Why Meta Is Renting Rival AI Models at Massive Scale

Picture this: a tech titan with its own data centers, crafting cutting edge models, still paying another company to run trillions of AI tasks every week. That is the odd, quietly expensive reality behind Meta's AI development rhythm.

Why rent a rival's brain?

Meta spends hundreds of millions of euros a year buying access to AI models through Microsoft's cloud. Reports say the company processes trillions of tokens on Azure each week, making Meta one of Microsoft's largest AI customers. Neither company has provided an official comment on the figures, yet the scale itself tells the story.

At first glance this seems counterintuitive. Meta builds proprietary models and operates its own infrastructure. Why would it funnel so much workload to an external platform? The answer is pragmatic and blunt: speed and benchmarking. Meta taps into third party models, including those distributed on Microsoft Foundry, to test, score, and validate the outputs of its own systems.

Put simply, Meta pays to benchmark. Engineers use rented models as reference points to measure quality, safety, and performance. Andrew Bosworth, Meta's chief technology officer, acknowledged in July that while the company invests heavily in native technology, it also leases leading external models as part of its development pipeline.

Foundry functions like a marketplace within Azure. It does not primarily sell Microsoft proprietary models. Instead it hosts models from many providers, OpenAI among them, and offered access to roughly 100,000 customers as of July. Those customers range beyond manufacturing or logistics names that Microsoft highlights in marketing; in practice, the biggest buyers are technology companies, including ByteDance, Adobe, Perplexity, and Sierra.

ByteDance currently tops Foundry in spend, and alongside Meta, the two social media giants lead the list of model buyers on Microsoft's platform. That pattern underlines a surprising dynamic: major social networks, which are also model builders, are among the largest consumers of a rival cloud's model marketplace.

Financially, the growth is stark. Foundry's revenue more than doubled by July, and the number of customers consuming at least one trillion tokens annually quadrupled. Meta's own token use is measured in trillions every week. OpenAI remains a dominant revenue source for Microsoft's AI business, accounting for about 70 percent of that income in the most recent fiscal year.

Here is the high level takeaway: cloud access and model ecosystems are now as strategic as building models themselves.

Azure's broader cloud business is also surging. In the year ending June 30 Azure revenue grew by 41 percent and crossed the threshold of roughly €93 billion. Critics point out a broader concern: if enormous AI costs are recycled primarily among technology giants, the economic case for heavy investment becomes harder to justify unless the benefits extend widely across industries.

So what changes? For Meta, renting rival models is not a permanent surrender, but a tactical choice. It buys time, comparative benchmarks, and a practical path to iterate faster. For Microsoft, having major model builders as customers proves the commercial pull of an open model marketplace. For the industry, this arrangement highlights an emergent truth: building AI and validating it often require two different toolsets, sometimes supplied by competing vendors.

Will this posture remain as in house models mature and costs shift? That is the question keeping engineers and CFOs awake at night.