Google goes head-to-head with Nvidia in chip wars

Google is not content just renting out its tensor processing units from Google Cloud. It now wants those AI chips sitting inside other people’s data centres, right next to the Nvidia hardware they are supposed to replace.

According to talks described in The Information, Google has been pitching Meta and big financial institutions on a plan called TPU@Premises. The idea is simple enough.

Instead of only accessing TPUs via Google Cloud, companies could buy racks of them and run the chips on site, using Google’s new TPU Command Center software to manage clusters.

Meta is said to be considering spending billions of dollars on TPUs for its own data centers in 2027, as well as renting TPUs from Google Cloud next year.

If it lands, that would be a big upgrade from Google’s current sideline as “the other AI chip in the cloud.” Inside Google Cloud, TPUs have mostly been a cheaper alternative to Nvidia GPUs.

Now Google is trying to turn that into a standalone chip business. Internally, some cloud leaders have talked about grabbing 10 per cent of Nvidia’s annual revenue with TPUs. Given Nvidia’s current size, that would work out to billions of dollars a year.