光ファイバーは数十年、GPUは3年：AIブームはドットコムブームとは違う（第318回）

Thank you for listening to Data Science at Home podcast with Francesco Gattoleta. You are about to get cutting-edge insights from the people who are reshaping the world of technology with machine learning, data science, and artificial intelligence. It's time for Data Science at Home. Welcome to the show. So let me tell you a story that if you follow this show, of course, you might know. On September 15th, 2026, so this year, a company called TypeSafe AI came out of stealth. This company has been funded by Diogo Almeida, who spent about four years at OpenAI working on the training methods behind ChatGVT. Together with Eric Gaffney and Sasha Scheng, they launched a model called Jav. For those who don't know what JEV is, I made a podcast episode about JEV. There is also on YouTube and on the official website, datascienceatom.com.

And now, JEV is, I will explain better what JEV is, you know, just to make this episode self-contained, though I believe that the episode I'm referring to is much more complete and gives you a really good grasp on what JEV does, can do, and potentially cannot do. Now, alongside this company, Diogo Meda and Type-C FAI, they raised something like $40 million seed round, right? The round was led by DCTC, and Forbes put the valuation at around $200 million, right? So, now, why am I laughing is because if you know what Jeff does, you know, he would laugh too. So, essentially, what does Jeff do? Well, you give it some input. It can be a support ticket, a database row, a fraud signal, and it gives you back a decision. Okay? That is a label, approve or reject, urgent or not urgent, if it's tickets, with a confidence

score attached to it. Right? Now, if you have done any machine learning in the last 20 years, you already know what that is. Okay? It's a classifier. Right? A very fancy, a very fast transformer-based classifier, but a classifier. And that's not even the clue of the story. The real punchline is that 24 days later, on October the 9th, several outlets reported that the same company raised a Series A of about, hear this well, $870 million, led by Andresin Horowitz, of course, with Sequoia joining at a valuation of, hear me well, $7.5 billion. So from $200 million to $7.5 billion in just 24 days for a classifier. Frame this, okay, for a model

that picks an option from a list. that's the technology and the sophistication of the investment then four days after that another company called Scale Down launches something that they call Constrained Decoding Models CDN another fancy name for classification extraction and summarization claiming to be up to 200 times faster and 467 times cheaper I really want to know how did they get to that number? Like, not 465, not 470, but 467 times cheaper than Frontier LLM models, right? Now, I'm smiling because, in fact, I should be crying. I hope you understand the... You caught the... It's not even sarcasm. It's like I gave up.

Alright, jokes that part. I've seen this movie before, and I hope you did too. I'm sure actually you've seen this movie before too. It was called Blockchain. And today I want to show you why the story is kind of repeating, why it's repeating for exactly the same reasons. And of course, this is the part that should matter to you, even if you never wrote a line of code in your life. What happens to money and to the people holding it when the music stops. Now, let me be fair first, okay? Because otherwise the episode is just me yelling at things. Jamf is probably a good piece of engineering, all right? The company says it answers in 70 to 500 milliseconds, while a Frontier LLM on the same structure tasks takes something from three seconds up to several minutes, depending on how big the context is. They price it at 4.2 cents per million input tokens,

and output is basically free. So Vercel, Cloudflare, Lankchain, they all integrated it within days, okay, which tells you a lot about how sophisticated and difficult this thing is. It's a classifier. If it can be integrated within days, it means that after all, it was not that, you know, crazy sophisticated. Developers signed up in the hundreds of thousands, and, of course, good for them, seriously. But look at what the actual innovation is here. The innovation is the admission that you shouldn't use a gigantic multi-billion parameter language model, running on a rack of GPUs that costs as much as a house these days and to decide if an email is spam or not, which is something every data scientist on the planet already knew from 2015. And if you're following this show from 2018, you should know that too.

We had logistic regression for that. We had gradient boosting. Then we had BERT. They all ran on a laptop, right? Now, these are very nice models that are way enough for the thing they have to do, the task they have to do, which is classification. So, here's the plot. for three years, the industry told everyone that the giant LLM is the answer to everything. Okay? Classifications, use an LLM. Extractions, use an LLM. Summarizing your grocery list, use an LLM. We got bombarded by these messages. Okay? I'm trying to sum it up here. It took us about three years and about a trillion dollar to figure out, or rather to convince the marketing departments and the marketeers of the world, that large language models are too expensive for most of what companies actually need. And now the solution to that problem is new startups with new names, new acronyms, new benchmarks, that by the way, they run themselves, and new billion dollar valuations.

So it's essentially nonsense added on top of nonsense. In Italy, we say it, and I quote, Pioves sul bagnato. It rains on water. Yeah, you got a bit visual people. And, you know, by the way, those 190, what was, 93 times faster and 447 times cheaper, these numbers, you know, they come from TypeSafe's own benchmark, not an independent audit, you know. So you have to keep this in mind as well, because it's the first thing that should remind you of blockchain. Right. Remember the white paper. OK, those who have survived that period. And I'm telling you this because I worked on blockchain projects. Let me tell you a bit of details what happened back in the days, which was like less than eight years ago. Actually, 10 years ago, 2017. Right. The year everything was a blockchain. So let's go back to, let's say, December. Actually, let's go back to 21 December 2017.

There is a company called Long Island Iced Tea Corp. Long Island Iced Tea Corp. They make, well, Iced Tea. I'm not kidding. Market value, $24 million. That morning, 21st December, they announced that they are changing their name to Long Blockchain Corp. That's it. same iced tea, and they said explicitly keeping doing iced tea, the stock immediately jumped more than 200% at the open and closed the day up to 180%, something like that. It was an amazing jump. A few weeks later, January 2018, Kodak, the camera company, the one that had already gone bankrupt once when they announced the market of digital cameras because they were doing negative films and stuff and chemicals and all that stuff.

Well, they announced Kodakoin, right? The cryptocurrency for photographer, whatever that means or that meant. Stock more than doubles in two days. Kodakoin, as far as anyone can tell, never really became a thing. So it disappeared. Then there were all the ICOs. Remember, the initial coin offerings and recipe was very simple. You write a white paper, typically a PDF with some diagrams and the word decentralized every three lines. And you announce a token and people send you money. It works like I wasn't in such things. So all the crypto folks want to leave bad comments, please do. I was there, guys. I've been in these teams. now billions of dollars went into this white paper uh you know scam at the end we can call it the right name in march 2018 a firm called satis group looked at the icos of that period

and classified roughly 81 of them as scams not failed scams ernst and young looked at the class of 2017 ICOs, initial coin offering, and found that if you had bought a basket of them on January the 1st, 2018, you would most likely have lost about two thirds of your money within the same year. And 71% of those projects had absolutely no product on the market at all. Okay? All right, so you'll say, okay, but that's crypto, that's the Wild West, and, you know, serious companies don't do that. I've heard this story. The drill is very well known to me, guys. So let me stop you right there because serious companies did exactly that. And I'm telling you, I'm giving you names now. The Australian Securities Exchange, right? The stock exchange of an entire country. They decided in 2017 to replace its settlement system, which is called CHESS, C-H-E-S-S,

with a blockchain-based one. five years later in November 2022 after an independent review by Accenture found the project only 63% complete and quote excessively complex no shit they scrapped it and wrote off roughly 250 million Australian dollars Maersk and IBM built TradeLens another company actually that was a blockchain platform for global shipping they shut down 2022 who remembers Facebook Facebook and Libra or Libra or whatever that's called the global currency it disappeared in 2019 they renamed it to Diem SaltBot was left in early 2022 nobody is ever talking about that pretty much the same with the metaverse a few years later and in research

well that's in fact the part that hurts me personally because I've been a researcher for a few years you could not get a grant, a paper or a conference talk without the word blockchain in it and now the same is happening for AI but we'll get there we had blockchain for healthcare we had blockchain for supply chains we had blockchain for voting systems we had blockchain for agriculture I'm not joking I've read some of those proposals I still do paper reviews for academic papers. Most of them were solving a problem that a regular database had solved already in the 70s. But they used this decentralized ledger, blah, blah, just to justify their white paper, climbing the new tech frenzy. But, you know, technologically, there was nothing that a regular database could have solved So where is blockchain today Well Bitcoin is still there sure Speculation is still there But blockchain will transform every industry statement if you remember

Nobody in the boardroom says that word anymore. Nobody ever talks about decentralization anymore. Why? Because there was a tiny kernel of real technology wrapped in a gigantic layer of narrative. And when the narrative ran out, well, the kernel turned out to be much, much smaller than the money that had been thrown at it. That was the sad reality, and nobody was there to tell you. So, why the story repeats? Because now I want to be very careful here, because, you know, this is where most people who make this comparison get lazy. and that's also a big exposure to attacks, you know, in my statements. So let me put a disclaimer here. AI is not blockchain, right? AI is far more useful than blockchain. That's my personal opinion. By having been in both, there was even AI and blockchain combination startups.

That was like really, really crazy. AI, blockchain and space. They started mixing. fuzzy words. But anyway, let me go back to reality. I do think that AI is far more useful than blockchain, as it was. I personally use these tools every day, and millions of people do. Language models are genuinely remarkable technology. That's not the point, okay? The point is, the investment behavior is the same, and that's what I'm discussing today. And it's the same for the same reasons. And that's also what I'm discussing today. So let me give you five of them. And if you're a manager, I want you to recognize these. And please get back in comments on YouTube or the official website or even on Discord where we can interact with each other. Okay? I'm really eager to have a conversation about these things because you're probably being exposed to all of them right now. So hear me out. The first reason is there

There is a real kernel wrapped in a story of inevitability. Bitcoin worked. LLMs worked. That's what makes these doubles very dangerous. It's not pure fraud. There is something real underneath. It's inflated, but there is something. There is a root. And the story you build on top is this will change everything. And if you're not now, you'll be left behind. Sylvester Stallone I have a lot of respect for that guy this will take everything and if you're not it will leave me behind it was like saying these things like you're an actor dude don't talk about things that you cannot understand like really I don't talk about acting and stuff talk about your shit anyway there are many you know these marketeers in the world and influencers that have been pumping that's my point they have been pumping very wrong messages and today we are where we are anyway once the story is about inevitability, nobody asks about

price anymore, right? That's a psychological trick. The second reason is the venture capital math rewards insanity, and years I've been saying this I mean, I'm sorry folks, I don't like VCs, I don't like venture capitalists because they pollute technology they pollute the messages to their own benefit. VC returns follow what's called a power law, all right? Out of 100 bets, 90-something die, okay? And one or two have to pay for everything. So that means for a fund, the risk is not losing money on a stupid bet. The risk is missing the one that becomes OpenAI, for example. So when a hot category appears, everyone piles in at any price, you know, it's rational for the individual fund, but it's completely irrational for the system. I hope this is clear because, you know, this is something that many people keep ignoring. And really, you don't need to be a financial expert to understand this.

Anyway, this episode is supported by OutShift, Cisco's incubation engine for frontier technology. Multi-agent AI is in pretty much every enterprise roadmap right now. But there is a problem with that, and we are all facing it. Agents can pass messages, but they cannot think together. So these systems silently underperform through cognitive failures, things like misreadings, unverified claims, and of course, false consensus. You know, all the stuff that conventional monitoring cannot detect. OutShift by Cisco is building this fix, the Internet of Cognition, which is an open source foundation for building multi-agent systems from design to production. With shared contacts, shared memory, and guardrails to drive the results that we actually expect from artificial intelligence. So, read the paper, experience the demo, grab the code from outshift.com.

That's O-U-T-S-H-I-F-T dot com. the third one, the third reason pedigrees the new white paper and here is something that I noticed personally in 2017 the white paper was the collateral, it's a PDF today the collateral is a resume a CV for the Europeans if you read through the lines and if you pay attention to who these people are, whenever they make a claim X open AI, X deep mind, Mira Murati, former CTO at OpenAI, founded Thinking Machine Lab, and in July 2025, they closed a $2 billion seed round, seed, a seed at $12 billion valuation, again, led by Horowitz, they're always, you know, when there is some shitstorm, they're always in that equation, with no product and no revenue at the time, $12 billion valuation. I mean,

this is crazy, guys, I mean, really. The investors were buying the people, right? X CTO, X OpenAI, X DeepMind, XXX. They don't say, who are these people? What did they do exactly? They just sit at the table of OpenAI and what did they do? We don't know. So the investors buy the people. Maybe that works out, but notice the lines. There's no number in the page you can check. That's exactly what the white paper was. Claims, typically unverified or verified internally. The fourth reason, self-reported benchmarks, speaking about self-verified things. In blockchain, it was transactions per second. In the white paper, they claimed essentially, you know, that was the only metric like, we have this side chain that can process this many contracts

or smart contracts per second or this many transactions. That was like the metric, like speed, which was a good heuristic. Now, in 2025, in AI, it's this 193 times faster on our benchmark. That's, you know, that's the new metric. Benchmarks that the company designed, ran, and published itself. If your accountant did that with your taxes, you and your accountant would be together in prison. The fifth reason, nobody gets fired for having an AI strategy. In 2018, every CEO needed a blockchain strategy. In 2025, every CEO needed an AI strategy. So boards ask for it, shareholders reward it, consultants sell it. And here's the receipt for that, you know. Now, MIT's NANDA Initiative, N-A-N-D-A, published a report in summer 2025 called the Gen AI Divide.

And their headline finding was that about 95% of organizations in their study were getting no measurable profit and loss impact from their generative AI initiatives. There are these reports. The problem is that nobody reads, nobody pays attention to them, or they just get submerged by noise. Submerged by noise. 30 to 40 billion dollars of enterprise spending and only about one pilot in 20 producing real revenue. You can argue about the methodology and people did, but the direction is pretty clear here. Right. Now, if you think TimeSafe is an isolated case, think twice because that's, you know, let me let me read a few more names here. And some are hilarious, some are scary, some are very well known. Probably you know some of them. So the first one is Builder.ai. Builder.ai is a London-based startup backed by Microsoft, among others,

valued at around $1.5 billion, billion, that promised to build your software app with AI. It was called Natasha. Natasha was their AI assistant. she or he, whatever, would assemble your app like Lego. In 2025, the whole thing collapsed into insolvency after a lender pulled about $40 million out of its accounts. And then came the report that much of the AI was, in fact, hundreds of human engineers, many of them in India, not paid well, for sure, doing the work by hand. And on top of that, Bloomberg also reported allegations of inflated revenue. Yes. U.S. prosecutors started asking for documents. This is not, of course, AI underperformed. This is AI in quotes. AI. Okay? The second name worth reading is Humane. You know, the AI pin.

That probably very few people remember because it was really a meteor. it was a little device you would clip to your shirt that would replace your smartphone I mean that was ideally the use case and it raised hundreds of millions of dollars sold for something like $700 to the $700 gadget that reviewers literally destroyed tried to sell to sell itself for somewhere between $750 million and a billion but you know what they got? in February 2025 HP bought what was left of that startup the patents, the team for 116 million a fraction of what they asked and then they of course in my opinion already too much but okay and then they switched the pins off because you know if you bought one you now have a very expensive piece of museum then we have the biggest of all these financial experts here

are really trembling the circular deals. And this is, you know, this really keeps economists up at night because when you try to explain this, people will usually tell me, like, come on, dude, you're kidding. That's not possible. And it is very possible. It's actually happening. It's happened and it keeps happening. In September 2025, NVIDIA announced it would invest up to $100 billion in OpenAI. So hear me out now, because this is hilarious. Now, what does OpenAI do with that money? Well, it buys NVIDIA chips. Around the same time, Oracle signed a deal to provide about $300 billion of computing to OpenAI over five years. What does Oracle do to provide it? It buys NVIDIA chips. and then OpenAI signs a deal with AMD that could make OpenAI one of AMD's biggest shareholders

in exchange for buying AMD chips so add it all up and OpenAI had computing commitments of about a trillion dollars for a company whose revenues are a very small fraction of that a very small fraction not a fraction a very small fraction of that so money goes from the chip maker to the customer, from the customer back to the chip maker, and everybody books revenue. Analysts started calling it the you know, it's like a snake that is hitting its own tail. And if that structure sounds familiar, it should. We'll come back to that in a second, towards the end of the episode. Now, the skill of OpenAI himself is another reason. In fact, August 2025, Sam Altman told The Verge, are we in a face where investors as

a whole and I quote are overexcited about AI My answer is yes And and I quote someone is going to lose a phenomenal amount of money He literally said that So when the guy at the center of the party tells you you know tells you the punch, in fact, maybe we should be listening. Now, one can say, as I already had, because I'm reading my notes here, I have some objections to my statements and observations of course which is fine, you know, I speak with friends and colleagues and other people and say, what do you think about this? Do you find it familiar? Do you find an analogy? And I found one of the most sophisticated defense of all this spending from very smart people and it goes typically like this say, sure, it's a bubble AI is a bubble, so was the dot-com bubble.

So they mentioned the dot-com bubble. So for those who don't know the dot-com bubble, in the late 90s, telecom companies laid fiber optic cables pretty much everywhere, right? Way more than anyone needed. And, you know, all network equipment, not just cables, but all the network equipment. Now, in 2001, they went bankrupt, right? So investors lost everything, but the fiber stayed in the ground. And a few years later, we got YouTube, Netflix, cloud computing, all of it ran on the same cheap, abandoned, let's say, abandoned fiber, right? That was like in a summary, really in a nutshell, what happened with the dot-com bubble, right? But of course, you know, check Wikipedia, Google, whatever source you find familiar, working with, and read this because it's a nice story of, you know, of the history of the internet, in fact. the bubble financed the future that's what people say and the same thing will happen with AI

that's what people say today even if the companies die the data centers will stay which when you hear this is a beautiful story and it's partly true that's why it's beautiful it's credible because it's partly true so let's look at it properly now because I think it's actually the most important part of the episode I keep saying that, but, you know, everything is important in this episode. Now, let me tell you what happened with the fiber. In the late 90s, network builders powered something like 30 billion dollar into laying around 19 million miles of fiber only in the United States. Why so much, you might ask? Well, partly because of a famous claim pushed by WorldCom's internet unit that internet traffic was doubling every 100 days. Then there was a researcher, by the way, Andrew Odlitsko, who actually checked and found that traffic was growing pretty much,

but more like doubling every year, not every 100 days. But okay, a tiny mismatch. But the whole industry was building for a demand curve that did not exist. That's why they went bankrupt. These are the numbers, folks. I'm not making shit up. This is what happened. This is history, right? Facts. Now, this should sound familiar to you. I'm driving you towards where I want you to go. Compute demand is infinite. Scaling loads. We need 10 gigawatts. Sound familiar now? So, by 2002, estimates put the share of long owl fiber that was dark. it's called in the jargon, dark fiber, which means installed but not carrying any traffic, at well over 90%. So 90% of the fiber back in the days was dark, unused, okay? It was buried in the ground, but nobody was ever using that. So more than 60 telecom companies went under between 2001 and 2003.

In just two years, 60% flashed or slashed, okay? Global Crossing, which had built a network spanning 27 countries, also filed for bankruptcy in January 2002 and was eventually sold for $250 million after investors lost tens of billions. WorldCom, you know, the guys of the claim, they also followed six months later, the biggest bankruptcy in American history at the time, with an accounting fraud on top. So, you know, they really didn't miss anything. And by the way, vendor financing was a big part of that bubble too, right? Equipment makers lent money to the telecom startups, so those startups could buy their equipment. Sounds familiar now? The city was financing again. Well, again, the first time back then. Then the startups died, and the loans died with the startups, right? So who pays for that? No one.

Cisco was the darling of that era, you know, the most valuable company in the world for the moment in March 2000, actually. That was my first year at university. I remember that very well. Cisco's stock took roughly a quarter of a century to get back to that peak, right? They never went back, if not after 25 years. Today, a lot of people call NVIDIA the Cisco of this cycle, right? So I'll let you draw your conclusions about these circular deals that I mentioned earlier, because, you know, I think you're on the right track now. Well, you should. So why the fiber eventually becomes useful? Well, there were three reasons, and they're really, really crucial. The first reason is physics. Glass doesn't rot, okay? A fiber optic cable in the ground is good for decades, 20, 30 years or more, right, if it's well-maintained. the fiber lady in 1999 is in many cases still carrying your traffic today and probably this video.

The second is that the expensive part of the work was digging the trench. It was not the glass, it was not the fiber per se, right? Once you have dug a hole across a continent, in the ocean, for example, throwing in 10 extra fibers costs pretty much nothing, right? That's why they overbuilt so much. The third reason is that, and it's quite beautiful, the same glass got faster over time. And you say, how? Well, there was a technology that was called Wavelength Division Multiplexing that basically consists of sending many colors of light down the same fiber at once. So this technology multiplied the capacity of a single strand of fiber by up to 100 times. you didn't have to replace the cable you just replaced the boxes at the two ends at the two edges the asset in the ground got even more valuable

every year so the dot com burst left behind an asset that last decades that was cheap to keep and that got better with age like cheese, like parmigiano I'm kidding that's why someone could buy for pennies after the crash and could build Netflix on top of it. Now, let's move back to the present, okay? GPUs are not fiber, right? Now, look at what the iBoom is actually buying. The biggest chunk of the money is not going into trenches, but it's going into GPUs. And GPUs are the opposite of fiber on every single one of those three points. GPUs rot economically, of course. NVIDIA now releases a new architecture every year. Hopper, then Blackwell then Rubin, God knows what, next and the speed up between generations is actually very consistent at GTC in March 2025

I call him Black Jacket, Black Leather Jensen Wang Jensen Wang the guy in the Black Leather Jacket introduced Blackwell and literally called himself I'm not kidding guys the chief revenue destroyer I mean, when you are at the apex of something, you become also a bit stupid that's, you know, kind of the rule Peter's principle, probably maybe chief revenue destroyer, well, anyway he just said, you know, that Blackwell Blackwell Hopper is fine in some circumstances but not many Blackwell is the news the future, blah, blah, blah, blah the usual thing that's the CEO of NVIDIA telling you essentially on stage that the chips that your customer that his customers bought I don't know, a year ago or two years ago are obsolete

and the market agrees because renting an H100 NVIDIA H100 the flagship chip of 2023 costs something like $8 an hour in early 2024. And by late 2024, not even one year, reports put it down to $2. Okay? So fiber didn't do that. That's what I'm saying. Like fiber actually got more valuable. The expensive part is the part that rots. This is another difference. In a data center, the building, the power connection, the cooling, those last for decades, like the trench, of course. but the bulk of the cost is the silicon and the silicon has a useful economic life of a few years so the analogy is backwards in the dot-com era, the expensive bit was durable and the cheap bit got upgraded while in the high era, the expensive bit is the one with the shortest life you get the point

and you cannot upgrade a GPU in place there is no equivalent for wavelength division multiplexing for an H100 You just have to replace the entire thing, right? When it's obsolete, it's obsolete. It goes in the bin or it gets sold off for a fraction of its price or it gets used for other tasks. So here's where it gets even more interesting for the managers, I promised at the beginning of the episode, and the finance people listening, which I believe there are. Well, at least I hope. So when a company buys an expensive asset, it doesn't put the whole cost of this year's books, right? It spreads over the asset's useful life. In finance, this is called useful life. And it's called depreciation, right? Now, the longer the useful life you assume, the smaller the yearly cost, okay? And the bigger your profits look, right? So what useful life are the hyperscaler assuming for these GPUs that their own suppliers say are obsolete in a year or two?

Let's look at some numbers. Microsoft and Alphabet, so Google, they say six years. Meta says five and a half years. And Meta's own filings showed that extending that assumption cut its depreciation expense by about $2.3 billion in just the first month of 2025. Amazon actually went the other way and shortened the life of some of its servers from six to five years in early 2025, which tells you that at least someone looked at it honestly. Michael Burry, the big short guy the one who saw the 2008 housing crash coming he estimated that between 2026 and 2028 this practice could understate depreciation across the big players by about 176 billion dollars in plain English this means that the profits look better than they are because the books assume the chips will be useful

for much longer than anyone believes, right? They should last less. And that, you know, history tells this guy is right. Now, to be fair, there's a counter argument to that. Older GPUs don't, you know, actually disappear. They get moved to less demanding work, like inference, for example, instead of training. That's real. But the rental prices I just told you about suggests that less demanding work also pays a lot less. You know? Easier tasks are also done with less money. Okay? Now, GPU as a collateral. This is the part that turns this from a stock market problem into a credit problem now. Because in August 2023, there was a company, very famous, you should know that, CoreWeave, a NeoCloud, as they call basically a landlord renting out GPUs, they borrowed $2.3 billion from Magnetar and Blackstone

using the NVIDIA H100 as collateral, right? And since then GPU lending has become an entire industry by itself okay Core weaves that went from under billion in 2024 to more than billion now So think about what that means. In the dot-com era, the collateral was fiber that lasts 30 years, right? And eventually improving. Today, the collateral is a chip whose rental value can drop by two-thirds in a year. So imagine a mortgage on a house that loses half its value every 18 months. Would you land on that? Apparently, lots of people would. And it's not only the new clouds. In October 2025, Meta financed its giant Hyperion data center in Louisiana through a separate vehicle with the Blue Owl capital, around $27 billion of debt

that sits outside Meta's balance sheet. PIMCO and BlackRock bought the bonds. Meta keeps about 20%, rents the facility, and guarantees part of the value. It's perfectly legal. It's actually rated A+. But remember, in October 2025, the IMF chief economist said, one reason an AI crash would be less dangerous than past crisis was that, and I quote, This is not financed by debt. A year later, more and more of it is. I'm laughing here, but there is, you know, this is dangerous. Especially when these things end up in stocks, if you know what I mean. Don't take my words as a financial advice, of course, but pay a lot of attention. Be careful. so is the dot-com comparison valid yes but not in the comforting way that people use it

at least the people i have been speaking with it's valid on the psychology that's the the thing as many other messages the same euphoria the same demand doubles every hundred days like statements the same vendor financing loops the same but this time the fundamentals are different every time is different than it's not. Believe me, it's really not. It's valid on the numbers. Aliens Research or Alliance Research estimated that the gap between AI capital spending and AI revenue growth is running at around 46%, higher than the 32% seen in 2001 telecom cycle. JP Morgan estimated that the hyperscalers' CAPEX would eat 90% to 100% of their operating cash flow in 2026. Goldman Sachs estimated that the hyperscalers need roughly $300 billion a year a year in AI revenue just to break even

on depreciation and operating costs. Just to operate the service they need $300 billion revenue. And back in 2024, David Kahn at Sequoia was already calling this AI's $600 billion question. now where the analogy is not valid it's not valid on the hangover when the fiber bubble burst what was left behind was a gift to the future when the GPU bubble burst if it does a big chunk of what's left behind will be a warehouse of chips three generations old the part of the legacy that will survive is the boring part, you know, the buildings, the substations, the grid connections, the cooling, the power contracts. That's the fiber optic of this cycle, not the silicon. So, let's talk about what could actually go wrong

and who pays for it, because, you know, bubbles sound like an abstract word now, but it's not. There are several risks, okay? Risk one, from equity to debt. When a bubble is financed with equity, with stock, the people who lose are the shareholders. And they knew they were gambling. So that was mostly the dot-com crash. It was painful, but contained. When a bubble is financed with debt, the loss is spread to whoever lent the money. Banks, private credit funds, insurance companies, pension funds, all that stuff. That's the difference between 2001 and 2008. GPU-backed loans, off-balance sheet data center vehicles, private credit, that's the road from the first time to the second. The Bank of England already warned in October 2025 that the risk of a sharp correction had increased and that stock valuations were, and I quote, comparable to the peak of the dot-com bubble.

The second risk is concentration. A handful of companies, think about the usual suspects here, Nvidia, Microsoft, Google Alphabet, Amazon, Meta, and a few others, make up an enormous share of the US stock market. So if you have a pension fund or an index fund, you are heavily invested in this AI bet, whether you know it or not. You didn't choose to buy GPUs, but you did anyway. That's why I said, pay attention. The third risk, you're probably already seeing the first cracks. In mid-October this year, the 20 most valuable AI and cheap stocks globally reportedly lost more than a trillion dollars in a single week on worries about higher borrowing costs and how the build-out is going to get financed. Yeah, no shit. black leather jacket Jensen Jensen Huang himself was quoted saying that demand from the neoclouts is strong

but the capital is scarce when the guy who sells the shovels says the miners are running out of money to buy shovels that's a data point the fourth risk, opportunity cost this is a silent one I would say every dollar, every engineer, every PhD student every megawatt of electricity going into LLM for X, LLM for Y, is not going somewhere else, right? So during the blockchain years, we burned a generation of research funding on bullshits like distributed ledgers for problems that definitely didn't need one. Today, we're doing the same thing. If your research doesn't have LLM or agent in the title, good luck publishing. robotics, energy, materials old school statistics, actual software engineering, they all lose talent and funding to the shiny thing of the moment risk 5, the vendor that

vanishes, for the managers this is the most practical because if your company builds a critical workflow on top of a startup value that 8 billion 7 billion with a few weeks of track record, ask yourself what happens if they get equi-hired, pivot or go bust. Ask the people who bought a humane pin, right? Ask the customers of builder.ai. And risk six is the overshoot. When bubbles burst, you know, they don't just kill the bad stuff. They kill the good stuff too, for a while. This is like cancer, okay? When you treat cancer, you kill the bad cells and the good cells, right? Unfortunately. After 2001, it was hard to raise money for anything with dot-com, including companies that later turned out to be brilliant places and brilliant companies.

An AI crash would trigger an AI winter in funding as well that would hit the genuinely useful projects along with the nonsense. and good luck for all the .ai websites or companies or startups or genuinely good products. They will be affected as well because they are part of that wave. So what should we expect? Folks, I don't have a crystal ball. I mean, I'm just a podcaster. I'm kidding. Well, I am a podcaster. I'm also a podcaster. No, honestly, I don't know. the problem is that anyone who tells you the exact date a bubble pops is telling you something so watch out there too what I can tell you is look at the patterns I hope I did a decent job showing you where the patterns are in my head of course and I'm open to discuss with you and ask you if you see the same pattern as I do

but definitely I can share what I would expect in the near future well one is the efficiency narrative becomes the new hype which is exactly what JEV and CDMs now these are just the latest ones but that's exactly what they do once everyone agrees that the giant models are too expensive or too slow or whatever or both the next wave is small cheap specialized which is great that's what we should have been doing all along in fact but watch the valuations You know, the pitch changed, the behavior investors didn't. And that's scary. So rebranding the classifier doesn't make it a $7.5 billion company, right? The second thing that I expect is consolidation and acquisitions, so-called acquisition. And we have already seen the playbook here. You know, Humane to HP, Inflection Steam to Microsoft, Characters.ai founder back to Google, Windsor's leadership back to Google.

You know, expect a lot more of that. The big players buy the people and the patents. The investors get a haircut, of course, and the product dies. You know, we have seen that. It's part of the deal. It's part of the game. I've been into startups enough to see this. This is really a clear pattern to me. The third thing that can happen is the write-downs. At some point, someone will have to admit that a GPU bought in 2023 is not worth what the books say. when the depreciation schedules get shortened, like Amazon already started doing, profits will look worse overnight, you know, without anything actually changing in the real world. Neo clouds with a heavy GPU-backed debt are the most exposed, of course. Another thing that I see happening, the name will change. Name changes. blockchain became Web3, became decentralized, and then there was a period with NFTs and all that other stuff,

and then there was silence, right? So watch the vocabulary with AI as well. We started with AI, then generative AI, then it became agents, okay? And now we have decision models, and now we have constrained decoding, okay? So when a sector keeps renaming itself, it's often because the old names have stopped raising money and they need new, fresh stuff, right? So watch out there. Another thing that could happen and probably is going to be my last one, the boring useful part. The optimistic note, because I don't want to leave you depressed with all this. After the dot-com crash, the Internet didn't go away, right? It actually became more important, not less. I think that the same will happen here. The useful applications of AI, and there are many, will survive, right? Compute will get cheaper for whoever is left standing. The power infrastructure built during this boom will be useful,

definitely for decades, but just the power infrastructure. The difference is just who pays for the lesson. so I guess that's it for today and I hope to see you soon and get in touch YouTube and official website datascience.com but especially YouTube guys you would make me a big big favor if you put a like if you share you know that's my currency and that's how I can do this without getting paid thank you so much a quick one I wrote a book The Mythical Token Month Here's the idea. AI coding agents feel like magic, but they don't remove the hard parts of building software. You know, they just clear out the busy work around it, as we all know. What's left is indeed the actual thinking. Now, I have to say that no AI wrote a word of this book. I have to say that. It's a self-claim, but you have to believe me. It's out now as an e-book, NoDRM.

The website, mythicaltokenmonth.com. Links in the show notes. Thank you. You've been listening to Data Science at Home Podcast. Be sure to subscribe on iTunes, Stitcher, or Podbean to get new fresh episodes. For more, please follow us on Instagram, Twitter, and Facebook, or visit our website at datascienceathome.com.