Every bubble in history shared one feature: anyone could buy in.

Dutch merchants bought tulip bulbs in taverns. Victorian clerks subscribed to railway shares from the back pages of newspapers, and your uncle bought Pets.com from his kitchen table. That's what a mania is: an open door and a queue.

Now try to buy a GPU.

The asset at the center of the biggest capital boom ever measured is sold out and is allocated years in advance. Jensen Huang decides who gets compute the way De Beers once decided who got diamonds, except De Beers never wrote equity checks to its biggest customers. NVIDIA put $30 billion into OpenAI and $10 billion into Anthropic; the labs spend billions on chips, and the spending books as demand.

So the speculation lives one layer up, in the equity, because the thing itself cannot be bought. Nebius turns customers away. Anthropic pays its fiercest rival $1.25bn a month because it can't get enough compute of its own.

Which brings us to Michael Burry's strange new problem. He got famous betting against a bubble nobody believed in. He spent 2005 being laughed out of rooms and 2008 being proven right. Now he has over a billion dollars of notional value bet against NVIDIA and Palantir, and his situation has inverted. Everyone believes him. The BIS believes him. Half your feed believes him.

The crash is the consensus position at every dinner party in finance.

And it keeps arriving everywhere except the center. Palantir fell more than 40% from its peak, and Burry took profits on half that position in June. His NVIDIA puts sit deep underwater, with the stock near all-time highs. The most anticipated crash in the history of capital is picking off software multiples at the edges, while the company in the middle of it all refuses to fall.

It refuses because you can't overbuild something that's sold out.

Maybe a bubble with no unsold inventory is a strange kind of bubble. Maybe it's taking its time.

Either way, the useful question is what to do with your dollars and your energy while we find out.

By any historical measure, the level of investment in AI is the biggest bubble we have ever recorded. This chart blew my eyebrows off the first time I saw it.

This tweet put it well:

The Bank for International Settlements (BIS) put out its annual report, in which this chart featured. They list rising inflation and a drop in AI sentiment driven by "opaque" accounting practices and increasing competition from open source as potentially driving a "sharp reversal" if AI payoffs disappoint. This could have a contagion effect on a global economy that already has high inflation, stocks at all-time highs, and is heavily reliant on non-bank sources of lending (emphasis added):

A capex slowdown could hit suppliers first, then credit markets, then households, because US stocks make up about 64% of the MSCI Global index and household equity exposure is higher than in past cycles.

This "accounting opacity" has been chronicled by the always entertaining Ed Zitron, who says the AI labs are structurally unable to be profitable. Reporting recently leaked 2025 financials, revealing that OpenAI lost approximately $38.5 billion (with an operating loss exceeding $20 billion).

For some context, $34bn of spending is more than the US Government spent in the FBI and NASA combined. Even if you remove the massive $19.2bn in training costs, there's still $14.8bn of costs, against $13.1bn of revenue. But you can't do that; you need to constantly raise new funding for the R&D to stay at the frontier. This is before Anthropic took a lead, and OpenAI started discounting to win back market share.

There was also a worry that AI hyperscalers were wildly overestimating the useful life of their AI chips and infrastructure. Hyperscalers depreciate the cost of their AI chip investments over 6 years instead of two, which many skeptics like Michael Burry (yes, that guy who predicted the financial crisis in 2008) have claimed is unreasonable, and will be found out.

And this is before Z.ai released GLM 5.2, a model that is almost equal to Opus 4.8, and GPT 5.5 on every single benchmark, but at half the price.

It's just very unevenly distributed.

The one thing that's majorly different about AI from previous bubbles is that historically, we had a lag between Capex and Revenue. That's less true with AI. The combined revenues of Anthropic and OpenAI have reportedly surpassed $100 billion. Microsoft, Google, and even Meta are seeing major business benefits (like boosting return on ad spend) from their AI investments. Even if unprofitable (for now), the revenue allows capex spenders to plan accordingly and more accurately; they don't just rely on assumptions and hopes.

Micron, the provider of high-bandwidth memory needed for AI, said in its latest earnings that its gross margin for future business is 84.9%for AI. Nvidia's gross margins are 74.9%, the hyperscalers like Microsoft closer to 68% in its business (bearing more of a capex burden). Then at the bottom of the pile are the labs, who are paying handsomely for AI capacity they don't own, and subsidizing end users on subscriptions (who use somewhere between $8 and $13 for every $1 of subscription).

The old saying is, your margin is my opportunity.

This is leading to a world in which everyone is trying to eat adjacent parts of the supply chain. Amazon and Google have their own AI chips, with OpenAI reportedly developing their own too. SpaceX is leasing out its data center capacity to other labs and technology companies, and NVIDIA has started to release a flood of small, open-weight models.

Another thing that's different is that AI chips are lasting the 6+ years hyperscalers tend to depreciate their assets over. Gavin Baker from Atredis told the All In podcast that specialists are able to re-use A100S and H100S. He gave the example of disaggregating compute between pre-fill and decoding.

Put simply, pre-fill is remembering everything generated so far; decoding is generating new tokens. An older chip like an H100 can hold the pre-fill, while a new specialist chip from Cerebrus or NVIDIA's Groq can then decode (generate). This means data centers become wildly more efficient.

It also makes data center payback realistic for the hyperscalers (and everyone else).

The CapEx for standing up a Gigwatt data center is staggering: it costs $35 billion in silicon (Nvidia GPUs) and another $25 billion for the power and cooling infrastructure, driven by human labor costs.

Against that $60 billion upfront cost, Baker notes that a 1-gigawatt facility can generate roughly $15 billion in cloud or token revenue. Because this hardware is expected to have a usable lifecycle of 7 to 12 years (especially with the disaggregation of inference where older chips handle decoding), the long-term returns are viable, but the upfront capital requirements are a massive barrier to entry.

On paper, at these prices, a fresh datacenter is one of the best trades in tech.

Of course, all of this assumes demand and revenue for AI continues to exponentially rise, which is not guaranteed.

Plenty could still break the demand assumption the whole payback rests on.

The payback math holds only if the tokens keep flowing. Baker thinks they will, just not toward where most people are looking.

No serious enterprise runs one frontier model for everything. They run a council. A local, fine-tuned open-weight model takes the routine 85% on proprietary data, and a router kicks the hard, high-stakes calls up to a frontier model acting as conductor (emphasis added).

That split snaps the link between token volume and value capture. Open weights run the overwhelming majority of tokens globally, a gift to whoever sells the picks and shovels. The labs keep the thin slice of complex work and capture most of the economic value inside it.

Santander already lives this. When they moved a process off a major-lab model onto a small local one, the cost fell from about €5 a run to €0.03 at the same output quality (I went deep on this last week). Multiply that across an enterprise, and you see why every heavy user is learning to route.

Token volume goes vertical while value capture concentrates. The infrastructure layer wins on volume, the frontier wins on margin, and the company in the middle that owns neither pays full retail for genius it never needed.

The big problem remains that 9 out of 10 companies haven't yet seen a productivity uplift.

I pulled this apart in The Token Economy. An NBER study of 6,000 executives across the US, UK, Germany and Australia found nearly 90% reported no impact on productivity or employment over three years, on an average of 1.5 hours of AI use a week. They had access to the same models as everyone else. What they lacked was the operating model to turn tokens into outcomes. Most firms bought seats and called it a transformation.

The diffusion of ROI will take a lot longer than the access to the technology. But ironically, the biggest beneficiaries of AI today are the most technical, engineering-capable organizations and individuals.

Hard work pays off, apparently. The people who compounded technical capability for a decade are the ones AI just made richer. It handed the already-capable a bigger lever, and they pulled it harder.

The spend data says the same thing. Ramp's top quartile of AI spenders more than doubled revenue since 2023, while the average firm poking at AI 1.5 hours a week felt nothing.

The ROI is real. It's just hoarded by the capable, for now.

AI is unlike any bubble in history because demand and revenue are immediate, and the technology can diffuse to anyone with an internet connection as soon as a new model is available.

Canal mania needed canals dug and railway mania needed track laid. This needs a browser tab.

That's why the bulls are right in the end, and why the road there will be rough. Demand is real, diffusion is slow, and the repricing in between can destroy an enormous amount of capital without un-inventing a single capability.

So where do you put your dollars and your energy?

For most of business history, you had two kinds of capital to allocate. Human and financial. You hired people, and you funded them. The operators who win the next decade allocate a third: AI. Compute, tokens, and intelligence, allocated on purpose the way you already allocate money and people.

Bet it gets faster, cheaper, and more capable, because every line of this points that way. Then build the company that compounds the bet. Route your workloads, own your data, train your own models, and control your own compute if you're big enough (the four ways are here).

The labs' moat is staying at the frontier, and it's expensive.

Yours is cheaper.

ST.

If you enjoy this kind of content, I can guarantee you’ll love being in a room of 1,500 other folks who love to go deeper into where finance meets AI. That’s a huge theme for us at this year’s Nerdcon in San Diego on the 19th and 20th November. I’m bringing my audience, the operators, the people who read this newsletter. And it’s the perfect place to find your next hire, client, or just get inspired. Let’s make events awesome again.

Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)

Want more? I also run the Tokenized podcast and newsletter.

(1) All content and views expressed here are the authors' personal opinions and do not reflect the views of any of their employers or employees.

(2) All companies or assets mentioned by the author in which the author has a personal and/or financial interest are denoted with a *. None of the above constitutes investment advice, and you should seek independent advice before making any investment decisions.

(3) Any companies mentioned are top of mind and used for illustrative purposes only.

(4) A team of researchers has not rigorously fact-checked this. Please don't take it as gospel—strong opinions weakly held

(5) Citations may be missing, and I've done my best to cite, but I will always aim to update and correct the live version where possible. If I cited you and got the referencing wrong, please reach out