Google restricts Meta's access to Gemini over compute shortage

Google has restricted Meta Platforms' access to its Gemini AI models. The cause is a compute shortage: Google could not allocate the capacity Meta requested. The Financial Times reports this, citing its own sources.

What happened

According to the Financial Times, Google has set strict limits on Meta's use of Gemini models. The restrictions are not tied to a commercial dispute or a breach of contract. The reason is purely technical: Google simply does not have enough computing infrastructure to meet Meta's requests in full.

The details of the limits have not been disclosed. It is unknown how much capacity Meta requested or how significant the cut turned out to be. Bloomberg, citing the FT, only confirms the fact that restrictions were introduced.

The situation looks paradoxical: one of the world's largest AI providers has to curb access for another tech giant, and not for strategic reasons but because of a basic lack of resources. It illustrates how quickly demand for compute has grown over the past two years, and how far the manufacturing base has fallen behind that growth.

GPU shortage as a systemic problem

The episode is a symptom of a structural crisis across the AI industry. Several key factors stand out:

• Demand for GPUs for training and inference of AI models is growing exponentially, outpacing the production capacity of chip fabs
• Even the largest cloud providers face an acute shortage of NVIDIA H100 and H200 accelerators
• Delivery queues for server hardware at leading manufacturers stretch for many months
• The constraints affect small startups and enterprise customers with multibillion-dollar contracts alike
• Companies increasingly compete not only on model quality but on access to basic infrastructure

It is telling that this comes amid record investment in AI infrastructure. Google, Microsoft and Amazon have announced combined spending of hundreds of billions of dollars on new data centers. But the fruits of that investment have not yet materialized as usable capacity: building and commissioning a large data center takes from two to four years.

Meta is developing its own AI chips, MTIA (Meta Training and Inference Accelerator), and is expanding its manufacturing cooperation with TSMC. The company has also released the Llama series under a conditionally open license, which lets partners run AI inference on their own hardware and reduce dependence on outside providers. Google's restrictions only reinforce this in-house course.

Google, Meta and a complicated interdependence

Google and Meta are both partners and competitors. Meta uses Google's cloud infrastructure for part of its computing workloads, yet it also competes with Google for the advertising market, for users of AI services and for the developer ecosystem.

For Google, access to the Gemini API is a strategically important source of B2B revenue and a tool for positioning Gemini as the enterprise standard. News that the company could not supply the needed volumes to one of its largest customers can potentially damage its reputation as a reliable provider, especially as Microsoft Azure OpenAI Services and Amazon Bedrock push aggressively, actively poaching enterprise customers by betting on guaranteed scalability.

The situation also raises the question of the limits of Google's B2B strategy in AI. If even Meta, with its negotiating leverage and scale, could not get the volumes it needed, that is a signal to thousands of smaller customers across the market.

What this means

The compute shortage is turning into a strategic resource that matters as much as the quality of the AI models themselves. Companies that manage to build their own infrastructure or lock in priority access to new production capacity will gain a long-term competitive advantage, regardless of the quality of their algorithms. The AI race is increasingly a race for hardware.

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