Google rations Meta's access to Gemini: compute capacity, not talent, is the new bottleneck

By Financial Times · July 2, 2026. The original article is behind a paywall and lets little more than the headline through: Google has begun limiting Meta's use of Gemini because demand for artificial intelligence is straining its capacity. There are no figures, timelines or technical details available in the source, so any further precision would be speculation. The honest thing here is to acknowledge that the material is minimal and that the value of the story lies in the fact itself, not in a development we cannot verify.

And yet the fact itself already says a lot. That Meta depends on Gemini —the flagship model of a direct rival— for part of its AI infrastructure, and that Google has to ration that access, confirms something we have been pointing out for months: the sector's real scarcity is not in algorithms or datasets, but in the hardware and energy needed to train and infer at the scale current demand requires. The labs compete on benchmarks, but they compete even more fiercely for GPUs, TPUs, data centers and electricity contracts. When not even Google —one of the three or four players with the most in-house computing capacity on the planet, thanks to its TPUs— can meet all the external demand, it becomes clear that the physical bottleneck rules above any other strategic consideration.

This fits a thesis we have been developing: the AI war is fought less and less on the pure quality of the model and more and more on who controls the infrastructure, the distribution and the access to scarce resources. If Google decides to prioritize its own consumption or that of other clients over Meta's, it is exercising de facto power over a competitor that, paradoxically, needs its services. It is an uncomfortable dynamic of interdependence: the very companies competing for AI leadership also need one another to sustain their products, because building your own capacity from scratch —chips, plants, power grids— takes years and billions that not even the tech giants can improvise from one quarter to the next.

In the short term, this translates into very tangible frictions: products that do not scale at the promised pace, AI features restricted by quotas, and an infrastructure race that consumes capital at a rate that is starting to raise reasonable doubts about the sector's financial sustainability, something already openly discussed in investment circles. In the long term, however, this tension is also a sign that demand for AI is real and massive, not a marketing bubble: the world wants more artificial intelligence than can physically be produced today. That pressure, over time, is precisely what pushes for building more capacity —more efficient chips, cheaper energy, architectures that are less costly to run inference on— and that is, ultimately, the path toward the computational abundance that will make it possible for AI to stop being a rationed resource and become a utility accessible to everyone.