Google is working with AMD on a 10th-generation Tensor Processing Unit project, according to a client note from semiconductor research firm SemiAnalysis cited by Tom’s Hardware.
The note said “market chatter suggests” Google is working with AMD on a TPU project in the v10 generation. Neither Google nor AMD has publicly commented, and the scope of AMD’s role remains unclear.
The reported collaboration would mark AMD’s first major involvement in a custom AI accelerator project. Analysts said AMD’s appeal lies in its CPU intellectual property, advanced packaging expertise, and experience building hybrid designs that combine x86 processors with AI accelerators in a single package.
Google has developed nine generations of TPUs with Broadcom as the primary silicon designer. Google also has a long-term supply agreement with Broadcom that runs through at least 2031.
Analysts said that history makes it unlikely AMD would design a conventional training or inference TPU. Instead, Google appears to be exploring a variant that places general-purpose CPU cores alongside tensor compute chiplets in the same package.
SemiAnalysis said the design rationale is tied to reinforcement learning and agentic AI workloads, which require more general-purpose compute than standard large language model training. Google has already increased CPU resources in TPU 8i systems by pairing one Google Axion CPU with every two TPUs, double the ratio used with seventh-generation hardware.
SemiAnalysis said some configurations may require a 1:1 CPU-to-accelerator ratio. AMD’s Instinct MI300A already integrates x86 and accelerator chiplets in a single package.
Featured image credit
