Loading page…
The reported arrangement shows a major model developer can also be a major buyer of outside models. Microsoft gains high-volume demand now, while Meta’s planned API service could eventually overlap with the marketplace it uses.
Listen to this story
Meta is simultaneously a heavy Microsoft Azure model customer and a potential competitor to Microsoft’s Foundry marketplace. Its developers reportedly consume trillions of tokens weekly through Azure, including OpenAI models used to benchmark Meta’s own systems, while annual spending reaches hundreds of millions of dollars. Meta is also building an API that could package access to multiple models, though its launch timing, lineup, customers, and any plan to replace Azure remain unknown.
Meta uses outside models for evaluation and development, not only as substitutes for its own systems.
Microsoft Foundry had 100,000 customers as of July, but its largest spenders remain mostly technology companies.
ByteDance is generally Foundry’s biggest spender; Adobe, Perplexity, and Sierra are also major customers.
Meta has reportedly become one of Microsoft’s largest AI customers, spending hundreds of millions of dollars annually for access to AI models through Azure. At the same time, Meta is developing an API service that could compete with Microsoft Foundry, the marketplace through which it has used outside models.
The reported arrangement is more than a purchase of computing capacity. Meta developers have used OpenAI technology through Foundry to evaluate the output of Meta’s own models, according to a person familiar with the matter. That places external models inside an evaluation workflow rather than treating them simply as replacements for Meta’s systems.
Meta also sources model access through multiple platforms based on availability and cost, according to people with knowledge of the matter. Its use through Microsoft’s platform reaches trillions of tokens each week, a measure of AI-computing consumption. Meta and Microsoft declined to comment on the reported arrangement.
Meta Chief Technology Officer Andrew Bosworth described the broader approach in July: alongside making its own models, Meta rents leading models as part of its development process. The reported Azure usage fits that division of labor, in which a company can build models while obtaining outside systems for particular tasks.
Microsoft’s Foundry offers models from multiple providers through a single marketplace and had 100,000 customers as of July. Meta’s planned service would also sell access to various models, creating a potential overlap in distribution even as Meta remains a buyer.
Foundry is one element of Microsoft’s AI business. Microsoft also offers its Copilot assistant and rents AI-focused computing power. A marketplace gives customers a route to models from several providers, while Meta’s proposed API would seek a comparable role on the selling side.
The service could reduce Meta’s spending on external providers, but that is a possible outcome rather than an announced migration plan. Its timing, model lineup and eventual customer base remain unspecified, leaving the competitive impact unresolved.
Meta’s spend adds to a customer base that is still heavily weighted toward technology companies. ByteDance has generally been Foundry’s biggest spender, while Adobe, Perplexity and Sierra are among other large customers, according to Bloomberg’s sources.
Microsoft’s promotional material often highlights clients in manufacturing and transportation, but its largest AI customers are mostly technology companies, according to people familiar with the business. The contrast separates Foundry’s broad customer count from the composition of its biggest spenders.
OpenAI contributed about 70% of Microsoft’s overall AI revenue in its most recent fiscal year, Bloomberg reported. That figure covers Microsoft’s overall AI revenue, rather than Foundry alone, so it is not a measure of the marketplace’s individual customer mix. Meta adds demand beyond that dominant relationship while pursuing its own API ambitions.
Meta has relied on Microsoft technology before. Bing supported web searches on Facebook beginning in the late 2000s, before Meta stopped using Bing by late 2014. Microsoft also supplied Meta with AI-development computing power before ChatGPT’s release, while Meta has since become one of the largest developers of AI data centers for its own model work.
Editorial analysis
Our take: The important signal is not that Meta buys outside AI services; it is that it appears to be buying them for a specific development function while keeping open the option to become a distributor itself. That makes token volume a less durable indicator of platform loyalty than it might seem. The next concrete evidence to watch is the scope of Meta’s API service: which models it offers, whether outside customers can buy them, and whether Meta’s use of external model platforms changes after launch. Model-routing options are widening elsewhere too, as cloud platforms add new model providers.
Citation desk / original work
Our take: The important signal is not that Meta buys outside AI services; it is that it appears to be buying them for a specific development function while keeping open the option to become a distributor itself.
/posts/meta-is-a-major-microsoft-ai-customer-and-a-potential-foundry-rival#finding-1
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.