• Compute
  • Customers
  • Pricing
Sign In
More Blog Posts
XDiscordLinkedInYouTube

Products

  • GPUs
  • Inference
  • Studio

Developers

  • Model library
  • Documentation
  • Glossary

Company

  • About Us
  • Blog
  • Events
  • Partnership
  • Scale
  • Career
  • Ambassador program
  • Mission & Vision

Popular models

    Stay in the loop

    By submitting, you acknowledge that we may collect and use the information you provide, which may include personal information.

    XDiscordLinkedInYouTube

    Copyright ©2026 All rights reserved.

    Privacy PolicyTerms of UseLegal Documentation
    More Blog Posts
    Other

    What Acceptance Criteria and Capacity Commitments Should You Agree Before Moving From a Small GB300 NVL72 Pilot to a Full-Rack Reservation?

    November 01, 2026

    A pilot can only pass if "pass" was written down as numbers before the first job ran, and the same document should say what passing triggers: how many full racks are held for you, for how long, at what configuration and price, and what either side may do if the pilot fails or a date moves. Agree four things before the pilot starts: measurable acceptance criteria, a capacity hold you can exercise by notice, a validity period for the configuration and the price, and exit terms. GMI Cloud takes GB300 NVL72 on pre-order through our sales team, and commitment options are agreed in that same conversation, so a buyer can table all four terms alongside the reservation request. More broadly, GMI pricing lets workloads move from on-demand to dedicated or committed deployments as they stabilize, without being pushed into a long-term commitment too early.

    Without these terms, a successful pilot leaves procurement with a good engineering report and still no claim on a rack, no agreed rate, and no date.

    Acceptance criteria written as numbers

    Each criterion needs a metric, a threshold, a measurement method, and a named person who records the result. The thresholds below are placeholders to show the format; yours come from your own capacity plan.

    Area Metric Example threshold (assumed) Recorded by
    Functional Agreed job list runs end to end on the pilot trays with the pinned image Every listed job completes Customer
    Throughput Tokens or samples per second per GPU at fixed batch size and precision At least 90% of the per-GPU figure in the capacity plan Customer, with provider witness
    Latency, for inference p95 time to first token at the target concurrency 800 ms or less at 64 concurrent sessions Customer
    Stability Continuous run on the pilot trays 72 hours with no hardware-caused interruption Both sides
    Data path Checkpoint restore time or dataset read rate Restore within 10 minutes Customer
    Operations A support request handled through the agreed process Opened, acknowledged, and closed as the support terms describe Both sides

    Four details keep the table from turning into a dispute later.

    • A fixed method. The model version, precision, dataset, container image, and load pattern are frozen when the criteria are signed, so the result measures the hardware and service, not a moving software target.
    • Evidence. Raw logs and the measurement scripts go into the acceptance report next to the numbers.
    • A retest window. If a run fails for a provider-side reason, a defined period to fix and rerun, for example 10 business days.
    • Scope. Criteria cover what the pilot allocation can actually measure. Behavior across the full 72-GPU NVLink domain is checked in a separate acceptance event on the rack itself.

    What each pilot outcome triggers

    The acceptance report is only useful if each possible result has a contractual consequence agreed in advance.

    Outcome Customer position Provider position Next step
    Pass May reserve the held racks within the exercise window Keeps the racks held until the window closes Reservation signed on the pre-agreed terms
    Fail, provider-side cause No obligation to reserve Fixes the cause, retests, and extends the hold by the retest time Rerun the failed criteria
    Fail, workload-side cause No obligation to reserve Hold lapses at the end of the window Re-scope, for example a different rack count, or stop
    Inconclusive Pilot extended Pilot trays and hold extended together Finish the remaining tests

    The line between provider-side and workload-side causes is where disagreements start, so agree beforehand on the evidence that settles it: the hardware and fabric logs the provider will share with you, the provider's incident records, and your own job logs on the same timeline. A throughput miss from a failing GPU and one from an untuned input pipeline look alike in a summary table and very different in the logs.

    From intent to commitment: what the capacity line should say

    A sales note that says full racks are "expected to be available after the pilot" records intent. It gives the provider nothing to hold and gives you nothing to exercise. A capacity commitment names a quantity, a hold period tied to the acceptance date, a way to exercise it, and what happens if you don't.

    Here is an example of a clause a buyer can propose for the contract. It is a drafting sample, not GMI's standard terms.

    Example clause: "Provider will hold two (2) GB300 NVL72 racks for Customer from the pilot start date until thirty (30) days after the Acceptance Date (the Hold Period). The Acceptance Date is the day both parties sign the pilot acceptance report, whether it records a pass or a fail, or, if no report is signed, fifteen (15) days after the scheduled pilot end date. A provider-side retest, or a pilot extension agreed for an inconclusive result, extends the Hold Period by its length. Customer may reserve the held racks by written notice during the Hold Period, on the term, rate, and configuration set out in Schedule B. If Customer gives no notice, the hold ends without charge."

    A right of first refusal works differently. It gives you priority when another buyer makes an offer, and unless the clause says more, it sets nothing aside before that happens. A fee or deposit doesn't strengthen a hold by itself either. The strength sits in what the clause obliges: whether named capacity is reserved for you, whether you can exercise without further consent from the provider, whether the provider may withdraw, and what happens to any fee on each outcome in the table above, including racks that can't be delivered. If a fee applies, the contract should also say whether it is credited against the reservation or can be applied to other usage when the pilot ends without a reservation.

    Freeze the configuration and the price for the decision window

    A pass on one configuration says little about a different one, so the reservation should describe the same system you piloted. Attach a configuration manifest at pilot end covering the system, the host architecture, the software baseline, the access model, the fabric type, the storage tier, and the region. Any change between pilot and rack, such as a different driver baseline or a different site, should trigger a rerun of the criteria it affects.

    Price needs the same treatment. Schedule B should state the committed rate together with the term it assumes and keep both valid until the hold period ends, because the rate depends on that term: GMI pricing prices on-demand and committed usage differently. The fabric and storage lines that come with the racks need their own rates and the same validity date in that schedule. Leave them out, and a held rack can arrive with a compute rate that held and a storage line that moved.

    A worked timeline makes the dates concrete. Assume a four-week pilot that starts in week 1. The acceptance report is signed in week 5, the 30-day hold runs to roughly week 9, and the reservation signed inside that window carries its own delivery window and acceptance test for the racks. If the report stalls, the clause's fallback date starts the 30 days anyway.

    If the rack date moves after the pilot passes

    Delivery clauses for any Blackwell reservation, such as slip notice and remedies, apply here as usual. A pilot-to-rack path adds a few items of its own.

    • Pilot continuation. The right to keep the pilot trays on the pilot terms until the full rack passes its own acceptance, so the team keeps working while it waits.
    • A long-stop date. A final date after which you may cancel the reservation, with any hold fee or advance payment returned.
    • Re-acceptance on the rack. The pilot criteria, scaled to the rack, plus the tests only a full NVLink domain can answer, run before the committed term starts. The clause should also say whether a rack-level miss opens a fix-and-retest period or moves straight to the long-stop rights.
    • Support across the transition. Which support terms cover the pilot trays while you wait, and when SLA coverage starts on the racks. GMI's GPU infrastructure comes with SLA-backed service delivery, and tying that start date to rack re-acceptance keeps the committed term and the coverage aligned.

    Handoff from the pilot team to procurement

    Once the pilot passes, the workload team hands procurement a package that lets the reservation be signed without reopening technical questions.

    1. The signed acceptance report, with measured values against each threshold and the raw logs
    2. The configuration manifest frozen at pilot end
    3. The capacity line: rack quantity, hold period, and how notice is given
    4. The price sheet: committed rate and term, fabric and storage lines, and their validity date
    5. The delivery window, rack acceptance test, and long-stop date for the reservation
    6. Support scope and escalation contacts for the racks
    7. A named owner on each side for the transition from pilot trays to racks

    FAQ

    Who should sign the pilot acceptance report?

    Two engineers who watched the tests run, one from your workload team and one from the provider, should sign the measured values. A commercial owner on your side then signs the decision to exercise or not. Splitting the signatures keeps the technical record separate from the buying decision, which helps if the hold or retest terms are ever questioned.

    Can the acceptance criteria change once the pilot is running?

    They can, through a written amendment both sides sign before the affected test runs. A threshold changed after a result is known turns the pilot into a negotiation. If your capacity plan shifts mid-pilot, update the plan, re-derive the threshold, and amend the criteria before rerunning.

    Bring your pass mark to the pre-order conversation

    Bring the rack count you expect to reach, a draft of your acceptance criteria, and the dates your project depends on. With those in hand, our sales team can discuss GB300 NVL72 pre-order and commitment options against your own pass mark, so the pilot result has a clear place to land. Contact GMI Cloud sales to start the conversation.

    Colin Mo

    Build AI Without Limits

    GMI Cloud helps you architect, deploy, optimize, and scale your AI strategies

    FAQ

    Two engineers who watched the tests run, one from your workload team and one from the provider, should sign the measured values. A commercial owner on your side then signs the decision to exercise or not. Splitting the signatures keeps the technical record separate from the buying decision, which helps if the hold or retest terms are ever questioned.

    Ready to build?

    Explore powerful AI models and launch your project in just a few clicks.

    Get Started