Google pitches TPU 8t for superpod clusters linking thousands of chips for large‑scale training

Google’s new TPU story is really a cluster story. The chip matters, obviously, but the bigger pitch is the system around it — thousands of chips wired together so one training run can behave like a single giant machine. That is what Google rolled out at Cloud Next ’26 on April 22, when it introduced TPU 8t as the training half of its new eighth-generation TPU family. The point is simple: frontier models are now so big that a “fast chip” is not enough. You need a fast pod, a fast network, and enough shared memory that the whole thing does not stall halfway through training. (blog.google) ### What is TPU 8t, exactly? TPU 8t is Google’s new custom chip for model training, while TPU 8i is the sibling aimed at inference and post-training work. Google is explicitly splitting those jobs now instead of pretending one design should do everything well. That matters because pre-training giant m(blog.google)n the model” chip. (blog.google) ### Why is the superpod the real headline? Because Google is selling scale more than silicon. TPU 8t is designed to link up to 9,600 chips in a single superpod, with 2 petabytes of shared high-bandwidth memory. Google says that lets customers train the biggest models on one massive memory pool instead(blog.google)tween chips. (cloud.google.com) ### What numbers is Google putting on it? The headline numbers are aggressive. Google says TPU 8t delivers 3x the processing power of Ironwood, its prior generation, plus up to 2x more performance per watt. On the cloud product page, Google also says TPU 8t can deliver up to 2.7x better performance per dollar than Ironwoo(cloud.google.com)y, one is economics — but together they tell you the pitch. Faster runs, lower waste, better utilization. (blog.google) ### Why split training and inference now? Because the AI workload split has become too wide to ignore. Training wants giant synchronized clusters and huge memory pools. Inference wants lower latency, lower cost, and often different model shapes like sparse MoE serving. Google says those requirements have (blog.google)ead of one compromise. (blog.google) ### Is this a direct Nvidia replacement play? Not really — at least not in the broad “replace every GPU everywhere” sense. The way Google is framing TPU 8t points at flagship training runs, embedding-heavy workloads, and giant AI Hypercomputer deployments. This is a hyperscaler saying: if you are trai(blog.google)tch. (cloud.google.com) ### Who is this for? Cloud customers building very large models, plus Google’s own ecosystem. Google has already been leaning on TPUs as a differentiator for major AI partners, and the company says more than half of its machine-learning compute investment in 2026 is expected to go toward the Cloud business. That is the te(cloud.google.com)nternal infrastructure flexing. (blog.google) ### So what is the bottom line? TPU 8t is Google admitting that the important unit of competition is no longer the chip by itself. It is the superpod. If Google can really make 9,600-chip clusters usable, efficient, and available to customers, then TPU 8t is less a new processor than a bid to become the place where the biggest training jobs happen. (cloud.google.com)