Google Rations Meta's Gemini Access as Compute Runs Short

The scarcest thing in artificial intelligence right now isn't talent or data — it's compute, and even Google doesn't have enough of it. According to the Financial Times, citing three people familiar with the matter, Google told Meta back in March 2026 that it could not sell the social-media giant all the Gemini capacity it wanted. The result: Meta had to tell its own engineers to start conserving tokens, and quietly delayed some internal AI projects that had come to depend on Google's models.

The detail is striking because of who is being rationed. Meta is one of the best-capitalized companies on earth, spending tens of billions of dollars a year building its own AI infrastructure — yet it had still been leaning on Gemini for internal workloads including coding, customer service, advertising tools, and content moderation. When Google capped how much it could buy, Meta didn't just shrug it off; it changed how its teams work, pushing them toward more efficient token usage rather than assuming near-infinite access to a rival's frontier model.

What makes the squeeze remarkable is that it is happening despite historic spending. Google is on track to pour more than $180 billion into AI infrastructure this year, and it still cannot meet all of the demand landing on its cloud. In June the company went so far as to agree to pay SpaceX roughly $920 million a month for about 110,000 Nvidia GPUs housed inside xAI's data centers — capacity Google openly described as a "bridge" to satisfy demand for its Gemini Enterprise product. When one hyperscaler is renting a competitor's GPUs by the month to keep up, the shortage is not a rounding error.

The episode reframes a rivalry that is usually told through benchmark scores. On paper, Meta, Google, OpenAI, and Anthropic compete on model quality. Underneath, they are all fighting over the same finite pool of leading-edge chips, power, and data-center space — and the companies that control that physical layer can throttle the ones that don't. Google selling Gemini to Meta is lucrative, but not as valuable as reserving that compute for Google's own products and highest-margin cloud customers when supply is tight.

For the broader industry, the takeaway is blunt: access to compute is becoming a strategic weapon, not just a line item. It helps explain why Anthropic is signing multi-billion-dollar, multi-year data-center leases, why OpenAI is designing its own inference chips with Broadcom, and why Anthropic is reportedly courting Samsung to build custom silicon of its own. Every lab now understands that owning — or at least locking up — the hardware underneath the model is the difference between shipping on schedule and telling your engineers to count tokens. Meta just learned that lesson from the wrong end.