Google limits Meta's access to Gemini amid compute shortage

Google has restricted Meta's access to its Gemini AI models. The company can't keep up with demand for compute and is imposing quotas on enterprise customers.

What happened

On Sunday, the Financial Times reported that Google had sent Gemini usage limits to a number of enterprise customers. Several large companies were affected, but Meta took the biggest hit: it had requested far more compute than Google was willing to allocate.

One detail stands out: Meta is a paying enterprise customer of Google, and even that didn't help. When a large enterprise buyer with budget to spend can't get the resources it needs, that shows how severe the shortage is. The constraint isn't economic but physical: there simply aren't enough servers, chips, and data center capacity.

This comes amid record growth in demand for AI inference. As an infrastructure provider, Google has to divide a limited resource among thousands of customers.

How it hit Meta

The restrictions are already affecting Meta's internal AI projects. Meta runs one of the largest AI programs of any tech company, with AI built into Facebook, Instagram, WhatsApp, and Ray-Ban Meta smart glasses. Meta has repeatedly said publicly that it plans to put AI into all of its products, and a disruption in compute supply directly threatens that strategy.

The blow hurts more because Meta is also developing its own family of open Llama models, which need compute for training and inference too:

• Delays in AI features could affect billions of Facebook, Instagram, and WhatsApp users
• Training new Llama versions requires enormous amounts of compute
• Meta's internal AI deadlines are at risk of slipping
• Dependence on an outside provider is becoming a strategic vulnerability

Why Google is short on compute

Demand for AI inference has surged over the past two years. Each new generation of models is much more capable than the last, but also much more resource-hungry: today's flagship models use several times more compute per request than their 2023 predecessors.

Google is building new data centers and ramping up production of its own TPU chips, but its infrastructure isn't keeping pace with demand.

Making matters worse, Gemini is now deeply integrated not only into Google's enterprise products for business customers but also into Google's own services: Workspace, Search, and Android. The company's internal needs compete with external customers' requests for the same physical resources.

What it means

Compute scarcity is becoming a new strategic bottleneck in the AI race. Companies without their own infrastructure risk ending up at the back of the line, even if they have the budget.

For Meta, it's another signal to speed up investment in its own data centers. For the market as a whole, it's a reminder that the fight for AI compute over the next few years will be just as fierce as the fight over the models themselves.

*Meta is designated an extremist organization and banned in Russia.