Meta pays Microsoft several hundred million dollars annually to lease AI models.

      Meta is investing hundreds of millions of dollars annually to access AI models via Microsoft’s cloud services. According to Bloomberg, which cited an unnamed source, the company processes trillions of tokens weekly through Azure, making Meta one of Microsoft's largest customers in the AI sector. Both companies have chosen not to comment on this information.

      The significance of these numbers lies in who is making the payments. Meta develops its own advanced models and operates its own data centers, yet it continues to purchase model access from a competitor at this magnitude. In June, TNW first reported part of this story.

      The same franchise was addressed two months prior under a different name. TNW stated that ByteDance is Microsoft’s largest AI customer, especially at a time when the U.S. government views Chinese AI as a strategic threat. Bloomberg indicates that ByteDance has consistently been the biggest spender on Foundry, Microsoft’s model marketplace, with Meta now joining it among the top buyers.

      Consequently, two major purchasers of AI models in Microsoft’s marketplace are social media companies that also develop their own models.

      Regarding Meta’s purchases, Foundry facilitates access to models from a variety of providers via Azure. Microsoft reported having 100,000 customers as of July. However, the marketplace does not sell Microsoft’s own models; instead, it offers those of others, including OpenAI.

      Meta is acquiring model access to aid in software development, sourcing across different platforms based on availability and pricing. A notable detail from Bloomberg’s report is that Meta developers have utilized OpenAI models purchased via Foundry to evaluate the performance of their own models. Essentially, Meta is compensating Microsoft, which in turn resells OpenAI’s models, allowing Meta to assess its own developments.

      Andrew Bosworth, Meta’s chief technology officer, confirmed this in July during a Big Technology podcast, stating that the company utilizes prominent external models as part of its development alongside crafting its own.

      The roster of customers appears predominantly to consist of tech firms. Microsoft’s Foundry marketing includes manufacturers and transportation companies, but its top AI customers remain technology-focused, including ByteDance, Adobe, Perplexity, and Sierra, the customer service startup co-founded by OpenAI chairman Bret Taylor.

      The concentration of spending extends further. OpenAI reportedly accounted for about 70% of Microsoft’s entire AI revenue in the last financial year. Bloomberg emphasizes that for AI spending to be justified, it must see broader adoption beyond transactions between tech companies. This concentration and interdependency have been discussed several times this year.

      In terms of platform performance, Implicator.ai gathered figures published by Microsoft. Foundry revenue has more than doubled year-on-year up to July, and customers utilizing models from various providers surged fivefold since early 2026. The number of inquiries run at an annual rate of a trillion tokens has increased fourfold.

      Meta falls within this last figure, processing trillions of tokens weekly, not annually. Azure’s revenue surpassed $100 billion in the year ending June 30, marking a 41% increase, which TNW reported in July. Chief financial officer Amy Hood mentioned at that time that there are still “constraints in the system” and that "demand continues to exceed available supply."

      TNW assessed that for the second quarter, Microsoft’s AI expenditure transformed into cloud revenue, while Meta’s expenditures became a cash-flow deficit. Some of Meta's AI spending contributes directly to Microsoft’s revenue, but neither company disclosed the specific amounts.

      In the quarter ending June 30, Meta invested $31.08 billion in capital expenditure, raising its annual forecast from a range of $115 billion to $135 billion in January to between $130 billion and $145 billion. The Azure costs are accounted separately as an operating expense, outside capital expenditure figures.

      Meta has expressed interest in developing its own version of this business model. In July, Zuckerberg confirmed that establishing an AI cloud service made sense for the company since it would utilize its existing resources. According to Bloomberg, Meta is also working on an API service to provide access to various AI models, directly competing with Foundry.

      There is historical context for this trend. Microsoft’s Bing powered Facebook’s web search from the late 2000s until Meta developed its own system by late 2014. Microsoft also provided computing resources for Meta's AI development prior to the launch of ChatGPT, and now Meta ranks among the world's largest builders of AI data centers.

      This pattern is consistent: Partner with Microsoft for initial acquisition, build independently, then cease purchasing from them.

      Unconfirmed points include the exact spending figures and token volumes; both companies chose not to comment. Bloomberg did not specify the models, pricing, or contract details, nor disclose what percentage of Foundry’s revenue Meta represents.

      Both the spending and token figures stem from an anonymous source, and Bloomberg did not provide an accounting period for either. Separately, Microsoft’s capacity has come under scrutiny; The Guardian reported on August 17 that it had 2.2 million AI chips