Meta might deploy Google TPUs from 2027, causing a decline in Nvidia’s stock value.
According to The Information, Meta wants to use Google’s new TPU in its own data centers. For Nvidia, which normally supplies to Meta, this is not pleasant. Especially now that the demand for Google’s Ironwood accelerators is growing strongly.
In response, Nvidia emphasizes that Google is making “great progress,” but also states that its own platform is “a generation ahead” and supports all AI models, regardless of workload or environment.
No Easy Switch to TPU Clusters
Even if Google were to sell TPUs, a switch for Meta would be anything but easy, writes The Register. TPU clusters use optical circuit switches, which are different from Nvidia and AMD-based GPU environments. PyTorch also doesn’t work on TPUs by default.
This makes it more likely that Meta is mainly talking about better inference optimization for Llama on Google hardware, not about full TPU adoption. Inference is less intensive, cheaper, and runs smoothly on rented TPUs.
Google TPUs Gaining Ground
Anthropic is employing up to one million Google TPUs for its next Claude generation, and Amazon is also investing in TPUs for Trainium. At the same time, the company signed a billion-dollar contract with Microsoft and Nvidia for Azure last week.