Google splits its eighth-generation TPU into two chips for training and inference

Google has split its eighth-generation tensor processing unit into two distinct chips, one dedicated to AI training and the other to inference, reflecting the industry's shift toward specialized hardware for different stages of model development. The move acknowledges that training and inference workloads have diverged enough to warrant separate silicon rather than a single general-purpose accelerator. The split matters because hyperscaler chip design choices ripple through the AI compute supply chain, influencing demand for memory, networking and advanced packaging. It also intensifies competition with Nvidia and AMD, which have similarly segmented their roadmaps around training and inference products. What to watch next is whether Google discloses performance and efficiency figures for each variant and whether rival cloud providers follow with comparable splits.