The competition in AI development has shifted: raw compute capacity is now as decisive as models' capabilities. Supporting that shift, Google has limited Meta's access to its Gemini artificial intelligence models after Meta requested more compute resources than Google could provide.
According to the Financial Times, Google informed Meta in March that it could not fully satisfy requests related to the Gemini models. The restrictions remain in place and have delayed several of Meta's in-house AI projects. The episode is notable because both Google and Meta spend tens of billions of dollars annually on data centers, AI chips and related infrastructure, yet even that investment has not kept pace with surging demand.
Meta uses the Gemini models across multiple internal systems: for fraud detection, automated removal of harmful content, customer-service chatbots, advertising systems and programming tasks. Meta originally opted for Google’s models because, in several areas, they outperformed Meta’s own open-source Llama model.
The Financial Times also notes that other corporate customers have experienced capacity limits from Google, although to a much lesser extent; Meta’s exceptionally large needs made it the most affected. Google’s management has acknowledged an overall shortage of compute capacity, and CEO Sundar Pichai has said Google Cloud revenue could have been higher if more infrastructure were available.
To address the shortfall, Google is actively trying to add data-center capacity: earlier this month it signed an agreement to rent additional compute from Elon Musk’s SpaceX worth $920 million per month (about 315 billion forints).
Because of these capacity constraints, Meta is increasingly relying on its own models. In recent months the company has prioritized its new Muse Spark model, which is regarded as a direct competitor to Google Gemini. CEO Mark Zuckerberg has previously announced that Meta will spend a total of $600 billion (roughly 205 trillion forints) on U.S. AI infrastructure by 2028 to reduce reliance on external providers and support next-generation AI with its own data-center network.
The case underscores that in the AI race, scalable and rapidly available compute has become a strategic bottleneck that directly affects product development and operations at major technology companies.
