Even Google doesn't have enough Google AI to go around.

Google has started limiting how much of its Gemini models Meta can use, Reuters reported over the weekend, citing the Financial Times. Meta had been leaning heavily on Gemini through Google Cloud to power internal workloads like content moderation and safety automation, and Google decided it was time to pull some of those tokens back.

The reason isn't strategic so much as it is mechanical. Google doesn't have the compute to keep selling Gemini at the rate Meta wants to buy it.

On the Q1 2026 earnings call, Sundar Pichai said the quiet part out loud. "Obviously, we are compute constrained in the near term," he told analysts. "Our Cloud revenue would have been higher if we were able to meet the demand." CFO Anat Ashkenazi added that Google is seeing "unprecedented internal and external demand for AI compute resources."

The numbers behind that statement are wild:

  • Google Cloud backlog hit more than $460 billion, nearly doubling in a single quarter. Just over half of that converts to revenue in the next 24 months, meaning Google has years of committed workloads it physically can't deliver yet.
  • Capex for 2026 was raised to $180 to $190 billion, with Pichai signaling 2027 will be "significantly" higher. Even with that kind of spend, the demand curve is still pulling ahead.
  • Gemini's API is now processing more than 16 billion tokens per minute, up from 10 billion the prior quarter. Meta is one of the largest of those customers.

A Google Cloud spokesperson described its existing arrangement with Meta as "a short-term, timely agreement to ensure we have bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise." In plain English, Google would rather sell those tokens to enterprise customers paying for Gemini Enterprise than keep subsidizing a competitor's product roadmap.

The Meta side of this is its own story. After the rough reception to Llama 4, Meta launched Muse Spark earlier this year as its new flagship, and it has been quietly using Gemini in the meantime to keep things like moderation and safety pipelines running at scale. Gizmodo's framing was that Meta got "too addicted to Google AI tokens", and that's basically what happened. When your supplier is also your competitor and your supplier runs short, you're the first one rationed.

This is also happening while Google is already behind on Gemini 3.5 Pro and watching researchers leave for rivals. So the company is rationing tokens externally while trying to free up enough compute internally to ship its own next model.

The squeeze isn't limited to hyperscaler-to-hyperscaler deals. Smaller developers paying for the Gemini API have been hitting unexpected quota limits and billing surprises for months, which is why Google finally rolled out spend caps in May. The whole stack is running hot.

And the bottleneck isn't really chips anymore. Steven Dickens, president of HyperFrame Research, put it bluntly: "The bottleneck has shifted to the grid-to-chip interface, especially transformer availability and local utility capacity." Google can buy all the TPUs it wants. It still has to get them powered and online.

INTO THE VALLEY

If Google is rationing one of the biggest AI labs in the world, nobody else is safe from the same call. The story of 2026 was supposed to be which model is smartest. The story it's actually becoming is which company can plug in enough of them to matter. Compute scarcity is starting to do what model benchmarks couldn't, which is decide who gets to build at scale and who has to wait in line. Meta just learned what that line feels like.