Two-Thirds of Microsoft's AI Revenue Comes From OpenAI: $24.1 Billion

The Line That Appeared in the Annual Report for the First Time

Microsoft's annual report, published after the close on August 5, contained one sentence that had never appeared in any prior filing: "For fiscal year 2026, we recorded revenue from commercial arrangements with OpenAI, inclusive of revenue sharing payments, of $24.1 billion, and accounts receivable from OpenAI as of June 30, 2026 was $6.0 billion."

There is nothing exotic about the sentence itself. Large companies are periodically required to disclose revenue from a single customer. What is unusual is where the number lands. Microsoft reported $331.8 billion of total revenue for the same fiscal year, so $24.1 billion is 7.3% of the whole. But the story the company has been telling investors for months is not about total revenue. It is about AI revenue. And there, the ratio looks very different.

Microsoft said in the March quarter that its AI business was running at "an annualized revenue run rate exceeding $37 billion." Divide $24.1 billion by $37 billion and you get roughly 65%. Bloomberg's arithmetic points slightly higher: annualize differently, use a $34 billion base, and the share rises to about 70%. Both methods arrive in the same neighborhood. Somewhere between two-thirds and four-fifths of Microsoft's AI revenue comes from a single customer — a customer Microsoft happens to own a piece of.

How the Two Companies Became This Entangled

The knot took years to tie. Microsoft began investing in OpenAI in 2019, and a meaningful portion of that investment took the form of Azure compute credits rather than cash. Microsoft put money in, the money was spent on Azure, and Azure booked it as revenue.

On October 28, 2025 the structure was rebuilt. OpenAI converted into a for-profit entity, Microsoft emerged as the largest corporate shareholder with roughly 27%, and on the same day OpenAI committed to $250 billion of incremental Azure services on top of existing spending. In April 2026 the exclusivity was loosened and the revenue share OpenAI pays Microsoft was capped.

Those three moves produce today's picture: Microsoft is simultaneously OpenAI's largest supplier, its largest shareholder, and — for a period — a partner taking a cut of its revenue. An investor pricing Microsoft as a cloud company is, without necessarily meaning to, also pricing the cash burn of a single AI lab.

A Third of the Order Book

The larger number is not on the income statement. It is in the order book. At the end of June, Microsoft reported commercial remaining performance obligations — contracted business not yet recognized as revenue — of $678 billion. CFO Amy Hood said the figure "increased 25% when excluding OpenAI."

Investors reverse-engineered the rest from that single clause. If the non-OpenAI book grew 25% to roughly $460 billion, the gap between that and the reported $678 billion belongs to OpenAI: approximately $218 billion. Close to one contracted dollar in three.

The fourth link in the chain carries the same name as the first. That is not, in itself, an accusation; strategic investment cycling back to the supplier is routine in cloud. But once the loop closes, a question follows: how much of this revenue was paid with money OpenAI earned, and how much with money Microsoft supplied? Analyst Jackson Ader put it in one line: "The more of that revenue comes from services to OpenAI rather than the benefits of investment, the more favorably I'm going to look at it."

The Mechanism: One Customer Feeding Both the Top Line and the Loss

There is a second channel here that most readers miss. Microsoft is not merely OpenAI's vendor; it is a part-owner. That roughly 27% stake is accounted for under the equity method. Which means: when OpenAI loses money, Microsoft must deduct its share of that loss from its own net income.

The figure is not abstract. In the first quarter of fiscal 2026, Microsoft recorded a $3.1 billion reduction in net income attributable to its OpenAI investment. Worked backwards, that implies a quarterly loss at OpenAI in the neighborhood of $11.5 billion. OpenAI's own internal forecast, meanwhile, projects a loss of about $14 billion for all of 2026. The two numbers do not reconcile; the gap most likely reflects one-off items tied to the restructuring. What is not in dispute is the order of magnitude.

The lesson here is accounting, not morality. Revenue growth and cash generation are not the same thing, and when a customer is also an equity investee, a single transaction shows up in two places on the income statement pointing in opposite directions. The fundamentals of reading financial statements are built on exactly this distinction.

There is a third line item as well: capital expenditure. Microsoft has guided to $190 billion of capex for calendar 2026, bringing the cumulative total since 2022 to $261.3 billion. Lease commitments that have not yet commenced stand at $329.1 billion. A significant share of that capacity is being built to serve the $218 billion of commitments that will be delivered over a period of years. Once a data center is built, depreciation begins running — not because the customer consumes the capacity, but because the calendar advances. The seller ends up carrying the risk on the buyer's balance sheet.

What Asia Did Overnight

Because the disclosure landed after the U.S. close, Asia paid the first bill. In the August 6 session the Kospi fell 302.82 points, or 4.59%, closing at 6,295.44. Memory manufacturers sat at the center of the damage.

Microsoft's disclosure was not the only cause. SanDisk and Western Digital reported after the close as well, feeding doubts about where the memory cycle has peaked. But both headlines connect to the same question: who ultimately pays for AI spending, and where does that payer's money come from? A memory maker's order book depends on hyperscaler capex, and that capex depends on cloud contracts. If the customer at the far end of the chain is losing something on the order of $14 billion a year, the entire chain depends on that customer finding its next round of capital.

The chart is context, not evidence; the argument above comes from a footnote, not from price. The path of the tech-heavy index over the same stretch is useful for comparison.

Timeline

Nine months is enough to see how the relationship reached its present shape.

What Each Side Says

Two conclusions can be drawn from the same footnote, and the market is currently pricing both.

The analyst split is just as clean. Tigress Financial raised its price target to $690 and Goldman Sachs added the stock to its conviction list with a $640 target; at the other end sit estimates near $450 from those uncomfortable with the pace of capital spending. The gap is not really a valuation argument. It is the same footnote, read two ways. Most valuation disputes start exactly here.

What Remains

Nothing happened on August 5. A relationship that had been guessed at for months was simply reduced to a number. That is precisely why markets distinguish between a guess and a figure: as long as a dependency is merely "probably large," everyone assumes their own number; the moment it is disclosed, everyone has to argue about the same one.

The practical question that follows is this: what happens to Microsoft's AI revenue if OpenAI cannot raise its next round? The contract stays on paper, the data center stays where it is, and depreciation keeps running. That is why concentration risk is not a question of whether a partner is good or bad. It is the oldest subject in risk management: cash flow tied to a single counterparty is an undiversified position, no matter how strong that counterparty looks.

This piece draws on the disclosures in Microsoft's fiscal 2026 annual report, Bloomberg's derivation of the AI revenue share from those disclosures, the company's quarterly earnings releases, and closing data from the August 6 Asian session. Figures for OpenAI's losses are estimates reverse-engineered from Microsoft's accounting rather than disclosed by OpenAI itself, and they vary across sources; they are presented here as a range.