Google reportedly told Meta in March that it could not supply the full volume of Gemini compute the social media company had requested, a shortfall that set back some of Meta’s internal AI projects and prompted Meta to push staff toward more careful use of AI tokens, according to the Financial Times.
The FT, citing people familiar with the matter, said the constraint affected other Google clients as well, though none as sharply as Meta. Google and Meta had not responded to the newspaper before publication.
The reported dynamic matters beyond its headline awkwardness. When one frontier company becomes a meaningful compute customer of a rival frontier company, the supplier holds an instrument that is, in practice, a throttle on its competitor’s internal AI roadmap. Whether Google intended this outcome is separate from the fact that it occurred: a capacity ceiling set by Google produced a delay in Meta’s AI development timeline.
Google Cloud’s compute constraints are not a secret. In the first quarter of 2026, Google Cloud revenue reached $20 billion, but CEO Sundar Pichai flagged on the earnings call that compute limitations had held back even faster growth, with the unit’s backlog nearly doubling quarter over quarter. That backlog represents demand Google cannot yet serve. Meta, apparently, was somewhere in that queue.
The episode surfaces a structural tension that has not been widely discussed. Several frontier AI companies, including Meta, are simultaneously building their own foundation models and purchasing capacity from competitors to run workloads that their own infrastructure cannot yet handle at scale. This is not unusual in enterprise cloud, but it becomes complicated when the supplier is also a rival in the same model market. A cloud provider rationing capacity to a customer who is also a direct competitor is a scenario that regulators focused on AI concentration will likely want to understand.
Meta’s response, reportedly encouraging staff to use AI tokens more efficiently, is a reasonable operational adjustment, but it is also a signal of how tight the compute market remains even for companies spending billions on their own chips and data centers. Efficiency pressure on internal tooling at a company of Meta’s scale suggests the shortfall was not trivial.
The reported compute ceiling also raises a question the FT story does not resolve: whether Meta’s demand for Gemini capacity reflects a gap in Meta’s own infrastructure, a deliberate multi-vendor strategy, or both. Meta has invested heavily in custom silicon and large-scale data center buildouts. The fact that it reportedly sought Gemini capacity at a volume Google could not fill suggests demand is running faster than any single company’s ability to build.
For teams at companies that rely on third-party frontier APIs for production workloads, this incident is a reminder that compute availability is a business risk, not just a technical one. Supplier-side constraints can ripple through a customer’s roadmap with no warning. Diversifying across providers, or maintaining fallback capacity, is now a legitimate risk-management question to put in front of engineering leadership before it becomes a project delay.
Reported by the Financial Times, via CNBC, on June 28, 2026.
