Google ‘caps Meta’s Gemini use’ as compute shortage bites

Meta’s usage of Google’s Gemini AI models has been restricted after it requested more compute than the tech giant could provide, the Financial Times has reported.

Citing three people familiar with the matter, the paper reported that Google told the social media and AI conglomerate around March that it was unable to provide all of the Gemini capacity it wished to purchase.

This has disrupted and delayed some of Meta’s projects, according to the people, a sign that even the largest companies in the world are not immune to the ongoing shortage of RAM, GPU and CPU capacity.

The restriction has reportedly led Meta to encourage staff to be more efficient with their token usage, two months after taking down an employee-created internal leaderboard that recorded over 60 trillion tokens over a one month period.

Several other Gemini clients have also been affected by the restrictions, according to one person familiar with the matter, but Meta has found itself particularly impacted due to its exceptionally high demand.

The FT reported that as a direct result of the increased demand, particularly from large customers such as Meta, Google has moved swiftly to secure increased capacity. Earlier in the month, it signed a $920-million-a-month contract with SpaceX to lease compute capacity from its tech rival.

In April, Google parent company Alphabet reported that Google Cloud had a backlog of demand for its cloud services, including Gemini, that reached $460 billion.

Other major AI companies are also grappling with compute constraints, and searching for ways to solve them. In May, Anthropic signed a contract with SpaceX worth $1.25 billion a month to buy compute power from its Colossus datacentre.