Google has placed limits on Meta’s access to its Gemini AI models after Meta requested more computing capacity than Google could provide. The restrictions took effect around March 2026. Several of Meta’s internal AI projects have been disrupted and delayed as a result.
Google informed Meta it would be unable to fulfil the full Gemini computing capacity the company had sought to purchase. Meta’s exceptionally high demand for access to Google’s AI models made it one of the hardest-hit customers, with the capacity shortfall disrupting several AI projects under development.
The constraints have also prompted Meta to urge employees to use AI resources more efficiently by reducing the consumption of AI tokens, the units used to measure usage of generative AI models.
Other Google Cloud customers have also faced similar, though less severe, capacity constraints.
Why Meta Was Using Google’s AI At All
The detail that raises immediate questions is why Meta — a company with its own AI research division, the open-source Llama model family and significant GPU infrastructure — was relying on a competitor’s AI models at scale.
Training your own models and having enough inference capacity to run every AI-powered feature across WhatsApp, Instagram, Facebook and the Meta AI assistant simultaneously are different problems at different scales. For at least some workloads, buying access to Gemini AI capacity made more sense than building additional in-house capacity.
Meta is not unusual in this regard. Across the industry, even the most well-resourced companies outsource portions of their AI inference workloads to third-party providers when in-house capacity runs short. The difference here is the scale of Meta’s appetite and the public visibility of the shortfall.
The Infrastructure Problem Behind The Story
Gemini API request volume doubled between March and August 2025, forcing Google to reassess how to allocate computing resources. Starting May 17, 2026, Google implemented compute-quota-based usage limits on Gemini applications, similar to mobile data plans, with users operating within weekly rolling windows subject to usage caps. These restrictions apply across all customers, not just Meta.
In the first quarter of fiscal 2026, Google Cloud reported revenue of $20 billion. However, Alphabet CEO Sundar Pichai noted at the time that constraints in computing capacity prevented Google Cloud from fulfilling additional customer demand, causing its order backlog to nearly double compared to the previous quarter.
The supply problem is physical. Building data centres takes years. Sourcing advanced AI chips — primarily Nvidia H100 and B200 GPUs — requires navigating supply chains that cannot be accelerated on demand. Power grid availability and electricity infrastructure add another constraint. The AI boom has triggered an unprecedented demand for data centre capacity, advanced chips and reliable power. The supply of physical infrastructure, including electricity grid availability and high-performance hardware, is struggling to keep up.
What This Signals For The Industry
The Google-Meta situation punctures the narrative that AI is an infinite, always-available resource. The assumption that cloud AI services will always scale — that you can simply throw more API calls at Gemini or GPT-4 and get results back in milliseconds — is being tested directly.
For businesses and developers building AI-powered products on top of third-party model APIs, the Google-Meta story is a practical warning. Capacity limits, usage caps and throttling are no longer hypothetical risks. They are happening to Meta. They can happen to anyone.
The AI race is no longer just a software competition. It is a physical infrastructure race, and the infrastructure is not keeping up.
