GPU capacity is sold three main ways: on-demand (pay by the hour, stop any time), spot or interruptible (cheaper, but can be reclaimed), and reserved (a committed term, usually billed around the clock). Choosing between them is mostly arithmetic about utilization and risk.
What the market looks like
Published H100 on-demand list prices on our board range from $1.99 to $6.16 per GPU-hour, median $3.94. Contract and transacted levels sit lower:
- SemiAnalysis put 1-year H100 contracts at $2.00–$2.70 per GPU-hour in March 2026.
- Ornn's OCPI-H100 transacted index printed $2.78 on Nov 19, 2026, after falling 18% in seven days; its three-month range was $2.46–$3.53.
- Silicon Data's H100 neocloud index was $3.08 on Nov 20, 2026, against $7.08 for its hyperscaler index.
Treat these as reference points, not offers. They mix products, terms and dates.
Spot and interruptible
Spot capacity is the cheapest way in when your job can survive interruption. Examples from providers' pages on Nov 20, 2026: Jarvis Labs lists H100 spot at $1.54 vs $3.49 on-demand (56% less) and H200 spot at $2.29 vs $4.59; Verda lists spot at half its on-demand price; Vast.ai says interruptible is 50%+ cheaper than on-demand. Use it for checkpointed training, batch inference and experiments, not for latency-sensitive serving.
The break-even rule
A reservation bills every hour of the term whether you use it or not. On-demand bills only hours used. So reserved is cheaper only when:
hours used per day > (1 − reserved discount) × 24
- At a 20% discount, break-even is 19.2 hours a day (80% utilization).
- At 30%, it is 16.8 hours (70%).
- At 50%, it is 12 hours (50%).
Our calculator does this math with your numbers.
Clusters cost differently
Multi-node clusters with high-speed fabric are often priced above single instances. Lambda, for example, lists 8-GPU on-demand H100 instances at $4.59 per GPU-hour, while its 1-Click Clusters (2 weeks to 1 year) list H100 at $6.16 for 16 GPUs, $5.85 for 64 and $5.54 for 256. Compare like with like: a cluster quote buys interconnect and guaranteed co-location, not just GPUs.
Price risk cuts both ways
Locking in a rate protects you if prices rise. Crusoe, for example, lists increases on H100, H200 and B200 from Dec 1, 2026. It costs you if prices fall, as OCPI did in mid-November. Shorter terms, break clauses, or the right to sublease unused capacity reduce that risk.
How to choose
- Unknown or spiky usage: on-demand, with spot for interruptible work.
- Steady usage above the break-even: reserved, after a short on-demand trial on the same provider.
- Large or long commitments: get several quotes and negotiate terms, not just price. Send a brief.
Sources
- GPU TRADER price board (Nov 20, 2026)
- SemiAnalysis: The Great GPU Shortage (Mar 2026)
- Ornn OCPI-H100
- Silicon Data H100 index
- Silicon Data index overview
- Jarvis Labs pricing
- Verda pricing
- Vast.ai pricing
- Lambda pricing
- Crusoe pricing
General information, not professional advice. Prices are published list prices as of Nov 20, 2026 and change often; confirm with the provider.