Entity — Google TPU
Status: active Owner: Finance / Charlie AGT-002 Entity ID: ENTITY-AI-GOOGLE-TPU Visibility: PUBLIC Last verified: 2026-07-10
Snapshot
Google's TPU line is the most mature hyperscaler in-house ASIC program (seven generations) and the only one sold broadly as a cloud product at scale. The latest generally-available chip is TPU v7 "Ironwood" (TPU7x): 4.614 PF FP8 per chip, 192 GB HBM3e, pods to 9,216 chips — the first TPU pitched head-on at frontier training and high-volume inference against Blackwell-class GPUs. Broadcom co-design of TPUs is widely reported and reflected in Broadcom's disclosed custom-AI-silicon revenue, though Google does not name partners on spec pages.
Chip Table
| Chip | Status | Key spec highlights | As-of | Source | Grade |
|---|---|---|---|---|---|
| TPU v7 Ironwood (TPU7x) | shipping (GA on Google Cloud) | 4.614 PF FP8 / 2.307 PF BF16 per chip; 192 GB HBM3e at 7.37 TB/s; ICI 1.2 TB/s bidirectional per chip; dual-chiplet (2 TensorCores + 4 SparseCores); pods to 9,216 chips (42.5 EF FP8 per pod, vendor) | 2026-07-10 | https://docs.cloud.google.com/tpu/docs/tpu7x | 🟢 |
| TPU v6e Trillium | shipping | 918 TF BF16 / 1,836 TOPS INT8 per chip; 32 GB HBM at 1.64 TB/s; ICI 800 GB/s; 256-chip pod | 2026-07-10 | https://docs.cloud.google.com/tpu/docs/v6e | 🟢 |
| TPU v8 (next gen) | rumored | press reports of an eighth generation split into separate training and inference chips on TSMC 2nm-class node; no official documentation | 2026-07-10 | https://thenextweb.com/news/google-ironwood-tpu-inference-cloud-next | 🟠 (single secondary; treat as rumor) |
Technical Trajectory
- Cadence: roughly annual — v5e/v5p (2023) → v6e Trillium (2024/25) → v7 Ironwood (announced 2025, GA into 2026). Google now leads with an inference-first framing ("age of inference", blog.google, 🟢).
- Compute jump: Trillium 918 TF BF16 → Ironwood 2,307 TF BF16 (2.5x) and 4,614 TF FP8 — first TPU generation with headline FP8 (🟢, official docs).
- Memory: 32 GB → 192 GB HBM3e per chip (6x) at 7.37 TB/s; closes the capacity gap with B200-class GPUs (🟢).
- Scale: pod size step from 256 (v6e) to 9,216 chips (TPU7x) with optical circuit switching heritage; Google's differentiation is pod-scale fabric, not single-chip FLOPs (🟢 docs; pod EF figure is vendor marketing 🟡).
- Node: process node undisclosed for both generations (n/a); external visibility comes via Broadcom earnings commentary (see custom-asic.md).
- Availability model: internal-first (Gemini training/serving) plus external Cloud GA — the only custom ASIC with meaningful third-party rental, making TPU the cleanest public test of CLAIM-AI-SILICON-001.
Watch Items
- Any official TPU v8 disclosure (Cloud Next / Hot Chips 2026) — upgrade or kill the 🟠 rumor row.
- MLPerf Training/Inference submissions using Ironwood (Google has submitted TPU results in past rounds; none verified for v5.1 this pass).
- Ironwood external capacity signals: region availability expansion, committed-use pricing changes on Google Cloud pricing pages (A4).
- Alphabet earnings commentary quantifying TPU fleet share vs GPU purchases.
Claims
- CLAIM-AI-SILICON-001 — primary evidence stream: TPU is the largest and most externally visible custom-ASIC program taking share of accelerator compute vs merchant GPU.
- CLAIM-AI-BOTTLENECK-001 — TPU HBM3e demand and OCS-based pod fabric are inputs to the memory/packaging bottleneck picture.
Sources
- 🟢 Google Cloud TPU7x (Ironwood) documentation — https://docs.cloud.google.com/tpu/docs/tpu7x (accessed 2026-07-10)
- 🟢 Google Cloud TPU v6e (Trillium) documentation — https://docs.cloud.google.com/tpu/docs/v6e (accessed 2026-07-10)
- 🟢 Google blog, "Ironwood: the first Google TPU for the age of inference" — https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/ (accessed 2026-07-10)
- 🟡 TrendForce, Ironwood 9,216-chip superpod — https://www.trendforce.com/news/2025/11/07/news-google-unveils-7th-gen-tpu-ironwood-with-9216-chip-superpod-taking-aim-at-nvidia/ (accessed 2026-07-10)
- 🟠 TNW, eighth-gen TPU split report — https://thenextweb.com/news/google-ironwood-tpu-inference-cloud-next (accessed 2026-07-10)
Changelog
- 2026-07-10: Created; first verification pass.