Who Pays for the AI Compute: OpenAI, Anthropic, Microsoft and Alphabet, Examined

• Four companies, and a question about who is paying for whom
• What the equity actually is, set against the spending
• The money comes back the other way
• What the earnings line now contains
• The depreciation that is running behind the spending
• Two private laboratories, and where the disagreement lives
• The Chinese open-weight question, and the difference between tokens and dollars
• Who is selling compute below cost
• What it means for the four companies
• The price question, three scenarios, and the markers

Two of the four companies are public and old enough to be boring, two are private and losing money on purpose, and each private laboratory is simultaneously the largest shareholder-relationship and the largest customer of one of the public ones — which is not a scandal, only a description of how the money moves, and it makes every number in this episode harder to read than it looks.

DeepDives · 2026 · An AI–human collaboration that weighs every source, mainstream or alternative, under strict rules of truth and logic.

§Four companies, and a question about who is paying for whom

Four companies, and a question about who is paying for whom. Two are public and old enough to be boring; two are private, enormous, and losing money on purpose. Each private one is simultaneously the largest shareholder-relationship and the largest customer of one of the public ones. Microsoft holds roughly a quarter of OpenAI. Alphabet holds a stake in Anthropic and has committed up to forty billion dollars to it. Both laboratories then spend heavily on the cloud businesses of the companies that own a piece of them. That is not a scandal; it is how the money moves, and it makes every number here harder to read than it looks.

The questions are narrow, and they are about cash rather than capability. How much of their own cash flow did Microsoft and Alphabet actually put into these laboratories? What do the two private laboratories’ forward earnings look like, read at the evidence rather than the announcements? Is there real uptake of Chinese open-weight models by American companies? Is anyone selling compute below cost, and if so who? And is there a defensible way to think about where these share prices go over two years without pretending to know?

Declare the stake first, as this series does: the audience for this show owns these companies, and this examination holds no position in anything it discusses.

§What the equity actually is, set against the spending

Start with the equity, because it is the number people quote and it is by far the smallest. Microsoft has committed thirteen billion dollars to OpenAI, of which eleven and eight tenths billion had been funded by the end of March this year, for approximately twenty-seven per cent of the company on an as-converted basis, held under the equity method. It has separately put up to five billion dollars into Anthropic. Alphabet had invested a little over three billion dollars into Anthropic and owned around fourteen per cent, and in April it announced a further commitment of up to forty billion dollars, in cash and compute. Together, the announced commitments come to something in the order of fifty billion dollars, spread over several years.

Then set the capital spending beside it. Microsoft’s capital expenditure and finance leases for the calendar year are running at approximately one hundred and seventy-five billion dollars. Alphabet has raised its guidance to between one hundred and ninety-five and two hundred and five billion. The equity stakes are therefore about a tenth of one year’s spending, and the spending is the actual bet: the equity is what gets the two laboratories onto the two platforms.

Put the same equity as a share of cash generation instead. Microsoft’s thirteen billion dollars is roughly a fifth of the sixty-seven billion dollars of free cash flow it produced in the year to June. Alphabet’s commitment of up to forty billion dollars is about the whole of the thirty-eight billion it generated last year, part of it in compute rather than cash. In both cases the equity is small against the same year’s capital spending.

§The money comes back the other way

And the money comes back the other way. Anthropic has committed at least thirty billion dollars to Microsoft’s cloud and, on the reporting, around two hundred billion to Google’s over five years. OpenAI has committed two hundred and fifty billion dollars to Microsoft’s cloud, three hundred billion to Oracle over five years — sixty billion a year from next year — and thirty-eight billion to Amazon. The shareholder supplies capital and compute; the customer returns the money as cloud revenue. Both directions are real, and both are booked.

Now the free cash flow, which is where the strain shows. Microsoft generated sixty-seven billion dollars in its year to June against a consensus of thirty-two billion for the current year, roughly half. Alphabet is the sharper case: thirty-eight billion dollars in the year just past against a consensus of under nine billion for this one, with a second quarter of negative five point nine billion against fifty-three billion on a trailing-twelve-month basis.

Set the commitment against the customer’s own revenue and the size becomes plain: OpenAI’s announced compute obligations of some six hundred billion dollars are about fifteen times its current run rate, and the mechanism supposed to fund them is a listing that has not happened.

§What the earnings line now contains

Alphabet’s quarter contained a second thing anyone reading the headline should sit with. Net income came in at one hundred and twelve billion dollars, and about ninety-nine billion of that was an unrealised gain on equity securities, overwhelmingly on its stakes in Anthropic and SpaceX. Strip it out and adjusted earnings per share were two dollars eighty-five against an expectation of two dollars eighty-nine, a miss. That is not improper; the accounting is required. It is a warning about what the earnings line now contains: when private valuations rise, the public company’s profit rises, without any cash arriving.

Microsoft’s version of the same mechanism is smaller but swings both ways. Its equity-method share of OpenAI produced a net gain of about five billion dollars across the year, adding sixty-seven cents to earnings per share, and a loss of eleven and a half billion dollars was implied in a single three-month window in earlier quarters. That is how marks behave: the direction is the message only when it is sustained.

§The depreciation that is running behind the spending

Which brings the third accounting fact, and it is the one to put at the top of the list. The depreciation is behind the spending. Microsoft is extending the assumed life of its data-centre and office buildings from fifteen years to twenty-five, starting in the new fiscal year, and it extended the life of its servers and networking equipment from four years to six back in fiscal 2023, a change it said at the time would help that year’s operating income by about three point seven billion dollars.

Across the industry, the assumed life of a server has drifted from three years toward six, and the aggregate picture for this year is capital expenditure of roughly four hundred and thirty-four billion dollars across the large American hyperscalers against about one hundred and forty-nine billion dollars of reported depreciation. The two are not meant to be equal in any single year, because the assets are still being installed and the depreciation follows. But the gap is why reported profits look better than cash flow, and it rests on the single most consequential assumption in the buildout: how long a graphics processor is assumed useful, and therefore how much of its cost is charged against this year’s earnings rather than the next five.

§Two private laboratories, and where the disagreement lives

Now the forward earnings of the two private companies, which is where most of the disagreement lives. OpenAI’s revenue run rate passed forty billion dollars in August, but that is a latest-month annualisation, while recognised revenue in the second quarter was six point seven billion dollars, which annualises nearer twenty-seven. Its internal target is sixty-two billion dollars by the middle of next year, against an independent forecast median of sixty-six. Its adjusted gross margin is reported at about thirty-three per cent against an internal target of forty-six, and in the first quarter it recorded five point seven billion dollars of revenue against three point seven billion of cash burn. One forecast puts its loss on a strict accounting basis near sixty-two billion dollars this year, against the fourteen billion figure that circulates. Its announced compute commitments total some six hundred billion dollars through 2030, and it has filed confidentially to list, leaning toward next year, with a reported preference for a valuation above a trillion dollars.

Anthropic’s numbers are, on the evidence available, better. Its revenue run rate went from forty-seven billion dollars in May to sixty-five billion at the end of July, and its second quarter came in above eleven and a half billion with its first operating profit, five hundred and fifty-nine million. Compute cost per dollar of revenue fell from seventy-one cents to fifty-six, a gross margin of about forty-four per cent against a negative ninety-four per cent in 2024. The forecast median for its annualised revenue next May is one hundred and thirty-eight billion dollars, and its listing is expected around late October at a valuation in the high hundreds of billions to a low trillion.

Two caveats belong beside that. Its second-quarter costs were flattered by a discounted compute arrangement. And it books its cloud-reseller end-customer spending on a gross basis, so full customer spend counts as revenue and the partner payout counts as an expense; a forced restatement to net during listing due diligence would cut the headline annualised number by between a fifth and two fifths in a single print. That is a real risk, and not the most likely outcome, but worth knowing before the number is used as a valuation anchor.

So the comparison comes down to this. One laboratory is larger on the headline, losing more, with a thinner margin. The other started smaller, is growing faster, has just crossed into operating profit, and is improving its unit economics quickly. Both are financed in part by the balance sheets of the two public companies that are also their landlords.

§The Chinese open-weight question, and the difference between tokens and dollars

Now the Chinese open-weight question, because it is the part where the numbers are most confidently asserted and least carefully read. On the developer gateway OpenRouter, Chinese models have taken more than thirty per cent of the tokens used by American companies every week since the eighth of February, peaking at forty-six per cent. On Vercel, another developer platform, DeepSeek’s share of token usage went from under one per cent to seventeen in May, and one of the newest Chinese models recorded the fastest adoption the platform has tracked.

Then there is who is actually using them. Airbnb has used Alibaba’s Qwen for customer-service work; a San Francisco company called Lindy moved all of its traffic to DeepSeek and said the saving ran into millions of dollars within months; and Microsoft has been evaluating DeepSeek as a lower-cost option inside one of its own products. The reported price difference is sixty to ninety per cent cheaper than the leading American models, with capability judged six to nine months behind the frontier.

The limits are just as real. Two of those companies are under investigation by a congressional committee for having used the models at all, and the chief executive of one had to clarify publicly that no customer data was being sent to the developers. Large companies in regulated industries are staying away on data-residency and censorship grounds. And most American usage is served through American clouds rather than by paying the Chinese laboratories directly, so the laboratories win the adoption and lose the customer relationship.

That is why the raw adoption number overstates the commercial threat. There is a difference between tokens and dollars, and the gap is the whole point: on Vercel, DeepSeek’s share of tokens reached seventeen per cent while its share of revenue stayed near one per cent. The models are used heavily where work is cheap and interchangeable, and barely where a customer pays for a relationship, a contract, a support obligation and a compliance posture. Chinese open-weight models have become a real and growing part of the cheap end of the American market, and they are not yet a substitute for the expensive end. They are, however, setting the price of the cheap end, which is a durable kind of power.

The asymmetry behind the strategy explains why open weights are the Chinese approach rather than a concession. The four largest American cloud companies are spending more than four hundred billion dollars this year on artificial-intelligence infrastructure; China’s major cloud providers together are spending under forty. Open weights are what a challenger does when it cannot win on compute, and it works: it commoditises the layer above the model, which is where the expensive companies earn their margin.

§Who is selling compute below cost

Which brings the question of whether anyone is selling compute below cost. OpenAI’s adjusted gross margin is reported at about thirty-three per cent against a target of forty-six, and Anthropic’s is about forty-four per cent and rising. Neither is covering the cost of building the next model, and a company at a thirty-three per cent gross margin with a rising inference bill is not obviously selling above its own cost of serving.

The distinction that makes this legible is between the two kinds of spending. Training is a choice: it buys a model that does not exist yet, and a company can slow it down. Inference is the cost of success: it rises when the product works and more people use it, and it cannot be grown out of, because growth is what makes it larger. A gross margin that falls as a company scales is therefore a more serious fact than a loss that is mostly research: it means the business is not yet selling at a price that covers its cost of serving, and the price is falling faster than the cost.

Now take the Chinese side, where the assumption runs the other way. The reported figure for DeepSeek’s interface margin is eighty-two point nine per cent, with a permanent seventy-five per cent cut to its prices, on revenue of roughly seventy million dollars across the first seven months of the year. So the cheap models are not the ones selling below cost: they are running a high-margin, low-volume business and using price as a weapon to take the commodity layer.

The genuine below-cost selling is one layer beneath all of this, in the renting of the machines. Rental rates for the previous generation of graphics processors fell between sixty-four and seventy-five per cent within fourteen months, to roughly three dollars an hour for on-demand use. A company that financed its hardware on the assumption of eight dollars an hour is now competing in a market at three, which is a business renting below its own cost of capital. It also tells you that the price of the compute the two public companies are buying and building will be set by competition rather than scarcity, even while they spend as though scarcity were permanent.

There is one more turn of the screw. The cost per token is falling, and the cost per task is rising. Agentic work, where a model plans and executes many steps, consumes between ten and a hundred times more tokens than a simple exchange, so a company can pay less for each unit and more for the job. One firm is reported to have modelled an extra six million dollars of inference cost over the next year not because its product was broken but because it needed to stay competitive. Cheaper intelligence and a rising bill are not a contradiction; they are what happens when the price falls and the quantity rises faster.

§What it means for the four companies

So what does all of this mean for the four companies? Take Microsoft first, because its position is the most misunderstood. Azure has passed one hundred billion dollars of annual revenue for the first time and grew forty-three per cent in the latest quarter. It is hosting Anthropic and it has its own model line. The April amendment to its OpenAI partnership dropped the revenue share it owed, ended the exclusivity, and left it holding the equity and the distribution.

The concentration is the thing to watch. Its commercial remaining performance obligations reached six hundred and seventy-eight billion dollars, and the company has said that around forty-five per cent of a somewhat smaller figure was tied to OpenAI. Sequential growth in those obligations came from customers other than the model developers, which is the right direction, and the arrangement means Microsoft earns from the buildout whether or not any particular laboratory wins. Its exposure is to compute demand, and to one counterparty’s ability to pay.

There is an asymmetry worth naming, because it is a consequence of the accounting rather than of any decision. Under the equity method, Microsoft captures about twenty-seven per cent of OpenAI’s upside while it has funded a substantial share of the infrastructure that makes that upside possible. If the laboratory succeeds modestly, the stake is worth a great deal and the earnings contribution stays small; if it struggles, the losses arrive in the same line, immediately.

Alphabet’s position is different in one decisive way: it makes its own silicon. The tensor processing units are its own design, which gives it the lowest cost per unit of compute of any company in this comparison, and that advantage compounds in a price war rather than being destroyed by it. Its cloud business grew eighty-two per cent to twenty-four point eight billion dollars with an operating margin of thirty-five point six per cent, and it carries a backlog of five hundred and fourteen billion dollars of signed contracts.

The tell for Alphabet is its financing. It raised about forty-nine point six billion dollars in June across share classes and convertible preferred and set up a programme to sell up to forty billion more shares at the market. A company with a search business that prints money is raising external capital for the buildout. That is the honest signal about the size of the commitment, and it is why its free cash flow went negative in the quarter.

For the two laboratories, the picture is a role rather than a business. OpenAI is the marginal buyer of compute and the system’s weakest link: its obligations run an order of magnitude above its revenue, its losses are large, its margin is thinner than its competitors’, and the mechanism that funds all of it is a listing that has not happened. Anthropic is closer to durable, with better margin, faster growth and an operating profit just reached, though its growth is booked gross of partner payouts and it depends on the same two balance sheets the whole system does.

The bull case deserves to be stated at full strength, because the bearish version would be easy and cheap. The money is not being burned; it is being converted into an asset that lasts years. Demand currently exceeds supply, on the word of one of these companies itself, and the contracts are signed rather than hoped for, which is what a five-hundred-billion-dollar backlog means. The two laboratories’ revenue is compounding faster than any enterprise software company in history, and the quantity of work routed to these machines is rising faster than the price is falling. If that continues, the depreciation catches up to a much larger revenue base and the whole argument evaporates.

And the bear case at the same strength. The same dollar is counted more than once: it leaves a hyperscaler as capital and a cloud cost, and returns as cloud revenue, and the market pays a multiple on the revenue twice, once at the laboratory if it lists and once at the landlord. The depreciation is running behind because the assets are new, and the assumed life of those assets has been extended twice, which is defensible in each instance and suspicious in aggregate. A meaningful slice of the two public companies’ reported profits this year came from writing up private stakes rather than from selling anything, and the price of the output is collapsing while the cost of the input — land, power and chips — is not.

The circularity deserves to be stated plainly and without innuendo. It is not fraud and it is not a scheme; every leg of it is a disclosed contract. But it is amplification, because growth in one place manufactures reported growth in another, and the amount of hyperscaler revenue funded by capital the same company supplied is not published anywhere. Microsoft’s own disclosure puts a large share of its commercial obligations on a single customer it part-owns. Nobody outside the companies can size the loop, and the loop makes the reported growth rate a less reliable signal than it would otherwise be.

§The price question, three scenarios, and the markers

So to the price question, which can be answered without pretending to a forecast. Microsoft trades around four hundred and ninety-four dollars against a consensus target of about five hundred and seventy-three, with a range from four hundred to eight hundred and seventy. On consensus earnings for its year to June 2028, that is roughly twenty-one times. On free cash flow for the year to this June, it is roughly one hundred and fifteen times. Those two numbers are the whole debate: the earnings are fine and the cash is not, and the difference is capital spending.

Alphabet trades around three hundred and fifty dollars against a consensus target of about four hundred and twenty-eight, with a range from three hundred and forty to five hundred and fifteen. Its current-year earnings are inflated by the paper gain, and next year’s consensus earnings fall to about fifteen dollars, which puts the shares at roughly twenty-three times. Its free cash flow for the year, on consensus, is under nine billion dollars, a rounding error against a four-trillion-dollar company. A range spanning more than a factor of two on either name is not a forecast but an admission that the answer depends on which scenario turns out to be true.

So here is the frame, in three scenarios rather than a target, and with the plain statement that this is arithmetic and not advice; the decision belongs to whoever is listening. The most likely outcome, as this examination reads it, is that capital spending keeps rising into next year, free cash flow stays compressed, earnings keep growing, and the multiple compresses as the market waits for the cash to arrive. That gives a flat-to-modest return with a lot of noise, and it is roughly what the last three quarters have already delivered.

The bull scenario requires that agentic demand fills the capacity as it arrives, that cloud margins hold as third-party capacity is used as a bridge, and that the depreciation catch-up lands against a revenue base large enough to absorb it. It requires that the demand already contracted for turns up on schedule. The bear scenario requires one thing only: that the laboratories cannot fund their commitments. If the listing stalls, or prices below what the private marks imply, then the equity marks in two public companies’ earnings reverse, the cloud growth that depends on those commitments slows, and the market discovers that it was paying for a revenue stream funded by the buyer’s own balance sheet.

The markers are what make this checkable. Each quarter, compare capital-expenditure guidance with the depreciation actually reported; if the gap widens again, the earnings are being flattered further. Watch free cash flow per share rather than earnings per share, because that is where the spending lands. Watch the token share and the revenue share of the Chinese models, and whether the gap between them narrows; if it does, the cheap end has started paying. Watch OpenAI’s monthly revenue against its own target and the price its listing actually achieves. And watch for another extension of assumed asset lives, because that is the clearest available sign that the assets are not earning their keep.

And the test that would prove the sceptical case wrong. If depreciation catches up over the next two years without earnings falling, then the assets are earning their keep and the lag was only the arithmetic of a buildout. If revenue per dollar of capital spending rises rather than falls, then the spending is turning into a business rather than a subsidy. If either laboratory reaches positive cash flow without new financing, then the loop is not load-bearing. And if the Chinese models’ share stops growing without a rule that requires it to stop, then the cheap end was not as durable as the traffic suggests.

So the examination reduces to one comparison and one distinction. The comparison is what the two public companies spend and commit against what the two laboratories pay them, and whether the difference is funded by earnings or by balance sheets and marks. Microsoft’s equity commitment is about a tenth of a single year of its capital spending, and Alphabet’s is comparable, while the cash generation of both has fallen sharply in the year the spending peaked. The distinction is between a company selling below cost to buy a market and a company selling at a margin into a market where the price of its output is falling. The first is a strategy that can work if it ends; the second is a business that has to keep improving faster than its own prices fall.

Two notes on the evidence. The figures for the two private laboratories come from investor documents reported second-hand and are not audited; the forecast medians are one forecasting firm’s estimates rather than a consensus; the DeepSeek margin is Chinese media reporting of internal figures; and the adoption numbers come from developer platforms, which measure the cheap end of the market well and the enterprise end not at all. Everything in this examination that is load-bearing is one of those.

This has been a deep dive into who pays for the AI compute.

Sources: this examination names its sources inline — the corporate disclosures of Microsoft and Alphabet on committed equity, capital expenditure and finance leases, free cash flow, net income and the unrealised gains on equity securities, adjusted earnings against expectations; the forecast medians, drawn from one forecasting firm rather than a consensus; developer-platform traffic and revenue data from OpenRouter and Vercel; the congressional committee’s investigations of two of the named users of the Chinese models; Chinese media reporting of DeepSeek’s internal interface margin and permanent price cut; and the rental-rate data for the graphics processors beneath all of it. Every source is weighed with the same skepticism regardless of politics or business model. The audio version of this article, read by DeepDives, is available on Wavlake.