Google releases new chips for AI training and inference, competing with Nvidia
Make us a preferred source
Make us a preferred source on Google
The news: Google is launching a new generation of chips, producing two distinct processors which will handle inference work and AI model training separately, as it seeks to compete with market-leader Nvidia in AI hardware.
The context: A Google blog post on Wednesday explained that the release will be made for the eighth generation of its tensor processing units (TPUs), with both chip types to become available later in 2026.
“TPU 8t and TPU 8i. These two chips are designed to power our custom-built supercomputers, to drive everything from cutting-edge model training and agent development, to massive inference workloads,” said Amin Vahdat, CVP and chief technologist, AI and infrastructure at Google.
TPU 8t “shines at massive, compute-intensive training workloads designed with larger compute throughput and more scale-up bandwidth,” while TPU 8i has more memory bandwidth to serve the most latency-sensitive inference workloads.
With the rise of AI agents, Google determined individually specialised chips would better serve the needs of customers: “Importantly, both chips can run various workloads, but specialization unlocks significant efficiencies and gains,” Vahdat wrote.
Meanwhile, at the company’s annual conference in Las Vegas, Google unveiled a slew of tools to help companies build AI agents to automate tasks.
“This isn’t about offering individual services that can be cobbled together; it is about providing a comprehensive backbone for innovation,” Google Cloud CEO Thomas Kurian said in a blog post. He said that Google’s full stack approach across AI infrastructure; leading models like Gemini; data management capabilities; enterprise-grade foundation; developer tools and platform; and agents and applications “make us unique in the market.”
The sources: Amin Vahdat Google blog post, Thomas Kurian Google blog post
