Google has placed limits on Meta’s use of its Gemini AI models because it could not supply as much computing capacity as Meta wanted, the Financial Times reported on June 28, citing three people familiar with the matter.
Google told Meta around March that it could not meet the full Gemini capacity the company sought to buy through cloud and API services, and the shortfall disrupted and delayed some of Meta’s internal AI projects. The restrictions remain in place. Reuters said it could not independently verify the report, and both companies declined to comment.
Several other Google clients were affected by the same capacity limits, but Meta was hit hardest because of its unusually high demand. In response, Meta has told employees to make more efficient use of AI tokens as part of a wider push to control compute costs. Meta had leaned on Gemini for tasks including coding, customer service, advertising tools and content moderation, in some cases because it performed better than Meta’s own Llama models. The episode is an awkward one, since it leaves Meta dependent on, and now constrained by, a direct competitor for capabilities it considers critical.
The disclosure is a vivid example of a problem running through the whole industry: demand for AI computing is outpacing even the most aggressive spending on chips and data centers. The constraint is showing up in Google’s own results. Revenue at Google Cloud grew 63% to $20 billion in the first quarter, but CEO Sundar Pichai said limited computing power held back even faster growth and helped push the unit’s order backlog to about $460 billion, nearly double the prior quarter.
The Bigger Squeeze
No company is immune, including the largest suppliers. Google has been so constrained that it agreed to pay SpaceX about $920 million a month for access to roughly 110,000 Nvidia chips as what it called bridge capacity, and Anthropic has rented data center space from SpaceX.
When a hyperscaler that sells compute is also buying it from a rocket company, the scarcity is structural rather than temporary. For customers, the lesson is that access to frontier models can be throttled with little notice, turning a cloud contract into a potential single point of failure for core operations.
Meta’s Response
For Meta, the limits accelerate a shift it was already making: reducing reliance on outside models by building its own. The company has begun moving workloads to a new internal model called Muse Spark, developed under its Superintelligence Labs division, after a major restructuring of its AI group that included roughly 8,000 job cuts in May and the reassignment of thousands of workers to AI roles.
Meta has guided to capital spending of $115 billion to $135 billion in 2026 and pledged hundreds of billions toward US data centers through 2028. The Gemini squeeze underlines why: depending on a rival for content moderation and other essential systems is a vulnerability Meta is now spending heavily to remove.
