Today's H100 prices, all three modes
There are three H100 packages on the sheet and one rule behind every number: the public marketplace median for that package × 0.70, rounded down, re-checked weekly (snapshot 2026-09-03). Live figures:
| Package | Memory | On-demand | Interruptible | Reserved (3 mo) | Market median |
|---|---|---|---|---|---|
| H100 SXM | 80 GB HBM3 | $1.587 | $0.793 | $1.031 | $2.27 |
| H100 PCIE | 80 GB HBM2e | $2.147 | $1.073 | $1.395 | $3.07 |
| H100 NVL | 80 GB HBM3 | $1.877 | $0.938 | $1.220 | $2.68 |
Two things surprise people. First, the SXM part — the one with NVLink and the fastest memory — is the cheapest of the three here, because its marketplace pool is by far the deepest and the median follows supply. Second, interruptible is not a fluctuating spot auction: it is a flat 50% of on-demand, on every H100, all the time.
What the rest of the market charges
Public list prices for a single H100 SXM GPU-hour, September 3, 2026, grouped by provider type. Ranges, because every provider packages the card differently (regions, node sizes, minimum commitments):
| Provider type | Typical H100 $/GPU-hr | What moves the number |
|---|---|---|
| Hyperscalers (list) | $4 – $8 | 8-GPU node minimums, region, committed-use discounts |
| Specialist GPU clouds | $2 – $4 | Secure vs community tiers, PCIe vs SXM, reservations |
| Public GPU marketplaces | $1.5 – $3 (median $2.27) | Per-host auctions; changes hourly, reliability varies by host |
| PowerGPU | $1.587 fixed | Median × 0.70, verified datacenters, re-checked weekly |
The gap between the top and bottom rows is 3–5× for the same silicon. Most of it is not margin: it is the cost of sales teams, card fraud, free tiers and idle capacity that a fixed-price, crypto-settled operator does not carry.
Per day, per month, per training run
Per-second billing makes the hourly rate the only number you need, but here are the shapes people actually budget:
| Scenario | Math | On-demand | Interruptible |
|---|---|---|---|
| One H100 SXM, one day | 24 h | $38.09 | $19.03 |
| One H100 SXM, one month | 730 h | $1,159 | $579 |
| 8× H100 node, one month | 8 × 730 h | $9,268 | $4,631 |
| A 72-hour fine-tune on 8× H100 | 8 × 72 h | $914 | $457 |
| Reserved H100 SXM, 3-month term | 3 × 730 h × $1.031 | $2,258 total | |
Run your own schedule through the cost calculator — it adds storage and bandwidth and compares the result with the marketplace median.
The costs that are not the GPU
- Storage — $0.08/GB/month here, per second, while the disk or volume exists. A 1 TB checkpoint volume is $80/month; on hyperscalers, $80–170.
- Egress — $0.01/GB flat, both directions. Hyperscalers charge $0.05–0.12/GB out: pulling 10 TB of results costs $100 here versus $500–1,200 there.
- Idle time — the biggest hidden line everywhere. Per-second billing and volumes let you destroy the GPU between runs and keep the data; hourly rounding and "minimum 1 hour" policies do not.
- Node minimums — many H100 offers are 8-GPU nodes only. If you need one card, you pay for eight. Here 1×, 2×, 4× and 8× carry the same per-GPU price.
Rent or buy: the break-even
A bare H100 SXM sells for roughly $27,000 in 2026, and it needs an HGX host, 700 W of power and cooling, a network port and someone to run it. Ignoring all of that, renting at $1.587/hr buys 17,013 GPU-hours for the sticker price — about 23 months of 24/7 use. Add hosting and a realistic 40–60% utilisation and the break-even moves past the card's useful life. Buying makes sense for permanent, saturated fleets with their own datacenter; for everyone else, a reserved H100 at $1.031/hr is the closest thing to owning one without the depreciation.
Five ways to pay less for H100 time
- Go interruptible for anything that checkpoints. Training does; it is half price and restarts from your volume. Keep on-demand for the deadline run and for serving.
- Reserve what runs every day. Above ~65% utilisation, $1.031/hr beats every other mode, and capacity is held for you during launch-week droughts.
- Check whether an A100 is enough. At $0.583/hr the A100 SXM4 still wins total job cost for ≤13B models and LoRA work — see H100 vs A100.
- Use FP8. The Hopper Transformer Engine is the discount most teams never enable: roughly 1.5–2× throughput on the same hourly rate.
- Destroy, keep the volume. Datasets and weights stay warm at $0.08/GB/mo; the GPU bills nothing until the next run.
Bigger than an H100? The H100 vs H200 vs B200 guide covers when the $3.058/hr H200 or a Blackwell card ends up cheaper per token.
Put the numbers to work
Every price in this guide is our live rate — fixed, ≥30% under the market median, billed per second. Deploy the exact setup above from the console in about 30 seconds, paid in crypto, no card and no KYC.
