Google's tensor processing unit is the most mature alternative to Nvidia and the only one with a decade of production use. Its advantage is the pod, not the chip: Ironwood links 9,216 chips over its own interconnect, against 72 GPUs in an Nvidia NVLink rack. In 2026 it became the chip the other labs want: Anthropic, Meta and, reportedly, OpenAI are all buying capacity.
Key Takeaways
- What it is — Hyperscaler custom silicon. ASIC built around systolic-array matrix units, with inter-chip interconnect (ICI) ports on every chip.
- Best for — Large, steady training runs and high-volume serving of standard transformer models, by teams already on Google Cloud.
- Watch out — No CUDA. Custom CUDA kernels need porting, and capacity outside Google Cloud is only starting in 2027.
At a glance
| Company | Google (Alphabet) |
|---|---|
| Current chips | Ironwood (TPU7x) generally available; TPU 8t (training) and TPU 8i (inference) announced April 2026 |
| Architecture | ASIC built around systolic-array matrix units, with inter-chip interconnect (ICI) ports on every chip |
| Memory | Ironwood: 192 GiB HBM at about 7.4 TB/s per chip. TPU 8i: 288 GB HBM plus 384 MB on-chip SRAM |
| Scale-up | Ironwood pod: 9,216 chips in a 3D torus over ICI and optical circuit switches, 1.77 PB shared HBM. TPU 8t superpod: 9,600 chips |
| Software | JAX and PyTorch via the XLA compiler, vLLM for serving; TorchTPU (native PyTorch) in preview. TensorFlow not supported on Ironwood |
| Where to get it | Google Cloud (Compute Engine, GKE). Alphabet began selling TPU systems to third parties in 2026; Oracle resale from 2027 |
| Pricing | On demand per chip-hour, US: v5e $1.20, Trillium $2.70, v5p $4.20, Ironwood $12.00 ($5.40 on a 3-year commitment) |
What changed in 2026
- Oct 2025. Anthropic agrees to use up to one million TPUs, with well over a gigawatt online in 2026.
- Nov 2025. Ironwood (TPU7x) becomes available to Google Cloud customers.
- Feb 2026. Meta signs a multibillion-dollar deal to rent TPUs through Google Cloud, Reuters reports.
- Apr 2026. Google announces TPU 8t and TPU 8i at Cloud Next; Anthropic contracts multiple gigawatts more from 2027 (about 3.5 GW per Broadcom's filing).
- Oct 2026. Oracle announces the first TPU resale deal outside Google Cloud, for 2027. OpenAI is reported to be in talks to buy TPUs directly.
Scores
Scored 1 to 5 on the same five axes as every vendor in the AI chip companies analysis.
| Performance | Scale-up | Software | Access | 2026 momentum | Total / 25 |
|---|---|---|---|---|---|
| 5 | 5 | 3 | 3 | 5 | 21 |
Verdict
The strongest second supply line for any team running large transformer workloads in JAX or PyTorch. Choose it when the job is big enough for pod-scale interconnect to matter and you can live inside Google Cloud.
Alternatives
In the same lane (hyperscaler custom silicon):
- AWS Trainium: teams already committed to AWS that want a cheaper path for large PyTorch training and inference.
- Microsoft Maia: nobody buys Maia directly. It matters because it lowers the cost of the Azure models you already pay for.
In other lanes:
- AMD Instinct (merchant gpu)
- Cerebras (inference specialist)
- Groq (inference specialist)
- SambaNova (inference specialist)
- Tenstorrent (open hardware)
Head-to-head comparisons
- Google TPU vs AWS Trainium vs Nvidia GPU: The three ways to train and serve frontier-scale models in 2026, scored on scale-up domain, software, access and price.
- Cerebras vs Groq vs SambaNova: The SRAM-first inference specialists compared on speed, token price, deployment options and company risk.
New to custom AI silicon? Our partner site CTAIO explains how TPUs, pods and GPU clusters differ in TPU vs GPU.
Frequently asked questions
What is Google TPU?
Google's tensor processing unit is the most mature alternative to Nvidia and the only one with a decade of production use. Its advantage is the pod, not the chip: Ironwood links 9,216 chips over its own interconnect, against 72 GPUs in an Nvidia NVLink rack. In 2026 it became the chip the other labs want: Anthropic, Meta and, reportedly, OpenAI are all buying capacity.
How do you get access to Google TPU?
Google Cloud (Compute Engine, GKE). Alphabet began selling TPU systems to third parties in 2026; Oracle resale from 2027. Pricing: On demand per chip-hour, US: v5e $1.20, Trillium $2.70, v5p $4.20, Ironwood $12.00 ($5.40 on a 3-year commitment).
What software does Google TPU run?
JAX and PyTorch via the XLA compiler, vLLM for serving; TorchTPU (native PyTorch) in preview. TensorFlow not supported on Ironwood.
Is Google TPU a good alternative to Nvidia?
The strongest second supply line for any team running large transformer workloads in JAX or PyTorch. Choose it when the job is big enough for pod-scale interconnect to matter and you can live inside Google Cloud. It fits best for: large, steady training runs and high-volume serving of standard transformer models, by teams already on Google Cloud. The main caveat: no CUDA. Custom CUDA kernels need porting, and capacity outside Google Cloud is only starting in 2027.
Sources
- Google Cloud TPU7x documentation
- Cloud TPU pricing
- Google Cloud: AI infrastructure at Next '26
- DCD: Google unveils eighth-generation TPUs
- Data Center Knowledge: Anthropic multi-gigawatt TPU deal
Facts verified 2026-10-31. Vendor performance figures are as stated by the vendor.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.