Google has capped Meta’s access to Gemini, leaving the social media giant’s AI plumbing looking a bit wheezy.
The search outfit told Meta around March that it could not supply all the Gemini capacity the company wanted to buy.
According to the Financial Times three people familiar with the matter said the move disrupted and delayed some of Meta’s internal AI projects.
The restrictions remain in place, while Meta has been nudging staff to be more careful with AI tokens, the units used to measure AI usage.
Several people said the company is trying to cut waste as AI bills start looking less like innovation and more like a bonfire.
Other Google customers have been hit by the squeeze, though not as badly, according to one person familiar with the matter.
Meta appears to have taken the worst of it because its demand for Google’s models is unusually large.
The decision offers a rare peek at the plumbing problems behind the AI boom, where even the biggest outfits are running short of compute.
Despite spending tens of billions of dollars on chips, data centres and power, the industry still cannot feed every hungry chatbot, agent and coding assistant.
Google has been chasing extra capacity to keep up with large corporate customers such as Meta.
Earlier this month, Google signed a $920 million-a-month deal to lease computing capacity from SpaceX boss Elon Musk’s SpaceX.
Google and Meta declined to comment, which is corporate speak for hoping the problem wanders off before anyone asks harder questions.
At Google’s first-quarter earnings in April, Google chief executive Sundar Pichai said cloud revenue topped $20 billion for the first time.
He said Google’s signed but undelivered cloud contracts nearly doubled quarter-on-quarter to more than $460 billion.
Pichai said: “Obviously, we are compute-constrained in the near term, and as an example, our Cloud revenue would have been higher if we were able to meet the demand.”
The trouble is being driven by inference workloads, the compute needed to run AI models after they have been trained.
AI lab Anthropic, maker of the Claude chatbot, struck a similar capacity deal with SpaceX last month.
Meta’s reliance on Gemini shows how much it has leaned on rival models while trying to make its own AI kit look less second-division.
Meta chief executive Mark Zuckerberg has been throwing billions at talent and infrastructure to build what he calls “personal superintelligence”.
Unlike Google, Meta lacks a cloud business and is racing to build its data centre fleet for training and inference. The company has committed to investing $600 billion in the US by 2028 as part of that push.
Gemini has been used within Meta to automate safety work, including scam detection and takedowns of harmful content.
It has also been used for customer service, advertising help chatbots, internal workflows and coding, alongside models such as Anthropic’s Claude.
Apparently, Meta picked Gemini because it outperformed its own open-source Llama models. More recently, Meta has shifted attention to its Muse Spark model, which some staff view as more competitive with Gemini.
That would reduce Meta’s dependence on external models for some applications, provided the latest in-house miracle does not need another miracle to run.
