GPU price sheet, 2026 edition

How much does a cloud GPU cost in 2026?

Prices vary by card, provider and billing model. Teams serving LLMs, speech, image or video models can pick from low-cost cards like the L4 and A10, mid-range cards like the L40S, and top cards like the H100, H200 and B200.

The newest card is not always the right one. Model size, memory, latency targets and traffic shape decide it.

Price per GPU-second

Providers publish hourly, per-minute or per-second prices. To compare them we converted every public price to US dollars per GPU-second. The figures below leave out idle billing, startup time and platform fees, which often matter more than the rate.

GPU	VRAM	Axonfield	Provider B	Provider C
B200	192 GB	$0.00161	$0.001739	$0.001497
H200	141 GB	$0.00112	$0.001210	$0.001042
H100	80 GB	$0.000912	$0.000985	$0.000848
RTX 6000 Pro	96 GB	$0.000671	$0.000725	$0.000624
A100 (80GB)	80 GB	$0.000562	$0.000607	$0.000523
A100 (40GB)	40 GB	$0.000531	$0.000573	$0.000494
L40S	48 GB	$0.000518	$0.000559	$0.000482
A10	24 GB	$0.000294	$0.000318	$0.000273

Which card for which model

Models under 10B parameters run well on an L4 or A10. Models from 20B to 70B want 80 GB of memory, so an A100 80GB or H100. Long-context and mixture-of-experts models are where the H200 and B200 pay for themselves.