Compute
TPUs, AI chips & CPU endpoints
Non-GPU hardware for AI: TPUs, Trainium/Inferentia and dedicated CPU endpoints.
What it is
Besides GPUs, you can rent specialised AI chips (Google TPUs, AWS Trainium and Inferentia) or plain CPU instances for small-model inference. They can be much cheaper per unit of work, but require supported frameworks and sometimes model conversion.
How pricing works
Billed per chip or per instance per hour, often by the minute. We list the price of one chip (TPU/Trainium) or one smallest endpoint instance; larger slices and pods are multiples. Committed use and reserved capacity are cheaper.
We compare everything in one unit: USD per accelerator- or instance-hour.
Typical uses
- Cost-optimised inference at steady volume
- Training with JAX or PyTorch/XLA on TPUs or Trainium
- Small embedding or classification models on CPUs
- Dedicated, always-on model endpoints
What to compare
- Price per unit of throughput for your model, not the hourly rate
- Memory per chip
- Framework and model support
- Availability and quota
- Interconnect for multi-chip training
Watch out for
- Hourly price alone is misleading: benchmark your model, since a cheaper chip with lower throughput can cost more per token.
- Models may need conversion (XLA, Neuron SDK) and not every architecture is supported.
- Quota and regional availability can limit what you can actually rent.
All 5 APIs, cheapest first
Free to browse| API | $ / accel-hr | Accelerator type | Memory | Frameworks | Quality | Price data | Compare |
|---|---|---|---|---|---|---|---|
| HF Endpoints CPU (4 vCPU) Hugging Face Inference Endpoints | $0.134 | cpu | 8 GB | onnx, pytorch | – | Verified Oct 6, 2026 | |
| HF Endpoints AWS Inferentia2 Hugging Face Inference Endpoints | $0.75 | inferentia | — | pytorch | – | Verified Oct 6, 2026 | |
| HF Endpoints Google TPU Hugging Face Inference Endpoints | $1.20 | tpu | — | jax, pytorch | – | Verified Oct 6, 2026 | |
| AWS Trainium2 (per chip) Amazon Web Services | $2.20 | trainium | 96 GB | pytorch, jax | – | Unverified sample | |
| Cloud TPU v6e (Trillium) Google (Gemini / Vertex) | $2.70 | tpu | 32 GB | jax, pytorch, tensorflow, vllm | – | Unverified sample |