Google launched two new AI chips built for AI agents, the bots that carry out tasks on their own, bringing its total to four new chip systems in a single year as it races to power the next wave of AI.

Key Takeaways

  • Google added TPU 8i Max and TPU 8tx to its eighth generation of TPU chips
  • TPU 8i Max is built for very long inputs and the most computing power per server rack
  • TPU 8tx is built to run AI models securely, keeping data protected while it is being processed
  • Google also made a feature that lets cloud customers reshape TPU clusters on the fly generally available, saying it cuts startup times by up to five times

Google announced the chips on X on Monday, saying the four systems give developers a purpose-built chip for every stage of an AI agent's work. TPUs, or tensor processing units, are the AI chips Google designs itself and rents out through Google Cloud.

Two Chips, Two Different Jobs

AI agents do more than answer a single question. They read long documents, call outside tools and work through many steps, which puts very different demands on hardware.

TPU 8i Max is aimed at the heavy end, handling huge amounts of text at once with as much computing power as Google can pack into a rack. TPU 8tx is aimed at banks, hospitals and governments that want to run AI on sensitive data, with protection built into the chip itself.

Also Read: Oracle Hands Google AI Chips To Cloud Buyers Eyeing Nvidia

Less Waiting, Less Wasted Chips

In a separate post on its cloud blog, Google said dynamic slicing for TPUs is now generally available on Google Kubernetes Engine, its service for running software across big clusters of servers.

Until now, customers had to set up a fixed group of chips for each job and tear it down and rebuild it when their needs changed. Dynamic slicing lets them join chips into bigger blocks or split them into smaller ones on demand.

Dynamic Slicing Google's Claim
Job startup time Up to 5x faster
Recovery after a failure 4.5x faster
Availability Generally available on GKE

Figures from Google Cloud's Nov. 9 blog post

For customers, that means less time waiting for hardware and fewer chips sitting idle. The next thing to watch is pricing for the two new systems, which Google has not yet shared.

Read Next: Google Sued After Gemini On School Laptop Helped Teen Plan Suicide

Originally published on Fathom News.