Three platforms carry frontier-scale AI in 2026: Google's TPU, Amazon's Trainium and Nvidia's GPUs. The clearest evidence that none of them has won is that Anthropic trains on all three. This is how they differ on the attributes a buyer can verify, and which one fits which workload.
Key Takeaways
- Google TPU — Largest fast-interconnect domain by far: 9,216 Ironwood chips per pod over ICI. Best for big, steady JAX or PyTorch jobs on Google Cloud. $12.00 per chip-hour on demand, $5.40 on three years.
- AWS Trainium — Trainium3 (3 nm, 144 GB HBM3e) in UltraServers of up to 144 chips. The cost play for AWS-native PyTorch teams, and the chip behind Anthropic's 500,000-chip Project Rainier.
- Nvidia GPU — 72 GPUs per NVLink rack, but CUDA, every framework and every cloud. Still the default unless one workload dominates your bill.
- Bottom line — Anthropic runs all three. Most teams should keep Nvidia as the default and move their single biggest workload to the custom chip of the cloud they already use.
The short answer
Pick by cloud first and workload second. If your data and identity perimeter live in Google Cloud and your models are standard transformers, the TPU is the strongest second supply line. If you live in AWS, Trainium is. If you need portability, custom kernels or the broadest hiring pool, Nvidia stays the default.
Side by side
| Feature | Google TPU (Ironwood) | AWS Trainium3 | Nvidia GPU (Blackwell) |
|---|---|---|---|
| Chip | |||
| Current generation | Ironwood (TPU7x), GA Nov 2025; TPU 8t / 8i announced Apr 2026 | Trainium3 (3 nm), GA Dec 2025; Trainium4 announced | Blackwell B200 / GB200 / GB300; Vera Rubin in 2H 2026 |
| Peak dense FP8 per chip | 4,614 TFLOPs | About 2.5 PFLOPs (362 PF per 144-chip UltraServer) | B200 in the same class (several PFLOPs) |
| HBM per chip | 192 GiB, about 7.4 TB/s | 144 GB HBM3e, 4.9 TB/s | B200: 192 GB HBM3e |
| Scale | |||
| Fast scale-up domain | 9,216 chips per pod (3D torus over ICI + optical circuit switches) | 144 chips per UltraServer (NeuronSwitch) | 72 GPUs per NVLink rack; 144 with Kyber in 2027 |
| Largest cluster | 134,000 TPU 8t in one fabric; 1M+ across sites (Google) | Hundreds of thousands (UltraClusters 3.0); Rainier ~500k Trainium2 | Hundreds of thousands over InfiniBand or Ethernet |
| Self-serve slice size | Up to 2,048 chips; larger via account team | Instances and UltraServers; larger via reservation | Varies by provider; neoclouds sell large reserved clusters |
| Software | |||
| Primary stack | XLA: JAX, PyTorch/XLA, vLLM; TorchTPU in preview | Neuron SDK: PyTorch, vLLM, Transformers, Ray; NKI kernels | CUDA plus every framework and library |
| Custom CUDA kernels | Must be rewritten (Pallas) or replaced | Must be rewritten in NKI or replaced | Run as is |
| Buying | |||
| Where to get it | Google Cloud; Oracle resale from 2027 | AWS only | Every hyperscaler, neocloud and on premises |
| Published price | $12.00 / chip-hr on demand; $8.40 1-yr; $5.40 3-yr | Per EC2 instance-hour; no per-chip list price | Per GPU-hour; spread of 2-3x between hyperscalers and neoclouds |
| Lock-in | High: one cloud, one compiler | High: one cloud, one SDK | Low: portable across providers |
Verified 2026-10-31. Peak FLOPs and performance claims are vendor figures.
Scale-up is the real difference
Per chip, the three are closer than the marketing suggests. Ironwood and B200 both carry 192 GB of HBM; Trainium3 trades some memory for price. The gap opens at the next level up. Ironwood links 9,216 chips with its own inter-chip interconnect and optical circuit switches. Trainium3's NeuronSwitch connects up to 144 chips. Nvidia's NVLink connects 72 GPUs per rack today and 144 with the Kyber rack in 2027. Beyond those domains, all three fall back to the data center network. For a training job whose tensor and expert parallelism spans hundreds of chips, staying on the fast domain is worth more than a few extra petaflops per chip.
Software is the switching cost
Nvidia's lead is CUDA, not silicon. Both custom chips reach the hardware through a compiler: XLA for TPUs, Neuron for Trainium. Standard PyTorch and JAX models port with sharding and configuration work. Hand-written CUDA kernels have to be rewritten in Pallas or NKI, or replaced. Inventory those kernels before comparing prices, because the porting project is usually the larger number in the business case.
Price
Google publishes per-chip prices: Ironwood is $12.00 per chip-hour on demand in us-central1, $8.40 on a one-year commitment and $5.40 on three years. AWS sells Trainium by EC2 instance and UltraServer and argues on cost per token. Nvidia GPU-hours vary by two to three times between hyperscalers and neoclouds. None of these numbers compare directly. Run the same short paid benchmark on each and compare cost per step or per million tokens.
Verdict
- Training a large model on Google Cloud: Google TPU. The pod-scale interconnect is the deciding factor.
- Training or serving PyTorch models inside AWS: AWS Trainium, once your models run cleanly on Neuron.
- Mixed workloads, custom kernels, multi-cloud or on premises: Nvidia GPU.
- High-volume inference: benchmark all three; TPU 8i and Trainium3 both target cost per token directly.
Related
- AI chip companies 2026: all eight Nvidia competitors scored.
- Cerebras vs Groq vs SambaNova: the inference specialists.
- TPU vs GPU on our partner site CTAIO: how TPU pods and clusters work.
- AI compute and neocloud providers 2026: where to rent Nvidia GPUs.
Frequently asked questions
Is Google's TPU better than Nvidia's GPU?
For large, regular training and serving jobs on standard transformer models, an Ironwood TPU pod often does more work per dollar because 9,216 chips share one high-bandwidth network, against 72 GPUs in an Nvidia NVLink rack. Nvidia wins on software breadth, custom kernels and availability across clouds. The answer depends on the workload and on whether your team can work in JAX or PyTorch/XLA.
Trainium vs TPU: which is cheaper?
It depends on your model and your cloud. Google publishes TPU prices per chip-hour (Ironwood $12.00 on demand, $5.40 on a three-year commitment); AWS prices Trainium per EC2 instance and positions it on cost per token. Compare cost per training step or per million tokens on a short benchmark of your own model, inside the cloud where your data already lives.
Does Anthropic use TPUs or Trainium?
Both, and Nvidia GPUs too. Anthropic agreed in October 2025 to use up to one million Google TPUs and in April 2026 contracted multiple gigawatts more from 2027. It also trains on AWS's Project Rainier, about 500,000 Trainium2 chips.
Can TPUs and Trainium run PyTorch?
Yes. TPUs run PyTorch through PyTorch/XLA, and Google's native TorchTPU is in preview. Trainium runs PyTorch through the AWS Neuron SDK. Standard models port with modest effort; custom CUDA kernels have to be rewritten.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.