The global race for generative AI is hitting a physical wall: hardware availability. According to the Financial Times, Google has placed limits on Meta's access to Gemini AI models, unable to provide the massive computing capacity requested by the social media giant. This rationing began around March and has affected several Google Cloud clients, though Meta was hit hardest due to its sheer volume of requests.
Operational Disruptions at Meta
Meta had been relying on Gemini to automate critical safety processes, including scam detection and the removal of harmful content, as Google's models outperformed Meta's own Llama systems in these areas. The restrictions have delayed internal AI projects and forced staff to use AI tokens more efficiently to stretch remaining capacity.The Pivot to Internal Infrastructure
This bottleneck has accelerated Meta's strategy to decouple from external AI providers. The company is shifting workloads to Muse Spark, a new internal model developed by its Superintelligence Labs. To support this independence, Meta is aggressively expanding its own infrastructure, with projected capex between $115 billion and $135 billion for 2026, including the reassignment of 7,000 workers to AI-focused roles.A Systemic Infrastructure Crisis
The situation reveals that even the world's largest hyperscalers are struggling. Despite spending over $180 billion this year, Google has resorted to paying SpaceX $920 million per month for access to 110,000 Nvidia GPUs, describing this as "bridge capacity" to meet the surging demand for Gemini Enterprise.The industry's primary bottleneck has shifted from algorithmic breakthroughs to physical infrastructure. As Western labs face pricing stress tests and capacity limits, the race is no longer just about who has the best model, but who owns the most silicon.