Google Cloud Launches Eighth-Generation TPUs Split into Training and Inference Chips

Google Cloud announced its eighth generation of tensor processing units on Wednesday, splitting the chips into TPU 8t for model training and TPU 8i for inference. The TPU 8t provides up to 3x faster AI model training compared to previous generations, while the TPU 8i delivers 80% better performance per dollar. Inference involves the ongoing usage of models after users submit prompts.

These advancements enable more than 1 million TPUs to work together in a single cluster. Google's TPUs are custom low-power chips originally named Tensor. The company uses its TPUs to supplement Nvidia-based systems in its cloud infrastructure.

Google promises Nvidia’s Vera Rubin chip will be available in its cloud later this year. Google agreed to work with Nvidia to engineer computer networking for Nvidia-based systems in its cloud. The two companies are working to improve the software-based networking tech called Falcon.

Google created and open-sourced Falcon in 2023 under the Open Compute Project. Nvidia has a nearly $5 trillion market cap. @techcrunch reported these details on the TPUs and partnerships. Separately, TechCrunch Disrupt 2026 will occur in San Francisco, CA, from October 13-15, 2026.

The event will gather 10,000+ founders, investors, and tech leaders for three days of 250+ tactical sessions. Registration for TechCrunch Disrupt 2026 saves up to $410. The StrictlyVC event will occur on April 30 in San Francisco, CA.

Julie Bort is the Startups/Venture Desk editor for TechCrunch.