Around March, Google informed Meta Platforms, Inc. that it could not supply the full amount of compute capacity Meta sought to purchase for access to the Google Gemini AI model. According to the Financial Times, which cited three employees familiar with the situation, these restrictions have delayed some of Meta's internal AI projects.
The limitations remain in effect, and facing those constraints and a need to rationalize AI spending, Meta has encouraged employees to use available AI resources more efficiently. The FT reported that Google implemented similar limits for other customers as well, though the degree of restriction varies by client.
The episode highlights that even the largest AI providers face hard limits on available computing resources. Despite multibillion-dollar investments in advanced chips, data center construction and power infrastructure, available compute still appears insufficient to meet the surge in AI demand.
Google seeks additional capacity
Sources told the Financial Times that earlier this month Google signed a $920 million contract for services offered by Elon Musk’s SpaceX to secure additional compute capacity. That amount is roughly equivalent to HUF 286.9 billion. Google did not respond to the FT’s inquiries about the reports.
In April, Google CEO Sundar Pichai said in the company’s first-quarter earnings call that cloud revenue had for the first time exceeded $20 billion, and that the backlog of signed but unfulfilled cloud contracts had nearly doubled from the prior quarter, topping about $460 billion. Pichai also warned then that the company faces short-term constraints in meeting demand.
Consequences for Meta
The restrictions underline how much Meta relies on rival models such as Gemini even as it spends aggressively to develop its own AI capabilities. CEO Mark Zuckerberg has previously invested billions to recruit engineering talent and secure infrastructure for what he has described as a form of "personal superintelligence."
Unlike Google, Meta does not have a commercial cloud business, and the company is racing to build its own capacity. Meta has committed to roughly $600 billion in investments in the United States through 2028.
According to the Financial Times, Meta initially chose Gemini because it outperformed the company’s Llama models. More recently, Meta has started prioritizing its new Muse Spark model, which is considered more competitive in some applications and may reduce the company’s exposure to external providers.
Conclusion
Google’s decision to limit access to Gemini compute, and its subsequent contract with SpaceX, illustrate the scale of infrastructural pressure created by rapidly growing AI demand. For Meta, the episode underscores the importance of expanding its own compute capacity and advancing its internal models to reduce dependence on external providers.
