Google Caps Meta’s Gemini Access, Forcing $600B Infrastructure Bet

Published by James Harris on June 29, 2026June 29, 2026

Google capped Meta’s access to its Gemini AI platform around March after Meta’s compute demands exceeded what Google Cloud could supply, forcing Meta to accelerate development of an internal model and shift workloads accordingly. For a company running content moderation, advertiser chatbots, scam detection, and coding assistance through a competitor’s infrastructure, that cap wasn’t a minor inconvenience.

Meta had been using Gemini because it outperformed its own open-source Llama models on specific tasks, which is a quietly uncomfortable admission for a company that has staked considerable credibility on open-source AI leadership. It also uses Anthropic’s Claude for similar workloads, meaning Meta’s internal AI operations depend on two rival foundation model providers simultaneously. Following the restrictions, Meta told employees to use AI tokens more efficiently and accelerated development of Muse Spark, an internal model built under its Superintelligence Labs division. The company has pledged $600 billion in cloud computing investments over the next two years specifically to reduce this kind of dependency, and the Gemini cap makes clear why that timeline feels urgent. Meta cut 8,000 jobs in May and reassigned 7,000 workers to AI roles, so the internal pressure to make Muse Spark viable is now organizational, not just technical.

The constraint here isn’t software or talent. It’s GPUs, cooling, and power, and no amount of engineering headcount fixes that in the short term.

Google Cloud posted $20 billion in Q1 revenue, but CEO Sundar Pichai acknowledged that compute constraints caused the unit’s backlog to double last quarter, which means even Google cannot build fast enough to serve its own customers. The pattern of enterprises exhausting AI capacity faster than providers can supply it is becoming a structural feature of this cycle, not an anomaly. Google has separately agreed to pay SpaceX $920 million per month for access to 110,000 Nvidia GPUs as bridge capacity, and Anthropic is renting an entire SpaceX data center for similar reasons. The companies nominally competing to build the most capable AI models are all, at the infrastructure layer, in the same supply queue.

The Cloud Layer Gets More Complicated

For cloud providers trying to position themselves in a tightening regulatory environment, the capacity crunch adds another variable. Google Cloud’s competitive standing relative to AWS and Azure is already under scrutiny on regulatory grounds in Europe, and now its largest customers are being told their demand cannot be met. A provider that cannot fulfill enterprise-scale orders loses negotiating leverage precisely when the market is deciding which two or three cloud relationships will anchor the next five years of AI infrastructure spending. Meta’s push toward Muse Spark and internal capacity is the most visible signal yet that depending entirely on external model providers is a strategic liability even the most well-resourced companies are now actively working to eliminate.