Google restricts access to Meta’s Gemini AI: Limited computing power.

The Financial Times reports that Google has placed limits on Meta’s use of its Gemini AI models in early 2026 because Meta used more compute resources than Google has available. Google notified Meta in March that it was reaching capacity, and then imposed a hard limit. Meta uses Gemini for code assistance, chatbots for customers and advertisers, ad fraud detection, and content moderation on its various platforms.

This is interesting because Meta has its own set of Llama open-weight models that are among the most popular in the world. However, Meta’s own models were not performing well enough for certain production tasks, so Gemini “outperformed its own Llama open-source models,” per the FT article, and Meta opted to use it for those purposes. Meta also uses Anthropic’s Claude for some AI tasks.

1The capacity crunch behind the headline

Meta doesn’t have its own cloud infrastructure, so it needs to use third-party clouds to run large-scale AI. That’s unusual for a company of its size. Meta plans to invest $600 billion in cloud computing over the next two years, which gives you a sense of its need to reduce its dependence on outside parties. Token prices have skyrocketed this year, and many companies are cutting back on their use of AI to control costs.

Google recently signed a deal to pay SpaceX $920 million per month for access to xAI data centers. That’s another sign of the tightness of compute resources. If one of the world’s richest companies has to pay almost a billion dollars per month to get access to extra servers, and another company has to tell its largest enterprise customer that it’s using too much, then the infrastructure shortage is more than just a problem for future planning. It’s a here-and-now problem that is having an immediate effect on what kinds of AI products get built and how they are priced.

2What this means for the AI services market

This is a real-world example of single source risk for companies and developers using AI APIs. The big AI players aren’t just tap water. They have finite resources, and those resources are getting constrained. Those constraints can affect enterprises of all sizes suddenly and unexpectedly. The idea that there will always be more tokens is being put to the test.

This is part of a broader trend away from the winner-take-all era of 2023–2024 and toward an infrastructure market where resources are scarce, allocations are important, and prices are high. Companies that viewed AI APIs as frictionless are finding out, as Meta found out, that there is friction, and it’s only becoming visible now that the check has come due.