How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
Ironwood, also called TPU7x, is Google’s seventh-generation TPU — the first Google explicitly built for inference rather than general use. Announced 9 April 2025, Google called it “the first Google TPU for the age of inference” 1.
Deep diveThe technical detail
Ironwood entered preview 24 November 2025 and reached general availability 31 March 2026 2 — roughly a year after announcement.
2.Launch and history
SimpleStart here
Ironwood marks a deliberate shift in how Google frames its chips: away from “faster, general-purpose AI compute” and toward serving specific, proactive AI agents at scale — what Google calls a move from “responsive AI” to “the age of inference.”
Deep diveThe technical detail
Google’s own numbers versus Trillium: 10 times the peak performance of TPU v5p and 4 times better performance per chip than Trillium for training and inference combined 3, with 2 times better performance per watt stated in Google’s original Ironwood announcement 1. A full 9,216-chip “superpod” delivers 42.5 exaflops of FP8 compute — described by Google as 24 times the compute of the El Capitan supercomputer 1.
3.What’s inside it
SimpleStart here
Two TensorCores and four SparseCores per chip, connected inside 64-chip “cubes” that link into a 9,216-chip superpod over a dynamically reconfigurable optical network 4.
Deep diveThe technical detail
Ironwood uses a 3D-torus interconnect within each 64-chip cube (8 chips per host), with those cubes linked into the full superpod via Optical Circuit Switches, plus a separate data-center-network layer for connecting multiple superpods together 5. Advanced liquid cooling is stated to deliver “twice standard air-cooling performance” 1. Google’s documentation explicitly states TensorFlow is not supported on Ironwood — only JAX and PyTorch are — and the older Cloud TPU API itself does not support TPU7x or later, requiring Compute Engine or GKE instead 46.
4.Full spec table
SimpleStart here
2,307 teraflops (bf16) and 4,614 teraflops (fp8) per chip, 192GB of HBM memory, connected up to 9,216 chips per superpod 4.
Deep diveThe technical detail
| Spec | Google Ironwood (TPU7x) |
|---|
| Peak compute (bf16) | 2,307 TFLOPs per chip |
|---|
| Peak compute (fp8) | 4,614 TFLOPs per chip |
|---|
| Memory | 192 GiB HBM, 7,380 GBps (~7.37 TB/s) bandwidth |
|---|
| Inter-chip interconnect | 1,200 GBps bidirectional |
|---|
| Data-center network link | 100 Gbps per chip |
|---|
| TensorCores / SparseCores | 2 / 4 per chip |
|---|
| Superpod size | 9,216 chips |
|---|
| Superpod compute | 42.5 exaflops (fp8), 1.77 PB shared HBM |
|---|
| Framework support | JAX, PyTorch — not TensorFlow |
|---|
Per-chip figures per Google Cloud’s TPU7x documentation 4; superpod compute and El Capitan comparison per Google’s original Ironwood announcement 1, memory-bandwidth context per Google’s Axion-VM blog 3. Note that almost all of Ironwood’s headline bandwidth lives inside the superpod’s optical fabric: the per-chip external network link is a comparatively modest 100 Gbps against a 1,200 GBps chip-to-chip link.
5.Where it’s made
SimpleStart here
Google does not disclose who manufactures Ironwood or on what process node. Broadcom, Google’s TPU co-development partner, has not named Ironwood specifically in its own public filings.
Deep diveThe technical detail
Broadcom’s April 2026 SEC filing on its long-term TPU agreement with Google discusses future TPU generations and networking components broadly, without naming Ironwood or TPU7x by name, and does not disclose a foundry 7. No official Google or Broadcom source names TSMC or any other specific fabricator for Ironwood; this page does not assert one.
6.Which systems use it
SimpleStart here
As of general availability, Ironwood is offered through Google Cloud in only two North American zones, both in us-central1 8 — a narrower footprint than earlier generations, likely reflecting its recent GA date.
Deep diveThe technical detail
Named adopters in Google’s own material include Lightricks (using Ironwood for its LTX-2 generative model) and Essential AI 3. The largest named commitment is Anthropic’s: on 23 October 2025, Anthropic announced access to up to one million TPUs, with compute capacity “well over a gigawatt” expected online in 2026, explicitly naming Ironwood as the chip involved 9.
7.Official pricing
SimpleStart here
Google Cloud’s own pricing page lists Ironwood on-demand at $12.00 per chip-hour in Iowa, rising to $13.20 in London 10.
Deep diveThe technical detail
Checked September 2026, from Google Cloud’s own pricing page 10: on-demand $12.00/chip-hour in us-central1 (Iowa) and $13.20/chip-hour in europe-west2 (London); Dynamic Workload Scheduler options at $6.00/chip-hour (Flex-start) or $8.40/chip-hour (Calendar mode); committed-use rates as low as $5.40/chip-hour (3-year, Iowa). No page-level “last updated” date or fixed spot rate is published.
9.What came before, what came next
SimpleStart here
Came before: Trillium (TPU v6e). Came after: TPU 8t and TPU 8i, announced April 2026 — the first time Google has split a TPU generation into separate training and inference chips.
Deep diveThe technical detail
Ironwood succeeded Trillium after roughly eleven months 1. Google’s eighth generation broke from Ironwood’s single-chip-for-everything approach entirely, shipping TPU 8t (training) and TPU 8i (inference) as two distinct chips rather than one — covered on their own pages.
10.Hidden in plain sight
SimpleStart here
Two of the biggest AI companies in the world are both, per Google’s own material, running on the same chip family — at very different scales.
Deep diveThe technical detail
Google’s own blog names Gemini 2.5 and AlphaFold as workloads run on Ironwood-generation infrastructure 1, while Anthropic’s own announcement, separately, commits to up to one million TPUs of Ironwood-generation compute 9 — meaning Google’s own flagship model and one of its largest external AI customers both depend on the same chip family, disclosed in two entirely separate official announcements rather than one joint one.