Google TPU
Google's custom Tensor Processing Unit — the largest non-Nvidia AI training fleet
Google's TPU is the most mature custom AI accelerator, now in its Trillium (v6e) and Ironwood (v7) generations, powering Gemini training and Google Cloud AI workloads at massive scale via Optical Circuit Switch pods. TPUs are co-designed by Google and Broadcom and fabricated at TSMC, giving Google the strongest in-house alternative to Nvidia GPUs. In 2025 external demand grew as labs like Anthropic committed to large TPU capacity, making TPU a genuine competitive counterweight.
Latest gens
Trillium (v6e), Ironwood (v7)
Design partner
Broadcom (ASIC), fabbed at TSMC
Scale
OCS-connected pods, powers Gemini
External use
Anthropic + Google Cloud customers
How it fits the stack
Google TPU with what it depends on (above) and what it feeds (below). The figure renders as a crawlable diagram and upgrades to an interactive 3D graph as it scrolls into view.
Google TPU in the AI stack. Google TPU with its immediate upstream dependencies (top) and downstream dependents (bottom) in the AI value chain. Hover a node in 3D, or read the full relationships below.
Graph data (text) — 12 entities, 12 relationships
- Google TPU —uses→ High-Bandwidth Memory (HBM)
- Google TPU —used by→ JAX / XLA
- Google TPU —depends on→ MLIR / OpenXLA / StableHLO
- Google TPU —depends on→ TSMC (Taiwan Semiconductor Manufacturing Company)
- Anthropic —used by→ Google TPU
- Google Cloud (GCP) —used by→ Google TPU
- JAX —depends on→ Google TPU
- MLPerf (MLCommons) —uses→ Google TPU
- Waymo / autonomous driving —uses→ Google TPU
- Google TPU —partners with→ Broadcom
- Google TPU —competes with→ Nvidia
- Google TPU —designs→ Broadcom
Depends on ↑ · 4
Context — capital, rivals, policy · · 3