How Should You Compare Managed B200, B300, and NVL72 Quotes When Delivery Dates, Minimum Commitments, and Capacity Guarantees Differ?

Two quotes at the same hourly rate produce different totals once billing starts at different events, one at contract signature and the other when your team can run jobs. Compare managed Blackwell quotes on one project timeline instead of on rate: place each delivery window against your need-by date, count the GPU-hours each minimum commitment obliges you to pay for, price the interim capacity a late delivery would force you to buy, and record what each provider actually guarantees. On GMI Cloud today, the Blackwell lineup (B200, B300, GB200 NVL72, GB300 NVL72) is taken as pre-orders while H100 and H200 can be run immediately, and our pricing covers both on-demand use and committed structures, so our sales team can price a Blackwell pre-order and the H100 or H200 capacity that would run before it side by side.

For teams that need Blackwell capacity on a firm project date, that pairing is what makes a provider worth shortlisting: a committed rate for the rack or nodes you plan around, plus capacity you can run on before they arrive. A lower rate tied to a later date or a larger minimum often costs more over the project.

Read delivery status as a date range

For each quote, record five things next to the rate:

• Status. Available now or pre-order.
• Delivery unit. An NVL72 rack is a single 72-GPU NVLink domain, while HGX B200 and B300 systems come eight GPUs to a node. Ask whether a late rack can be handed over in part or only as a whole, and whether node orders can arrive in tranches.
• Earliest and latest handover date. Use the latest date in every calculation.
• Slack. Your need-by date minus the latest handover date and the validation time you plan after handover. Negative slack is a number of weeks you'll have to cover with other capacity.
• Quote validity. The date the rate and terms expire. A pre-order quote that expires before the provider can confirm allocation needs its validity extended in writing.

Billing start and minimum commitment set what you pay for

The billing-start clause decides how many paid hours fall before your workload runs. Ask every bidder which of these events starts the meter:

For rack-scale capacity, write the acceptance test into the contract as a full-load run across every node, including multi-node collective bandwidth, held for a stated number of hours without hardware errors.

Minimum commitments come in a few forms: a GPU count held for a term (take-or-pay), a total spend over the term, or a ramp in which the committed count rises on set dates. For a fixed GPU count:

Committed GPU-hours = committed GPUs _ hours in the term

A ramp is summed stage by stage, each stage's GPU count multiplied by its hours. A total-spend commitment converts to GPU-hours only at a fixed system and rate, so a spend pool shared across B200, B300, and NVL72 is tracked in spend rather than in hours. Then compute what each used hour really costs:

Effective rate = committed spend ÷ GPU-hours your workload actually uses

Committing to 72 GPUs (one NVL72 rack) when the job needs 64 raises the effective rate by 72 ÷ 64, or 12.5%, so a 10% lower list rate becomes a slightly higher effective one. A ramp schedule that tracks your real growth curve does the reverse: it keeps committed hours close to used hours, so the effective rate stays near the committed rate.

Price delay and substitute hardware before they happen

A delivery slip has two costs: the interim GPU-hours you buy on hardware available now, and whatever a later launch costs your business. Keep the second as its own line rather than folding it into the GPU math.

Size interim capacity against the same acceptance metric, not the same GPU count. If the interim hardware needs more GPUs to hit your step time or p95 latency, the interim GPU-hours rise in proportion. Ask, too, whether interim spend can be credited toward the Blackwell commitment.

Substitute hardware offered at handover gets the same treatment. When a provider offers B200 nodes in place of B300, work out how many B200 GPUs it takes to pass your acceptance test. B300 carries more HBM per GPU, so a memory-bound model may need extra B200 GPUs. Multiply the GPU-hours accordingly and compare that total with the original quote. Agree in writing how the substitute is billed: at the committed rate per GPU, or at a rate adjusted so that the extra GPUs needed to pass the test don't raise your bill. Settle as well whether the committed term pauses until the original system arrives.

What a capacity guarantee actually covers

Score each quote only on what its clauses commit the provider to:

A quote with no term and a short notice period can be stopped soon after you give notice. A twelve-month take-or-pay commitment may leave the remaining months due if the project is cancelled halfway through, and that figure belongs next to the rate when the project itself is uncertain.

A worked comparison on one timeline

The inputs below are hypothetical, chosen to show the method. Costs are expressed in cost units (GPU-hours _ a rate index), not prices.

• The project needs 64 GPUs for 52 weeks (8,736 hours), starting at week 0.
• Quote A. Available now, 64 GPUs, no term or exit fee, billed until released, rate index 1.00.
• Quote B. Pre-order NVL72 rack, 72-GPU minimum, 52-week take-or-pay billed from handover, rate index 0.85.
• Both systems pass the same acceptance test at the GPU counts shown, so the comparison covers spend for equal work.
• Any interim capacity before Quote B's handover runs on Quote A terms.

Quote B's 15% rate advantage shrinks to about 4% on time, because the rack bills 72 GPUs for a 64-GPU job. Each week of slip adds interim capacity worth roughly 2% of Quote A's total, so B's lead is gone after a little over two weeks. Exposure moves the same way. If the project stops at week 26 and B arrived on time, half of B's future committed charges remain, depending on its termination clause. Because B bills from handover, a six-week slip leaves about three-fifths of its term still ahead at that point, and a 14-week slip more than three-quarters, while Quote A owes nothing further.

Change one input and the answer flips. If the workload is shaped for 72 GPUs, Quote A at 72 GPUs comes to about 629 thousand units against Quote B's 535 thousand, and an on-time rack is 15% cheaper. A lower rate wins when its shape matches the job and its latest delivery date falls inside the slack.

Blackwell pre-order and the capacity that bridges to it

Our Blackwell lineup is in pre-order, and commitment options for it are set with our sales team. Reserved capacity and sustained deployments lower unit GPU cost. H100 and H200 are available now on dedicated, single-tenant GPUs, which is the interim capacity in the method above.

Managed GPU Cluster (early access) is requested through the console and reviewed by GMI support before resources are provisioned, as the GPU Compute documentation describes. A cluster can run on per-minute pay-as-you-go billing or on a prepaid plan, and growing an existing cluster is a Node Request.

If part of the commitment is for serving, Prime Inference has no minimum contract: on-demand billing is hourly per GPU, reserved capacity is available on a seasonal or annual basis at lower per-hour rates, and B300 is one of its hardware options. Qualified prospects receive free GPU-hour trial credits, so model performance can be measured on your own workload before you pick a term.

FAQ

How should a pre-order deposit enter the comparison?

Treat it as cash paid before handover, and have the contract state whether it is credited against the committed spend, refunded, or forfeited if delivery slips past a stated date. A deposit credited against the commitment changes timing but not the total, while one that isn't credited raises the effective rate.

Can quotes with different term lengths be compared fairly?

They can, once each is converted to cost units over your project window rather than over its own term. Hours a longer term bills after your project ends count as cost unless you have a planned workload to put on them.

Bring your need-by date to sales

Contact our sales team with your need-by date, the GPU count your workload is shaped for, and the longest term you can commit to. Ask for the Blackwell pre-order and any interim H100 or H200 capacity to be quoted together, so both sit on the timeline you use for every other bid.