Flexible Compute Pricing
Pay by the second for exactly the GPU capacity you use, or reserve ahead for a lower rate. Scheduling overhead, queue time, and failed placements are never billed.
Three ways to buy GPU capacity
Rates are per GPU-hour in USD. Reserved pricing reflects a 12-month commitment — see the commitment terms below for shorter and longer options.
| Resource Type | Spec | On-Demand Price | Reserved Price | Availability |
|---|---|---|---|---|
HGX B300 NVL8New 8× NVIDIA Blackwell Ultra B300 |
| Contact Sales | Contact Sales | Q4 2026 Reservation open |
Virtualized GPU slice A100 / H100-class |
| $0.90/GPU·hr | $0.60/GPU·hr | Available |
Dedicated GPU node Full 8-GPU server, single tenant |
| $2.40/GPU·hr | $1.70/GPU·hr | Available |
Prices are in USD and exclude VAT, GST, and local taxes. Figures shown are indicative list rates and may change; contact sales for a binding quote covering your region, term, and volume.
The longer you commit, the less you pay
Per-second metering with no commitment. Start and stop whenever you need capacity.
Capacity held for you at a fixed rate. Suited to recurring training cycles and steady inference traffic.
The rate shown in the Reserved column above. Best value for production workloads with predictable baseline demand.
Negotiated rates, dedicated capacity blocks, and co-located deployments. Includes an enterprise support plan.
What the rate covers
Included in every GPU-hour
- Hardware-isolated GPU slices with enforced QoS
- Per-second metering with signed usage records
- Slice-level telemetry and performance dashboards
- Private networking between AGGX nodes — no inter-node egress charges
- Container and Kubernetes-native scheduling
- Standard support with a 99.5% monthly uptime target
Billed separately
- Persistent block and object storage, billed per GB-month
- Public internet egress beyond the monthly free allowance
- Premium support plans with named engineers and faster response targets
- Custom co-location, bare-metal, or air-gapped deployments
Everything billable appears as a separate line item on your invoice. There are no platform fees, orchestration fees, or minimum spend requirements.
Before you talk to sales
- How is usage measured?
- Every slice is metered per second from the moment it is scheduled until it is released. Scheduling overhead, queue time, and failed placements are not billed. Each billing period is backed by signed usage records you can reconcile against your own telemetry.
- What is the difference between a slice and a dedicated node?
- A virtualized slice is a hardware-isolated partition of a GPU with a guaranteed share of memory and compute — ideal for inference and smaller training jobs. A dedicated node is a full single-tenant 8-GPU server with root access and no co-tenancy, for large distributed training or workloads with strict compliance requirements.
- What does “Reservation Open” mean for HGX B300?
- B300 capacity is being deployed for Q4 2026. You can reserve a block now to lock allocation and pricing ahead of general availability. Reservations are confirmed in writing and do not bill until capacity is delivered.
- Can I mix on-demand and reserved capacity?
- Yes. Reserved capacity covers your baseline and on-demand absorbs peaks. Usage above your reserved allocation is billed automatically at the on-demand rate, with no separate contract required.
- Do you offer credits for research or early-stage teams?
- We run a limited allocation program for academic research groups and early-stage AI companies. Contact our team with a short description of your workload and we will review eligibility.
Get a quote for your workload
Tell us your model sizes, throughput targets, and region requirements. We will size the deployment, quote a binding rate, and run a benchmark on your own workload before you commit.