Google Caps Meta's Gemini Access as the AI Compute Crunch Bites

The scarcest thing in artificial intelligence this year is not talent, and it is not data. It is the raw computing power to run the models. The squeeze has now reached a point where even Google is telling one of its biggest customers no.

Google has put limits on Meta's use of its Gemini models after the social network sought to buy more computing capacity than Google could supply, the Financial Times reported on Sunday in an account carried by Reuters. Google told Meta around March that it could not meet the full Gemini capacity the company wanted to purchase, the report said, and the shortfall delayed some of Meta's internal AI projects. Reuters said it could not independently verify the account, which cited people familiar with the matter. Neither company responded to requests for comment.

Strip away the corporate names and the detail that lands is this: a firm pouring tens of billions into its own data centers still had to ration what it could rent from a rival. Several other Google clients were affected too, according to the report, though Meta felt it most because its demand was so high. Inside Meta, the response has been to tell staff to be more careful with "tokens," the units that meter how much AI work gets done. That is a telling phrase. Tokens are now budgeted the way a factory budgets electricity.

Why even Google is short

The bottleneck is easier to understand from Google's own books. Over the past year, more than 375 Google Cloud customers each processed over a trillion tokens, and the company's monthly token processing jumped roughly sevenfold to more than 3.2 quadrillion, figures Sundar Pichai laid out at Google I/O in May. Demand like that does not leave much slack to resell.

Pichai has been blunt about it. We are compute constrained in the near term, he told investors, adding that our Cloud revenue would have been higher if we were able to meet the demand. Google Cloud booked about $20 billion in the first quarter, and its order backlog nearly doubled from the prior three months. To dig out, Alphabet has committed somewhere between $175 billion and $185 billion in capital spending for 2026, close to double the $91.4 billion it spent the year before, much of it aimed at the data centers and custom chips that train and serve models like Gemini.

The appetite is real. Google's Gemini app has crossed 900 million monthly users, up from about 400 million a year earlier, according to figures Google shared alongside its developer conference. Serving that many people leaves a thinner margin to sell wholesale to a competitor that is racing to train its own systems.

The episode rhymes with the wider hardware shortage now rippling through the market. The same scarcity of high-end silicon has whipsawed chipmakers and rattled equity markets from Seoul to New York, and it is one reason Apple and others have warned that memory prices are climbing. When the company that builds the most AI infrastructure on Earth is rationing it, the constraint is no longer hype. It is physics, power, and supply chains.

For Meta, the practical effect is a nudge toward efficiency and toward leaning harder on its own Llama models and its own buildout rather than a neighbor's spare capacity. For everyone watching the AI race as a contest of clever models, the Gemini cap is a reminder that the contest is increasingly being decided one rung lower, in the unglamorous business of who can pour enough concrete and plug in enough chips.

Meta got the message in the simplest possible form: a bill it could not fully pay, from a rival that could not fully fill it. The next phase of the buildout — Google's racing rivals included — will be measured less in benchmarks than in megawatts.