Compute
GPU capacity, defined by your workload.
Start with the hardware and time you need. Bring the details that matter into one capacity request.
The specification matters.
A GPU model and an hourly price only tell part of the story. A useful offer defines the complete resource and how it will be delivered.
- Hardware
- GPU model, variant, memory, and count.
- Topology
- GPUs per server and connectivity between nodes.
- Location
- Region, data locality, and network access.
- Time
- Start date, duration, and delivery window.
- Environment
- Bare metal or virtualized access, software, and scheduler requirements.
- Commercial terms
- Full commitment, payment schedule, support, and remedies.
Different workloads. Different requirements.
Training
Connected GPUs, sustained performance, and a capacity window long enough to complete the run.
Fine-tuning & research
Enough memory for the model, an appropriate environment, and room for iteration.
Inference
Capacity that fits serving needs, latency requirements, and the expected demand profile.
Before you reserve.
What should I include in a capacity request?
Include your preferred GPU model, GPU count, memory needs, region, start date, duration, and workload. For multi-node training, include networking and cluster requirements. If you are unsure, describe the workload and the constraints you already know.
How is GPU compute priced?
Capacity is often quoted per GPU-hour, per node-hour, or for a fixed reservation term. Compare the full commitment, GPU configuration, networking, storage, support, and other charges—not just the headline hourly rate.
Is a reservation the same as a running machine?
No. A reservation is a commercial allocation of capacity. A running machine is an operational resource that has been provisioned and made accessible. An agreement should state when and how access will be delivered.
Start with the compute you need.
Tell us about your hardware, workload, and timing.