Scheduling
On-demand GPUs are a queue, not a hope. The scheduler places jobs on free nodes, respects reservations, and can land farm frames on spot when the pool is idle. Median time to a schedulable on-demand GPU is under two minutes.
Pick a SKU, a mode and a run length
Illustrative USD before tax. Spot is typical, not a guarantee. Reserved discounts are the published 18 / 24 / 28 / 31% schedule.
How this actually runs
Scheduling is a queue, not a hope. Reserved and bare metal are out of the pool; the scheduler will not place a stranger on your term. On-demand is FIFO with fair-share so one account cannot starve everyone else. Spot is whatever is still idle, reclaimed with five minutes’ notice if a higher class needs the node. Farm frames use the same free-node logic, with priority as a surcharge not a separate fabric.
H100 / A100 density on InfiniBand for training and HPC. L40S / 4090 density on 100 GbE for render and inference. One invoice, one rate card. Data stays where you provision it unless you ask otherwise.
Support has no tier-1 layer. Median first response on a production incident is under fifteen minutes. 99.9% SLA and 24/7 on-call attach to six- and twelve-month reserved and all bare metal. One- and three-month reserved are business hours. On-demand and spot are best effort: we replace failed hardware, we do not guarantee a free slot existed.
Billing is USD. On-demand in arrears, per second, sixty-second minimum. Reserved and bare metal at the start of each term. NVMe, object storage and static IP are separate lines. Engineering and advertising retainers are a different invoice. Taxes follow the entity on the contract and applicable local law.
Who gets the next free card
Those GPUs are out of the pool. The scheduler will not place a stranger’s on-demand job on them.
FIFO with fair-share so one account cannot starve the queue. Priority render surcharge jumps farm frames, not arbitrary compute.
Whatever is still idle after the above, with a five-minute reclaim if a higher class needs the node.
Illustrative month · time to first response (min)
What changes when you change the commercial model
| On-demand | Reserved | Spot | Bare metal | |
|---|---|---|---|---|
| Term | None | 1–12 months | None | 1–12 months |
| Price | Public $/hr | 18–31% off | ~70% off typical | Reserved + isolation |
| Kill / reclaim | You stop it | You stop it | 5-minute notice | You stop it |
| SLA | Best effort | 99.9% on 6–12 mo | None | 99.9% + 24/7 |
| Isolation | Hypervisor | Hypervisor | Hypervisor | Single tenant |
| Good for | Burst, tests | Steady baseline | Checkpoint batch | Train / prod inference |
Published SKUs, both sites
Full table including reserved discount math lives on Pricing. Spot is typical. Fabric is InfiniBand at InfiniBand fabric, or 100 GbE (25 GbE on smaller cards) depending on the SKU.
| Instance | Site | VRAM | CPU / RAM | Fabric | On-demand | Spot typ. |
|---|---|---|---|---|---|---|
| H100 PCIe 80GB | InfiniBand | 80GB HBM3 | 24 / 200 GB | InfiniBand | $2.19 | $0.66 |
| A100 PCIe 80GB | InfiniBand | 80GB HBM2e | 16 / 180 GB | InfiniBand | $1.24 | $0.37 |
| L40S 48GB | 100 GbE | 48GB GDDR6 | 12 / 128 GB | 100 GbE | $0.86 | $0.26 |
| RTX 4090 24GB | 25 GbE | 24GB GDDR6X | 8 / 64 GB | 25 GbE | $0.41 | $0.12 |
| RTX A6000 48GB | Multi-site | 48GB GDDR6 | 12 / 96 GB | 25–100 GbE | $0.95 | $0.29 |
| CPU render node | Multi-site | 64 / 256 GB | 25 GbE | $0.68 | $0.20 |