· Solveion · Perspectives · 7 min read
A trillion dollars needs a reason
There are only three things an AI lab can be worth a trillion dollars for: the models, the products built on them, or the thing that does not exist yet. Two of those legs hold more weight than people assume, and neither holds enough.
The frontier labs now carry valuations in the same range as the largest listed companies on earth. Reported figures put OpenAI around $852 billion and Anthropic close to a trillion after its most recent round. These are private marks and should be read with the usual caution, but the order of magnitude is not in dispute.
So it is worth asking the plain question. What exactly is being bought? There are only three candidate answers, and it is a useful discipline to force any argument about AI valuations into one of them.
One: the models. The labs are worth this because they build the best models and will sell access to them.
Two: the products. The labs are worth this because of what sits on top — the coding agents, the desktop agents, the enterprise deployments.
Three: the thing that does not exist yet. The labs are worth this because one of them arrives at something categorically beyond current systems, and that is worth more than any of the rest.
Each answer implies a completely different company, and a completely different procurement posture for anyone buying from them.
The models are commoditising, and the labs know it
The first answer is the weakest, and the evidence is not close.
Open-weight models now trail the closed frontier by roughly four months. Stanford’s 2026 index put the leading closed model just 3.3 per cent ahead of the leading open one. On the workloads most organizations actually run — coding, agentic tasks, extraction, summarisation — the open models are within a few points, and some of them cost around fifty times less per million output tokens. Inference cost per query has fallen roughly 95 per cent since GPT-4 launched.
If your thesis is that a lab is worth a trillion dollars because its model is the best available, you are betting on a four-month lead over a free alternative in a market where the price of the underlying commodity has fallen by 95 per cent in three years. That is not a moat. That is a head start.
The most persuasive evidence for this view is what the labs themselves are doing. A company confident that raw model access would hold its value would sell raw model access. Instead, every serious lab has spent the past eighteen months building applications, agents, desktop products and enterprise deployment machinery. That is precisely the behaviour of a firm that expects the layer beneath it to commoditise. Their product strategy is a revealed belief, and it is more informative than anything in a keynote.
The second answer is far stronger than “just a SaaS company”
Here is where we would push back on the common dismissal, which holds that if the value is in the peripheral products then a lab is merely a software business wearing a laboratory coat.
Consider what has actually happened. Anthropic’s coding product went from nothing to a reported $8 billion annual run rate in under a year, and by early this year accounted for more than half of all enterprise spending on the company’s products. That is not a side business. On the numbers available, it is one of the fastest-scaling software products ever built.
More importantly, it is the only part of the stack with switching costs. API access is a commodity by construction: swapping one endpoint for another is an afternoon’s work, which is exactly why the price collapses. A coding agent embedded in your repositories, your review process and your engineers’ daily habits is not an afternoon’s work to replace. If there is a durable business anywhere in this industry, it is here — in the part often dismissed as the least interesting.
So we would invert the usual ranking. The model layer is the impressive part and the weak business. The product layer is the unglamorous part and the strong one.
But — and this is the pivot — strong is not the same as sufficient. Anthropic’s reported $47 billion annualised revenue against a valuation near a trillion is roughly twenty times revenue. OpenAI’s $25 billion run rate against $852 billion is closer to thirty-four times. For a company growing five-fold in a year, twenty times revenue is not absurd; it is roughly what excellent hypergrowth software has always commanded. Which means the product answer, taken seriously, gets you to a very large number. It does not obviously get you to a trillion, and it certainly does not get you there twice over.
The question that decides it is margin, not growth
If you want one number to watch, it is not revenue. It is gross margin, and the picture there is genuinely improving: Anthropic’s has gone from around minus 94 per cent in 2024 to roughly 40 per cent in 2025 to about 60 per cent now, with a stated ambition of 77 per cent by 2028. That is the shape of a research programme turning into a business.
The complication is that falling unit costs are not translating into banked margin. OpenAI’s inference bill is projected to rise from $8.4 billion to $14.1 billion this year, while cost per query falls. Cheaper tokens are being consumed faster than they get cheaper. Efficiency is being spent on capability rather than saved as profit, which is a rational choice while the race is on and a genuine problem if the race does not end. OpenAI is reported to expect a $14 billion loss this year and no profit until the end of the decade.
So the trillion is the third answer
Work backwards. Value these as outstanding software businesses — generous multiples, durable growth, margins landing where the labs say they will — and you arrive somewhere in the high hundreds of billions. The remainder, and it is a large remainder, is not paying for the coding agent or the API. It is paying for the possibility that one of these organizations builds something categorically different.
That is a coherent thing to buy. It is simply important to be honest that it is what you are buying: an option on an event with no agreed definition, no timetable and no way to verify progress from the outside. Priced as an option it may well be cheap, because the payoff if it lands is not a large software company but something closer to a claim on economic output itself. Priced as a multiple of a coding-tools business it makes no sense at all. Most commentary confuses the two, which is why the arguments never resolve.
We should be clear that we hold this loosely. Our own record on predicting this industry is no better than anyone’s, and any of these legs could be reweighted by a single result.
What this means if you are buying, not investing
Most readers are not allocating capital to these companies. They are buying from them, and each answer implies a different posture.
If you believe the model layer is the value, act as though it will be commoditised anyway: keep an abstraction between your systems and any one provider, resist architectures that assume a specific model, and negotiate short. Prices in a commoditising market move in your favour if you have not pre-committed.
If you believe the product layer is the value — and on the evidence, we think it is the strongest of the three — then take seriously that its worth to the vendor is exactly the switching cost it accumulates in your organization. That cost is real and it accrues quietly, in habits and embedded workflows rather than on any invoice. Being deliberate about which tools you allow to become load-bearing is not paranoia; it is the same discipline as knowing which of your systems you are unwilling to stop understanding.
And if you believe the third answer, then very little in your procurement process is the important variable, and you should be spending your attention on capability rather than contracts.
The useful exercise is to notice which of the three you have implicitly assumed, because most organizations have assumed one without ever saying so, and it is quietly determining decisions they think they are making on the merits.