Meta Platforms Inc. has become one of Microsoft Corp.’s biggest AI customers, spending hundreds of millions of dollars a year to access artificial-intelligence models through Microsoft’s Azure cloud, according to a person familiar with the matter. The disclosure, reported by Bloomberg on Aug. 20, has the look of a small irony: a company that builds its own frontier models, operates its own data centers and publishes its own open-source AI is paying its biggest rival for the same thing it makes.
The scale is the surprise. Meta runs trillions of tokens a week through Azure’s AI platform, according to the person, who asked not to be identified discussing an internal matter. Tokens are the unit of consumption for AI computing, and trillions per week puts Meta near the top of Microsoft’s AI customer list — in the same league as ByteDance Ltd., which has generally been the biggest spender on Foundry, Microsoft’s marketplace for third-party AI models.
The relationship makes sense at the margin, even if it looks odd at first glance. Meta’s own models, the Llama family, are open-weight and free to download, but running them at Meta’s scale requires infrastructure the company largely builds itself. For certain workloads — testing competitors’ models, benchmarking, specific product features — renting capacity on Azure is faster and cheaper than building it. The purchase is not a confession that Meta’s in-house AI effort failed; it is a procurement decision.
What makes the deal notable is the direction of the money. Microsoft’s AI revenue is concentrated in OpenAI, which accounts for roughly 70% of the company’s AI business and buys massive amounts of compute through Azure. Foundry, where Meta and ByteDance spend, is a smaller slice, but it is the part of the business that sells access to models other than OpenAI’s. Meta’s spending there means Microsoft is now profiting from both sides of the AI market: the models it co-owns and the models it merely hosts.
The numbers put the relationship in context. Azure revenue passed $100 billion in the year to June 30, up 41%, making it the growth engine of Microsoft’s financial results. AI is the reason. The company said it had about 100,000 AI customers as of July, and Meta’s trillions-of-tokens weekly usage makes it one of the heaviest. For Microsoft, a customer that large is a revenue line worth protecting; for Meta, an AI bill that large is a cost worth managing.
The arrangement also says something about the structure of the AI economy. The biggest companies in the industry are increasingly selling to each other: Meta buys models from Microsoft, Microsoft buys compute from OpenAI’s investors, Google rents TPUs to Anthropic, Amazon supplies chips to Anthropic while competing with it. The money circulates inside a small group of firms, and the question that follows, analysts said, is how much of the industry’s growth is real demand and how much is internal churn.
There is real demand underneath, they point out. Meta runs AI across Facebook, Instagram, WhatsApp and its advertising business, and its models serve billions of users. The company’s capital spending on AI infrastructure is among the largest in the industry, and it is simultaneously the biggest builder of new data centers and a buyer of cloud capacity. The two are not contradictory: building ahead of demand is slower and lumpier than renting, and the rental fills the gap.
The practical trigger for the Azure relationship, according to people familiar with the matter, was a combination of speed and access. Meta wanted to evaluate and deploy models from multiple providers, including OpenAI’s GPT family and other Foundry listings, without building separate pipelines for each. Azure’s model marketplace gave it one interface. The spending grew from a pilot into a standing relationship, and at some point the totals crossed into the hundreds of millions.
The arrangement also reflects how far the adversarial relationship between Meta and Microsoft has softened. The two companies have competed fiercely in AI — Meta’s open-source Llama models were built partly as a counterweight to Microsoft’s alliance with OpenAI, and Zuckerberg has publicly criticized the closed, licensing-heavy approach Microsoft championed. Yet the rivalry has not stopped the purchasing. In business, the person familiar with the matter said, the model that wins is the one that works.
For investors, the disclosure cuts both ways. For Microsoft, it is evidence that AI demand is broadening beyond OpenAI and that Foundry can stand on its own. For Meta, the disclosure carries a cost message: AI costs are rising even as the company builds its own infrastructure, and the path to AI profitability runs through expenses like this one. Wall Street has rewarded Meta’s AI spending because its models have driven engagement and ad revenue; a $500-million-a-year cloud bill fits inside that story.
The bigger question is what happens when the AI economy’s biggest players stop buying from each other. If Meta’s in-house chips and data centers eventually cover the workloads it now rents, Microsoft loses a large customer. If OpenAI’s models remain the best available, Meta keeps paying. The circularity that critics worry about — AI companies buying each other’s services and calling it growth — has a natural resolution: the internal consumption ends when the internal buildout is complete.
For now, the deal stands as one of the largest cross-purchases in the industry, a transaction between rivals that each side can explain and neither side advertises. Meta pays, Microsoft earns, and the AI economy’s ledger gets a little more complicated.
