AI Bubble Theory or AI Future Theory?

AI has been running hot for years now, lifting practically every stock it touches — and when things sit at such a height, it's natural to worry about the fall. That's why roughly every six months, a fresh round of "AI bubble" debate breaks out.

With Michael Burry — the real-life inspiration behind a character in The Big Short and a well-known fund manager — once again warning of an AI stock bubble, the question of whether the AI market is heading for a crash has become the newest hot topic. Burry has spent the past few years pushing the AI-bubble thesis, going so far as to put his money where his mouth is by shorting Nvidia and Palantir stock (which earned him a public rebuke from Palantir's CEO). Given that these two companies are the seemingly unstoppable leaders in AI hardware and software respectively, Burry evidently really does believe AI is a bubble.

Accounting Optics Under Fire — Doubts Grow Over an "AI Perpetual Motion Machine"

Why the skepticism? For one thing, there's the now-familiar setup where Nvidia invests in foundation-model companies, those companies buy cloud services, and cloud providers turn around and buy Nvidia GPUs — a circular arrangement that props up everyone's stock price and revenue simultaneously. Critics have taken to calling this the "AI stock perpetual motion machine," and plenty of stock analysts have slammed the whole arrangement as a suspiciously convenient three-way trade, one where a lot of the "revenue" hasn't actually happened yet — it's future revenue, still on paper. There's a more fundamental complaint too: Burry, for instance, argues that depreciation is being understated, meaning the financial statements are effectively flattered.

Why would depreciation be understated? Because after a GPU has been in service for two or three years, Jensen Huang rolls out a new one with several times the compute power, and those older GPUs essentially turn into e-waste — worthless scraps, like used betel nut husks. So the depreciation, critics argue, should be much steeper. But the Neoclouds — the new breed of cloud providers at the center of this "perpetual motion machine" — push back, saying customers aren't actually rejecting the older GPUs the way people assume; given how compute-starved everyone is, customers keep signing on regardless, so there's nothing wrong with the current depreciation math. So it's very much a case of each side telling its own story.

The engine driving this AI stock perpetual-motion machine is mainly the foundation-model labs and Nvidia. Conceptually: OpenAI signs a mountain of long-term contracts with cloud providers, those cloud providers turn around and buy Nvidia chips, and Nvidia turns around and invests in OpenAI and in data centers — and the cycle repeats. Throughout this loop, plenty of investors remain eager to fund the cash-burning engine at the center of it all, OpenAI, and every stock caught up in the cycle rises together. It really does look like a perpetual motion machine.

As things stand now, the Neoclouds — the new cloud providers most entangled with Nvidia in this perpetual-motion setup — are seen as the riskiest link in the chain. Nvidia itself doesn't show much sign of a bubble yet; whether its stock price is too high is a separate question, but GPU demand is real and tangible, and the pattern of investors buying their own customers' products has actually happened plenty of times before across other industries. Then there's Oracle — even though its stock, as a cloud provider, has been whipsawing up and down, the company itself continues, quite steadily, to collect fees from enterprise customers. Foundation-model companies like OpenAI, while still deeply in the red, genuinely do have huge user bases — it's just that what those users pay is far less than what's being invested. But isn't investing ahead of the curve, by definition, getting there before your customers do? Isn't that exactly what "ahead of the curve" means?

Similarly, Meta has been going all-in as well. Its stock recently plunged, partly due to a one-time deferred tax expense, but also because AI is driving up capital expenditures. Typically, when a company faces a one-time expense hit like that, it wouldn't want capital investment rising in the same period too — you'd expect them to smooth things out, maybe push some of that spending into the next quarter to keep the current-period numbers cleaner. But Meta showed zero interest in optimizing this quarter's financials — it simply doesn't care what the market thinks, treating AI investment as too urgent to delay, with no intention whatsoever of trimming costs to make the books look better.

Big Tech's Bet on Getting There First — Winner-Take-All Logic, Once Again

So here's where you can see the fundamental disagreement: the core difference between the skeptics and the AI-future believers is that the AI-future camp sees AI as the foundational infrastructure of tomorrow, meaning investing ahead of the curve and accepting a slow payback period is simply table stakes — do you really worry about whether a highway or a railroad pays for itself within three years? The skeptics, by contrast, believe this is money that simply won't come back. The network effects and payback speed of, say, mobile internet adoption or smartphones were far faster than what AI is showing now — and it's still an open question whether AI can actually generate productivity gains across as many domains as claimed.

There are also lingering doubts about whether LLMs represent AI's ultimate form at all — and these skeptical arguments aren't baseless, especially given that the big labs are running this stock perpetual-motion machine quite openly, in broad daylight.

But at minimum, America's tech giants are, as of right now, all committed to going all-in — whether you believe in it is your own business. The core logic behind their investment decisions is that today's enormous capital outlays are analogous to laying internet infrastructure or undersea cables in the 1990s — extremely large, forward-looking investment was simply necessary back then too. And in fact, the world is full of precedents for massive forward investment of this kind — China's EV and solar industries follow a similar logic, to name two concrete examples. None of it seems to have stopped those companies from surviving; meanwhile, competitors who didn't invest ahead of the curve often got pushed clean out of the market in this kind of winner-take-all game — losing the whole business, let alone worrying about a bubble.

Coming back, finally, to the reality on the ground: is AI really as useless as the skeptics claim? It's true that the number of companies that have actually turned AI into real, deployed, revenue-generating business models still isn't that large. But from my own personal experience, AI is genuinely useful and genuinely can be deployed in practice — what's slow is that the existing shape of the world simply wasn't built around AI to begin with, so deployment takes time.

AI, in other words, is a transformation on the same scale — and the same kind of long timeline — as the internet itself, and given the even larger scope of its impact, the transition will naturally take even longer. I lean toward believing AI is tomorrow's infrastructure, not a passing bubble. Long term, given how many domains humanity intends to apply AI to, compute is probably still nowhere near sufficient — we're still at the starting point of exploring the full range of possibilities. But whether the AI stock perpetual-motion machine actually makes sense, or whether various AI tech companies are overvalued, is a separate question — the two shouldn't be lumped together and judged as one.