Reserved GPU pricing
Compare reserved rates by platform, GPU count, contract term and advance. Calculate the initial payment, monthly cost and total commitment.
B300 base rates by term
Indicative base rates. Final pricing depends on the agreed configuration and service scope.
| Term | Per GPU-hour | Per GPU-month¹ | Advance |
|---|---|---|---|
| 12 months | US$5.25 | US$3,833 | 20% |
| 24 months | US$4.95 | US$3,614 | 20% |
| 36 months | US$4.75 | US$3,468 | 20% |
| 48 months | US$4.60 | US$3,358 | 20% |
¹ 730 hours. Minimum 32 B300 GPUs. Volume adjustments: 128+ GPUs −1%, 256+ −2%, 512+ −4%. Advance adjustment: 30% −2%, 10% +2%, none +4%. Base rates assume a 20% advance.
Estimate your cluster cost
Choose a platform, GPU count, term and advance to calculate the estimated cost.
US$4.66/ GPU-hour
Indicative · fixed-term modelIllustrative comparison based on the assumptions selected.
On-demand is billed only for the hours used. Reserved capacity is billed for all contracted hours.
730 hours per model month. Excludes tax, additional storage, transfer and separately scoped services. This is not a quotation or reservation.
Storage and additional services
| Service | Illustrative planning rate | Basis |
|---|---|---|
| On-demand B300 | US$6.90 / GPU-hour | Whole eight-GPU nodes · after reserved deployment |
| Parallel file storage | 2 TB / GPU included; then US$0.04 / GB-month | Final baseline and performance scoped in agreement |
| Object storage | US$0.02 / GB-month | Metered use |
| Data transfer | Scoped by connection and contract | Connection and included allowances agreed in advance |
| Managed inference | Published at launch | Planned product |
Commercial assumptions
- Term
- 12, 24, 36 or 48 months from acceptance in the planning model.
- Advance
- Model options of 0%, 10%, 20% or 30%; credited evenly across the term. Refund conditions belong in the signed contract.
- Billing
- Reserved GPU-hours are billed whether used or idle. Additional services are separately scoped.
- Acceptance
- Hardware, workload tests and the service window are agreed before handover.
- Service level
- Targets, exclusions, response windows and remedies are defined in the service schedule.
- Changes & exit
- Scaling, renewal, transfer and termination require agreed contract terms.
This page describes an illustrative commercial model. It does not establish binding service terms.
Pricing questions
Is reserved capacity always cheaper?
No. A reservation pays for all contracted GPU-hours. Low utilization may favor on-demand, and ownership has a different upfront cash and residual-value profile.
Does the advance add to the contract value?
In this model, no. It is part of the total commitment and is credited against monthly invoices. The planner shows both the initial advance and the net monthly amount.
Can I model GB300 NVL72?
Yes, as a separate roadmap scenario. The planner uses 72-GPU rack increments and platform-specific assumed rates. It is not currently available capacity.
Can I use this as a quote?
No. Use it to prepare a configuration for discussion. Hardware, delivery, scope, eligibility and final pricing must be confirmed together.