YT YouTube2Text

YouTube2Text → Transcripts → Warum die KI-Blase bald platzt & was du tun solltest

Warum die KI-Blase bald platzt & was du tun solltest — Transcript

by Julian Ivanov | KI-Automatisierung · 4,717 words · 715 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00For a change, there’s no tutorial on
0:02the channel today, because we need to
0:03talk about something currently hanging
0:05over the entire AI world that many
0:07people working with AI aren't even
0:08aware of. If you look at developments
0:11over the last few months and years,
0:13more and more people are asking if we
0:16are actually in a massive AI bubble
0:18that threatens to burst. The figures
0:21behind this concern are truly extreme,
0:24as OpenAI, the company behind GPT, says
0:27it has committed around 1.4 trillion
0:29dollars for the next 8 years for
0:31computing power, data centers, and
0:34chips. That is more than 30 times what
0:37the company brings in annually, and
0:38they are incurring billions in losses
0:40in the process. Oracle, the corporation
0:42building the data centers for OpenAI,
0:44has taken on so much debt that experts
0:46often doubt whether that money will
0:48ever be recouped. And almost the entire
0:51growth of the American economy is
0:52currently dependent on these AI
0:54investments. So if that tips over, it
0:56will hit much more than just a few tech
0:58stocks, let me tell you. At the same
1:00time, business is booming like crazy,
1:02as Anthropic, for example, the company
1:04behind Claude, has grown from a billion
1:06to 65 billion dollars in annual revenue
1:09in less than 2 years, faster than any
1:11software company before. And naturally,
1:13you think, okay, how does that even fit
1:15together? Are we experiencing the next
1:17dot-com bubble just before it bursts,
1:18or simply the next expensive build-out
1:20of a foundational technology, so to
1:22speak? So I did some serious research.
1:24I looked at the real numbers, the debts
1:26, the contracts, and the cash flows
1:28behind the scenes, and some of it
1:30really surprised me. I will also
1:32provide all the sources for this video
1:34in the video description. And I am now
1:36going to present everything on this
1:37topic in this video in a compact way,
1:39so that by the end you will know
1:41exactly how to assess the situation and
1:42what it means for you if you work with
1:44AI. So let's dive right in. Let's start
1:46with the scale to understand what kind
1:48of territory we are dealing with here.
1:50Microsoft, Amazon, Google, Meta, and
1:53Oracle alone are spending almost 700
1:56billion dollars on AI data centers this
1:58year combined. That is more than the
2:01entire economic output of Austria, and
2:03that is being handled by just five
2:04companies in a single year. You
2:07naturally ask yourself where this money
2:09comes from, and until recently, these
2:10companies paid for their data centers
2:12almost entirely from their own profits,
2:14as they earn hundreds of billions every
2:15year from ads, cloud, and office
2:17subscriptions without having to borrow
2:19a single cent. But this year, for the
2:21first time, that’s no longer possible
2:23because the expansion now costs more
2:25than the companies earn in a year. So
2:27now, all profits are flowing into
2:29expansion, and the rest is being
2:30borrowed. Every third dollar for AI
2:32expansion, or rather for more data
2:34centers, is now borrowed. Two years ago
2:36, it wasn't even every tenth dollar. In
2:38the first five months of this year
2:40alone, these five companies borrowed
2:42159 billion dollars, more than in the
2:45five years prior combined. Even the
2:47International Monetary Fund, the
2:49organization that monitors the
2:51financial system for over 190 countries
2:53, is now warning us, as the major tech
2:55giants are starting to take on debt—
2:57specifically through bonds that last
2:59longer than the chips themselves. An AI
3:01chip is technically obsolete in two
3:03years. However, the debt taken out for
3:05them runs for 10 years or more. And now
3:06, a quick estimation question for you.
3:08What percentage of total American
3:10economic growth in the first quarter of
3:13this year do you think was tied to AI
3:16investments? It’s around 50%, and
3:18depending on how you calculate it, as
3:19much as 75%. That means, without the AI
3:22expansion, the world's largest economy
3:24would have grown only half as fast.
3:26This means America itself is now hooked
3:28on this boom. Do you get that point?
3:30We’ll come back to that in a moment.
3:31And it gets really uncomfortable when
3:33you look at where all this money is
3:35actually going: surprisingly often,
3:36it’s just moving in circles. Nvidia,
3:39the manufacturer of AI chips, invests
3:41up to 100 billion in OpenAI, while
3:43OpenAI, in return, promises to buy
3:46Nvidia chips. Furthermore, OpenAI is
3:48renting 300 billion in computing power
3:51from Oracle—the largest contract of
3:53its kind ever. And for that, Oracle is
3:56building data centers filled with, you
3:58guessed it, Nvidia chips. The media
4:01company Bloomberg traced these circular
4:03transactions and found over 800 billion
4:05dollars being shifted around in circles
4:07. This means that from the outside, it
4:09looks like gigantic growth. From the
4:11inside, however, it’s partly the same
4:13money just spinning around on a
4:14carousel. And the carousel has a second
4:17level, because the cloud revenue that
4:18Microsoft and Google are currently
4:20boasting about on the stock market
4:22comes, to a surprising extent, from
4:23just two customers. The major bank UBS
4:26calculated that OpenAI and Anthropic
4:28will account for a good quarter of
4:29Google’s cloud revenue this year and
4:32almost half next year. And according to
4:34analyst Citron, 69%of Microsoft’s
4:37cloud growth last year came from OpenAI
4:40alone. Without Anthropic, without
4:44OpenAI. Well, I have serious questions
4:46about that. So, in cal 2025, according
4:49to my own reporting about Open AI
4:51numbers, 69%of the year over year
4:53growth of Microsoft's intelligent cloud
4:55segment was actually from Open AI.
4:58Without that, it would have only grown
5:008%year over year, which is just
5:02inflation. The celebrated AI growth of
5:10these corporations is, therefore, to a
5:12large extent, money from two firms that
5:14pay for their own expansion with
5:16investor cash, partly provided by these
5:17same corporations. It’s as if a car
5:21dealer were to lend you the money for
5:22his own car, and if your bank hesitates
5:24with the loan, the dealer even vouches
5:26for you. On August 17th. Nvidia
5:28announced via a stock market filing
5:30that it would stand for up to 105
5:33billion dollars to ensure a gigantic
5:35data center can be built in Ohio for
5:37Open AI. It will be built and operated
5:40by SB Energy, a Softbank subsidiary in
5:42which Nvidia is simultaneously
5:44investing 1.5 billion, and it will be
5:46stocked exclusively with Nvidia chips.
5:49Open AI is leasing the entire facility
5:51for 20 years, and if Open AI cannot pay
5:53the rent, Nvidia is liable. So, the
5:55chip seller is vouching for the debts
5:57of its largest customer so that they
5:58can continue to buy chips. And to the
6:01obvious question of whether this isn't
6:03circular financing, Nvidia CEO Jensen
6:05Huang responded on X: No, Open AI is
6:07paying the rent. And that may be true,
6:09as long as Open AI can actually afford
6:11to pay that rent. At the same time,
6:12Open AI is currently preparing one of
6:14the biggest IPOs of all time. The goal
6:17is a valuation of one trillion dollars.
6:20And since July, we also know that Open
6:22AI has proposed to the US government to
6:24transfer 5%of the company to them,
6:26officially as a partnership for America
6:29as an AI hub. Some say you can also
6:32view this as: whoever has the state as
6:34a co-owner will be bailed out in an
6:36emergency. Because companies whose
6:39business speaks for itself don't
6:40actually need guarantees from their own
6:42suppliers or the state on board. And if
6:44you now ask yourself how this even
6:46works, how long this can keep going,
6:48and where this chain might eventually
6:50snap, then you naturally look at the
6:52weakest link. And in this case, there
6:55are two: namely, Oracle and Open AI.
6:57Let’s start with Oracle. There are
6:59rating agencies, like S&P for example,
7:01that evaluate how securely a company
7:03can repay its debts. In principle,
7:05it’s the credit score for
7:07corporations, and Oracle received a
7:09rating of BBB-in July. This is the
7:11lowest grade at which a company is
7:13still considered a secure debtor. One
7:16step lower, Oracle would officially be
7:17a risky borrower, and pension funds and
7:19insurance companies would no longer be
7:20allowed to lend Oracle money under
7:22their own rules. And the reason for
7:25this is that half of all future Oracle
7:27orders depend on a single customer,
7:29namely OpenAI. For this one customer,
7:32Oracle has increased its debt in one
7:34year from 85 to around 125 billion
7:37dollars. On top of that, the contract
7:39timelines don't match up at all.
7:41Because Oracle has leased these data
7:42centers for nearly 20 years. OpenAI, on
7:45the other hand, has only booked them
7:46for 5 years. If OpenAI doesn't renew,
7:49Oracle will be stuck with empty halls
7:50for 15 years, unless they find someone
7:52else. Financial market professionals
7:54have long since been hedging, and
7:56insurance policies that would pay out
7:58in the event of an Oracle bankruptcy
8:00are currently at an all-time high. The
8:03second weak link is OpenAI itself; in
8:05June, OpenAI's financial figures were
8:07released and confirmed by the Financial
8:10Times. 13 billion dollars in revenue,
8:1321 billion in operating losses. For
8:16every dollar OpenAI brings in, the
8:17company loses a dollar sixty. HSBC bank
8:21estimates that OpenAI will need to
8:23raise an additional 200 billion dollars
8:25externally by 2030 just to fulfill its
8:27own contracts. generating revenue model
8:39intelligence And the CFO announced to
8:51employees in August that OpenAI would
8:53be on the stock market by 2027. The
8:57money has to last until then, and in
8:59March 2027, Softbank—one of OpenAI's
9:01most important backers—has to repay a
9:0340 billion dollar loan of its own.
9:07Which means things are going to get
9:08pretty uncomfortable for them. OpenAI
9:11is also under significant pressure
9:13internally, as within just a few days
9:15in August, COO Brad Lightcap left after
9:188 years and sales chief Denise Dresser
9:21left after less than a year, shortly
9:23after employees sold a total of 7
9:25billion dollars 'worth of OpenAI shares
9:28. CNBC cites industry experts who call
9:31this a huge warning sign ahead of the
9:34IPO, because normally no one leaves
9:36voluntarily right before their own
9:38shares go public. And raising investor
9:41capital isn't as easy as it used to be,
9:43as evidenced by how aggressively OpenAI
9:45is now courting money and customers.
9:48For a joint vehicle with private equity
9:50firms, OpenAI is guaranteeing investors
9:52a minimum annual return of 17.5%for 5
9:54years. No one gives those kinds of
9:57guarantees if investors are lining up
9:58on their own. This means that as long
10:01as new money keeps coming in, the party
10:03continues, but if the cash flow dries
10:05up, the first domino falls, followed by
10:06the data centers, and then the cloud
10:08revenue of the world's largest
10:10corporations. And of course, one has to
10:12wonder, if it’s all so shaky, why is
10:14everyone participating? Nvidia,
10:16Microsoft, Google—they know the
10:18numbers much better than we do. Why are
10:20they still pumping hundreds of billions
10:22more into this merry-go-round? That is
10:24because the other side of the equation
10:26looks just as extreme as the debt side.
10:29The demand for AI is currently growing
10:31faster than for any technology before
10:32it. Just take a look at Anthropic. At
10:35the end of 2024, they were at one
10:37billion dollars in annual revenue, 9
10:40billion by the end of 2025, 47 billion
10:43in May 2026, and 65 by the end of July.
10:48No company in software history has ever
10:50grown as fast as Anthropic, and in the
10:52second quarter, Anthropic reported that
10:54for the first time, their operating
10:56income exceeded their expenditures.
10:58They still need billions from the
11:00outside because training the next
11:02models is so expensive, and they have
11:04secured 200 billion dollars in
11:05computing power from Google alone for
11:07this purpose. The difference from
11:09OpenAI, however, lies in the ratio,
11:10because Anthropic promised much less
11:12and is also generating significantly
11:13more revenue. And it is precisely this
11:15ratio that ultimately determines who
11:17survives a crisis. And this is not an
11:18isolated case either. On OpenRouter, a
11:21platform that allows you to use any AI
11:23model from any provider, usage has
11:25increased fivefold in 6 months. This
11:27means more and more people want to use
11:29AI for their use cases. And by the way,
11:32you have long been feeling this demand
11:34yourself, because 32 GB of RAM for your
11:36PC cost 80 € a year ago, and today
11:39you are at 400. This is because memory
11:41manufacturers are now putting almost 70
11:43%of their production into AI servers.
11:46Production for 2027 is even already
11:48sold out, and Samsung itself is warning
11:50of shortages until at least 2027. If
11:54you want to buy a PC sometime soon, you
11:55are indirectly paying for this AI
11:57expansion as well. That is obviously
11:59annoying for you, but it is a strong
12:00signal for our question. Bubbles burst
12:02when too much supply meets too little
12:05demand. But right now, it is the other
12:06way around. Demand is outstripping
12:09supply. So demand is extremely high,
12:11and that is the reason why everyone is
12:13joining in. The crucial question is
12:15just whether this money will also end
12:17up with the companies that are
12:18currently spending the 700 billion and
12:20promised the 1.4 trillion. And this is
12:22exactly where China comes into play.
12:24You could see how nervous the market is
12:26about this question in January 2025. A
12:28then-barely-known Chinese lab called
12:30DeepSeek released a model that could
12:32keep up with the best American models,
12:34supposedly at a fraction of the
12:36training costs. On the next trading day
12:39, Nvidia lost 17%of its market
12:40capitalization, meaning nearly 600
12:42billion dollars in market value in a
12:44single day. That was the biggest
12:47single-day stock loss in history at the
12:49time, and not because of any financial
12:50problems, but simply because the
12:52question suddenly arose as to whether
12:54you really need so many expensive chips
12:56for top-tier AI and whether you really
12:58have to rely on these American models.
13:00And that question has remained ever
13:01since. The American government tested
13:04DeepSeek's current model themselves in
13:06May and concluded that it is about 8
13:09months behind the US leaders. So, only
13:118 months. But in return, it is freely
13:14available. Anyone can download it and
13:16run it on their own server. You need
13:18the hardware, of course, but
13:19nonetheless. And in operation, very
13:22importantly, it costs 60 to 90%less
13:23than the models from OpenAI and
13:25Anthropic, depending on the task. And
13:28DeepSeek is not alone. Alibaba,
13:30Moonshot, and Jina AI are releasing
13:33similar models weekly for the same low
13:35costs. And that is precisely why the
13:39demand for Chinese models has grown
13:40extremely in recent months, because for
13:42most everyday tasks, a model that is 8
13:44months behind but costs a fifth is
13:46simply the better choice. You don't
13:49always have to take the newest Bugatti
13:50just to drive somewhere. Many companies
13:53are therefore already relying on and
13:55using Chinese AI models. On OpenRouter,
13:58the share of American models has fallen
14:00from 70%to 30%within a year. A year and
14:04a half ago, Chinese models there were
14:06still at around 1%. And if you look at
14:09the leaderboard on OpenRouter under the
14:11most used models, you actually see
14:12almost nothing but Chinese models now.
14:15I still see GPT-4o, Claude, and Llama
14:17here, and Nemotron down there. But
14:19otherwise, those are all Chinese models
14:21, and I've also listed a few examples
14:22of companies that use these Chinese
14:24models. Coinbase, for example, the
14:27largest crypto exchange in the US,
14:29switched its developers to Chinese
14:31models by default in the summer and
14:33halved its AI costs. With more usage
14:36than before. Cursor, one of the world's
14:39most famous AI programming tools or
14:41development environments, built its own
14:43model on a Chinese model, and in March
14:45of this year, there was a huge scandal
14:47because they initially hid the fact
14:49that they were using Kimi 1.5 Pro for
14:51it, but then someone noticed it, and
14:53everyone was shocked again that it was
14:55even possible—that Chinese models are
14:57so good that Americans are now secretly
14:59using them in the background. And
15:03Airbnb, for example, operates its
15:04customer service with Alibaba Qwen
15:06because, according to their boss, it is
15:07good, fast, and cheap. Nonetheless,
15:10there are of course some security
15:11concerns, but the models are all
15:13open-weight. That means you can host
15:15them yourself, which ensures no data is
15:16sent to China, and that is exactly what
15:18many companies are doing now. Chinese
15:20models have also been hosted in the
15:22Ollama Cloud for a long time now. And
15:24here we can see. For example, Qwen 2.5:
15:25You can use that here now, and the data
15:27doesn't end up in China. Or DeepSeek V3
15:29, a very good model, and there will
15:31also be many data centers in the EU
15:33that will host these Chinese models.
15:36And this is the point where the US
15:38companies' math falls apart, because
15:39the old tech strategy—conquer the
15:41market first, then raise prices, and so
15:43on—no longer works because we have
15:45cheap alternatives. And you could
15:47really see that in the pricing of
15:49American models as soon as a new AI
15:51model was released by the Chinese.
15:53OpenAI cut its prices by up to 80%at
15:56the end of July. Anthropic had
15:58announced a 50%price increase for
16:00September and canceled it again in
16:01August. Even Claude 3.5. For example,
16:05Anthropic's most powerful model was
16:07originally supposed to be available
16:09only via API after one month, meaning
16:11billing per token. But that was exactly
16:14when the latest Chinese models, like
16:16Qwen 2.5, were released, and Anthropic
16:17thought: "No, we can't do that now." We
16:19have to keep Claude in the regular
16:21subscription plans so that people
16:23continue to use and buy it. This means
16:25that the prices per request are falling
16:27instead of rising. And for us as users,
16:29that is of course good news. But not
16:31for a company that now has to earn back
16:331.4 trillion. And China is attacking on
16:36a second front, namely with AI chips.
16:38Nvidia once had a 95%market share in
16:41China. Today it is zero because the US
16:44banned the export of AI chips to China,
16:46so China is now building its own chips,
16:49and Huawei is even doubling production
16:51this year, with DeepSeek's new model
16:54already running on Huawei chips. The
16:57second-largest AI market in the world
16:59is no longer buying from Nvidia, and so
17:01Nvidia's growth depends all the more on
17:03the carousel from before. The four
17:05largest Chinese tech giants are
17:07spending a combined 100 billion dollars
17:09on AI this year. And that is only a
17:11fraction of what the Americans are
17:13spending. This means they are paying
17:15only a fraction of that, yet they are
17:17only 8 months behind the leaders. And
17:19that is exactly a huge reason why the
17:21entire valuation of these companies is
17:23so shaky. So, let's get back to the
17:25question of whether or not we are
17:26currently in an AI bubble. It must be
17:28said quite clearly: people are using AI
17:29, they are paying for it, and the
17:30numbers are growing every month. AI is
17:32also here to stay. The product itself
17:35is not a bubble, but what companies
17:37spend and promise doesn't match their
17:39revenue. OpenAI has signed for 1.4
17:42trillion and brings in 40 billion.
17:45Nvidia is essentially financing its own
17:47customers. Companies are borrowing one
17:49in every three dollars, which only
17:50works if almost nothing goes wrong in
17:52the next few years. And as we have just
17:54seen, quite a bit is already going
17:55wrong. So you could say the bubble is
17:58in the price, not the product;
18:00companies are overvalued and
18:01over-leveraged for what they earn today
18:04, while Chinese models are driving down
18:06the prices they hoped to catch up with.
18:09At some point, it will balance out
18:11again, either because revenues keep
18:13growing like with Anthropic, or because
18:15valuations fall. And this isn't a new
18:16pattern either. In the year 2000, the
18:19Nasdaq tech index lost 78%, Amazon
18:21stock over 90%, yet the internet didn't
18:24disappear back then, and neither did
18:26Amazon. Investors may have lost their
18:29money, but the technology remained, and
18:31it was the same story 100 years earlier
18:32with the railroads. Shareholders went
18:34bust, yet the tracks were still there
18:36and were used for another century. And
18:38this isn't just what I think. In July,
18:41an internal report from the US Treasury
18:43leaked, written for the Treasury
18:45Secretary and top financial regulators.
18:47It states that AI companies are
18:49currently much more deeply embedded in
18:51the economy than dot-com firms were
18:53back then, and that a crash would
18:55therefore hit stock markets, credit
18:56markets, and even power utilities. But
18:59it also says that the big AI firms are
19:01financially much healthier than the
19:04dot-com busts of 2000, which could
19:06cushion a crash. But publicly, the US
19:09Treasury naturally continues to promote
19:11the growth story and the big AI boom.
19:14So, in conclusion, let's look at the
19:15question of what exactly is going to
19:16happen now? How can things proceed from
19:18here? And, of course, no one can answer
19:20that for you right now. Neither can I,
19:22but there are basically three ways this
19:23whole thing can play out. The first way
19:26is a slow correction, I would say. I
19:28actually consider that quite likely.
19:30That means there won't be a big bang,
19:31there won't be a bubble that suddenly
19:33bursts, but rather the weakest
19:34companies and projects will simply fall
19:36away one by one. This means a portion
19:38of the planned data centers will simply
19:40never be built. Of those scheduled by
19:422027, 16%haven't even started
19:44construction yet. And a few providers
19:47that built everything on credit will
19:49likely go bankrupt, and AI stocks on
19:51the market will drop by 20 or 30%once
19:53or twice. Oracle, for example, even
19:55more so. This trend has already
19:58partially begun, because the Stargate
20:00project, which Trump announced in
20:02January 2025 with $ 500 billion, still
20:04has no staff today and isn't building
20:06anything itself. Oracle and OpenAI
20:09already halted expansion at the
20:11flagship site in Texas in March, and in
20:14the first quarter of this year alone,
20:16public protests in the U.S. blocked or
20:18delayed $ 130 billion worth of data
20:20centers. This means that at the end of
20:23this path, maybe a few big providers
20:25plus open models from China will remain
20:27, and the expansion will continue, just
20:29slower. Let’s move to the second
20:31possible path, and that is the chain
20:33reaction. It begins with a single event
20:35; for instance, if the OpenAI IPO fails
20:38or if a Chinese lab delivers the next "
20:40Deepsea" moment right beforehand,
20:42OpenAI won't get any fresh capital.
20:45Then OpenAI won't be able to pay its
20:47contracts. Then Oracle will lose half
20:49its orders. Then Nvidia will have to
20:51foot the bill for the rent in Ohio. And
20:53now the thought from the beginning
20:54comes back. Three-quarters of U.S.
20:56growth depends on this expansion; if
20:58this chain breaks. Then we're not just
21:00talking about a few tech stocks, but
21:02perhaps a recession that hits many
21:04people who might not even use AI or
21:06haven't written a single AI prompt. And
21:09that is exactly why I consider this
21:11path somewhat less likely, because a
21:13government whose growth depends on this
21:15expansion won't just stand by and watch
21:17the chain break, right? So, I believe
21:19they will definitely take their
21:20measures. And the fact that OpenAI
21:22offered the state 5%also fits into this
21:24picture for me. And the third path is
21:26what’s called "growing into it." This
21:28means revenues continue to grow. Demand
21:30remains extremely strong for American
21:32AI models, and eventually, these firms
21:34truly earn as much as they are
21:36currently valued at. According to
21:38Anthropic's figures, that is possible.
21:40But it is the only path where very
21:42little can go wrong over the next few
21:43years and everything has to play into
21:45American hands. So, what does all this
21:47mean for you concretely, and what do
21:49you do with it now? A financial
21:51YouTuber would tell you at this point
21:52which stocks you should avoid. I’d
21:54rather tell you what this means for
21:56your work with AI. Perhaps a small
21:58sentence on stocks anyway. I’m no
21:59financial advisor, but if an industry
22:02needs record-breaking IPOs just to pay
22:04its bills, the retail investor who gets
22:07in at the very end is rarely the winner
22:09. For your work with AI, however, the
22:11direction remains the same in all three
22:13scenarios, because models are getting
22:15better and prices per query are falling
22:17due to competition from China, open
22:19source, and more efficient chips. And
22:21even in the event of a crash, the data
22:23centers aren't just going to be torn
22:25down if an investor goes bankrupt. They
22:28will be sold and keep running, just at
22:30a lower cost. There are three specific
22:32things I would advise you to do now.
22:33First, build your automations so that
22:36you can swap out the AI model at any
22:38time. Don't make yourself dependent on
22:41a single provider. We saw how necessary
22:43that is back in June, when Claude 3.5
22:46was supposed to be shut off for
22:47everyone outside the US per government
22:50order, and Anthropic just turned it off
22:52for everyone for two weeks, and nobody
22:54knew what exactly was going to happen.
22:57And if your entire workflow relies on
22:59one single provider, you have the same
23:00problem as Oracle, just on a smaller
23:02scale. By the way, if you can swap your
23:05model, you can also do what many big
23:07companies are currently doing: offload
23:09standard tasks to a cheap, open-source
23:11Chinese model and only use the
23:13expensive one where it truly makes a
23:15difference. Second, be careful with
23:18long, expensive contracts for any AI
23:20platforms, because as you can see,
23:22things change daily and prices are
23:24falling extremely fast. This means a
23:26three-year contract signed today might
23:28be completely overpriced in a year. And
23:30third, and this is the most important
23:32point, in the end, it doesn't matter
23:34which model you use or if you are using
23:36the very best one. What matters is
23:39whether AI is being used in your
23:40company in a way that actually saves
23:42time, cuts costs, or generates revenue.
23:45This means it must be properly
23:46implemented within the company, and you
23:48may even need to completely rethink
23:50your own processes to do so. And that
23:52is exactly where many companies are
23:55currently failing; in a major survey,
23:57for example, where over 2,500 corporate
24:00decision-makers were polled, 74%said
24:02they had to shut down an AI agent after
24:04launch. And not because the model was
24:07too poor, but because the rules,
24:09processes, and oversight were missing.
24:11What’s scarce isn't the AI, but the
24:13ability to actually implement it
24:14correctly within the company. And
24:16whoever can do that, whoever rethinks
24:18processes and adapts accordingly, will
24:19be needed in every scenario, especially
24:21in a crash scenario, because that is
24:23exactly when companies watch their
24:25money. If you would like to learn more
24:27about this, you are of course welcome
24:28to drop by my community at any time.
24:30The link is in the video description.
24:31Here, we cover exactly how to really
24:33use AI in a corporate context and not
24:35just automate processes, but rethink
24:37them as well. And how can you approach
24:40all these topics, especially if you are
24:41not a programmer? We cover all of this
24:43very intensively here. That means if
24:44you are interested, you are always more
24:46than welcome. That is my assessment as
24:48of today. I hope you were able to take
24:49something away from this. Now, I would
24:51be very interested to hear your opinion
24:53on this and which scenario you think is
24:55the most likely. Feel free to write it
24:57in the comments so we can continue the
24:58discussion there. Otherwise, thank you
25:00very much for watching. And if you are
25:02specifically interested in the
25:03practical side, meaning automations
25:05that pay off even without
25:06trillion-dollar valuations, you will
25:07find plenty of content here on this
25:09channel. So feel free to take a look,
25:11and leave a like and a subscription so
25:12you don't miss any future content. I
25:14would say we will see each other again
25:15in the next video.

About this transcript

This page contains the full transcript of Warum die KI-Blase bald platzt & was du tun solltest by Julian Ivanov | KI-Automatisierung, generated from the public captions YouTube serves with the video. The transcript has 4,717 words across 715 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.