very powerful one. Bulls argue
11:55that the AI revolution is the largest
11:57technology build out in history. Bulls
12:00also argue that this is only the third
12:03inning of a nine-inning game. Capital
12:06keeps flowing into compute, power, and
12:08data centers because the demand signal
12:12is overwhelming. The hyperscalers have
12:14around two trillion of multi-year AI
12:17cloud related backlog. Currently, the
12:20binding constraints have shifted from
12:22capital to power, permitting and memory
12:25supply. It's a wonderful story with one
12:28simple underlying assumption [music]
12:31that AI is going to be transformative.
12:34It's going to transform software and
12:36medicine and everything else you can
12:38think of. AI will change it. Why else
12:41would great companies like Google,
12:42Amazon, and Microsoft be spending so
12:44much on AI CapEx? That's the bull case
12:48and for the last two years it has
12:50dominated the markets [music] and for
12:52two years or so AI bulls have been
12:55right, but that does not mean that they
12:57will be right forever. Let me first
12:59point out that everything that is being
13:01said about AI was said about the build
13:04out of the internet, that the internet
13:05was going to transform the world and
13:08that prediction proved to be totally
13:10true. What also turned out to be true
13:12was that the first generation of
13:14internet companies largely failed and
13:16failed spectacularly [music] during the
13:18bursting of the dot com bubble. It was
13:21only the second generation of internet
13:23companies like Google that truly
13:25succeeded. Thus, AI may be
13:28transformative, but that does not mean
13:30that all the money currently being spent
13:32will generate an adequate return. I want
13:35to touch on a theme that does not
13:37involve the risks to AI, but the risks
13:40to software companies from AI. This
13:43story has its own meme. It's called the
13:45SaaS pocalypse. What a great name. And
13:47what does it mean? It means that the
13:49cost of creating software because of AI
13:52has collapsed. As a result, AI is going
13:55to replace, so the story goes, all
13:58existing software with the exception of
14:00cybersecurity. Cybersecurity remains
14:03necessary because AI makes it easier to
14:05hack. The SaaS pocalypse narrative
14:07dominated software stocks from the
14:09summer of last year through the early
14:12summer of this year. Thereafter, most
14:14software stocks had a nice
14:16as the market realized that the problem
14:18with the SaaS pocalypse narrative is
14:21that it is overly simplistic. Being a
14:23successful software company requires
14:25more than just creating software. In
14:28fact, some software companies are so
14:29embedded in the operations of
14:31enterprises that the switching costs are
14:33too prohibitive. I think that probably
14:35applies to companies like Service Now
14:37and Salesforce. [music]
14:39However, many software companies do not
14:41have that kind of moat. Take Adobe, for
14:44example. [music] The perception is that
14:46Adobe is particularly vulnerable to AI.
14:49Adobe hit an all-time high of $630 in
14:52February 2024. [music]
14:54Today, the stock is around 245. That's a
14:57decline of 60 percent. The threat to
15:00Adobe is not just idle speculation.
15:03Recently, it was announced that software
15:05engineer Brandon Thomas has built and
15:07released an open source free alternative
15:11to Adobe's Creative Cloud Suite. Now, I
15:13have no idea how good this Adobe
15:15alternative is, but free is free. The
15:17tentacles of the SaaS pocalypse extend
15:19beyond just the risk to software stocks.
15:23Between 2017 and 2022, private equity
15:26went on a buying spree of software
15:29companies, specifically SaaS, software
15:32as a service. Private equity was the
15:35buyer, and private credit was often the
15:37lender to the buyout. Most transactions
15:40took place with 50 percent debt and 50
15:42[music] percent equity. That's an
15:44important figure because of the decline
15:46in public software stocks. Let's use
15:48Adobe's decline of 60 percent as a
15:50proxy, but make it a tad more
15:52conservative and say that the value of
15:54most private equity held software
15:56companies is down 50 percent with
15:59cybersecurity software companies being a
16:01notable exception. A 50 percent decline
16:04in value means that the equity is wiped
16:07out. This will start to become clearer
16:09when these loans start to be refinanced
16:12next year. Now, I don't think this
16:13private equity, private credit problem
16:15is anywhere close to being large
16:18to seriously hurt the [music] economy.
16:20But it does mean that private equity
16:23investors are likely going to get hurt
16:25and hurt a lot. And that leads me to the
16:28risks. What are the major economic risks
16:32to the AI story? Well, there are five
16:35predominant issues. I'm going to examine
16:37them in depth and then rank that risk in
16:40terms of its ability to derail the AI
16:42story. Issue one from capital light to
16:46raising capital. By tradition, the
16:48hyperscalers, Google, Microsoft, Amazon,
16:51Oracle are mostly software companies.
16:54Yes, they created cloud businesses years
16:56ago that required building out data
16:59centers. But the cost of those data
17:02centers was not prohibitive. So these
17:04companies were cash flow generating
17:08behemoths. They generated so much cash
17:10they barely knew what to do with it. All
17:12excess cash was used to buy back stock.
17:15Call these mostly capital light business
17:17models. AI has changed the game. The
17:19compute needs of AI are exponentially
17:21bigger than pre AI needs. Also, AI
17:25requires GPU chips and CPU chips and
17:28every other kind of chip. But the cost
17:31of a GPU chip is very very high. The
17:34cost to buy Nvidia's latest GPU chip,
17:37Blackwell, can exceed $5,000 per chip.
17:41An AI data center can require at least
17:43100,000 Blackwell chips, sometimes more.
17:47That means that the cost of just the
17:49GPUs is around $500 million.
17:52The cost of building the new AI data
17:53centers and filling them with the
17:55necessary equipment is prohibitive even
17:58for the hyperscalers. Pre AI, Google,
18:00for example, [music] generated over $80
18:04in annual free cash flow. That cash flow
18:06is now gone. In fact, Google raised $85
18:10billion in equity capital in 2026 to
18:13help fund its AI needs. This is an
18:15enormous change. The last time Google
18:18raised equity capital was 2005. Take
18:21Oracle is another example. Oracle is a
18:24large company, but at a market cap of
18:26390 billion, it is much smaller than its
18:29hyperscaler competitors, but it is still
18:31trying to play in this game, expensive
18:34game. To do so, it has had to take on a
18:37considerable amount of leverage. In the
18:39past two years, Oracle's total debt has
18:41increased from 94 billion to 155
18:44billion. The ratings agencies are not
18:46amused. S&P downgraded Oracle to triple
18:49B minus, which is only one level above
18:51junk. Bottom line, AI capex requirements
18:54have altered the business models of the
18:56hyperscalers from capital light to
18:58capital heavy. However, I don't think
19:01that this change in and of itself is
19:04enough to sink the AI story. If [music]
19:06the AI story succeeds, then all this
19:08leverage will simply look like good
19:10investments. One more point, the
19:12increase in leverage of the hyperscalers
19:14has fostered some financial shenanigans.
19:17In the days of Enron and the GFC, there
19:19were multiple examples of companies
19:21using off balance sheet techniques to
19:24hide their leverage. The goal was to get
19:27the ratings agencies to believe that
19:30these off balance sheet vehicles were
19:31not the responsibility of their creator,
19:34and so the debt in the vehicle should
19:36not be counted as part of the company's
19:39leverage. The hyperscalers are starting
19:41to take a page out of that old book for
19:43their own benefit. For example, Meta is
19:46using an off balance sheet technique to
19:48fund one of its data centers. Now, Meta
19:50is not a hyperscaler. It is creating its
19:52own LLMs and agentic AI apps, so it is
19:55building data centers for its own use.
19:57Last year, Meta dipped its toe in the
20:00world of off balance sheet financing in
20:02order to fund a 27 billion dollar data
20:05center in Louisiana. The data center is
20:08financed with debt, but the debt is not
20:11on Meta's balance sheet. Instead, an off
20:13balance sheet special purpose vehicle,
20:16SPV, was set up called Benyaye Investor
20:19LLC SPV rolls off the tongue. Hat tip,
20:22though, to the name, as anyone who has
20:24spent time in New Orleans knows that
20:26beignets are really delicious, but the
20:28financing is less appealing.
20:30Technically, Meta owns only 20% of
20:33beignet, with the rest owned by Blue Owl
20:35and Pimco. So even though Meta is
20:37building the data center [music] and has
20:39promised to rent the data center for
20:41beignet for 20 years and bears the cost
20:44of any project delays and overruns, the
20:46debt will not sit on Meta's balance
20:48sheet. That's some pretty fancy footwork
20:50that's reminiscent of bad times past and
20:53doesn't pass the smell test. At least
20:55not to me. Perhaps with the memory of
20:58what happened to Arthur Andersen in
20:59mind, Ernst & Young, Meta's auditor, was
21:02not totally comfortable with this
21:04beignet masterpiece of off-balance sheet
21:07financial engineering. [music] In Meta's
21:082025 10-K, Ernst & Young flagged it as a
21:13quote critical audit matter, [music]
21:15unquote. Now, I'm not going to quote the
21:18section from the 10-K because it is
21:20quite long, boring, and written in
21:22accounting language designed to
21:24obfuscate. What I'm going to do is
21:27translate that language into plain
21:29English, or at least the English that I
21:31think is what they're really saying. We
21:33at Ernst & Young are very nervous about
21:35this transaction. We are really not sure
21:38at all whether this should be
21:40off-balance sheet. We relied on the
21:42company's description of the
21:43transaction, but we did not do any real
21:46work of our own to determine whether
21:47that what Meta said about the
21:50transaction is really accurate. After
21:52all, we want to get paid and Meta is a
21:55big client. That's my slant on the
21:58accounting language. Admittedly, this
22:00off-balance sheet transaction is small.
22:02It's only 27 billion, but you should
22:04expect many more such transactions in
22:06the future, and not just at Meta. There
22:09is another aspect to AI financing that
22:11bears discussion, circular financing.
22:13Here, Nvidia seems to be the biggest
22:15culprit. It is either investing in or
22:17lending money to other tech companies,
22:20and those tech companies then take that
22:22money and buy Nvidia chips. Here too, I
22:25don't think off-balance sheet or
22:27circular financing is enough on its own
22:30to sink the AI story. It just makes it
22:32more dangerous. Now we talk about issue
22:35number two, power and data center build
22:37outs. Even assuming that the AI story
22:40will work, potential binding constraint
22:42is the lack of sufficient power and
22:44delays in building data [music] centers.
22:47I needed to hammer this point home. The
22:49AI story cannot work unless AI data
22:53centers are built. Now the sums here are
22:55insane. Hyperscalers will spend 850
22:58billion on building AI data centers just
23:01in 2026 and purchasing all the equipment
23:03to fill the data centers as well. But it
23:06takes time to build AI data centers.
23:08Here politics plays a role. The AI
23:12industry is not popular. There is now a
23:15backlash against AI in many communities.
23:17[music] People are afraid of the massive
23:19needs of AI data centers, needs for
23:22water and energy. [music] So the process
23:24of permitting AI data centers is being
23:27delayed in many parts of the country.
23:30Energy is also a potential binding
23:32constraint. AI data centers utilize
23:35enormous amounts of energy. Now the
23:36first choice is to hook up the data
23:38center to the electrical grid, but the
23:41grid is already stressed. Utilities are
23:43investing in new plants, but that
23:46construction also takes time. It takes a
23:49few years, for example, to build a gas
23:51turbine that powers a utility plant.
23:54That is why a stock like GE Vernova has
23:56done well because GE V is a large
23:59builder of gas turbines. Because
24:01utilities cannot satisfy all the needs
24:04of AI data centers, the hyperscalers are
24:06looking for alternatives. They are
24:08looking at nuclear or alternative gas
24:10turbine technology such as those
24:12provided by Bloom Energy. My suspicion
24:15is that AI data center construction is
24:17taking longer than hoped, but how much
24:19of an impact this will have is something
24:21I just don't know. Now we turn to the
24:24third risk, concentration risk. Before
24:27we get to this
24:28risk topic, let me scaffold it. Imagine
24:31there is a software company that is
24:33growing very rapidly, but only has three
24:36enterprise customers. What would you say
24:38about such a company? I know what I
24:40would say, that this software company is
24:43a very risky investment. Should one of
24:46its customers go bankrupt or go
24:48elsewhere, growth would collapse. There
24:50is something like this going on
24:53throughout the AI food chain. Let's
24:55start at the top of the food chain.
24:57[music] NVIDIA, in its most recent
24:59quarter, NVIDIA's revenue grew in excess
25:02of 100%.
25:04That's quite an accomplishment. That
25:06speaks to the bull case of AI. However,
25:08in NVIDIA's 10Q, it states that 70% of
25:12NVIDIA's accounts receivable, which is a
25:14proxy for revenue, was from five
25:17customers. That's a lot of concentration
25:19risk for the largest company in the
25:21world. It's unclear who these five
25:23customers are. I assume it's the
25:25hyperscalers or someone adjacent to the
25:28hyperscalers. Now let's go down the
25:30chain one level to the hyperscalers.
25:33According to reports from various Wall
25:34Street firms, something like 70% of
25:37hyperscalar AI revenue comes solely from
25:40Anthropic and OpenAI. 70% equates to
25:43around 25 to 35% of total hyperscalar
25:46cloud revenue. I cannot confirm these
25:49statistics, but they sound right given
25:51what we actually know about Oracle.
25:53Oracle has a backlog of around 600
25:55billion, and we know, because the
25:57information is public, that around 50%
25:59of that backlog is from OpenAI alone.
26:02And by the way, OpenAI accounts for over
26:0420% of CoreWeave's total contracted
26:07order book. Hyperscalers are spending
26:09about 850 billion in CAPEX this year and
26:12even more next year, and that spending
26:14accounts for around half of the 2% GDP
26:17growth projected for 2026. Therefore,
26:19one must conclude that the health of the
26:22US economy is extremely dependent on
26:24hyperscalar CAPEX, and hyperscalar CAPEX
26:27is highly dependent on the health of
26:29Anthropic and OpenAI. Risk number four,
26:32OpenAI's risk of failure. What would it
26:34mean? Between the two, OpenAI is the
26:37weaker entity. Remember OpenAI owns
26:39ChatGPT, which is rumored to be losing
26:42market share to Anthropic's Claude. The
26:44company has lost several key senior
26:46employees. Let's list them because the
26:48list kind of illuminating. Number one,
26:50Denise Dresser, Chief Revenue Officer
26:53resigned in August 2026 after only eight
26:56months on the job. That probably means
26:58she gave up all her stock options as
27:00those usually don't vest for at least a
27:02year. Two, Brad Lightcap, former COO,
27:05resigned in August 2026 after eight
27:07years with the company. Three, Caitlyn
27:10Kalinowski, Head of Robotics and
27:12Consumer Hardware resigned in March 2026
27:15over governance concerns regarding an
27:17agreement with the Pentagon. And four,
27:19Kate Rouch, Chief Marketing Officer and
27:21Kevin Weil, Science Research Officer
27:24both departed in April 2026. The most
27:27concerning departures are the ones that
27:28took place this August, Denise Dresser
27:30and Brad Lightcap. Supposedly, OpenAI is
27:33getting closer to an IPO. That's the big
27:36payday for employees because it means
27:37that eventually they can sell some of
27:39their shares. That two such senior
27:41employees would leave now, that's an
27:44important data point. More importantly,
27:46OpenAI's financials, as far as we can
27:49tell, don't look so great. The Wall
27:51Street Journal reported that OpenAI's
27:53June quarter revenue reached 6.7
27:55billion, up only 18% versus the March
27:58quarter. Compare that to Anthropic's
28:00revenue of 11 plus billion in the June
28:03quarter, which was up over 100% versus
28:05the March quarter. What's even worse
28:08than the relatively weak revenue growth
28:11is the explosion in costs. Again,
28:14according to the Wall Street Journal,
28:15OpenAI's costs reached 12.3 billion, up
28:183 billion versus the March quarter. So
28:21in three months, revenue increased 1
28:23billion, but costs surged 3 billion.
28:26[music] Things are not going in the
28:27right direction. Now what is the problem
28:29at OpenAI? Part of the problem is that
28:31OpenAI made a strategic mistake, I
28:33think. [music] It chose to chase the
28:35business to consumer market, B2C. It did
28:38so by giving its product away or
28:40underpricing it
28:41with the hope that later on it would
28:43raise prices. That strategy proved to be
28:46flawed and now OpenAI is chasing the
28:48business-to-business market, B2B. The
28:51problem is that it is late and playing
28:53catchup with Anthropic, which chose the
28:54B2B market at the outset. Again, I think
28:57that OpenAI is potentially in trouble
28:59and this bears continuous monitoring.
29:02According to Ed Zitron, a commentator on
29:04Substack, to survive, OpenAI will have
29:07to raise $100 billion per year for the
29:09next several years. Now, when you lose
29:12billions upon billions, appearances
29:15matter a lot. OpenAI is completely
29:17dependent on the kindness of strangers
29:19funding its cash flow needs. When a
29:21company is growing and very profitable,
29:24appearances don't matter nearly as much.
29:26Such a company can survive quite well,
29:28but when a company is not profitable and
29:30has an insatiable need for capital,
29:33appearances matter more than anything
29:35because if the narrative turns negative,
29:37raising capital becomes much more
29:39difficult. This is just me reading the
29:41tea leaves, but OpenAI's press is no
29:43longer all positive. The press has
29:46actually focused on the high-level
29:47departures and on the weak numbers that
29:50I just described. That's why I'm not
29:52predicting, at least not yet, that
29:54OpenAI will fail. I still think it's too
29:56early to make that kind of prediction,
29:58[music] but it's not too early to think
30:00about it. So if OpenAI fails, Oracle is
30:03in immediate trouble because of the
30:05large increase in its debt. Oracle's
30:07debt rating is barely above junk. Now,
30:10what would happen to Oracle's stock
30:11price if OpenAI failed? Well, we have
30:13already had something of a preview. Last
30:16year, when Oracle reported its backlog
30:19for the August quarter of 2025, what it
30:21calls remaining performance obligations,
30:25RPO, it showed an RPO explosion to $455
30:30billion. It's now over $600 billion. In
30:32that quarter, the RPO was up over $300
30:35billion in just 3 months. Now, prior to
30:37the earnings report, the stock was $230.
30:39[music] In just a few days, it jumped to
30:42$330.
30:44And then analysts published reports
30:46pointing out that 50% of the RPO was
30:48from OpenAI, and the stock gave back all
30:51of its gains plus in a few months.
30:54Today, the stock is get this around 130,
30:58well below where it stood before it
31:00reported the August 2025 quarter.
31:02[music] From the peak, the stock is down
31:05around 70-ish percent. That decline is
31:07because the market perceives an
31:09overreliance on OpenAI. Imagine what the
31:12market would do to Oracle's stock if
31:15OpenAI fails. The ramifications of an
31:17OpenAI failure extend beyond [music]
31:19Oracle. Remember I said that AI CAPEX
31:22accounts for 50% of US GDP growth, while
31:24the other hyperscalers are not quite as
31:26dependent on Anthropic and OpenAI as
31:29Oracle. They are dependent enough. If
31:32OpenAI failed, the hyperscalers, I am
31:34sure, would cut back on their CAPEX. And
31:37I'm starting to think that the demise of
31:39OpenAI would push the US into an
31:42immediate recession. Here's an important
31:44point. Timing is everything. Eventually,
31:48the AI industry may become more
31:50diversified and less dependent on
31:52Anthropic and OpenAI. If either company
31:54fails within, I believe, a year, the
31:57ramifications would be severe. But if
31:59that failure happens a few years from
32:02now, the ramifications could be much
32:04less important. For now, the health of
32:06Anthropic and OpenAI is crucial. So what
32:09would happen to particular stocks and
32:11sectors if Anthropic or OpenAI failed
32:13soon? First, the hyperscalers would go
32:16down. Failure of OpenAI would mean they
32:18would pull back on CAPEX. Cloud revenue
32:21growth would slow. These stocks, Amazon,
32:23Google, Microsoft, and Oracle, just to
32:25name a few, they'd all correct. Also,
32:28the whole tech sector would correct, but
32:29the ramifications are much broader than
32:31just that. Investment banks and large
32:33banks, these stocks are at fairly close
32:36to peak valuations. Also, the investment
32:39banking cycle is super strong, partly
32:41because of the financing needs of AI.
32:43Should those needs lessen, the
32:45investment banking cycle would weaken
32:47and these stocks would correct from
32:48their peak valuations. Industrials.
32:51There is a subset of industrial
32:53companies that are major beneficiaries
32:55of AI. They are in the power space like
32:57GE Vernova and Quanta. Or they are in
33:01the electrification or automation spaces
33:03like Eaton and Rockwell, and these
33:05stocks would decline as well. So what
33:07will do well? This is not a stock
33:09picking question, but a reallocation
33:11question. Investors will reallocate to
33:13safety sectors and subsectors. [music]
33:16In the safety sector category, think
33:17about healthcare and consumer staples.
33:19I'd also point out that within almost
33:21every sector there exists safety [music]
33:24subsectors. So for example, within
33:27financials, the property and casualty
33:28sector is considered the safety
33:31subsector. Since we are talking about
33:32capital reallocation and not stock
33:35picking, let me flag three safety ETFs.
33:37One, LVHD. It's the Franklin US low
33:40volatility high dividend index ETF.
33:43Number two, the SPLV,
33:45Invesco S&P 500 low volatility ETF. And
33:49last, KBWP, the Invesco KBW property and
33:53casualty insurance ETF. [music] And now
33:57we come to the last and probably the
33:59most important risk. Number five, there
34:01are no long-term [music]
34:03moats. Until recently, Google had an
34:06insurmountable moat around its search
34:08business. No one could touch it. There
34:10don't seem to be any moats around AI.
34:12One day ChatGPT is on top, the next day
34:15it's Claude, and the switching cost
34:16between models seems to be perhaps zero,
34:19but certainly very low. I can't
34:21overstate the importance of this risk.
34:24Today, I hear a lot of positive energy
34:26towards Claude. People literally email
34:28me that their new company or university
34:31is endorsing Claude. The argument is
34:34that enterprise reliance is a moat. To
34:36this I say, the minute ChatGPT or Gemini
34:39or the Chinese models offer the same
34:41quality and a lower price, pressure will
34:43be huge to switch. There goes the moat.
34:45The costs of AI are expressed in tokens.
34:49Think of tokens as three quarters of a
34:51word. Every time you query Claude or
34:53ChatGPT, it enters your query and there
34:56is a cost to the tokens for the query.
35:00Last year, LLM providers dramatically
35:02undercharged for their services. They
35:03charged a subscription and the cost of
35:06that subscription was dramatically lower
35:08than the cost of the tokens. This
35:10fostered something called token maxing,
35:13which is a fancy term for enterprises
35:15telling their software employees to use
35:17AI 24/7 and to hell with the costs.
35:20Earlier this year, prices were raised
35:22[music]
35:22and enterprises found their AI costs
35:25soaring. Uber, for example, blew through
35:27its entire AI budget in just a few
35:30months. Sometime this summer,
35:32token maxing ended as all enterprises
35:35became [music]
35:36more cost conscious. While the token
35:38maxing saga was going on, Chinese open
35:41weight models started to take market
35:43[music] share. The important point here
35:45is that open weight models are much
35:47cheaper than the models of Anthropic and
35:49OpenAI. The flagship models of Anthropic
35:51and OpenAI can cost $2 to $10 per 1
35:55million tokens. Open weight models can
35:57cost around 20 cents per 1 million
36:00[music] tokens, so much cheaper. The
36:02models of Anthropic and OpenAI are
36:04superior, but you don't need a
36:06Mercedes-Benz for every trip and this
36:08leads to the risk, potential [music] for
36:10a price war because open weight models
36:13are just so much cheaper. Add in the
36:15fact that there are no moats and the
36:17probability of a price war explodes.
36:20This is the key risk. Anthropic and
36:22OpenAI do not have the balance sheets to
36:24withstand a price war and should a price
36:26war break out, fears will grow that the
36:28entire AI story will just slow. And in
36:31fact, there are reports that both
36:33Anthropic and OpenAI are cutting prices
36:34right now at a fear of losing market
36:37share to the cheaper models. Stay tuned.
36:40And now we come to our conclusion. As
36:42you can see, there is a lot going on.
36:44The AI story is complex. It is also
36:46driving the US economy right now. The
36:48risks I have outlined are real and
36:50could, I emphasize, could derail the AI
36:54story. If Anthropic or OpenAI get into
36:56trouble in the next year, that would be
36:58telling. However, if it is years from
37:00now, the AI story may become so
37:02diversified by then that it won't
37:04matter. It is still too early to tell,
37:06but we will be watching very closely. So
37:09with that said, what should investors be
37:12doing right now? Well, the first thing
37:14I'd say is that being cautious about
37:16parts of the AI trade is not the same
37:18thing as being bearish on AI. I think AI
37:21is real. I think it's already started
37:22changing the economy, and I think an
37:24enormous amount of money will be made
37:25from it. The much harder question is who
37:28actually makes that money. Personally,
37:30I'm much more interested in the picks
37:31and shovels. NVIDIA obviously sits at
37:34the center of the compute build out.
37:36You've got companies like Micron
37:37supplying memory, Arista and Cisco
37:40supplying networking, and then on the
37:42power side, you've got companies like
37:43GE, Vernova, and Eaton. These businesses
37:46don't necessarily need to predict which
37:48LLM ultimately wins. They're selling the
37:50infrastructure required to build the
37:53entire thing. Where I'd be much more
37:55careful is in businesses where investors
37:57are assuming enormous future profits
38:00despite very low switching costs.
38:02[music] Who are we talking about here?
38:04Mainly the LLM providers themselves, but
38:06potentially to parts of software too as
38:08agentic AI becomes more capable. The
38:11question I keep coming back to
38:12personally when analyzing these
38:14businesses is very simple. How difficult
38:17is it for the customer to leave? [music]
38:19If the answer is not very, you'd better
38:22be sure about the price you're paying.
38:24And ultimately, that's how I'm looking
38:26at this entire AI boom. I'm certainly
38:28not trying to call a top. I'm not
38:30telling you this is 1999 or 2008 all
38:33over again, at least not yet anyway.
38:35[music]
38:36Maybe these companies deliver everything
38:38the market is expecting and more. But
38:40when valuations assume an extraordinary
38:42future, you don't necessarily need the
38:44technology to fail for the stocks
38:46disappoint. You just need reality to
38:49fall short of expectations. And right
38:51now, between Chinese competition, the
38:53lack of moats, the enormous commitment
38:55sitting on OpenAI and Anthropic, and the
38:57physical limitations of the power grid,
39:00there are a plethora of reasons to keep
39:02watching those expectations very
39:05carefully. And that's the end of our
39:06lecture. But before I go, let me just
39:08say that if you want to support the show
39:10further and get additional exclusive
39:12content, you can check out my Substack
39:14at
39:15Premium membership to Substack offers a
39:18premium only Wednesday episode, which
39:20can be a masterclass or deep dive
39:22interview. And premium subscribers
39:24receive early access to our Monday
39:26interviews and Friday weekly wraps, as
39:30well as full access to my daily chats
39:33and the ongoing conversations among
39:35members. Free membership in Substack
39:38offers access to my daily notes and
39:40previews of our Wednesday premium
39:42interview and previews of all
39:44newsletters. The link steve eisman.com
39:47is in the description. Thanks for
39:49watching.
39:51[music]
39:53This podcast is for informational
39:55purposes only and does not constitute
39:57investment advice. The hosts and guests
39:59may hold positions in stocks discussed.
40:01[music]
40:02Opinions expressed are their own and not
40:03recommendations. Please do your own due
40:05diligence and consult a licensed
40:07financial advisor before making any
40:09investment decisions.
40:11>> [music]

About this transcript

This page contains the full transcript of My Brutally Honest Thoughts About the AI Bubble. by Steve Eisman, generated from the public captions YouTube serves with the video. The transcript has 6,489 words across 1,009 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 choose a caption track available on the video.

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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.