Google Splits Its Eighth-Gen TPU in Two: One for Training, One for Reasoning

Google broke its own habit. At Google Cloud Next on April 22, 2026, the company unwrapped its eighth-generation Tensor Processing Units as two separate designs rather than a single catch-all chip — the first time any TPU generation has been split this way. The TPU 8t is built for large-scale pre-training; the TPU 8i is aimed at inference and high-concurrency reasoning workloads.

The 8t is the heavyweight. Google says it scales into superpods of up to 9,600 chips pulling 121 ExaFlops at FP4, with each individual chip delivering 12.6 FP4 PFLOPs. It pairs 216 GB of High Bandwidth Memory with 128 MB of SRAM.

The 8i trades raw scale for memory and cost. It carries 288 GB of HBM and 384 MB of SRAM — triple the SRAM of the previous generation, per Google — and computes at 10.1 FP4 PFLOPs per chip. Google's headline pitch: up to 80% better performance-per-dollar than Ironwood, the seventh-gen TPU that landed in 2025.

Both chips are built on TSMC's N3 process and pack Google's own Arm-based Axion CPUs, slotting into the company's wider AI Hypercomputer architecture. The 8i also gets a new Boardfly topology controlling how chips are wired to talk to each other, plus a CAE engine Google says lifts performance. On the software side, Google is pushing tighter hardware-tooling coordination through frameworks like JAX and Pathways.

The split tracks where AI has gone. Google framed the launch around the "agentic era," in which systems juggle many tasks at once instead of answering a single prompt — a workload that stresses hardware very differently than training does. General availability is expected later in 2026, with more technical details promised at Hot Chips 2026 in August.