Google Launches Separate AI Chips for Training and Inference, Boosting Performance 2.8x

Gate News message, April 23 — Google announced on April 22 that it will release separate eighth-generation TPU chips for training and inference later this year, replacing its previous combined design. The move targets AI agent workloads and offers Google Cloud customers an alternative to Nvidia hardware.

The training chip delivers 2.8 times the performance of Google's seventh-generation Ironwood TPU at the same price, while the inference chip is 80% faster and features 384 MB of SRAM, triple the amount in Ironwood. The separation of training and inference capabilities reflects a shift in how companies optimize for different computational demands.

The initiative is backed by a long-term partnership with Broadcom and Anthropic. Anthropic plans to use approximately 3.5 gigawatts of TPU computing through Broadcom starting in 2027, with Broadcom handling chip manufacturing and networking components through 2031. Anthropic, the AI startup behind Claude, has seen annualized revenue recently exceed $30 billion. Meanwhile, Apple, Microsoft, Meta, and Amazon are also expanding custom AI chip efforts to reduce reliance on Nvidia, which remains the market leader.