GPU compute and render capacity, without the hyperscaler contract
We own 312 GPUs and rent them by the second: on-demand, reserved, bare metal, or as a managed render farm. One account, one rate card, whether you need a single card for an afternoon or a cluster for a quarter.
One rate card for every account
A single GPU pays the same per-hour rate as a sixty-four-GPU reservation. There is no sales-negotiated pricing tier and no volume discount you have to ask for; reserved terms apply automatically the moment you commit to one.
Every instance is billed per second with a 60-second minimum, in USD, and stops accruing the moment you stop it. Bare metal nodes bill the same per-GPU-hour rate as their on-demand equivalent, with a flat provisioning fee waived on terms of one month or longer.
Reserved discount ladder
Storage and networking
Your scene file is the only integration you need
The render farm runs on the same fleet as everything else, so pricing never leaves the per-GPU-hour rate card. There is no separate per-frame black box; a render job is billed exactly like any other workload on the node types it lands on.
Studios usually run render jobs in the Priority or Reserved block queue against L40S or RTX 4090 nodes; CPU-bound renders and simulation caches use the CPU render node pool. Output writes straight to object storage so a farm job and a training job can share the same bucket.
Four ways to buy the same hardware
A single GPU is schedulable in under two minutes, billed per second and stopped whenever you stop it.
- No minimum term or upfront commitment
- Spin up from a base image or your own snapshot
- Same rate whether you rent one GPU or sixty-four
One to twelve month terms at 18 to 31% off list, with the node held for you whether or not you use it.
- Discount applies automatically at checkout
- Convert an on-demand instance to reserved without a restart
- Renews on the same terms unless you cancel with 30 days' notice
Whole nodes with root, your own kernel and network policy, no hypervisor between you and the hardware.
- Data residency stated in the contract
- Bring your own OS image or use ours as a starting point
- InfiniBand-connected clusters available on reserved terms
Submit a scene, pay per frame: the same node rates as everything else, with a managed queue on top.
- Blender, Houdini, Maya, Cinema 4D and Redshift supported natively
- Standard, Priority and Reserved-block queues
- Output writes directly to object storage as frames complete
Hardware we own, in power contracts we hold
Every GPU in the fleet sits in a site we have a direct power contract for; nothing here is colocated capacity resold at a markup.
An engineer answers, not a queue
Every account, on-demand or bare metal, is supported by the same team that runs the fleet. There is no tiered ticket system and no offshored first line; the person who answers already knows how the scheduler behaves under load.