Google Capped Meta's Gemini Access Because It Ran Out of Compute
Google restricted Meta's Gemini API usage due to compute constraints, forcing staff to ration AI tokens and highlighting the supply crunch hitting enterprise AI.
Around March 2026, Google told Meta it could not supply all the Gemini capacity Meta wanted to buy. As of late June, those restrictions are still in place, and some of Meta’s internal AI projects have been delayed as a result.
This is not a minor billing dispute or a policy technicality. Google, one of the largest AI infrastructure operators on the planet, looked at one of its biggest customers and said: we simply cannot give you what you need. That tells you something important about where the enterprise AI market actually is right now.
What Meta Was Using Gemini For
Meta had been purchasing access to Google’s Gemini models through cloud and API services for internal use across a range of workloads. Content moderation was a core application, specifically automating the detection of scams and harmful content at the scale Meta operates. Gemini was also running customer service and advertising support chatbots, plus internal code development workflows.
The reason Meta turned to Gemini in the first place is telling: its own open-source Llama models were not performing well enough for these workloads. So even as Meta publicly champions open-source AI and invests heavily in its own model development, it was quietly buying significant amounts of capacity from a direct competitor.
When Google applied the restrictions, Meta’s response was to direct employees to be more token-efficient. Tokens are the unit of measure for how much text an AI model processes, both input and output. Asking your staff to use fewer tokens is the AI equivalent of telling people to stop leaving the lights on. It works, but it is not a strategy.
Why Google Is Supply-Constrained at This Scale
The uncomfortable reality is that even Google cannot keep up with demand for its own models. Gemini API requests more than doubled between March and August 2025. Google Cloud revenue hit $20 billion in Q1 2025, but CEO Sundar Pichai acknowledged on earnings calls that capacity constraints were actively holding back growth and swelling the division’s backlog.
The clearest signal of how acute the problem is came earlier this month, when Google signed a $920 million per month agreement with SpaceX for access to 110,000 Nvidia GPUs. Google described this as “bridge capacity.” To be paying nearly a billion dollars a month to rent GPUs from Elon Musk’s rocket company, while simultaneously capping usage for enterprise customers, gives you a sense of the gap between current supply and what customers are actually trying to consume.
Google also implemented compute-quota-based usage limits across its Gemini applications from May 17, 2026, moving customers onto a rolling weekly usage window rather than unlimited access. The shift from unlimited to rationed is a meaningful change in how enterprise AI contracts are going to work.
What This Means for Anyone Building on Third-Party AI APIs
This is the part that should get your attention if your organisation uses AI APIs from any major provider.
Meta is not a small company testing a proof of concept. It has a 2026 capex budget of $115 to $135 billion, has reassigned 7,000 workers to AI-focused roles, and cut 8,000 jobs partly to redirect resources toward AI infrastructure. If Meta cannot get the Gemini capacity it needs, the implicit assumption that enterprise AI supply scales cleanly with demand deserves serious scrutiny.
The risk for businesses building products or internal tooling on top of third-party AI APIs is vendor concentration. When the vendor is itself supply-constrained, your reliability is contingent on where you sit in their allocation queue. For Meta, the answer to that question turned out to be uncomfortable.
A useful contrast is Apple’s arrangement with Google. In January 2026, Apple struck a multi-year deal worth roughly $1 billion a year for a custom version of Gemini to power its overhauled Siri, with the model weights running on Apple’s own Private Cloud Compute infrastructure. That is a fundamentally different kind of dependency than calling an external API and hoping the capacity is there.
Meta’s Longer-Term Response
The restrictions are accelerating a shift Meta was already making. In April 2026, Meta Superintelligence Labs, led by Alexandr Wang, unveiled Muse Spark, Meta’s first proprietary frontier model. It is a multimodal system with reasoning and multi-agent orchestration capabilities, and it now powers Meta AI across WhatsApp, Instagram, Facebook, and Messenger.
Interestingly, Muse Spark shows meaningful token efficiency in benchmarks, using 58 million output tokens for a full Intelligence Index evaluation. That compares to 157 million for Claude Opus 4.6 and 120 million for GPT-5.4. Given that Meta has been actively telling staff to reduce token consumption, building a more efficient proprietary model is a direct operational response to the Gemini constraints.
The irony is that Meta still needed to buy external capacity at all. Its own model development, impressive as it is, could not keep pace with the inference demand its internal teams were generating. That gap is what made the Gemini dependency necessary, and it is what makes the supply restriction genuinely disruptive rather than just a minor inconvenience.
The Bigger Picture
The Gemini capacity crunch affecting Meta is the most visible example yet of AI providers being forced to make hard allocation decisions among paying enterprise customers. It will not be the last. Compute demand across the sector is consistently outpacing supply, even as companies commit billions to new data centres and chip procurement.
For enterprises planning AI infrastructure strategy, a few practical takeaways stand out. Assuming unlimited API scalability is not safe planning. Diversifying across providers reduces the risk that one vendor’s supply constraints become your product’s reliability problem. And where workloads are critical enough, there is a strong argument for building or licensing model capacity you actually control, rather than renting it by the token from a competitor who may be juggling dozens of other large customers ahead of you.
The fact that Google had to cap Meta of all companies is not embarrassing for either party exactly. It is just an accurate reflection of where AI infrastructure supply stands right now, and it is worth building your plans around that reality.