Google's TPU & Blackwell tests
- Google unveiled TPU v8t and v8i chips and expanded GB300 Blackwell GPU access for research on its AI Hypercomputer. - Early GB300 tests for Thinking Machines Lab showed roughly 2x improvements in training and serving speed versus prior setups. - The moves show hyperscalers pursuing their own silicon and mixed‑infrastructure strategies that affect enterprise AI economics. ( )
Google on April 22, 2026 announced two eighth‑generation TPUs — TPU v8t and v8i — and expanded GB300 (Blackwell) GPU access for Thinking Machines Lab. (blog.google) TPUs are Google‑designed tensor processing units built for large‑matrix math in model training and inference; Nvidia’s GB300 NVL72 is a rack‑scale Blackwell system that integrates 72 Blackwell GPUs and 36 NVIDIA Grace CPUs. (cloud.google.com) Google says TPU 8t is optimized for frontier training and can scale to 9,600 chips in a single superpod, while TPU 8i targets low‑latency agentic inference; Google and press materials say both chips will be available later this year. (cloud.google.com) Google Cloud awarded Thinking Machines Lab priority access to A4X Max VMs powered by GB300 NVL72, and early internal tests reported about a 2x improvement in training and serving speed versus prior‑generation GPUs. (googlecloudpresscorner.com) A4X Max combines GB300 NVL72 racks (72 GPUs per NVL72, multi‑terabyte HBM and high NVLink bandwidth) with Google’s Jupiter network fabric and Titanium ML adapter to speed weight transfers and scale large‑model workloads. (cloud.google.com) Analysts and coverage note this is part of a broader hyperscaler trend: build custom silicon for some workloads while offering Nvidia Blackwell access for others, a mix that VentureBeat says can create a “structural cost advantage” over shops that buy only Nvidia hardware. (venturebeat.com) Google executives, including SVP Amin Vahdat, framed the two‑chip roadmap as a deliberate 2024 decision to separate training and inference needs; other firms continue to purchase GB300 hardware directly from Nvidia and partners. (venturebeat.com) Both Google’s TPU 8t/8i lineup and expanded GB300 access will begin rolling out to customers later in 2026, leaving enterprises to weigh Google’s TPU‑based AI Hypercomputer against GB300‑powered A4X Max instances as each option becomes generally available. (blog.google)
