Google Unveils 8th-Generation TPU Chips for the Agentic Era
What Happened
Google introduced two new chips at Cloud Next: TPU 8t for training, with a 124% performance-per-watt improvement, and TPU 8i for inference, at 117%. Both are designed for "agentic" workloads involving long-running model calls. The chips anchor Google's infrastructure pitch against Nvidia and underpin Anthropic's recently announced multi-gigawatt compute commitment.
My Take
Nvidia's moat isn't silicon — it's CUDA and the developer lock-in built over 15 years. What changes today is that Anthropic and Google are together large enough to justify building a second software stack, and Broadcom is making custom chips viable for anyone with a billion-dollar chip budget. Nvidia will still dominate 2026 revenue, but the 2027 story is a two-vendor world. Investors holding only Nvidia should start hedging into Broadcom and Alphabet infrastructure exposure.
Read Original Source