Meta spends $500M+ yearly renting Microsoft AI models

Meta is reportedly pouring over half a billion dollars annually into renting AI models through Microsoft’s Azure cloud, a figure that underscores the scale of its dependence on third-party infrastructure. The company is said to process trillions of tokens each week on the platform, positioning it as one of Microsoft’s most significant AI clients, according to Bloomberg, which cited an unnamed source. Neither firm has offered an official response.

The real story here is who is writing the check. Meta develops its own frontier models and operates its own data centers, yet it still finds it necessary to purchase model access from a competitor at this volume.

This development echoes reporting from TNW back in June, though the spotlight has shifted. At that time, ByteDance was identified as Microsoft’s top AI spender, a detail that carried geopolitical weight given Washington’s stance on Chinese AI. Bloomberg now indicates ByteDance has generally kept that top spot on Foundry, Microsoft’s model marketplace, with Meta climbing to join it near the summit. The result is that the two largest buyers of AI models on that platform are social media companies that also build their own technology.

So what exactly is Meta purchasing? Foundry serves as a reseller for models from various providers, not just Microsoft’s own offerings. The marketplace includes OpenAI’s models among others, and Microsoft has claimed 100,000 customers since July. Meta’s acquisition strategy appears focused on software development support, with purchases spread across platforms based on price and availability.

One particular detail from Bloomberg’s report stands out: Meta developers have used OpenAI models, bought through Foundry, to evaluate the performance of Meta’s own creations. In effect, Meta is paying Microsoft, which resells OpenAI, to grade its own homework. Andrew Bosworth, Meta’s chief technology officer, essentially confirmed this approach in July on the Big Technology podcast, noting the company rents leading external models as part of its development cycle while continuing to build its own.

The customer base for Foundry is telling. Microsoft markets the platform to manufacturers and transport firms, but its actual biggest buyers are tech companies. Bloomberg lists ByteDance, Adobe, Perplexity, and Sierra, the customer service startup co-founded by OpenAI chairman Bret Taylor. The concentration runs even deeper: OpenAI alone accounted for roughly 70% of Microsoft’s entire AI revenue in its most recent financial year. Bloomberg’s underlying point is that for AI spending to be justified, adoption must spread across the broader economy, not just circulate among technology firms. That circularity and concentration have been recurring themes throughout this year.

Platform data collected by Implicator.ai shows Foundry revenue more than doubled year over year as of July. Customers using models from multiple providers jumped fivefold since the start of 2026, and the number running at an annual rate of a trillion tokens quadrupled. Meta sits squarely in that last category, though it operates at trillions of tokens per week, not per year.

Azure revenue itself crossed $100 billion in the year ending 30 June, a 41% increase, a milestone TNW covered in July. Chief financial officer Amy Hood acknowledged at the time that constraints persist and demand continues to outpace available supply.

TNW’s earlier analysis of both companies’ second quarters suggested Microsoft’s AI spending was converting into cloud revenue while Meta’s was becoming a cash-flow drain. Part of that drain is now identifiable: some of what Meta spends on AI lands on Microsoft’s books as income. The exact size of that transfer remains unknown since neither company will disclose figures.

Meta’s capital expenditures reached $31.08 billion in the quarter ending 30 June, with its full-year forecast raised to between $130 billion and $145 billion, up from the $115 billion to $135 billion projected in January. The Azure bill sits outside those capex numbers entirely, as third-party cloud spending counts as an operating cost.

Meta has been transparent about its ambitions to replicate this business model for itself. Zuckerberg confirmed in July that an AI cloud offering makes sense for the company, renting out compute it already owns. Bloomberg also reports Meta is developing an API service to sell access to various AI models, which would put it in direct competition with Foundry.

There is historical precedent for how this plays out. Microsoft’s Bing powered web search on Facebook from the late 2000s until Meta replaced it with its own solution by late 2014. Microsoft also supplied Meta with computing power for AI development before ChatGPT existed, and Meta is now among the world’s largest builders of AI data centers. The pattern repeats: buy from Microsoft, build your own, then walk away.

Several details remain unconfirmed. Neither company has verified the spending figure or the token volume, and both declined to comment. Bloomberg did not identify the specific models, prices, or contract terms, nor did it specify what share of Foundry’s revenue comes from Meta. The spending and weekly token numbers both trace back to a single anonymous source, with no accounting period provided for either.

Separately, Microsoft’s own capacity claims have come under scrutiny. The Guardian reported on 17 August that its count of 2.2 million AI chips installed at Microsoft fell short of the company’s stated power capacity. Microsoft dismissed the calculation as based on faulty assumptions. Shaolei Ren, a professor at the University of California, Riverside, reviewed the estimates and offered a measured take: Microsoft could be correct by its own metrics, but the lack of context around its datacenter capacity claims leaves room for doubt.

Satya Nadella has faced repeated questions about whether this constitutes a bubble. He has avoided that word but has described, in his own terms, how a downturn would unfold.