tweak logo

Google TPU v8: Significantly faster AI chips

Skærmbillede 2026-04-23 084851
Translate from : Google TPU v8: Markant hurtigere AI-chips
At its Cloud Next conference, Google unveiled the 8th generation of its Tensor Processing Unit (TPU). The new chips are the heart of Google's supercomputers and are built for one thing: making training and using large AI models both faster and cheaper.
By Kenneth Johansen
23 apr. 2026 08:49 am
f

Computing power for training and interaction

The new TPU v8 comes in two versions, each with their own focus area:

  • TPU 8t (Training): Optimized for training the really heavy models, Google promises it can shave development time from months to weeks, which is a huge leap for companies building their own AI systems.
  • TPU 8i (Interaction): Acts as the engine that delivers answers to users. With three times more memory than its predecessor, it can handle complex instructions with almost no latency.

Overall, the new generation is 80% faster in terms of performance-per-dollar and twice as energy efficient as the previous model.


From simple chatbots to autonomous agents

In addition to the hardware, Google also unveiled a new platform to help companies build "AI agents." The vision is to move away from simple chatbots and instead create agents that can independently reason and perform tasks to create concrete results.


An open platform for managing AI

The new platform acts as an all-in-one control center. Here, companies can build and manage their agents across different models. While Google is of course pushing their own Gemini models, the platform is open, so you can also integrate external models such as Claude. This should make it easier for organizations to maintain an overview of their AI systems as they become more complex.

The new TPU chips are expected to be rolled out to enterprises later this year.

Latest hardware

Most read hardware

Latest hardware

Our Partners