Google Warns AI Compute Is Sold Out Into 2027

Alphabet chief executive Sundar Pichai said on the Big Technology podcast this week that the AI industry cannot secure enough computing capacity to meet demand, and that Google itself would struggle to serve a customer asking for compute in 2027. In his words, everyone is short, not just us, and everyone wishes they had twice what they planned for. He noted that the capacity arriving today reflects investment decisions taken in 2024, which is why a 2027 request is, in his phrasing, incredibly difficult to fulfil.

Pichai pushed back on comparisons with the fibre overbuild of two decades ago. Compute is already being used, he argued, and Google is capacity constrained on a daily basis, with long term customer contracts carrying committed revenue behind the build. That framing matters to suppliers because it describes a demand curve anchored in signed contracts rather than in forecasts, which is the difference between a capex cycle that extends and one that gets cancelled.

The same week brought two hard infrastructure announcements. On 22 October Google expanded its agreement with Voltus to develop a 1 gigawatt virtual power plant across the PJM grid region, up from an earlier 100 megawatt arrangement. It is not new generation. It aggregates batteries, smart thermostats and industrial loads so that demand can flex, which is a workaround for interconnection queues that now run for years. Google also confirmed Project Suncatcher, a two to three year research effort with Planet Labs and SpaceX to test solar powered satellites carrying TPUs in sun synchronous orbit.

On the software side, Google open sourced TorchTPU, a native PyTorch path for its tensor processing units, built with Meta, Hugging Face, the vLLM team at Inferact and the SGLang team at RadixArk. Amazon shipped a comparable TorchNeuron for Trainium a day earlier. Both projects exist to make non NVIDIA accelerators usable without rewriting code, and both expand the pool of buyers who can specify TPU or Trainium capacity instead of GPUs.

For procurement teams the message is that accelerator supply is diversifying while total compute stays scarce. Two practical consequences follow. A second source on AI silicon now means a software migration project with a real timeline, not a paper alternative, so start the port before you need it. And because power availability, not floor space, sets the schedule, the electrical and thermal content of each rack keeps rising: busbars, high current connectors, fan trays, cold plates and DC distribution all sit on the critical path. Buyers sourcing those parts should expect allocation language in contracts through 2027. Our component catalog and quote desk and PCBA assembly service teams can quote alternates and build options while you re-plan.