Marketplace clouds rent an H100 for $1.49–$2.24 an hour, while the big three sit at $2.25–$3.69 — and missing egress data blocks a universal all-in winner

Public-web scan of 2026-08-29: 18 GPU cloud providers with usable NVIDIA H100 pricing (generally H100 80GB), in USD per GPU-hour. Effective price is the lowest observed of sticker, spot, or committed; it does not price interruption risk, minimum spend, or hardware differences. A 10 GB/hour egress add-on was computable for only 3 providers.

effective price comes from a spot rate effective price is the on-demand sticker sticker reference (where above effective) effective with 10 GB/h egress — only where a fee was numerically resolved
Spot rates set the floor: Vast.ai $1.49, DataCrunch/Verda $1.65 and Spheron $1.66 — the three cheapest effective prices in the scan are all interruptible tiers gridstackhub.ai.
The cheapest guaranteed on-demand stickers cluster tightly: Vast.ai $1.50, Hyperstack $1.90, then Lambda, RunPod and Together AI all at $1.99 gridstackhub.ai.
Egress was numerically resolved for only Spheron, Lambda and RunPod — each reported at $0/GB, so 10 GB/hour adds nothing to their price; every other provider's egress is unresolved, not free spheron.network.
Hyperscaler discounts narrow the gap but do not close it: Azure spot at $2.25 undercuts its $6.00 sticker by 63%, Google Cloud spot $3.69 vs $5.97 sticker, while AWS's observed spot ($3.83) was above its $2.97 on-demand rate thundercompute.com.

All 18 providers, every observed field

Method: public-web market survey run 2026-08-29 over provider pages and recent specialist comparison pages; 18 provider rows with usable H100 pricing, generally H100 80GB, USD per GPU-hour. Effective compute-only price is the minimum of observed sticker, spot or committed rates; effective-with-egress adds 10 GB of public-internet egress per GPU-hour only where a fee was numerically resolved — a blank egress field means unresolved, not free. Availability is the supplied page claim at scan time, not a capacity check or launch test; blanks are marked not verified. Excluded from all figures: host CPU/RAM, storage, support, taxes, idle time, minimum cluster size or term, and interruption/restart costs — true workload-level cost can be higher. This is not an authenticated quote from every provider; some offers require contracts, clusters, specific regions or marketplace hosts, and spot can be interrupted. Nothing was cut: all 18 rows are shown.

Keenable · made with SELECT · 4,656 pages in 6m 13s · Open the conversationShare: XLinkedInReddit