Google TPU

Overview

Google has been building AI accelerators longer than anyone else in this market. The first Tensor Processing Unit was deployed internally in 2015, years before the current cycle began, and the programme has run continuously since. That accumulated experience — across silicon, compilers, interconnect, data-centre design and the models themselves — is the closest thing to a genuine peer of Nvidia's full-stack position.

The seventh generation, Ironwood, reached general availability for cloud customers at Google Cloud Next in April 2026. Google positioned it explicitly as “the first Google TPU for the age of inference”, a notable framing from a company whose earlier TPU generations were justified largely by training economics. Ironwood delivers 4.6 petaFLOPS per chip, and a 9,216-chip superpod reaches 42.5 exaFLOPS.

Two things have changed the strategic picture in 2026. Google has begun selling TPU systems externally rather than only renting capacity through Google Cloud, and it has committed to splitting future generations into separate training and inference designs.

Chip generations

Price-performance figures for the eighth generation are Google's own projections for unreleased silicon, measured against its own prior generation. They should be read as roadmap positioning rather than measured results.

The training and inference split

The eighth-generation split is the most consequential architectural decision Google has announced in years, and it deserves attention beyond Google itself. Until now every major merchant and hyperscaler accelerator has been a single device expected to serve both training and inference reasonably well. Google is the first at this scale to conclude that the two workloads have diverged enough to justify separate silicon.

The logic is the one running through this entire site. Training is throughput-bound and tolerant of latency; inference — particularly decode — is memory-bandwidth-bound and latency-critical. A chip optimised for both is optimal for neither. Every inference specialist tracked here has made that argument. Google is now making it too, with the difference that it has the volume to justify two separate designs.

It is also striking that Google has given the inference chip to a different partner. Broadcom has implemented Google's TPUs for years and takes the training part, but TPU 8i is being designed with MediaTek. That is meaningful supplier diversification on the half of the roadmap Google expects to grow fastest, and it reduces dependence on a partner that has become a significant AI-silicon power in its own right.

The Anthropic agreement

Anthropic is the anchor external customer, on a scale that changes what the TPU programme is. The agreement gives Anthropic access to up to one million TPU chips and more than a gigawatt of capacity during 2026, and has since expanded to around 3.5 gigawatts of compute coming online in 2027, making Anthropic the anchor customer for the eighth generation as well.

The structure is unusual and worth understanding. Roughly 400,000 Ironwood units are supplied as finished racks by Broadcom — the arrangement behind the approximately $10 billion order from the customer Broadcom identified as Anthropic in December 2025 — while the remaining 600,000 or so are consumed as rented capacity through Google Cloud. Anthropic is therefore both buying TPU hardware and renting TPU capacity, in parallel.

This matters for the market in two ways. It establishes that TPUs are genuinely competitive for a frontier lab that is not Google, which was the open question for a decade. And it shows Google willing to sell systems rather than only rent them — a change that puts TPU into direct competition with Nvidia in a way that captive cloud silicon never was.

Strengths and key questions

Technical positioning

TPUs are systolic-array accelerators built around dense matrix multiplication, with large HBM capacity and an interconnect designed for pod-scale operation. What distinguishes them is less any single specification than the fact that the compiler, the interconnect topology and the model architectures were developed together.

Ironwood's co-packaged optics are a signal worth noting. Moving to optical interconnect at the package level addresses the same constraint OLIX is attacking from a startup position and that Nvidia backed with a $4 billion investment in photonic networking suppliers. At pod scale, communication energy and latency between chips increasingly determine achievable performance, and Google is deploying optics in volume rather than discussing it as a roadmap item.

What to watch

• External sales beyond Anthropic. Whether other labs or enterprises buy TPU systems, which would mark a genuine shift to merchant silicon.
• TPU 8i specifications. The first purpose-built TPU inference chip is the most interesting silicon on any 2027 roadmap.
• MediaTek execution. A new implementation partner on a 2nm inference design is an unproven combination.
• PyTorch parity. Time-to-support for new model architectures on TPU compared with GPU.
• Anthropic ramp. How much of the committed gigawatt capacity actually comes online during 2026.

Sources

1. Google Cloud — TPU7x documentation. Per-chip memory and pod configuration.
2. Anthropic — Expanded partnership with Google and Broadcom for multiple gigawatts of compute. Customer statement of the agreement and its scale.
3. The Next Web — Google launches Ironwood and previews the eighth-generation split at TSMC 2nm. Ironwood availability and the 8t/8i split.
4. NAND Research — Google's eighth-generation TPU family splits training and inference. Analysis of the architectural decision and partners.
5. CNBC — Broadcom reveals its mystery $10 billion customer is Anthropic (11 December 2025). The rack supply arrangement.
6. SemiAnalysis — Google TPUv7. Independent technical analysis of Ironwood.

Google does not disclose TPU unit economics, production volumes or programme revenue. Performance claims for unreleased eighth-generation parts are company projections. See the editorial methodology.

Related pages