Google Capped Meta's Gemini Access: Not Enough Compute

Google limited Meta's access to its Gemini AI models in early 2026 after Meta consumed more computing capacity than Google could supply, according to reporting by the Financial Times. Google first warned Meta in March that it was approaching its limits; a formal cap followed. The restrictions cover Meta's use of Gemini for coding assistance, customer-facing chatbots, advertiser-facing chatbots, ad fraud detection, and flagging harmful content across its platforms.

What makes the story striking is why Meta was using Google's model in the first place. Meta publishes its own Llama family of open-weight models, one of the most widely distributed AI model families in the world. But for specific internal production workloads, its own models underperformed. Gemini "outperformed its own Llama open-source models," according to the FT's reporting, which is why Meta chose it for these systems. Meta also uses Anthropic's Claude for some of its AI workloads.

The capacity crunch behind the headline

Meta does not operate its own cloud infrastructure, which means it relies on external providers for AI workloads at scale — an unusual position for a company of its size. Meta has pledged $600 billion in cloud computing investment over the next two years, which is one indicator of how urgently it needs to close that dependency gap. Token prices have risen sharply across the industry this year, and a number of companies have begun pulling back on AI usage to manage costs.

Google, meanwhile, recently agreed to pay SpaceX $920 million per month to access xAI data centers — an arrangement that itself reflects how stretched compute supply has become. When one of the world's most valuable companies has to pay nearly a billion dollars a month just to access additional server capacity, and when another has to tell its biggest enterprise customer it's using too much, the infrastructure bottleneck is no longer a planning concern. It's a present-tense constraint that is actively shaping which AI products exist and how they are priced.

What this means for the AI services market

For businesses and developers building on AI APIs, this is a concrete illustration of single-provider risk. The major AI providers are not utilities with unlimited throughput. They are navigating genuine capacity constraints, and those constraints can reach enterprise customers — at any scale — without warning. The assumption that more tokens are always available is exactly the assumption this story tests.

There is also a longer structural story here. The AI services market is shifting from the land-grab dynamics of 2023 and 2024 toward something that looks more like a constrained infrastructure market: capacity is limited, allocation decisions matter, and price reflects scarcity. The companies that treated AI APIs as frictionless utilities are now learning, as Meta did, that the friction is real — it was just invisible until the bill arrived.