GPUs, cloud compute, and the infrastructure powering AI. Compare data-center accelerators, cloud GPU pricing, and the best options for local AI development.
FP8 TFLOPS (peak theoretical). Prices are approximate retail/grey-market estimates as of early 2026.
| GPU | Arch | VRAM | FP8 TFLOPS | Est. Price | Best For |
|---|---|---|---|---|---|
NVIDIANVIDIA B200 | Blackwell | 192GB HBM3e | 4,500 | ~$35,000 | LLM training & inference |
NVIDIANVIDIA H200 | Hopper | 141GB HBM3e | 1,979 | ~$30,000 | LLM inference, fine-tuning |
NVIDIANVIDIA H100 SXM | Hopper | 80GB HBM3 | 1,979 | ~$25,000 | General AI training |
NVIDIANVIDIA A100 80GB | Ampere | 80GB HBM2e | 624 | ~$10,000 | Budget training & inference |
AMDAMD MI300X | CDNA 3 | 192GB HBM3 | 2,600 | ~$15,000 | NVIDIA alternative, large VRAM |
On-demand rates per GPU-hour (single GPU). Spot/preemptible instances offer substantial savings but no availability guarantee.
| Provider | H100 (on-demand) | A100 (on-demand) | Spot / Savings | Notes |
|---|---|---|---|---|
CloudAWS | $32.77/hr | $19.50/hr | Up to 90% off | p5 instances, best availability |
CloudGCP | $31.22/hr | $19.82/hr | Up to 91% off | a3-highgpu, TPU alternative |
CloudAzure | $33.50/hr | $20.00/hr | Up to 80% off | ND H100 v5, strong enterprise |
IndieLambda Labs | $2.49/hr | $1.29/hr | N/A | Best indie pricing |
IndieRunPod | $3.89/hr | $1.64/hr | Community cloud cheaper | Serverless GPU option |
IndieTogether AI | API pricing | API pricing | N/A | Inference API, not raw GPU |
FP16 TFLOPS (peak theoretical). Best options for running local models with Ollama, LM Studio, or llama.cpp.
Best consumer AI GPU
Fine-tuning 7B models
Budget local inference
macOS, large unified memory
Join thousands of AI professionals. The week's most important stories, every Monday.