What an AI company is worth

Our principles say we buy businesses with honest accounts and cash flows we can inspect, and that we stay away from what we cannot value. Artificial intelligence is where those two rules strain hardest against each other. The companies are real, the revenue is real, and the spending on them will pass $700 billion this year. What is missing is the thing a valuation rests on: a view of what the business will earn once the race is over, and of who will still be standing.

This essay sets out how we think about that. It separates the sector into three layers that value differently, applies the reverse test we use elsewhere, asking what the price assumes rather than what the asset is worth, and says where a margin of safety can and cannot be found. Figures are as of October 8, 2026, unless stated, and we date the essay because in this sector the facts age in months.

Three layers, three kinds of asset

"AI company" describes three businesses that share a technology and little else. They sit in a chain: one sells the picks, one digs, one sells what is dug up. Each is valued by a different rule, and most errors in this sector come from applying the rule for one layer to another.

Infrastructure is the chips, the memory, the networking, the power and the buildings. It sells to a handful of customers who pay in cash today. Its revenue, margins and capital employed can be read from audited accounts. It can be valued. The question is not what it earns but how long it earns it, because its demand is another layer's capital spending, and capital spending is cyclical.

Model labs train the frontier models and sell access to them. They have revenue that has grown several times over in a year, and losses to match. Their product is matched by a competitor within months and by open models within a year or two. What a lab owns is a position in a race, and the value of that position is the chance of winning times the prize, which is the same shape as the problem in our essay on Bitcoin, with one difference: a lab can lose by being made ordinary, not only by failing.

Applications put a model inside a product that a customer already pays for: a coding tool, a legal assistant, a customer service desk. They are ordinary software businesses with an unusual input cost, and ordinary valuation applies. The question is whether the margin survives when the input supplier raises its price or sells the same thing directly.

The layers are not independent. The labs' revenue is a small fraction of the infrastructure spending made on their behalf; the applications' margins are the labs' pricing power in reverse. A valuation of any one layer that does not ask where the others' money comes from is a valuation of a part as if it were the whole.

Infrastructure: the money that is actually being spent

The four largest American cloud and platform companies, Amazon, Microsoft, Alphabet and Meta, have guided to between $720 and $745 billion of capital spending in 2026 [1][2][3][4], up from about $410 billion in 2025 and $250 billion in 2024, counting finance leases [5]. Oracle plans about $70 billion more in its current fiscal year [6]. UBS estimates that Amazon, Alphabet and Microsoft will spend about 102% of their cloud revenue on capital expenditure this year, and projects $4.1 trillion of spending across a wider group of cloud providers from 2026 to 2028 [7]. Oracle spent 83% of its revenue on capital expenditure in its last fiscal year, and 147% in its latest quarter [8][9].

These are the customers. The supplier that captures most of the margin is Nvidia, which reported revenue of $96.2 billion in the quarter to July 26, 2026, up 106% on the year, of which $89.0 billion was data center; gross margin was 75% and operating income $63.7 billion [10]. It guided to $108 billion for the following quarter [11], and told analysts to expect revenue growth of about 70% in its next fiscal year [12]. Its market value was about $5.7 trillion in early October [12]. The company added some $160 billion of supply commitments in a single quarter [13], and its senior officers' bonus plan for the year is tied to one metric, revenue [14].

This layer can be valued, and we have done the sum. Nvidia at $5.7 trillion, on a run rate near $400 billion of revenue and perhaps $250 billion of operating profit, trades at about 22 times operating profit. That is not an absurd multiple for a business growing 70% a year with a 75% gross margin. It is an absurd multiple for a business whose revenue is a handful of customers' capital budgets, if those budgets halve. Between 2000 and 2002, investment by American telecommunications carriers fell from $121 billion to $49 billion [15]. Cisco, which remained the leader in its field, lost 89% of its share price from the 2000 peak [16]; Lucent, Nortel and JDS Uniphase, each once worth more than $200 billion, lost nearly all of their value [17].

So the valuation of infrastructure reduces to one question: is the $730 billion a new level or a peak? Three facts argue for a peak. The spending exceeds the cloud revenue it is meant to serve. The labs that consume most of the compute are, by their own accounts, loss-making, so the end demand is funded by their shareholders rather than their customers. And the revenue each unit of compute earns is falling by design: each generation of chips does the same work for less, so sustaining revenue needs volume to grow faster than price falls, forever.

Three facts argue for a level. Revenue at the labs is growing several times faster than the spending, so the gap between spending and end demand is closing. The largest customers are among the most profitable companies in the world and can carry the spending for years, although they no longer pay for it from cash flow alone: Oracle borrowed $43 billion and raised $5 billion of equity in its last fiscal year [8], and Meta has sold $55 billion of bonds since October 2025 and financed its largest data center through a joint venture that keeps the debt off its balance sheet [18][19]. And the spending is a race between them, in which none can afford to stop first, so it persists for longer than any one company's economics would justify.

We do not know which is right. What we know is that a supplier is worth its earnings through a cycle, not at the top of one. The infrastructure layer is the only part of the sector where we can estimate value, and the estimate says the price assumes the peak is the floor.

Model labs: a position in a race

Two private companies now carry valuations that fewer than twenty listed companies exceed [20]. OpenAI closed a $122 billion round in March 2026 at a post-money valuation of $852 billion [21], and by the end of September its revenue run rate was about $50 billion [22]. Anthropic raised $65 billion in May 2026 at $965 billion, up from $380 billion in February [23][24]; its run rate passed $47 billion in May and reached $65 billion by the end of July, against a little over $5 billion a year earlier [23][25][26]. Both filed confidentially for public listings in June [27][28]. OpenAI is in early talks to raise about $30 billion more [22], and Anthropic is reported to be seeking a valuation of more than $2 trillion at its listing [29].

On run-rate revenue, OpenAI trades at about 17 times sales and Anthropic at about 15. For comparison, the best software businesses of the last decade listed at 15 to 25 times sales with gross margins near 80% and growth of 40 to 60%. The labs grow faster. They also spend more: neither has reported a profit, Anthropic's prospectus is reported to show an operating loss of more than $8 billion on revenue of $4.6 billion in 2025 [29], and both rely on the infrastructure layer's capital spending, part of which their own shareholders pay for, to deliver the compute they sell.

We are not able to value a lab, and we think it is worth being precise about why.

The product does not stay scarce. A frontier model leads for months, not years. Each lab has at times held the best model, and each has lost the lead to the other, to Google, and to open models whose weights anyone may run. A business whose product is matched within a year has no pricing power beyond that year, and a valuation at 15 to 17 times sales assumes pricing power for a decade.

The cost of staying in the race rises with the prize. Training runs that cost hundreds of millions of dollars now cost billions, and the next generation will cost more. A lab that stops spending falls behind; one that keeps spending is funding the race with its shareholders' money until, at some point not yet reached, revenue covers it. This is the capital-intensive, fast-depreciating, winner-takes-most structure that destroyed value in airlines and in telecommunications, with better gross margins and worse product durability.

The customer and the supplier are the same companies. Microsoft, Amazon and Google are shareholders in the labs, suppliers of their compute, resellers of their models and competitors with models of their own. Revenue that flows from a lab to a cloud and back as investment is counted twice and earned once. We cannot see through those arrangements from outside, and what we cannot verify we do not count.

The terminal state is unknown. The price of either lab assumes an end state in which it is one of two or three providers of a product the world depends on, with margins to match. That could be true. It is also consistent with an end state in which models are a commodity sold by cloud providers at cost, the way bandwidth became, and the value sits with whoever owns the customer. We do not know which, and nobody does.

What we can do is the reverse sum. A lab at $950 billion needs, at a required rate of 10% a year over ten years, an exit value near $2.5 trillion. At a mature software multiple of 25 times earnings that is $100 billion of annual profit; at a 30% net margin, $330 billion of revenue, five times Anthropic's current run rate. That is more than a fifth of what the world will spend on all software this year [30]. It is possible. The price does not pay us to find out whether it is likely.

Applications: where ordinary valuation applies

An application company sells a product a customer already understands, with a model inside it. A coding assistant, a contract reviewer, a support desk, a search tool for a company's own documents. These are software businesses, and software businesses can be valued: recurring revenue, gross margin after the cost of serving it, the cost of winning a customer and how long the customer stays. The accounts can be read, and the sum is one we have done many times.

Two things are new. The first is the input cost. A conventional software business has a gross margin near 80% because serving one more customer costs almost nothing. An application that calls a frontier model pays for every token, and the bill scales with use. Gross margins in this layer run from about 25% for the fastest-growing companies to about 60% for the next tier, by one venture firm's count [31], and the supplier sets the price. A business whose largest cost is bought from three vendors who also compete with it has a margin that is lent, not owned.

The second is the lab's reach. Every lab now sells coding tools, document tools, search and agents directly, and each release absorbs features that application companies built a year earlier. An application survives this when it owns something the lab does not: the customer relationship, proprietary data, a workflow embedded in how an industry operates, or a niche the law protects and the lab cannot easily enter. It does not survive by being a better interface to the same model.

The reverse test here is simple. Take the revenue multiple, which for the better-known private application companies has run from about 30 to almost 60 times recurring revenue [32][33], and ask what gross margin the price assumes once model prices and competition settle. If the answer requires 80%, the price assumes the input cost falls to nothing while the lab does not compete, and both are the lab's decision, not the application's.

This is also the only layer where we expect to find businesses we might own. A company with a durable customer base, a product that would be worth paying for with no model inside it, and a margin that survives a doubling of model prices is an ordinary good business that happens to use AI. We would value it as one, and pay for it as one, which is a good deal less than the sector currently asks.

Methods that do not hold up

Revenue multiples. The sector's common currency is a multiple of annualized run-rate revenue: 15 to 17 times for the labs, 30 to 60 for the applications. A run rate is one month's revenue multiplied by twelve, and in a business growing 200% a year it is already stale when published. The multiple itself carries no information about margin, which is the whole question. Two businesses at 20 times sales, one with an 80% gross margin and one with 40%, are not comparable, and the sector treats them as if they were.

Total addressable market. The argument runs: AI will displace a share of a labor market worth $30 trillion a year, a company with a tenth of that is worth so much. We have seen this method before. It valued the telecommunications carriers of 1999 on the whole future of the internet, which did arrive, and whose value went mostly to companies that were small or did not yet exist. The size of a market says nothing about who captures it or at what margin.

Compute as value. Some value a lab by the compute it controls, as a mine is valued by its reserves. But compute is rented, not owned; it depreciates in three to five years; and the next generation makes the last one worth less, not more. A lab's compute is a cost of staying in business, not an asset that holds value when the business falters.

Discounted cash flows with a terminal value. The method is right and the inputs are not available. For the labs, more than 90% of any DCF value sits in the terminal value, which rests on a margin and growth rate ten years out that nobody can estimate within a factor of three. A model that is 90% assumption does not reduce uncertainty. It hides it in a cell.

Strategic value to the acquirer. A company is worth what Microsoft or Google would pay for it. That is a price, set by a buyer whose own reasons may be defensive, and it depends on the buyer's share price holding. It is also the argument that was made about many companies bought in 2000 and written off in 2002.

A frame we can use: what the price assumes

As with Bitcoin, a value that cannot be estimated can still be tested. We start from the price and recover three assumptions from it: the revenue the business must reach, the margin it must earn on that revenue, and how long it must keep both. Then we ask whether each is reasonable, and whether the three are consistent with each other.

The table shows the sum for a lab priced at $950 billion, at a required rate of 10% a year over ten years and exiting at 25 times earnings. Each row is a net margin; each cell is the revenue the lab must earn in year ten.

The table turns a price into a pair of assumptions, and the pair is what we judge. A 40% net margin is what the best software monopolies have earned with no serious competitor. A 10% margin is what a commodity provider earns. Five times revenue in ten years is growth of about 17.5% a year, which is fast for a business already at $65 billion, and slow for one whose run rate grew more than tenfold in the past year. The price is consistent with a lab that becomes one of the most profitable companies in history. It is not consistent with one that becomes a utility. At the $2 trillion Anthropic is reported to be seeking, every revenue figure in the table roughly doubles.

The same frame runs for infrastructure, with time replacing margin as the uncertain input. Nvidia at $5.7 trillion needs about $250 billion of operating profit a year, sustained; the question is for how many years the four customers keep spending at a level that produces it. If the answer is three, the price is wrong. If it is fifteen, it is cheap. That answer depends on whether the labs' revenue catches up with the spending made for them, which closes the loop: the infrastructure price is a bet on the labs, and the labs' price is a bet on the applications' customers paying enough to make it all cash-flow positive.

The frame has limits. It assumes the business survives to year ten. It takes no account of dilution, and OpenAI has raised more than $180 billion and Anthropic more than $100 billion, and both will raise more [27][26][24][23]. And it treats the exit multiple as given, when a sector that has disappointed trades at 12 times earnings, not 25. Each of these makes the implied assumptions harder, not easier.

Where the margin of safety comes from

Graham's margin is the gap between a conservative estimate of value and the price. In the labs there is no conservative estimate, so there is no gap to measure. In infrastructure the estimate exists and the price sits above it. In applications the gap can be found, business by business. Where a margin exists in this sector, it comes from five places.

From position, not prediction. The businesses that earned through every previous technology cycle were the ones that owned something scarce whatever the technology did: the customer, the distribution, the data, the land under the data center, the power contract. A company that owns the grid connection earns whether the model inside the building is this year's or last year's. We prefer to own the toll road rather than guess which car wins the race.

From not paying for the race. The labs' value is a call option on winning; an option pays best when it is bought cheaply, and worst when it is bought after everyone agrees it will pay. At 15 to 17 times sales the option is priced for a win. A margin would exist at a price that assumed a draw, and no lab has offered that price since 2023.

From earnings through a cycle. An infrastructure supplier valued on peak revenue has no margin; one valued on what it would earn if its customers' spending halved and then recovered has one. For the leading suppliers we can make that estimate, and it sits well below today's prices.

From size. Whatever is held in this sector should be held in an amount whose loss would be survivable, because for most of these companies zero is a possible outcome and a 70% drawdown is a likely one in any disappointment. That is a rule about the holder, and it is the same rule we apply to any asset with a failure state.

From the balance sheet. The hyperscalers fund most of their spending from operating cash, and a growing share with bonds; the labs fund theirs from equity; some of the newer data-center operators fund theirs with debt against the chips. The last is the structure we would not hold at any price, because the collateral depreciates faster than the loan amortises.

The risks we watch

The capex-to-revenue gap. Spending of $730 billion a year, much of it built for AI workloads, compares with combined run-rate revenue at the two largest labs of about $115 billion. If the gap closes because revenue rises, the whole chain is justified. If it closes because spending falls, the suppliers reprice first and furthest. We watch the ratio of hyperscaler capex to cloud revenue; above 100% it is not sustainable, and it is above 100% now [7].

Circular funding. Investment flows from clouds to labs, comes back as compute purchases, and is booked as revenue at the cloud and as cost at the lab. Vendor financing inflated the telecommunications bubble in the same way. The larger the share of a lab's revenue that comes from, or its spending that goes to, its own shareholders, the less the accounts tell us.

Commoditization. Open models that match the frontier within a year, and cloud providers that sell inference at cost to win the rest of the workload, push model prices toward the cost of compute. Epoch AI estimates that the price of a given level of model capability has fallen about thirteenfold a year since 2023 [34]. Revenue grew anyway because volume grew faster, and the price assumes that continues.

Depreciation. A chip bought today earns at full rate for perhaps two years and is obsolete in five. Microsoft and Alphabet have lengthened the accounting lives of their servers to six years and Meta to five and a half, while Amazon shortened some back to five in 2025 [35][36][37][38]. Longer lives raise reported profit and do not change the physics. If economic life is shorter than book life, the sector's earnings are overstated now and will be understated later.

Power and permission. The constraint on the buildout is moving from chips to electricity, land and grid connection. Those are owned by utilities, governments and landowners, and the margin can migrate to them. It is also the part of the chain where a government can say no.

Governance. Nvidia pays its senior officers on revenue alone [14]. The labs are run by founders through structures that limit shareholder control. Neither is a reason not to own a company, and both are reasons to want a wider margin.

The listings. Both large labs have filed to go public. A listing turns a private mark into a public price, with daily liquidity and the obligation to publish accounts. We regard that as the most useful event in the sector's history, because for the first time the numbers will be verifiable, and we expect them to settle at least some of the questions in this essay.

What we conclude

• The technology is real and the revenue is real. Neither tells us what the businesses are worth, because worth depends on margins and durability that are not yet observable.
• Infrastructure can be valued, and the valuation says the price assumes the peak of a capital cycle is its floor. We would own the best of these companies at a price set by earnings through a cycle, which is far below today's.
• Model labs are options on winning a race, priced as if the race were won. We cannot value them, and we do not pretend to. Every price implies a revenue and a margin; at $950 billion the pair implied is one of the most profitable companies in history.
• Applications are ordinary businesses with a borrowed margin. The ones worth owning would be worth owning with no model inside them. We value those as software, and pay for them as software.
• The margin of safety comes from position, from not paying for the race, from earnings through a cycle, from size and from the balance sheet. It does not come from the size of the market.
• What we are most confident about is that the spending is unprecedented. What we are least confident about is who ends up with the money.

We would change our view on the labs if a lab published audited accounts showing operating profit at scale, with gross margins that survive falling token prices, or if a lab's price fell to a level that assumed a draw rather than a win. We would change our view on infrastructure if lab revenue caught up with capital spending, so that the buildout was funded by customers rather than shareholders. We would change our view on the sector as a whole if open models stopped matching the frontier, since that would make the labs' product scarce in a way it is not today. If any of these happens, we will revise this essay and date the change.

Sources

1. CNBC, Amazon second-quarter 2026 earnings, July 30, 2026.
2. Bloomberg, Google boosts 2026 spending estimate to as much as $205 billion, July 22, 2026.
3. Meta Platforms, second-quarter 2026 results, Form 8-K, July 29, 2026.
4. Microsoft, fiscal 2026 fourth-quarter earnings call, July 29, 2026.
5. Amazon, Alphabet, Meta and Microsoft, fourth-quarter 2024 and 2025 earnings releases; our sum, including finance leases.
6. CNBC, Oracle fourth-quarter fiscal 2026 earnings, June 10, 2026.
7. UBS estimates, as reported by 24/7 Wall St. on Yahoo Finance, August 22, 2026.
8. Oracle, fourth-quarter fiscal 2026 results, Form 8-K, June 10, 2026.
9. Oracle, first-quarter fiscal 2027 results, Form 8-K, September 10, 2026.
10. NVIDIA, financial results for second quarter fiscal 2027, Form 8-K, August 26, 2026.
11. NVIDIA, CFO commentary, second quarter fiscal 2027, August 26, 2026.
12. Bloomberg, Nvidia heads toward $6 trillion value, via Yahoo Finance, October 6, 2026.
13. NVIDIA, Form 10-Q for the quarter ended July 26, 2026.
14. NVIDIA, FY2027 Variable Compensation Plan, exhibit 10.1 to Form 8-K, March 6, 2026.
15. M. Doms, The Boom and Bust in Information Technology Investment, FRBSF Economic Review, 2004.
16. Yahoo Finance, Cisco daily closing prices, March 2000 to October 2002; our calculation.
17. CNN/Money, report on Lucent, Nortel, JDS Uniphase and Corning, September 24, 2002.
18. Meta, joint venture with Blue Owl Capital to develop the Hyperion data center, October 21, 2025.
19. Meta, bond pricing terms (Form FWP) of October 30, 2025 and April 30, 2026, on SEC EDGAR; our sum.
20. companiesmarketcap.com, market capitalizations on October 8, 2026.
21. OpenAI, announcement of its March 2026 funding round, March 31, 2026.
22. CNBC, report on OpenAI's revenue and funding talks, October 8, 2026.
23. Anthropic, Series H announcement, May 28, 2026.
24. Anthropic, Series G announcement, February 12, 2026.
25. CNBC, Anthropic says annualized revenue climbed to $65 billion in July, August 17, 2026.
26. Anthropic, Series F announcement, September 2, 2025.
27. CNBC, OpenAI confidentially files for IPO, June 8, 2026.
28. CNBC, report on Anthropic's IPO filing, June 1, 2026.
29. Reuters, report on Anthropic's prospectus, as published by The Korea Times, September 29, 2026.
30. Gartner, worldwide IT spending forecast, July 27, 2026.
31. Bessemer Venture Partners, The State of AI 2025, August 13, 2025.
32. CNBC, report on Harvey's funding round, March 25, 2026.
33. CNBC, Disruptor 50 list (Cursor), May 19, 2026.
34. Epoch AI, The plunging price of thought, September 22, 2026.
35. Microsoft, Form 10-K for fiscal 2022, July 28, 2022.
36. Alphabet, fourth-quarter 2022 results, February 2, 2023.
37. Meta, fourth-quarter 2024 results, January 29, 2025.
38. Amazon, Form 10-K for 2024, February 7, 2025.

Figures for private companies are as reported by the companies, their backers and the press; none has published audited accounts. Our own sums are marked as such.