Deepest rate
On-demand × 0.65 — roughly 55% under the market median once the 30% floor is stacked in. The cheapest legitimate way to run a permanent GPU.
Products · reserved
Reserved GPU capacity: −35% and a held seat
For workloads that run every day, reservations stack a further 35% discount on top of prices already 30% under market — H100 SXM drops to $1.031/hr, locked for the term, with capacity guaranteed even when a region runs hot.
Capacity held
Reserved capacity is fenced from the on-demand pool. Your deploys never queue, even during launch-week GPU droughts.
Rate locked
Market up or down, your price is signed for the term. Budget a quarter to the dollar.
Billed monthly
No giant upfront invoice: the commitment draws from your crypto-funded balance month by month.
Reserved rates on the big cards
The 3-month figure is the full commitment cost per GPU — compare it with four weeks of your current invoice.
| GPU | On-demand | Reserved | Per month (730 h) | 3-month commit | Action |
|---|---|---|---|---|---|
| B200 | $4.204 | $2.732 −35% | $1,994 | $5,983 | Reserve |
| H200 | $3.058 | $1.987 −35% | $1,451 | $4,352 | Reserve |
| H100 SXM | $1.587 | $1.031 −35% | $753 | $2,258 | Reserve |
| A100 SXM4 | $0.583 | $0.378 −35% | $276 | $828 | Reserve |
| L40S | $0.466 | $0.302 −35% | $220 | $661 | Reserve |
| RTX 5090 | $0.318 | $0.206 −35% | $150 | $451 | Reserve |
6-month terms take a further −5%, 12-month −10% — quoted per request from the console.
When reserving beats everything
- Production inference — endpoints with steady traffic 24/7.
- Training programmes — research groups with a standing compute budget.
- Render pipelines — studios with a predictable nightly queue.
- Migration off hyperscalers — an H100 reserved here costs less than a quarter of typical list price.
Rule of thumb: if your utilisation is above ~65%, reserved beats on-demand; below that, stay on-demand or go interruptible.
Need more than 16 GPUs?
Fleet reservations, InfiniBand clusters, custom regions and invoicing conversations live on the enterprise page. Everything is still crypto-settled and still priced from the public sheet.
Median time from ticket to running fleet: 3 business days.
Reserved GPUs: FAQ
Terms live in the terms of service; capacity questions in a console ticket.
How does a reservation work?
You commit to a GPU model and quantity for 3, 6 or 12 months and pay the reserved rate (−35% off on-demand) for the committed hours, drawn from your credit balance month by month — not all upfront. The capacity is held for you: deploys against a reservation never wait.
What if I stop using it halfway through?
The commitment bills for its term whether instances run or not — that is what buys the discount and the held capacity. You can upgrade a reservation to a bigger one at any time; we credit the remainder.
Does the reserved price ever change mid-term?
No. The rate is locked at signature for the entire term, even if market prices — and therefore our on-demand prices — move. Renewal re-prices at the then-current sheet.
Can I reserve multi-node clusters?
Yes — reservations are how most InfiniBand cluster capacity is sold. See the clusters page or open a ticket from the console with your target topology; provisioning takes 1–3 days.