Google has restricted Meta’s use of its Gemini models because it does not have sufficient computing power available. The move is slowing down internal AI projects at Meta and demonstrates that even the largest tech companies are reaching the limits of their AI infrastructure.
According to the Financial Times, Google informed Meta around March that it could not provide the full capacity requested for Gemini. The restrictions are still in place and have forced Meta to postpone or adjust some AI projects. In addition, the company is said to have asked employees to use AI tokens—the unit by which the use of AI models is measured—more sparingly.
Meta is not the only one to have faced these limitations. According to insiders, other Google customers are also feeling the effects of the shortage, although the impact is less severe for them. Meta stands out due to the exceptionally high demand for Gemini capacity.
The situation makes it clear that the massive investments in AI infrastructure still cannot keep up with demand. Major tech companies are pouring tens of billions of dollars into chips, data centers, and power supplies, but are seeing the need for computing power grow even faster. Inference—the execution of AI models after they have been trained—in particular is placing an ever-greater strain on available capacity.
Google is trying to alleviate that pressure by leasing additional infrastructure. Earlier this month, the company signed a deal worth $920 million per month with SpaceX for additional AI capacity. During the presentation of the quarterly results in April, CEO Sundar Pichai already indicated that Google is currently constrained by a lack of computing power. According to him, the cloud division could have generated more revenue if all customer demand could have been met immediately.
Muse Spark Aims to Reduce Dependency
Meta’s reliance on Gemini is striking. Although the company is investing billions to secure a leading position in AI, it relies heavily on a model from a direct competitor for its internal operations. Gemini is used for tasks including fraud detection, moderation of harmful content, customer service, advertising functionality, and programming tasks.
According to sources, Meta initially chose Gemini because it outperformed its own open-source models from the Llama family. The company is now reportedly switching to the new Muse Spark model more and more frequently. That model is viewed internally as a competitor to Gemini and is intended to reduce dependence on external AI providers.
At the same time, Meta is working on a major expansion of its own infrastructure. Unlike Google, the company does not have a public cloud environment on which its AI investments can run. That is why Meta is accelerating the construction of new data centers for both training and running AI models. To this end, a $600 billion investment program has been announced through 2028.
Google and Meta have not commented substantively on the reports. The capacity issues do, however, illustrate that the battle for AI is no longer just about the quality of models, but also about access to sufficient computing power. Even market leaders do not always have that immediately available.