Microsoft counts Meta among its largest AI customers, and the annual bill runs into the hundreds of millions of dollars, according to Bloomberg. That money buys Meta access to AI models on Microsoft’s Azure cloud, and Meta’s engineers reach for OpenAI models on Microsoft’s Foundry marketplace, alongside other tooling, when they want a yardstick for Meta’s own systems. The detail is not scandalous. Every serious lab tests its models against rivals. What makes it notable is the channel: Meta, which trains its own frontier systems under the Llama banner, is paying its cloud competitor for the privilege of measuring itself against OpenAI.

The scale is the real story. Weekly consumption on Azure climbs into the trillions of tokens, per Bloomberg’s reporting, a volume that puts Meta among Microsoft’s largest AI customers even as Meta operates its own massive data center buildout. That combination suggests Meta’s in-house infrastructure has not yet displaced its need for outside model access, at least for evaluation work.

In parallel, Meta has an API service of its own under construction, and Foundry is the obvious thing it would eventually take business from. Bloomberg frames this as a repeat of what happened with Bing: Meta once relied on Microsoft’s search engine to power results, then built and shipped its own search technology to replace it. The Foundry relationship looks like it is following the same arc, dependency now, substitution later.

Microsoft’s Foundry customer list underscores how concentrated this business is among a handful of large technology buyers. ByteDance tops the list, followed by Adobe, Perplexity, and Sierra. But Foundry itself is a small piece of Microsoft’s AI revenue picture. Roughly 70 percent of that revenue comes from OpenAI alone, which mainly buys compute through Azure rather than routing traffic through the Foundry marketplace. Meta’s spending, notable as it is, sits inside a business still dominated by Microsoft’s own investment partner.

The arrangement says something about where competitive advantage currently sits in AI infrastructure: not in any one lab’s models, but in who controls the compute and marketplace layer those models get evaluated on. Meta’s own Llama models are frontier-caliber by most public measures, yet the company still routes evaluation traffic through a competitor’s marketplace rather than building that layer first. That ordering rarely survives once the internal product is ready to stand alone.

For operators watching the Meta-Microsoft relationship, the signal to track is not the benchmarking itself but the API service Meta is building. If it matures into a real Foundry competitor, expect Meta’s Azure spending to taper the way its search traffic once did after Meta replaced Bing internally. Expect Microsoft, meanwhile, to keep defending the roughly 70 percent of AI revenue still anchored to OpenAI’s Azure compute purchases, since that concentration, not Foundry’s customer roster, is what actually underwrites Microsoft’s AI business today.

Reported by Maximilian Schreiner for The Decoder on August 21, 2026, citing Bloomberg.