GPUs held for you, at a rate that cannot move.
A reservation takes a number of GPUs of one model out of the pool and keeps them for your account for one to twelve months. They are never reclaimed, the rate is fixed for the whole term, and the capacity is there whether you use it or not.
- Term
- 1–12monthsLonger terms quoted individually
- H100 rate
- $0.65/h27% below our on-demand price
- Break-even
- 73%Utilisation above which it beats on-demand
- Reclaims
- NeverThe hardware is out of the spot pool for the term
Four steps, no sales cycle.
A reservation is a capacity allocation, not a contract negotiation. Most are confirmed the same day.
-
Tell us the shape
Model, number of GPUs, region, start date, term. If you are unsure of the model, tell us the workload and the model sizes — we will point at the cheapest card that fits.
-
We confirm the allocation
We check the hardware is free for the whole term and send back a confirmation naming the model, the count, the region, the dates and the monthly amount. Nothing is charged before you accept it.
-
The capacity appears in your console
Reserved GPUs show up in the deployment screen as your own pool. Launch, stop and relaunch as often as you like: the hardware stays allocated to your account.
-
Renew, resize or stop
We remind you two weeks before the end of the term. Renew as is, change the count, or let it lapse — the instances keep running on the on-demand rate until you stop them.
The monthly rate, and when it beats paying by the hour.
Monthly cost is the fixed rate × 730 hours per GPU. Break-even is the share of the month you would have to run an on-demand instance to spend the same amount — above it, reserving is cheaper.
| GPU | Reserved | Per GPU per month | On-demand | Saving | Break-even utilisation | Action |
|---|---|---|---|---|---|---|
| B200 SXM180 GB HBM3e | $1.85/h | $1,350.50 | $2.49/h | 26% | 74% | Request |
| H200 SXM141 GB HBM3e | $0.79/h | $576.70 | $1.09/h | 28% | 72% | Request |
| H200 NVL141 GB HBM3e | $0.69/h | $503.70 | $0.99/h | 30% | 70% | Request |
| H100 SXM80 GB HBM3 | $0.65/h | $474.50 | $0.89/h | 27% | 73% | Request |
| H100 NVL94 GB HBM3 | $0.59/h | $430.70 | $0.79/h | 25% | 75% | Request |
| H100 PCIe80 GB HBM2e | $0.49/h | $357.70 | $0.69/h | 29% | 71% | Request |
| GH20096 GB HBM3 | $0.59/h | $430.70 | $0.79/h | 25% | 75% | Request |
| A100 SXM 80GB80 GB HBM2e | $0.19/h | $138.70 | $0.29/h | 34% | 66% | Request |
| A100 PCIe 80GB80 GB HBM2e | $0.09/h | $65.70 | $0.19/h | 53% | 47% | Request |
| A100 40GB40 GB HBM2 | $0.09/h | $65.70 | $0.17/h | 47% | 53% | Request |
| MI325X256 GB HBM3e | $1.19/h | $868.70 | $1.59/h | 25% | 75% | Request |
| MI300X192 GB HBM3 | $1.05/h | $766.50 | $1.45/h | 28% | 72% | Request |
| L40S48 GB GDDR6 | $0.15/h | $109.50 | $0.25/h | 40% | 60% | Request |
| L4048 GB GDDR6 | $0.09/h | $65.70 | $0.19/h | 53% | 47% | Request |
| A4048 GB GDDR6 | $0.06/h | $43.80 | $0.09/h | 33% | 67% | Request |
| A3024 GB HBM2 | $0.06/h | $43.80 | $0.08/h | 25% | 75% | Request |
| A1024 GB GDDR6 | $0.09/h | $65.70 | $0.15/h | 40% | 60% | Request |
| L424 GB GDDR6 | $0.06/h | $43.80 | $0.08/h | 25% | 75% | Request |
| T416 GB GDDR6 | $0.05/h | $36.50 | $0.07/h | 29% | 71% | Request |
| V100 32GB32 GB HBM2 | $0.06/h | $43.80 | $0.09/h | 33% | 67% | Request |
| V100 16GB16 GB HBM2 | $0.06/h | $43.80 | $0.08/h | 25% | 75% | Request |
Workstation and consumer models can be reserved too — their rates are in the full price list. Storage, at $0.08 per GB-month, is billed separately.
Reserve when the load is steady. Not before.
A reservation is the cheapest way to run a GPU that is busy most of the month, and the most expensive way to run one that is busy twice a week.
- Running most of the month
- Reserve. Above 73% utilisation on an H100, the fixed rate costs less than paying on-demand for the same hours — and the capacity is guaranteed.
- Bursty, but cannot be interrupted
- On-demand. You pay only for the minutes you run, with the same 99.9% monthly target as a reservation.
- Interruptible at any time
- Spot. 38% below on-demand on an H100, with a 2-minute notice before a reclaim.
- A fixed date and a fixed count
- Reserve. It is the only tier where the hardware is guaranteed to exist on the day you need it.
A worked comparison
H100 · 1 GPU · 30 days- Reserved, whole month$474.50
- On-demand, running 100% of the month$649.70
- On-demand, running 73% of the month$474.28
- Spot, running 100% of the month$401.50
- Reserved against full-month on-demand$175.20
Spot is cheaper than a reservation per hour on this model, and always will be — it is the interruption you are paying to avoid. Many teams reserve the baseline and put the rest on spot.
Multi-node reservations, private networking between them and a dedicated interconnect are arranged individually. Tell us the shape of the cluster and we will come back with what is possible and what it costs.
Reserved capacity, in detail
What exactly is reserved?
A number of GPUs of one model, in one region, held for your account for the whole term. They are never offered to the spot pool and never reclaimed. You launch and stop instances on them as often as you like; the capacity stays yours until the term ends.
How long can a term be?
From one month to twelve. Shorter than a month is what the on-demand tier is for. Longer terms and multi-node reservations are quoted individually — tell us the model, the count, the region and the start date.
How is it billed?
Monthly in advance, at the fixed rate in the table above multiplied by 730 hours per GPU per month. The rate does not move during the term, whatever happens to the public price list. Persistent storage is billed separately at $0.08 per GB-month.
Do I still pay when the instance is stopped?
Yes — you are paying for held hardware, not for running minutes. That is the trade: the capacity is there at 3 a.m. on a Sunday whether you use it or not. If your usage is bursty, the spot and on-demand tiers are cheaper.
Can I mix reserved with spot?
That is the usual shape: reserved capacity for the steady load, spot for everything that can wait. Your reservation does not reduce what you can take from the spot or on-demand pools.
What if you cannot supply the capacity?
Then we do not sell it. A reservation is only confirmed once the hardware is allocated, and the confirmation names the model, the count, the region and the dates. If we fail to deliver a confirmed reservation, the term is cancelled and the month is refunded in full.
Is there a discount for longer terms?
For terms beyond a month and for large counts, yes — tell us the shape of what you need and we quote it. The published rate is the one-month rate, which is already below our own on-demand price.