Meta pays Microsoft hundreds of millions a year to rent AI models

Meta is spending hundreds of millions of dollars a year buying access to AI models through Microsoft’s cloud.

The company runs trillions of tokens a week through Azure, Bloomberg reported, citing a person familiar with the matter who asked not to be identified. That makes Meta one of Microsoft’s largest AI customers.

Both companies declined to comment.

The number matters because of who is paying it. Meta builds its own frontier models, owns its own data centres, and is still buying model access from a rival at that scale.

TNW reported the first half of this in June

The desk covered the same franchise two months ago, with a different name in it.

ByteDance is Microsoft’s biggest AI customer, TNW reported in June, at a moment when Washington was treating Chinese AI as a strategic threat.

Bloomberg says ByteDance has generally remained the biggest spender on Foundry, Microsoft’s model marketplace. Meta has now joined it near the top.

Two of the largest buyers of AI models on Microsoft’s marketplace are therefore social media companies that build their own.

What Meta is buying

Foundry sells access to models from several providers through Azure. Microsoft said it had 100,000 customers as of July.

The marketplace is not selling Microsoft’s own models. It resells other people’s, OpenAI’s among them.

Meta is buying model access to help with software development. It purchases across multiple platforms depending on availability and price.

One detail in Bloomberg’s account stands out. Meta developers have used OpenAI models bought through Foundry to assess the output of Meta’s own models.

Meta is paying Microsoft, which resells OpenAI, so that Meta can grade its own homework.

Andrew Bosworth, Meta’s chief technology officer, said as much in July on the Big Technology podcast. The company rents leading external models as part of its development process, alongside building its own.

The customer list is almost entirely tech companies

Microsoft features manufacturers and transport firms in its Foundry marketing.

Its biggest AI customers are not those. Bloomberg names ByteDance, Adobe, Perplexity and Sierra, the customer service startup co-founded by OpenAI chairman Bret Taylor.

The concentration goes further up. OpenAI supplied about 70% of Microsoft’s entire AI revenue in its most recent financial year.

Bloomberg’s point is that for AI to justify the spending, it has to be adopted across the economy rather than sold between technology companies. The concentration and the circularity have both been raised repeatedly this year.

What the platform numbers show

Implicator.ai collected the figures Microsoft has published around this.

Foundry revenue more than doubled year on year to July. Customers using models from several providers rose fivefold from the start of 2026. The number running at an annual rate of a trillion tokens rose fourfold.

That last figure is the one Meta sits inside. It runs trillions of tokens a week, not a year.

Azure revenue passed $100bn in the year to 30 June, up 41%. TNW covered that milestone in July.

Chief financial officer Amy Hood said at the time that “there are still constraints in the system” and that “demand continues to exceed available supply”.

The loop this closes

TNW’s reading of the two companies’ second quarters was that Microsoft’s AI spending became cloud revenue while Meta’s became a cash-flow hole.

Part of the hole is now identifiable. Some of what Meta spends on AI arrives on Microsoft’s books as revenue.

Neither company will say how much, so the size of that transfer is unknown.

Meta spent $31.08bn on capital expenditure in the quarter to 30 June. It raised its full-year forecast to between $130bn and $145bn, up from $115bn to $135bn in January. The Azure bill sits alongside that, not inside it. Third-party cloud spending is an operating cost, so it sits outside both capex figures.

Meta is building the thing it is buying from

Meta has been open about wanting its own version of this business.

Zuckerberg confirmed in July that an AI cloud business makes sense for the company, renting out compute it already owns. Bloomberg reports Meta is also building an API service to sell access to various AI models, which would compete directly with Foundry.

There is precedent for how that ends. Microsoft’s Bing powered web search on Facebook from the late 2000s. Meta stopped using it by late 2014 and built its own replacement.

Microsoft also supplied Meta with computing power for AI development before ChatGPT launched. Meta is now among the largest builders of AI data centres in the world.

The pattern repeats: buy it from Microsoft, then build it and stop.

What is not confirmed

Neither company confirmed the spending figure or the token volume. Both declined to comment entirely.

Bloomberg did not name the models, the prices or the contract terms, and did not say what share of Foundry’s revenue comes from Meta.

The spending figure and the weekly token figure both come from a single anonymous source. Bloomberg gave no accounting period for either.

Microsoft’s own capacity is under question separately. The Guardian reported on 17 August that it counted 2.2 million AI chips installed at Microsoft, fewer than its stated power capacity implied. Microsoft rejected the calculation as based on incorrect assumptions.

Shaolei Ren, a professor at the University of California, Riverside, reviewed those estimates. “According to their own metrics, Microsoft could be correct,” he said. “But it isn’t clear what they mean when they say they have added datacentre capacity. They are giving insufficient context.”

Satya Nadella has been asked repeatedly whether this is a bubble. He has declined to use the word, and described instead how it would end badly.