GPU price changes

Per GPU-hour moves among tracked providers, 2026-10-27 to 2026-10-29 (3 daily snapshots). Same data as the get_price_changes MCP tool.

Subscribe: /gpu/changes.xml (RSS) · JSON: gpu-prices.json · gpu-price-changes.json (RSS) · history.json

Read before comparing

Rankings are the cheapest among tracked providers on our check date, not across the whole market. Vendors we don't track may cost less.

GPU-hours vs wall-clock time

Prices are per GPU-hour: one GPU for one hour. Four GPUs for 10 wall-clock hours is 40 GPU-hours. Per-second billing (serverless) charges only while your code runs; hourly pods bill from start to stop, idle time included; monthly servers bill a flat price whatever you use. A faster GPU can finish the same job in fewer GPU-hours, so the lowest hourly rate isn't always the lowest cost per job.

Hardware differences

The same GPU name can mean different hardware. SXM modules link GPUs over NVLink and run at higher power limits than PCIe cards, so an H100 SXM is usually faster than an H100 PCIe, especially for multi-GPU training. Interconnect between GPUs and nodes (NVLink, InfiniBand, plain Ethernet) is rarely listed on pricing pages; treat it as unknown unless the vendor states it. VRAM decides which models fit at all. Hardware as listed: A10 24 GB (not stated); A100 80 GB (not stated); A100 NVLink 80 GB (not stated); A100 PCIe 40 GB (PCIe); A100 PCIe 80 GB (PCIe); A100 PCIe 80 GB (not stated); A100 SXM 40 GB (SXM); A100 SXM 80 GB (SXM); A100 SXM 80 GB (not stated); B200 180 GB (SXM); GH200 96 GB (not stated); H100 80 GB (not stated); H100 NVLink 80 GB (not stated); H100 PCIe 80 GB (PCIe); H100 PCIe 80 GB (not stated); H100 SXM 80 GB (SXM); H100 SXM 80 GB (not stated); H200 141 GB (not stated); H200 SXM 141 GB (SXM); H200 SXM 141 GB (not stated); L40 48 GB (not stated); L40S 48 GB (not stated); Quadro RTX 6000 24 GB (not stated); RTX 3090 24 GB (not stated); RTX 4090 24 GB (not stated); RTX 5090 32 GB (not stated); RTX 6000 Ada 48 GB (not stated); RTX A5000 24 GB (not stated); RTX A6000 48 GB (not stated); V100 16 GB (not stated).

Egress and storage extras

Totals on this page exclude these. Each cell cites the vendor page it came from; "Not published" means the vendor doesn't publish it there.

Interruptible vs on-demand

Interruptible (preemptible or spot) capacity can be stopped by the vendor mid-job and is usually cheaper; on-demand runs until you stop it. Interruptible (preemptible) rates listed here: Hyperstack A100 PCIe, Hyperstack H100 PCIe, Modal A100, Modal H100 SXM, Modal H200 SXM, Modal L40S, SaladCloud RTX 4090, Verda (formerly DataCrunch) A100 SXM, Verda (formerly DataCrunch) H100 SXM, Verda (formerly DataCrunch) H200 SXM, Verda (formerly DataCrunch) L40S, Verda (formerly DataCrunch) RTX A6000. All other rates are on-demand: the vendor doesn't publish preemption for them.

How to watch GPU prices from an agent