Runs until you stop it.
The on-demand tier is the same GPU on the same node as spot, with one promise added: we will not take it back. It carries the 99.9% monthly uptime target, is billed by the minute like everything else, and still lands about 33% below the cheapest on-demand rate we could find elsewhere.
- Interruptions
- NoneCapacity pressure never stops an on-demand instance
- Uptime target
- 99.9%Monthly, with service credits — see the SLA
- From
- $0.05/hBilled per started minute
- Against the market
- −33%Average gap to the cheapest on-demand price found
One question decides it: can the job survive a 2-minute warning?
If it can, spot costs 38% less on an H100. If it cannot, the difference is the price of certainty.
| Spot | On-demand | Reserved | |
|---|---|---|---|
| H100 SXM, one GPU | $0.55/h | $0.89/h | $0.65/h |
| Can be reclaimed | Yes, with 2 min notice | Never | Never |
| Uptime target | None | 99.9% monthly | 99.9% monthly |
| Commitment | None | None | 1 to 12 months |
| Capacity guaranteed in advance | No | Subject to availability | Yes, held for the term |
| Billing | Per minute | Per minute | Monthly, in advance |
| Best for | Training, batch, sweeps, CI | Serving, interactive work, deadlines | Steady production load |
All three tiers share the images, the disks, the regions and the API. Moving between them is a setting, not a migration. How spot reclaims work · How reserved works
Every model, at the never-reclaimed rate.
Per GPU-hour, billed per minute, identical in 3 regions. The last column is the cheapest on-demand price we found for the same GPU anywhere else.
| GPU | Memory | On-demand | Per minute | Spot | Cheapest elsewhere | Action |
|---|---|---|---|---|---|---|
| Data center | ||||||
| B200 SXMBlackwell | 180 GB | $2.49 | $0.0415 | $1.69 | $3.12 | Deploy |
| H200 SXMHopper | 141 GB | $1.09 | $0.0182 | $0.75 | $1.39 | Deploy |
| H200 NVLHopper | 141 GB | $0.99 | $0.0165 | $0.69 | $2.45 | Deploy |
| H100 SXMHopper | 80 GB | $0.89 | $0.0148 | $0.55 | $1.15 | Deploy |
| H100 NVLHopper | 94 GB | $0.79 | $0.0132 | $0.49 | $1.11 | Deploy |
| H100 PCIeHopper | 80 GB | $0.69 | $0.0115 | $0.45 | $1.49 | Deploy |
| GH200Hopper | 96 GB | $0.79 | $0.0132 | $0.49 | $1.99 | Deploy |
| A100 SXM 80GBAmpere | 80 GB | $0.29 | $0.0048 | $0.19 | $0.40 | Deploy |
| A100 PCIe 80GBAmpere | 80 GB | $0.19 | $0.0032 | $0.09 | $0.27 | Deploy |
| A100 40GBAmpere | 40 GB | $0.17 | $0.0028 | $0.09 | $0.47 | Deploy |
| MI325XCDNA 3 | 256 GB | $1.59 | $0.0265 | $1.09 | $2.00 | Deploy |
| MI300XCDNA 3 | 192 GB | $1.45 | $0.0242 | $0.99 | $1.85 | Deploy |
| L40SAda Lovelace | 48 GB | $0.25 | $0.0042 | $0.15 | $0.35 | Deploy |
| L40Ada Lovelace | 48 GB | $0.19 | $0.0032 | $0.09 | $0.34 | Deploy |
| A40Ampere | 48 GB | $0.09 | $0.0015 | $0.06 | $0.12 | Deploy |
| A30Ampere | 24 GB | $0.08 | $0.0013 | $0.05 | $0.10 | Deploy |
| A10Ampere | 24 GB | $0.15 | $0.0025 | $0.09 | $0.20 | Deploy |
| L4Ada Lovelace | 24 GB | $0.08 | $0.0013 | $0.05 | $0.11 | Deploy |
| T4Turing | 16 GB | $0.07 | $0.0012 | $0.04 | $0.10 | Deploy |
| V100 32GBVolta | 32 GB | $0.09 | $0.0015 | $0.06 | $0.12 | Deploy |
| V100 16GBVolta | 16 GB | $0.08 | $0.0013 | $0.04 | $0.10 | Deploy |
| Workstation | ||||||
| RTX PRO 6000 BlackwellBlackwell | 96 GB | $0.39 | $0.0065 | $0.19 | $0.54 | Deploy |
| RTX PRO 5000 BlackwellBlackwell | 48 GB | $0.29 | $0.0048 | $0.15 | $0.66 | Deploy |
| RTX PRO 4500 BlackwellBlackwell | 32 GB | $0.19 | $0.0032 | $0.09 | $0.29 | Deploy |
| RTX PRO 4000 BlackwellBlackwell | 24 GB | $0.15 | $0.0025 | $0.08 | $0.20 | Deploy |
| RTX 6000 AdaAda Lovelace | 48 GB | $0.19 | $0.0032 | $0.09 | $0.34 | Deploy |
| RTX 5000 AdaAda Lovelace | 32 GB | $0.17 | $0.0028 | $0.09 | $0.27 | Deploy |
| RTX 4500 AdaAda Lovelace | 24 GB | $0.15 | $0.0025 | $0.09 | $0.20 | Deploy |
| RTX 4000 AdaAda Lovelace | 20 GB | $0.09 | $0.0015 | $0.06 | $0.16 | Deploy |
| RTX A6000Ampere | 48 GB | $0.17 | $0.0028 | $0.09 | $0.28 | Deploy |
| RTX A5000Ampere | 24 GB | $0.09 | $0.0015 | $0.06 | $0.16 | Deploy |
| RTX A4500Ampere | 20 GB | $0.09 | $0.0015 | $0.06 | $0.13 | Deploy |
| RTX A4000Ampere | 16 GB | $0.05 | $0.0008 | $0.03 | $0.07 | Deploy |
| Consumer | ||||||
| RTX 5090Blackwell | 32 GB | $0.19 | $0.0032 | $0.09 | $0.25 | Deploy |
| RTX 5080Blackwell | 16 GB | $0.09 | $0.0015 | $0.06 | $0.15 | Deploy |
| RTX 5070 TiBlackwell | 16 GB | $0.09 | $0.0015 | $0.06 | $0.12 | Deploy |
| RTX 4090Ada Lovelace | 24 GB | $0.09 | $0.0015 | $0.06 | $0.16 | Deploy |
| RTX 4080 SUPERAda Lovelace | 16 GB | $0.09 | $0.0015 | $0.06 | $0.16 | Deploy |
| RTX 4080Ada Lovelace | 16 GB | $0.09 | $0.0015 | $0.06 | $0.14 | Deploy |
| RTX 4070 Ti SUPERAda Lovelace | 16 GB | $0.09 | $0.0015 | $0.06 | $0.12 | Deploy |
| RTX 4070 TiAda Lovelace | 12 GB | $0.08 | $0.0013 | $0.05 | $0.10 | Deploy |
| RTX 4070Ada Lovelace | 12 GB | $0.06 | $0.0010 | $0.04 | $0.08 | Deploy |
| RTX 3090 TiAmpere | 24 GB | $0.09 | $0.0015 | $0.06 | $0.16 | Deploy |
| RTX 3090Ampere | 24 GB | $0.08 | $0.0013 | $0.05 | $0.10 | Deploy |
| RTX 3080 TiAmpere | 12 GB | $0.07 | $0.0012 | $0.04 | $0.10 | Deploy |
| RTX 3080Ampere | 10 GB | $0.06 | $0.0010 | $0.03 | $0.09 | Deploy |
| RTX 3070Ampere | 8 GB | $0.05 | $0.0008 | $0.03 | $0.07 | Deploy |
Work that cannot be paused politely.
Serving a model
An inference endpoint with a latency budget cannot absorb a 2-minute warning and a cold start. Keep the baseline on-demand and put the overflow, the batch scoring and the nightly fine-tune on spot.
Interactive sessions
A notebook you are actually sitting in front of, a debugging session on a large model, a pair of hours with a customer on a call. Losing the process costs more than the hourly difference.
Deadlines
The last run before a release, a benchmark that must finish tonight, a demo at a fixed hour. Spot is cheaper on average; on-demand is cheaper when a missed slot costs a day.
The same persistent disk can be attached to a spot instance today and an on-demand instance tomorrow, with the same template and the same entrypoint. Teams usually start a job on spot and switch it to on-demand when a deadline appears.
On-demand, in detail
Is an on-demand instance ever stopped by you?
Only for a reason we would have to explain: an unpaid balance, a breach of the acceptable use policy, or an emergency affecting the physical node. Capacity pressure is never a reason — that is the entire difference between this tier and spot.
What uptime do you commit to?
A monthly availability target of 99.9% per instance, with service credits below it. The measurement, the exclusions and the credit table are in the service level agreement.
Can I move a running spot job to on-demand?
Yes, and without copying data. Stop the spot instance (or let the reclaim stop it), then launch an on-demand instance with the same persistent disk and the same template. The job resumes from its checkpoint on a GPU that will not be taken back.
Is on-demand billed differently from spot?
No. Per started minute, at the hourly price divided by 60, from the moment the instance is reachable until you stop it. No minimum duration, no hourly rounding, no reservation fee.
Why is your on-demand cheaper than other providers' on-demand?
Because we buy installed capacity rather than build it, and because the catalogue is priced against the cheapest published rate for the same GPU across 137 providers — at least 20% below it, checked by the script that publishes prices.
Do I get the same hardware as on spot?
The same node, the same GPU model, the same interconnect, the same images, the same regions. The tier is a commercial promise about keeping the machine, not a hardware difference.
An H100 that nobody takes back, for $0.89 an hour.
Pay as you go — no contracts, no minimum commitment. Add credit, launch, stop whenever you want.
Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription.