Google has placed restrictions on Meta Platforms’ use of its Gemini AI models because the company could not provide the computing capacity Meta had sought, the Financial Times reported on June 28, citing people familiar with the matter.

Google notified Meta around March 2026 that it could not fulfill the requested capacity, disrupting several of the company’s internal AI projects. In response, Meta instructed staff to make more efficient use of AI tokens and began accelerating a shift to Muse Spark, its own internal model developed under its Superintelligence Labs division.

What Meta Was Using Gemini For

Meta relied on Gemini for several functions: customer service chatbots, tools for advertisers, code generation for engineers, content moderation, and scam and harmful-content detection. The company had turned to Gemini in part because it outperformed its own open-source Llama models on these tasks. Meta also uses Anthropic’s Claude for some workloads.

A Compute Crunch Hitting the Whole Industry

The episode illustrates a widening gap between surging AI demand and the physical infrastructure required to serve it.

Google is investing more than $180 billion in capital expenditure in 2026, yet still cannot meet all client demand. To close part of its own capacity gap, Google signed a deal with SpaceX paying roughly $920 million per month for access to around 110,000 Nvidia GPUs. Anthropic struck a similar arrangement with SpaceX, committing approximately $1.25 billion per month for compute resources.

Google Cloud’s signed backlog of customer commitments now exceeds $460 billion — nearly double earlier levels — signaling that demand pressure will intensify further.

Meta has raised its 2026 capital expenditure guidance to $115–135 billion and reassigned 7,000 employees to AI-focused roles. The compute friction with Google is pushing Meta toward self-sufficiency: less reliance on external AI providers, and a faster build-out of its own infrastructure and proprietary models.

Several other Google Gemini clients are facing similar constraints, though Meta’s case has been the most publicly disclosed.