What Happened
Bloomberg reported on August 20 that Meta is spending hundreds of millions of dollars a year buying access to AI models through Microsoft’s cloud, running trillions of tokens a week through Azure. That volume places Meta among Microsoft’s largest AI customers. The report rests on a person familiar with the matter who asked not to be identified; both companies declined to comment.
The thing to understand is what Meta is actually buying. Azure AI Foundry is not a storefront for Microsoft’s own models — it is a marketplace that resells other providers’ models, OpenAI’s prominently among them. Microsoft said the platform had 100,000 customers as of July. Meta, according to the report, purchases across multiple platforms depending on availability and price, primarily to support software development work.
One detail travelled further than the rest. Meta developers have used OpenAI models bought through Foundry to assess the output of Meta’s own systems. Andrew Bosworth, Meta’s chief technology officer, described something close to this in July on the Big Technology podcast, framing external model rental as a normal part of the development process rather than an admission of anything. Meta pays Microsoft, which resells OpenAI, so that Meta can grade its own work.
Why It Matters
The figure is not remarkable in absolute terms. Hundreds of millions of dollars is close to a rounding error against Meta’s balance sheet. What makes it interesting is the identity of the payer: a company that builds its own frontier models, owns some of the largest data centres in the world, and spent $31.08 billion on capital expenditure in the quarter to June 30 alone.
Meta raised its full-year capex forecast to between $130 billion and $145 billion, up from a January range of $115 billion to $135 billion. The Azure bill sits alongside that rather than inside it — third-party cloud spending is an operating cost, so it appears in neither capex figure. Which means the headline infrastructure numbers, enormous as they are, understate what these companies spend on AI overall.
The platform figures Microsoft has published sketch the shape of the market underneath. Foundry revenue more than doubled year on year to July. The number of customers drawing on models from several providers rose fivefold from the start of 2026. The number running at an annualised rate of a trillion tokens rose fourfold. Meta sits well past that last threshold: it reportedly runs trillions of tokens weekly, not annually.
Bosworth’s framing deserves to be taken at face value. If you are building models, you need reference points, and the fastest way to get them is to rent whatever is currently best and measure against it. That is procurement, not surrender. But it does quietly settle a question the industry has argued about for three years: model capability is turning out to be less a moat you defend than an input you buy.
Reaction
The reaction focused less on Meta than on the customer list it joined. Bloomberg names ByteDance, Adobe, Perplexity and Sierra — the customer service startup co-founded by OpenAI chairman Bret Taylor — among Microsoft’s largest AI customers, with ByteDance generally the biggest spender on Foundry. Microsoft’s own marketing for the platform features manufacturers and transport companies. Its actual top accounts are technology firms, two of which are social media companies that build competing models.
The concentration runs further up the stack. OpenAI supplied roughly 70% of Microsoft’s entire AI revenue in its most recent financial year. Set that beside a top-customer list drawn almost entirely from the technology sector and the critique writes itself: for AI spending to be justified by anything other than momentum, adoption has to spread into the rest of the economy rather than circulating between the companies building it.
Satya Nadella has been asked more than once whether this is a bubble and has consistently declined the word, while describing in some detail how such a thing would unravel. A separate question hangs over capacity. The Guardian reported on August 17 that it counted 2.2 million AI chips installed at Microsoft, fewer than the company’s stated power capacity implied; Microsoft rejected the calculation as resting on incorrect assumptions. Shaolei Ren, a professor at the University of California, Riverside, reviewed the estimates and said of the disclosures: “They are giving insufficient context.”
What’s Next
Meta is building the thing it is currently renting. Mark Zuckerberg confirmed in July that an AI cloud business makes sense for the company — renting out compute it already owns — and Bloomberg reports Meta is also developing an API service to sell access to a range of AI models. That product would compete directly with Foundry, the marketplace Meta is presently one of the largest customers of.
There is a precedent, and it is not subtle. Microsoft’s Bing powered web search on Facebook from the late 2000s until Meta dropped it around 2014 and built a replacement. Microsoft also supplied Meta with computing power for AI development in the years before ChatGPT. The pattern in both cases was the same: buy it from Microsoft while the internal version is immature, then build and stop buying. Nothing in the current arrangement suggests a different ending.
What is worth watching in the meantime is supply. Azure revenue passed $100 billion in the year to June 30, up 41%, and chief financial officer Amy Hood said at the time that demand continues to exceed available supply. If capacity stays tight, the interesting question is not whether Meta keeps paying Microsoft, but whether Microsoft keeps having room to sell — and which customers get served first when it does not.
Closing Thoughts
It is worth holding this story a little loosely. Neither company confirmed the spending figure or the token volume, and both declined to comment entirely. Bloomberg did not name the models, the prices, or the contract terms, did not specify what share of Foundry’s revenue comes from Meta, and gave no accounting period for either number. Both figures come from a single anonymous source. That is a normal way for this kind of information to surface, and it is still a thin basis for anything load-bearing.
If the reporting holds, though, the implication is more interesting than the dollar amount. The prevailing story of the last three years has been that frontier models are the scarce asset and everything else is commodity. An arrangement where one of the largest builders of AI infrastructure on earth routinely rents a competitor’s models through a third company’s marketplace points the other way. Models look increasingly like electricity: you can generate your own, and you will still buy from the grid when it is cheaper or faster or simply available.
Which leaves the question the whole sector keeps circling. A market where the biggest buyers are also the biggest sellers can grow impressively for a long time without proving very much. The number that would actually settle the argument is not Meta’s Azure bill. It is how much of Foundry’s revenue comes from companies that do not build AI at all — and that is precisely the figure nobody has published.
한글 요약
블룸버그는 8월 20일, 메타가 마이크로소프트 애저(Azure)를 통해 AI 모델 이용료로 연간 수억 달러를 지출하며 주당 수조 개의 토큰을 처리하고 있다고 보도했습니다. 이 규모라면 메타는 마이크로소프트의 최대 AI 고객 중 하나가 됩니다. 두 회사 모두 확인을 거부했습니다. 메타가 사는 것은 마이크로소프트 자체 모델이 아니라, 애저 AI 파운드리(Foundry) 마켓플레이스에서 재판매되는 오픈AI 등 외부 모델의 접근권입니다.
주목할 점은 금액이 아니라 지불 주체입니다. 메타는 자체 프런티어 모델을 만들고 세계 최대급 데이터센터를 보유하며, 6월 말 분기에만 310억 8천만 달러를 설비에 투자했고 연간 전망을 1,300억~1,450억 달러로 상향했습니다. 그럼에도 경쟁사 모델을 빌려 씁니다. 클라우드 지출은 운영비로 분류돼 이 설비투자 수치 밖에 있으므로, 겉으로 드러난 AI 지출은 실제보다 작게 보입니다. 메타 CTO 앤드루 보즈워스는 외부 모델 임차가 개발 과정의 일부라고 설명한 바 있습니다.
논란의 초점은 고객 명단입니다. 바이트댄스, 어도비, 퍼플렉시티, 시에라 등 마이크로소프트의 주요 AI 고객이 대부분 기술 기업이고, 오픈AI가 마이크로소프트 AI 매출의 약 70%를 차지합니다. AI 지출이 정당화되려면 기술 기업끼리의 거래를 넘어 산업 전반으로 확산돼야 한다는 지적이 나오는 이유입니다. 한편 메타는 자체 AI 클라우드와 모델 API 사업을 준비 중이어서, 지금 이용 중인 파운드리와 직접 경쟁하게 될 전망입니다.