Google quietly capped Meta’s access to its Gemini models around March, the Financial Times reported Sunday, after one of the world’s largest AI buyers asked for more compute than the world’s largest AI seller could deliver. The cap is reportedly still in place. Both companies declined to comment.

The detail that matters isn’t the rationing. It’s why Meta was buying Gemini at all. According to the FT, Meta had quietly shifted safety workloads, content removal, scam detection, onto Gemini because it outperformed Meta’s own Llama models. The company doing the largest open-source AI push on earth was running its trust and safety stack on a competitor’s API.

Then it ran the bill up. Tech Brew reports Meta employees burned through 60 trillion tokens in a 30-day window, a consumption pattern staff internally christened “tokenmaxxing.” Once the cap landed, Meta urged staff to use tokens more sparingly and accelerated work on Muse Spark, the in-house model launched under its Superintelligence Labs division.

Google’s predicament is stranger. Even as cloud revenue topped $20 billion for the first time and the contract backlog nearly doubled to more than $460 billion, Sundar Pichai told investors the company was “compute-constrained in the near term.” The fix arrived in the form Silicon Valley reserves for true emergencies: Google agreed to pay SpaceX $920 million a month for 110,000 Nvidia GPUs, characterized internally as “bridge capacity” for Gemini Enterprise demand. Alphabet is renting silicon from a rocket company.

The capex arms race tells the rest. Google raised 2026 guidance to $180–$190 billion and expects 2027 to “significantly increase.” Meta guided $115–$135 billion, cut 8,000 jobs in May, and reassigned 7,000 workers to AI roles. A JPMorgan analysis cited by the Wall Street Journal found that over 60% of data centers slated for 2027 completion hadn’t broken ground as of May, blocked by power, permitting, and supply chains.

The hyperscalers spent a decade selling abundance as the product. They’re now rationing it to each other.

Sources