NovaPodThe H100 rental price is the number the rest of the market is still quoted against, and it is the one that moved. I checked this marketplace on November 4, 2026, then set the asks next to three public tapes from November 2 through November 4. They do not agree, and the disagreement is the story.
A weekly percent-change table is not a search trend. The fastest one-week mover I found was an RTX PRO 4500, up 26%, which almost nobody is pricing a cluster against. The H100 is. Executed trades in that contract rose 43% in a month.
What the H100 rental price is today
On NovaPod the H100 SXM floor was $1.21 an hour. The median was $1.68. There were 65 offers, each with 80 GB. The H100 NVL was cheaper at the floor, $1.10, with a $1.62 median and only 9 offers. The H100 PCIe sat in a tight band, $1.57 to $1.57, across 5 offers. All three are 80 GB. Billing is per second from a crypto balance.
That floor is a host's ask, not the price of the H100 market. GPU Economy's listed-rate index stood at $3.39 per H100 GPU-hour on November 4, the median of 40 on-demand quotes from 38 providers, read between November 2 at 17:40 UTC and November 4 at 12:58 UTC. Their November average so far was $3.87. October's finished average on the same panel was $4.20.
Ornn's compute index, which settles executed trades rather than listings, printed the H100 SXM at $3.53 on November 3. That was up 4.4% on the day and 43.5% over one month. A day earlier the same settlement was $3.38. Bloomberg carries it as ORNNH100.
| Source | Figure | What it counts |
|---|---|---|
| NovaPod SXM, Nov 4 | $1.21 floor, $1.68 median | 65 offers, 80 GB |
| NovaPod NVL, Nov 4 | $1.10 floor, $1.62 median | 9 offers, 80 GB |
| NovaPod PCIe, Nov 4 | $1.57 floor and median | 5 offers, 80 GB |
| GPU Economy, Nov 4 | $3.39 | Median of 40 listed on-demand rates |
| Ornn trades, Nov 3 | $3.53 | Executed H100 SXM, +43.5% in a month |
| Silicon Data, Nov 3 | $2.96 | Neo-cloud benchmark, +5.3% in 7 days |
| Provider median, week of Nov 2 | $3.81 | 46 providers, not a bookable quote |
Why this H100 rental price trend is the one that matters
GetDeploying's week of November 2 put the on-demand H100 median at $3.81 across 46 providers, up 9% from October 5 and up 28% from a year earlier. Over the same week the biggest single-model jumps were an RTX PRO 4500 from $0.92 to $1.16 and an A40 from $0.85 to $1.05. Those percents are larger. The dollar of compute people actually reserve is still the H100.
Ornn's one-month board makes the same point in trades, not listings. The H100 SXM was +43.5%. The H200 was +34.9% to $7.31. The B200 was +23.5% to $9.41. The A100 SXM4 was +17.8% to $1.06. If you are picking the price series with the strongest recent trend, it is the H100 rental price.
Silicon Data's neo-cloud reading is the quiet tape. It went from $2.81 on October 5 to $2.96 on November 3, and the last seven days were +5.3%. That index standardizes term, cluster size, and interconnect. It is not a screenshot of the cheapest host. A slow grind and a 43% trade move can both be true because they are not measuring the same thing.
An H100 price per hour on a hyperscaler is a third product
GetDeploying's split for the same week had dedicated GPU clouds at a $4.50 H100 median and hyperscalers at $7.89, about 75% more. They warn that a provider median is not a rate you can book. I am repeating that warning. AWS, Azure, and Google are selling a different bundle: account terms, region, and a support contract, not an 80 GB card by the second.
Marketplace medians sit under both. On November 3, one public read of single-GPU on-demand offers had Vast.ai at a $2.40 median across 7 H100 SXM listings, RunPod Community at $2.69, RunPod Secure at $3.99, and Lambda at $4.29 for one GPU. Our $1.21 floor is below that Vast median. It is the low ask on a 65-offer book, not a claim that every H100 hour costs $1.21.
GPU Economy's own hourly series moved only between $3.27 and $3.39 in the 24 hours to November 4 at 12:58 UTC, and four of those changes were one provider's listed rate. A listed index can look calm while the trade print is ripping. Read the label on the number before you put it in a budget.
When 80 GB is the wrong H100 rental price
The H200 on this list started at $2.76 an hour, median $2.76, with 82 offers and 141 GB. The H200 NVL started at $2.29, median $2.91, 76 offers. If the job is memory-bound, the H100 rental price is the wrong comparison. You are buying 80 GB against 141 GB.
The A100 SXM4 started at $0.30, median $0.46, 158 offers, also 80 GB. That is the cheaper 80 GB Hopper-generation neighbor, not a Blackwell card and not an H100. The B200 started at $4.38, median $6.30, 51 offers, 192 GB. Use it when the model needs the newer box, not because the H100 tape looks expensive this month.
An RTX 4090 here started at $0.09, median $0.35, 419 offers, 24 GB. It is not an H100 substitute. I include it only so a low hourly number does not get mistaken for the same product.
How to read an H100 rental price
Pick one tape and stay on it. A $1.21 marketplace floor, a $2.96 neo-cloud benchmark, a $3.39 listed median, a $3.53 trade print, and a $7.89 hyperscaler median are five sentences. Averaging them produces a sixth number that nobody charges.
On this site, open the H100 SXM if you want the deep book, the NVL if the $1.10 floor is still there when you look, and the PCIe only if you have already checked that five offers is enough. Stop the instance when the job ends. The 43% move is in the trade index. It is not a coupon on the card in front of you.