Meta, the company behind the Llama family of open-weight AI models, has quietly become one of Microsoft's biggest AI customers, according to a Bloomberg report relayed by The Decoder — spending hundreds of millions of dollars a year to access AI models through Microsoft's Azure cloud and burning through trillions of tokens every week on the platform.
The arrangement is a striking snapshot of the AI industry's tangled economics: the owner of one of the world's most widely used model families is simultaneously one of the largest consumers of its chief rivals' technology.
Benchmarking the Competition
According to the report, Meta's engineers use OpenAI models available through Microsoft's Foundry marketplace — among other tools — to benchmark the performance of Meta's own models. In other words, the company routinely rents access to the competition to find out how its own AI stacks up, and it pays Microsoft for the privilege at industrial scale.
The trillions-of-tokens figure underscores just how routine that workflow has become. Token consumption has emerged as one of the clearest measures of real AI usage in the enterprise market, and Meta's weekly volume places it among the heaviest users of Microsoft's AI infrastructure.
For a sense of scale: a single modern model query can consume anywhere from hundreds to many thousands of tokens, and enterprise API programs at large tech firms run continuous evaluation suites against rival models. At trillions of tokens per week, Meta's usage is less a marginal convenience than a standing operational dependency — one that now costs the company hundreds of millions of dollars every year.
The Bing Precedent
At the same time, Meta is building its own API service that could eventually compete directly with Microsoft's Foundry marketplace. Bloomberg's reporting draws a pointed historical parallel: Meta once relied on Microsoft's Bing search engine, and then replaced it with its own technology once it had built sufficient in-house capability. The Azure relationship, the implication goes, may have a similar shelf life.
The dynamic puts Microsoft in a familiar position. Its cloud franchise has long benefited from powering services that later grow into competitors, and AI is accelerating that cycle. Every year Meta spends honing its models against OpenAI systems on Azure is also a year Meta spends building the expertise — and the infrastructure — to make that spending unnecessary.
Microsoft's AI Business Runs on Tech Companies
Meta is not an outlier among Microsoft's biggest AI customers — most of them are technology companies themselves. According to the report, ByteDance tops the list of Foundry marketplace customers, followed by Adobe, Perplexity and Sierra.
The bigger picture is even more concentrated. Roughly 70 percent of Microsoft's AI revenue comes from OpenAI alone, which mainly purchases massive amounts of compute through Azure as it trains and serves its models. Foundry, the marketplace through which Meta and others rent model access, accounts for only a slice of that business.
That concentration cuts both ways for Microsoft. The OpenAI relationship anchors its AI franchise, but marquee external customers like Meta, ByteDance and Adobe diversify the revenue base — even if many of those customers are actively building the infrastructure that could one day replace it.
The figures also offer a rare glimpse into a market that is usually opaque. Cloud providers rarely disclose how AI revenue splits between training compute and inference services, and the finding that most of Microsoft's largest AI customers are themselves AI companies suggests the enterprise AI boom is, to a significant degree, technology firms selling to one another — a detail investors watching AI capital expenditure cycles are increasingly scrutinizing.
Coopetition at Scale
The pattern illustrates what analysts have taken to calling coopetition at internet scale. The same hyperscalers that compete for AI workloads also depend on one another: Meta buys rival models via Azure, Microsoft hosts the models Meta competes against, and OpenAI — whose compute spending underwrites much of Microsoft's AI revenue — is simultaneously the product Meta benchmarks Llama against.
For Meta, the spending also reflects an uncomfortable reality of frontier AI development: to prove your model is competitive, you need continuous, systematic access to the models you are competing with. Manual, occasional comparisons no longer suffice in a market where leadership positions shift with each release.
The strategic question is what happens when Meta's own API service matures. If the Bing precedent holds, hundreds of millions of dollars of annual Azure spending could gradually migrate in-house — making Meta's current status as a top Microsoft AI customer a snapshot of a market still in transition rather than a permanent arrangement.
For continuous coverage of the deals reshaping the AI economy, bookmark AI Buzz Wire, our daily hub for breaking AI news.
