Analysis
Meta Platforms is spending hundreds of millions of dollars a year and consuming trillions of tokens weekly through Microsoft's Azure AI Foundry marketplace, making it one of Microsoft's largest AI customers, Bloomberg reported Thursday, citing people familiar with the arrangement.
The awkward part: Meta and Microsoft are direct competitors
What makes the relationship notable isn't the spend itself -- plenty of companies buy AI compute through Azure -- it's that Meta builds and open-sources its own competing frontier models, and Microsoft's own AI ambitions through Copilot and its OpenAI partnership put the two companies in direct competition for enterprise AI customers. Bloomberg's reporting specifically notes that Meta developers use OpenAI models purchased through Foundry to assess the output quality of Meta's own models -- meaning Meta is paying for access to a competitor's model specifically to benchmark itself, rather than relying purely on internal evaluation tools.
Microsoft's Foundry marketplace works as a multi-vendor model bazaar rather than an exclusive Microsoft product: it hosts models from OpenAI, Meta itself, and other labs, letting Azure customers mix and match rather than committing to a single vendor. That structure is exactly what makes Meta's participation possible in the first place -- Meta isn't buying Microsoft's own models, it's buying compute and cross-vendor model access through Microsoft's infrastructure, which is a different and more defensible-sounding relationship than "Meta pays Microsoft to use Copilot."
Where Meta ranks
ByteDance remains Microsoft's single largest Foundry customer by spend, according to Bloomberg's sourcing, with Meta now sitting near the top of that list alongside it. Neither the specific dollar figures nor Meta's exact rank have been confirmed by either company -- Microsoft and Meta both declined to comment to Bloomberg -- so the numbers should be read as directionally accurate reporting rather than disclosed fact.
Why this matters for the AI infrastructure market
Microsoft closed its 2026 fiscal year with 30 million Copilot seats and Azure revenue growth of 43%, and OpenAI becoming a subprocessor within Microsoft 365 -- context that shows Microsoft's AI business is increasingly a platform layer serving many labs' models, not solely a vehicle for its own OpenAI partnership. A rival like Meta becoming one of its biggest customers reinforces that Azure's AI infrastructure business can grow by hosting competitors' workloads, similar to how AWS has long hosted workloads for companies that compete with Amazon's own retail business.
The counterweight worth naming directly: this is Meta buying incremental compute and cross-model access for specific evaluation and workload needs, not a signal that Meta's own infrastructure buildout -- which spans its own massive data-center investments -- is insufficient or being scaled back. Companies at Meta's scale routinely run hybrid infrastructure strategies, and using a competitor's marketplace for specific benchmarking tasks doesn't necessarily mean Meta trusts Microsoft more than it trusts its own systems generally; it may simply be the fastest way to get an independent, non-Meta-controlled comparison point.