Google reportedly taps AMD to design its next-generation TPU

SemiAnalysis reports Google is working with AMD on a 10th-generation TPU that could pair its tensor cores with on-package CPU cores — AMD's first foray into custom AI silicon, aimed at reinforcement-learning workloads.

Google is reportedly bringing in AMD to help design one of its upcoming 10th-generation TPUs, according to a SemiAnalysis research note cited by Tom's Hardware. If accurate, the move would mark AMD's first real involvement in a custom AI ASIC — a notable shift after nine TPU generations in which Broadcom served as Google's de facto silicon design partner.

The more interesting question is why Google would need AMD at all. Google already builds its own accelerator architecture, so SemiAnalysis analysts think the draw is AMD's CPU cores rather than its accelerator know-how. Google and its cloud customers are pushing for TPUs that place general-purpose CPU cores on the same package as the tensor cores, tuned for reinforcement-learning and agentic workloads — which demand far more surrounding general-purpose compute than conventional LLM training.

A hypothetical design could combine Google's TPU compute chiplets with AMD CPU and HBM in a single tightly integrated package, echoing AMD's Instinct MI300A data-center APU. Firstpost frames the report as a shift in Google's TPU strategy. For now the report remains speculative — but if it holds, the real story isn't another TPU. It's Google quietly betting on a CPU-heavy member of its chip family for the agentic era.