Google restricts Meta’s access to Gemini AI models as compute demand exceeds supply

Google has put a definite cap to how much of its Gemini AI platform Meta is allowed to make use of after the tech and media company’s demand for AI compute exceeded available capacity, according to a Financial Times report.

Alphabet’s cloud division informed Meta around March that its Gemini capacity demands could not be fulfilled, a situation that ended up delaying and disrupting some of Meta’s internal AI projects.

Meta and Gemini: why?

Meta had turned to Google’s Gemini models because they outperformed its own open-source Llama models for specific tasks. The company was using Gemini for content moderation processes that included harmful content removal and scam detection.

The AI model was also used for customer service automation, advertiser chatbots and coding. Engadget reported that Meta also uses Anthropic’s Claude for similar workloads.

Meta, which does not operate its own cloud business, has pledged $600 billion in cloud computing investments over the next two years to reduce that dependence.

Following the restrictions, Meta has instructed employees to use AI tokens more efficiently, according to FT. The company has also accelerated development of Muse Spark, an internal model built under its Superintelligence Labs division, and has started to shift workloads away from Gemini to this platform.

Meta cut 8,000 jobs in May and reassigned 7,000 workers to AI-focused roles.

AI industry in compute supply crisis

Several other Google customers are also facing reduced access, though to a lesser degree, according to the FT report. Even though Google operates one of the world’s largest AI infrastructure pools, the tech giant still cannot meet the massive demand for AI compute.

Google Cloud revenue reached $20 billion in Q1, but CEO Sundar Pichai acknowledged that compute constraints prevented higher growth and contributed to the cloud unit’s backlog doubling over the last quarter.

Google has also agreed to pay SpaceX $920 million per month for access to 110,000 Nvidia GPUs as “bridge capacity” for the Gemini Enterprise, according to Reuters.

Google’s competitor, Anthropic, is also separately renting an entire data center from SpaceX, proving that the AI industry’s biggest constraint currently is the physical infrastructure to run AI models and not the talent to build them.