Google limited Meta’s access to its Gemini AI models due to constraints in computing capacity, according to a report from the Financial Times. This restriction has significantly affected Meta, compelling the company to instruct employees to utilize AI tokens more efficiently. Meta is also shifting workloads from Gemini to its own Muse Spark model to reduce its reliance on external AI providers.
Meta had initially depended on Gemini for tasks such as content moderation and safety processes, owing to its superior performance compared to Meta’s Llama open-source models. With the capped access to Gemini, Meta is accelerating its transition to Muse Spark, which it launched under its Superintelligence Labs division. The adjustments signal Meta’s efforts to develop internal alternatives for essential workloads.
In response to growing demand for Gemini Enterprise, Google has paid SpaceX $920 million per month for access to 110,000 Nvidia GPUs, referred to as “bridge capacity.” This partnership underscores the compute shortages that are reshaping relationships in the tech industry. Despite owning a significant amount of AI infrastructure and projecting over $180 billion in capital expenditures for 2023, Google still cannot meet all client demands and is rationing access to its models.
Meta previously cut 8,000 jobs to focus on AI initiatives and has since reassigned 7,000 employees to roles concentrating on artificial intelligence. The restrictions on Gemini have pushed Meta to enhance its internal capabilities at a crucial time when demand for AI computing resources outpaces available infrastructure. Other companies, such as Anthropic, are similarly seeking resources from SpaceX to support their operations, highlighting a broader issue of supply constraints in the AI sector.
The current landscape reflects a significant bottleneck in the AI boom, where the growth in demand for computational power is outpacing infrastructure developments. This trend illustrates that the limitations faced by major companies in accessing AI models are not merely a result of algorithmic challenges, but stem from the physical infrastructure needed to support increasing consumption.
