Meta owns the models and the data centres, and still pays Microsoft to rent someone else's

The evaluation detail is the one that explains the bill. Meta developers have used OpenAI models purchased through Foundry to assess the output of Meta's own models ^[9]. A scoring model you trained yourself carries your own blind spots into the marks, so the reference has to come from outside the building. That requirement does not shrink as Llama improves. It is a permanent reason to hold an account with a competitor's reseller.

The rest reads as ordinary procurement. Meta buys model access to support software development and shops across multiple platforms depending on availability and price ^[8]. Andrew Bosworth, Meta's chief technology officer, said in July on the Big Technology podcast that the company rents leading external models as part of its development process while building its own ^[10]. That is a supply desk behaving like a supply desk, not a company caught short of compute.

The scale is worth putting against Microsoft's own yardstick. Microsoft counts a cohort of Foundry customers running at an annual rate of a trillion tokens, and says that group grew fourfold, while Foundry revenue more than doubled year on year to July and customers using models from several providers rose fivefold from the start of 2026 ^[14]. Meta's volume is trillions of tokens a week ^[2]. At two trillion a week, that is roughly a hundred times the annualised threshold Microsoft uses to describe its heaviest users ^[23].

Set the money against Meta's own build. Capital expenditure was $31.08bn in the quarter to 30 June, and the full-year range was raised to $130bn to $145bn from $115bn to $135bn in January ^[17], a midpoint increase of $12.5bn ^[24]. Third-party cloud is an operating cost, so the Azure bill sits outside both ranges ^[18]. Hundreds of millions a year is on the order of 0.2% of that midpoint ^[25]. Renting is not the alternative to building here. It is a small line beside a very large one, and it survives precisely because it is small.

Microsoft's side of the ledger is less comfortable. It markets Foundry with manufacturers and transport firms ^[13] and claimed 100,000 customers as of July ^[6], but the large buyers Bloomberg names are ByteDance, Adobe, Perplexity and Sierra, the customer service startup co-founded by OpenAI chairman Bret Taylor ^[11]. ByteDance has generally remained the biggest Foundry spender, with Meta now near the top ^[5]. Two of the marketplace's largest buyers therefore build their own frontier models ^[4]. Above that, OpenAI supplied about 70% of Microsoft's entire AI revenue in its most recent financial year ^[12]. Bloomberg's argument is that the spending only justifies itself if adoption spreads beyond technology companies selling to each other ^[26].

There is also a known ending shape. Zuckerberg confirmed in July that an AI cloud business makes sense for Meta, and Bloomberg reports it is building an API service to sell access to various models, which would compete with Foundry directly ^[19]. Microsoft's Bing powered web search on Facebook from the late 2000s until Meta stopped using it by late 2014 and shipped its own replacement ^[20]. Neither company would comment on the current arrangement ^[3], and neither will say how much of Meta's AI spending lands on Microsoft's books ^[22].