This page updates periodically. The numbers below are a snapshot taken in the last week of September 2026, on-demand (no commits), single-node, US/EU regions, with the standard caveat that list prices move and enterprise quotes don't match list. Treat this as a map, not a quote.
The snapshot
| Provider type | H100 on-demand | What you're actually getting |
|---|---|---|
| Hyperscaler A | $3.20-$3.60 | Pristine everything. Five nines, premium interconnect, support tickets answered by adults. You pay for the whole building. |
| Hyperscaler B | $2.90-$3.30 | Similar story, occasionally better regional pricing, equally proud of itself. |
| Hyperscaler C | $2.80-$3.20 | The "value" hyperscaler. Still roughly double decentralized rates. |
| Neo-cloud (GPU specialist) | $2.20-$2.60 | Same data centers, thinner margins, no sales org. Commits get much cheaper. On-demand inventory swings. |
| Decentralized marketplace 1 | $1.80-$2.20 | Verified data-center suppliers, escrowed payments, checkpoint/retry tooling. The sane middle. |
| Decentralized marketplace 2 | $1.40-$1.90 | Wider supplier mix, variable specs, read the node details carefully. |
| Decentralized marketplace 3 | $1.30-$1.80 | Cheapest credible hours, community-run verification, YMMV on reliability. |
Reading the spread correctly
Three things people get wrong when they see a table like this:
1. The cheapest column isn't the story
Anyone can find a $1.10 H100 if they tolerate a consumer board in a garage in a jurisdiction they can't spell. The story is the credible cheap tier: verified, data-center-grade suppliers at $1.40-$1.90 with real tooling around them. That tier didn't exist two years ago. Its existence repriced everything above it.
2. On-demand vs list price is a rigged comparison
Hyperscalers will tell you nobody pays list — and for training-scale spend, that's true. Commits cut 30-50%. But commits demand volume, credit checks and negotiations, which is exactly what a team of five doing a fine-tune doesn't have. For the long tail of buyers, list price is the market. And list price says 2.5x.
3. Price per hour isn't price per result
A $1.60 hour that dies halfway through your training job and costs you four hours of debugging is not cheaper than a $3.00 hour that just works. This is the decentralized markets' real remaining deficit, and it's an engineering problem, not an economic one — which means it's a shrinking deficit. Verification (zk proofs of execution, attestation, slashing) is the frontier, and the networks that nail it get to collapse the reliability discount to near zero.
What we're watching this quarter
- B200/GB200 trickledown. As next-gen cards absorb the top of the market, last-gen H100 supply flows down to open markets first. Watch decentralized rates soften before hyperscaler list prices do — list prices are sticky; market prices aren't.
- Verification wars. Expect at least two major decentralized networks to ship cryptographic compute verification this year. When that's table stakes, the $1.40 tier becomes the default tier.
- Hyperscaler spot markdowns. The big clouds are quietly discounting GPU spot capacity. When the incumbents start matching market prices on a visible SKU, you'll know the pressure is real.
Methodology note: rates collected from public pricing pages and marketplace order books, normalized to single H100 (80GB, PCIe or SXM as available), on-demand, US or EU region, taxes excluded. Providers knew we were watching. They always do better when nobody's watching — that's rather the point of this column.
Comments