Google has limited the amount of its Gemini AI models that Meta can use, telling the social-media company it could not meet the full scale of computing capacity Meta had requested, according to a Financial Times report described by CNBC.
What was reported
Google informed Meta as early as March that it was unable to provide all the capacity Meta sought to access Gemini, and the shortfall disrupted and delayed some of Meta's internal AI work, the FT reported, citing people familiar with the matter. The constraint was significant enough that Meta urged its own staff to be more sparing in their use of "tokens" — the units that measure how much computing an AI model consumes. The report is based on sources rather than official disclosures, and neither Google nor Meta confirmed the details publicly.
Rivals who are also customers
The episode underscores an unusual feature of the AI business: companies that compete fiercely on models also buy access to one another's technology. Google's Gemini and Meta's open-weight Llama models are direct rivals, yet Meta appears to have drawn on Gemini for certain workloads — the kind of arrangement that becomes awkward when the supplier is also a competitor and capacity runs short. Google sells access to Gemini through its cloud business, putting it in the position of rationing a scarce resource between outside customers and its own products.
A compute crunch across the industry
Underlying the dispute is a broader shortage of AI computing power. Demand for the specialized chips and data centers that train and run large models has outrun supply, despite the tens of billions of dollars that Google, Meta, Microsoft, Amazon and others are pouring into new facilities. Power availability, chip supply and the sheer growth of AI workloads have all become bottlenecks, leaving even the largest firms managing limited capacity. Meta has been expanding its own infrastructure and AI hiring aggressively as it tries to reduce its reliance on others.
Why it matters
For Meta, the limit is a reminder that building its own models and data centers does not fully insulate it from depending on rivals when demand spikes. For Google, prioritizing its own services and paying cloud customers reflects the hard choices that scarcity forces. And for the wider industry, the standoff highlights that access to compute — not just talent or model quality — has become one of the decisive constraints in the race to deploy artificial intelligence. As with any single news report citing unnamed sources, the precise terms of the arrangement may become clearer if the companies comment.
