Google Cloud Debuts Two New AI Chips, Claims 3× Training Speed Over Prior TPUs
Google's new TPU 8t for training and TPU 8i for inference promise 3× faster training and 80% better price-performance vs. prior generation, supplementing — not replacing — Nvidia GPUs.
Google Cloud unveiled two new custom AI chips at Google Cloud Next on April 22: the TPU 8t, optimized for training large models, and the TPU 8i, tuned for inference workloads. The company claims both deliver up to 3× faster training and 80% better performance per dollar versus its previous TPU generation.
Cluster-scale design
A single cluster can connect more than one million TPUs, Google said, positioning the chips for frontier-scale workloads. The company also said it continues to supplement — not replace — Nvidia infrastructure, and will offer Nvidia's upcoming Vera Rubin chip to customers later in 2026. Google and Nvidia are jointly developing Falcon networking technology.
What Google did not say
Notably, Google did not publish direct benchmarks against Nvidia's current H200 or B200 GPUs; the performance claims compare only to Google's own prior TPU line. Pricing was not disclosed.
Part of a larger capex push
The launch is part of Google's broader $175–$185 billion 2026 capex push to build out AI capacity, announced at the same event. For hyperscalers, custom silicon is a margin lever and a hedge against Nvidia supply constraints.
Market implication
For Nvidia, Google's ecosystem positioning — neither pure rival nor pure customer — suggests the AI-chip market will stay multi-vendor rather than winner-take-all.
Want every AI × Web3 signal the moment it breaks? Subscribe to the BlockAI News daily brief.
