AI Is a Sure Thing. Its Biggest Companies Are a Gamble.

Let me be clear up front about what I'm not saying. I'm not saying AI is a bubble. AI is not going anywhere. In a few years it'll be in your phone, your car, your fridge, your work tools, your kids' homework — most of it running quietly in the background, most of it cheap or free. That part is settled.

What I'm saying is narrower, and I think most people watching from outside the industry are missing it: "AI is huge and here to stay" and "therefore OpenAI and Anthropic are great businesses" are two completely different claims. The first is obviously true. The second is a genuine gamble. And the gap between them is where I'd keep my eyes.

The wrong argument everyone is having

The public debate is stuck on a binary: pop or no pop. Is it a bubble or isn't it. That's the wrong question, because it treats "AI" and "the AI companies" as the same thing. They're not.

The technology can be a world-changing success and the companies that built it can still be terrible places to have put your money. This is not exotic. It's the default outcome. Railroads transformed the world; most of the railroad companies went bankrupt. Electricity changed everything; naming the companies that first generated it is a pub quiz question. Being the pioneer of a technology and being the one who keeps the profits from it are rarely the same story.

So the interesting question isn't "will AI win." It's "who captures the money when it does." And right now the honest answer is: probably not the people we're currently valuing at a trillion-plus.

The model is becoming the commodity

Here's the mechanism. The frontier model — the thing OpenAI and Anthropic sell — is losing its status as the moat.

Open-weight models are roughly twelve to eighteen months behind the closed frontier, and the gap narrows with each release. By mid-2026, on OpenRouter — the largest neutral router of model traffic — Chinese open-weight models were the majority of tokens consumed, and most of the top models by usage were Chinese. You can download several of them and run them yourself. That's not a footnote. That's the moat leaking.

Add the "good enough" threshold. For a large share of real work — document analysis, triage, extraction, code review — a cheaper model that's a year behind the bleeding edge is simply good enough, and it costs a tenth to a thirtieth per token. Once a task crosses that line, nobody pays the premium for it again.

I want to be fair here, because this is where lazy versions of my argument fall apart: the frontier still wins the top end. For genuinely hard, high-value work — the most demanding agentic coding, serious R&D — being eighteen months behind is a real handicap, and people will pay for the best. "Good enough" is not universal; it's segment-by-segment. But the premium segment is a shrinking island, and it's surrounded by a rising sea of good-enough.

The part almost nobody talks about: the premium is now a compute bill, not a brains bill

This is the piece I think is genuinely underappreciated, and it's the one I'd actually build a thesis around.

The most impressive current models aren't winning mainly because they're categorically smarter architectures. A large part of the frontier edge now comes from letting the model think longer — spending far more compute per answer at the moment you ask the question. Being able to turn extra compute into a better answer is a real, trained capability; that's not nothing. But look at what it does to the economics.

A traditional software company has near-zero marginal cost. Once the code is written, the millionth user is almost free. That's why software prints money. But if your product's advantage is "we burn more compute on every single query," then your advantage is a variable cost. It scales with every answer you give. That's an operating expense, not defensible intellectual property.

Say that out loud and the picture changes. A premium tier whose edge is "more compute per question" structurally cannot hold software-like margins, because the cost never goes away — it recurs on every request, forever, and it competes directly with an open model that a customer can run for a fraction of the price. You're not selling a magic ingredient. You're reselling electricity with a markup, and the markup is under attack from below.

"Buy the infrastructure instead" — but be precise about which infrastructure

The obvious rebuttal is: fine, don't bet on the model, bet on the picks and shovels. Somebody has to sell the compute.

True — but "infrastructure" is not one thing, and the distinction matters enormously.

The chip vendor (Nvidia and its emerging competitors) actually benefits from hardware ageing fast. Short useful life means faster replacement cycles means more sales. The company holding the chips — the hyperscaler or the neocloud that spent the billions — is the one carrying the depreciation risk. Those two are on opposite sides of the same trade. Lumping them together as "infrastructure" hides the actual bet.

And the depreciation question is live and ugly. Michael Burry's late-2025 argument was blunt: the big buyers are depreciating AI GPUs over five to six years while the real economic life is closer to two or three, which by his math understates depreciation by something like 176 billion dollars across 2026–2028 and flatters profits accordingly. The buyers push back — older chips hold resale value, they say, and stay useful for inference under multi-year contracts. That dispute is unresolved, and it sits underneath a large chunk of the reported earnings in this whole complex.

A quick word on a bad intuition here, because I nearly fell for it myself. Gamers will tell you a great GPU lasts forever — the GTX 1080 Ti launched in 2017 and still runs modern games fine. That's true, and it's exactly the wrong lesson. A chip that's perfectly good in your living room can be completely uneconomic in a datacenter, where power draw and efficiency-per-watt decide whether it's worth switching on at all. "It still works" and "it's still worth running at scale" are different sentences. Don't let consumer nostalgia talk you into thinking datacenter hardware ages gently.

The one bull case I take seriously

If I only gave you the bear side, I'd be doing the same lazy thing I'm criticising. So here's the argument that could make the gamble pay off, and it's the real one.

As raw intelligence commoditises, a different moat gets dug: access and distribution. Who can afford the premium compute. Who is legally allowed to call which model. Who owns the surface where users actually arrive — the operating system, the browser, the default assistant, the enterprise contract. Intelligence leads decay in months, because competitors copy them. But distribution moats compound — through habit, defaults, contracts, and switching costs. That asymmetry is the strongest reason the current leaders might still win: they can spend a temporary intelligence lead buying permanent distribution.

That's a serious argument. It's also, notice, an argument that the frontier labs stop being model companies and become distribution companies — which is a very different, much more competitive business than the one their valuations are implicitly pricing. They'd be knife-fighting Google, Microsoft, Apple and Amazon on those companies' home turf. Winnable. Not a sure thing.

Weak moat is not the same as overpriced — but it's not nothing either

I'll hold one honest line. A thin moat doesn't automatically mean a stock is a bubble; the market may already have priced the commoditisation in. I'm not going to pretend I can value these companies, and none of this is investment advice.

What I am saying is a claim about uncertainty, not price. Three of the most talked-about companies of our era are reportedly heading toward public markets at combined valuations in the multiple trillions, and the case for those numbers rests on assumptions — a durable moat, premium pricing that holds, a defensible margin — that are each, individually, contested right now by serious people. Stack contested assumptions on top of each other and "obvious winner" quietly becomes "high-variance bet."

That's the whole point. From the outside it looks like a foregone conclusion: these are the AI companies, AI is the future, therefore. But the "therefore" is doing enormous work, and it's exactly the part that isn't obvious.

What I'd actually watch

Not a buy or a sell. A lens for the next eighteen months:

• Gross margins on the premium tiers. If the frontier edge is a compute bill, the cost does not disappear after the product is built. It comes back on every query. Watch whether margins behave like software — or like electricity with a markup.
• Where the token volume goes. The open, cheap, good-enough tier eating the volume while the labs keep only the shrinking premium is the commoditisation story playing out in real time.
• Whether the labs win distribution or just rent it. Owning the default surface is the difference between a durable business and a very expensive supplier.
• When the depreciation argument gets settled — by an accounting change, an auditor, or a wave of write-downs. That's the moment a lot of flattering earnings could turn honest.

AI wins. That was never the question. Whether these particular brands are the ones who get rich from it — that's the coin-toss dressed up as a certainty. Most people watching don't see the coin. I'd keep my eye on it.