What changed
Google announced Ironwood, its seventh-generation TPU, with configurations scaling to 9,216 chips and a stated 42.5 exaflops of FP8 compute per pod for demanding inference workloads.
Why it matters
The design reflected the industry's shift from training-only scaling toward serving reasoning, retrieval and agent workloads at fleet scale.
What to do
Benchmark price, latency, throughput and portability on representative serving traces once capacity is available, not on peak pod specifications.
Caveat
The figures were Google specifications and roadmap claims; they did not establish application performance or customer availability at announcement.