Skip to content

Reserved GPU capacity agreement

A customer reserves dedicated bare-metal GPU nodes for a committed term and pays for them whether used or not. Covers delivery and acceptance testing, take-or-pay pricing, prepayment, failed-node replacement, secure wipe and hardware substitution.

For operators · Updated Nov 2, 2026 · Example to adapt

An editable .docx with the fill-ins highlighted.

How this model works

How it works

  • A customer reserves a fixed set of GPU servers (nodes) in your facility for a fixed term, usually a year or more. The nodes are dedicated to that customer alone: no one else runs on them, and the customer gets direct, bare-metal access to the operating system.
  • You, as provider, deliver the cluster, prove it works through agreed acceptance tests, keep the hardware healthy, replace failed nodes quickly and keep the facility powered and cooled. The customer installs and runs its own software and is responsible for what it does with the cluster.
  • The customer pays for the whole reservation for the whole term, whether it uses the nodes or not. This committed revenue is what makes a cluster financeable: lenders and hardware suppliers look at the signed reservation before they release money or hardware.
  • On AICapacity, buyers describe the cluster they need (GPU model, count, interconnect, location, start date). The reservation agreement is usually signed after a letter of intent and once both sides have agreed the cluster specification and acceptance tests.

Good for

  • AI companies training large models, which need many GPUs connected by a fast network fabric for months at a time.
  • Operators buying or financing a new cluster who want a committed customer before ordering hardware.
  • Enterprises and cloud providers that need dedicated, isolated capacity in a particular country, for example for data residency.

How the money works

  • Prices are quoted either per GPU-hour or per node per month. They are the same money expressed differently: a month has about 730 hours, so a node with eight GPUs costs eight times the GPU-hour rate times 730 hours. Check which unit a buyer is comparing against before you quote.
  • Take-or-pay means the customer pays for every reserved hour, used or not. You carry no utilisation risk during the term, but you carry the customer’s credit risk instead, and the utilisation risk returns at renewal. That is why prepayments, deposits and credit checks matter.
  • Because the hardware is bought up front, providers commonly ask for some months of charges in advance, a security deposit, or a bank or parent-company guarantee. Agree how the money is held and when it is returned or set off.
  • For reserved compute the price usually includes electricity, so you carry the energy-price risk. A GPU rack draws many times the power of an ordinary server rack. Work out your energy cost per GPU-hour at full load, and use an energy adjustment or pass-through if your own power price is not fixed for the term.
  • Longer terms usually earn a lower rate, but GPU prices tend to fall as new generations arrive. Expect buyers to ask for upgrade options or repricing at renewal, and price the term with that in mind.

Watch out for

  • Do not promise a delivery date you do not control. Set the target delivery date and long-stop date from your hardware supplier’s confirmed dates; this draft does not treat supplier delays as force majeure.
  • The acceptance tests decide when billing starts. Agree the benchmarks, pass thresholds and reference figures before signing, not when the hardware arrives.
  • Failed-node replacement time is the commitment buyers watch most closely. Keep spare nodes on site; times that rely on shipping parts from abroad are hard to meet.
  • High-end GPUs are export-controlled goods. Know who your customer is, who its end users are and what the cluster is for, and keep the right to suspend if the answers change.
  • Drives must never leave your site with customer data on them, including failed drives returned under warranty. Plan secure wipe and destruction into your maintenance process.

An example of how this kind of deal can be structured, to adapt with your own advisers. Not legal advice and not a final agreement. AICapacity is not a party to agreements made from it.

Reserved GPU capacity agreement

This agreement is dated [date] and is made between:

(1) [Provider legal name], a company registered in [country] under number [registration number], whose registered office is at [address] (“Provider”); and

(2) [Customer legal name], a company registered in [country] under number [registration number], whose registered office is at [address] (“Customer”),

each a “Party” and together the “Parties”.

Background

(A) Provider operates the Facility and owns or controls GPU compute hardware installed in it.

(B) Customer wishes to reserve dedicated GPU compute capacity for a committed term on the terms of this agreement.

The Parties agree as follows:

  1. 1Definitions and interpretation

    1. 1.1

      In this agreement:

      1. (a)“Acceptance Tests” means the tests in Schedule 3, and “Acceptance Date” has the meaning given in clause 3.2;
      2. (b)“Business Day” means a day other than a Saturday, Sunday or public holiday in [country];
      3. (c)“Charges” means the charges for the Reserved Capacity and for any additional services set out in Schedule 2 (“Additional Services”);
      4. (d)“Cluster” means the Nodes, interconnect, storage and network equipment described in Schedule 1 (the “Cluster Specification”), as replaced or substituted under this agreement;
      5. (e)“Contract Year” means each period of twelve months starting on the Service Commencement Date or an anniversary of it;
      6. (f)“Customer Data” means all data, software, models, weights and other content that Customer or its users store on, process with or transmit through the Cluster;
      7. (g)“Facility” means the datacenter at [address] in which the Cluster is installed;
      8. (h)“Failed Node” means a Node that is unavailable, or cannot perform to the Cluster Specification, because of a hardware, firmware or Provider-side network fault;
      9. (i)“Node” means one GPU server in the Cluster, including its GPUs, processors, memory, storage and network adapters, and any Node that replaces it;
      10. (j)“Reserved Capacity” means the Nodes and associated resources in the Cluster Specification dedicated to Customer;
      11. (k)“Sanctions and Export Control Laws” means the sanctions and export, re-export and transfer controls of the European Union, its member states, the United Nations, [the United Kingdom], [the United States] and any other applicable jurisdiction, and “Sanctioned Person” means anyone targeted by them, directly or through ownership, control or location;
      12. (l)“Service Commencement Date” has the meaning given in clause 3.5, and “Term” means the Initial Term together with any Renewal Term under section 13; and
      13. (m)“SLA Schedule” means the service level schedule signed by the Parties and attached as Annex A.
    2. 1.2

      The Schedules and Annexes form part of this agreement. If there is a conflict, the clauses prevail, except that Schedule 2 prevails on amounts payable and the SLA Schedule on the measurement of service levels and service credits.

  2. 2Reserved Capacity

    1. 2.1

      From the Service Commencement Date and for the Term, Provider shall make the Reserved Capacity available to Customer on a dedicated basis. No Node shall be shared with or allocated to any other customer, and the Cluster’s network and interconnect shall be isolated from those of other customers. Provider shall not move any part of the Cluster outside the Facility without Customer’s consent.

    2. 2.2

      The services comprise the Reserved Capacity; space, power and cooling in the Facility; hardware monitoring, maintenance and Failed Node replacement under section 7; support and remote hands as described in Schedule 2; and any Additional Services Customer orders (together, the “Services”).

    3. 2.3

      Customer shall have administrative (root) access to the operating system of each Node. Provider retains access to the out-of-band management interfaces of the Cluster and shall use it only for monitoring, maintenance and support, in accordance with section 8.

  3. 3Delivery and acceptance

    1. 3.1

      Provider shall use [all] reasonable endeavours to have the Cluster ready by the target delivery date in Schedule 1 (the “Target Delivery Date”), reporting progress [weekly]. When the Cluster, or a batch of Nodes under Schedule 1, is ready, Provider shall run the Acceptance Tests and send Customer the results (a “Ready Notice”). Customer may observe and repeat the tests within [five] Business Days after the Ready Notice (the “Acceptance Period”).

    2. 3.2

      The “Acceptance Date” is the earliest of:

      1. (a)the date Customer confirms in writing that the Acceptance Tests are passed;
      2. (b)the end of the Acceptance Period, if Customer has not given a Rejection Notice under clause 3.3; and
      3. (c)the date Customer first uses the Cluster for production workloads.
    3. 3.3

      If any Acceptance Test is failed, Customer may, within the Acceptance Period, give written notice identifying the failed tests and Nodes (a “Rejection Notice”). Provider shall remedy the failure and repeat the failed tests within [ten] Business Days, and clauses 3.1 and 3.2 apply again.

    4. 3.4

      Customer may accept the Nodes that have passed while others are remedied. Charges for accepted Nodes start on the date of acceptance, pro rata.

    5. 3.5

      The “Service Commencement Date” is the Acceptance Date of the whole Cluster, and the Initial Term runs from it.

    6. 3.6

      If the Acceptance Date has not occurred within [number] days after the Target Delivery Date, other than because of Customer or Force Majeure, Customer is entitled to the delay credits in Schedule 2. If it has not occurred by the long-stop date in Schedule 1 (the “Long-Stop Date”), Customer may terminate this agreement by notice before the Acceptance Date, and Provider shall refund the Prepayment and return the Security Deposit within [30] days. These are Customer’s sole remedies for late delivery.

  4. 4Charges and payment

    1. 4.1

      Customer shall pay the Charges. Charges for the Reserved Capacity are calculated, as stated in Schedule 2, either at a rate per GPU-hour applied to every reserved GPU for every hour of the Term, or at a rate per Node per month.

    2. 4.2

      Customer commits to paying for the full Reserved Capacity for the whole Initial Term (the “Committed Minimum”), whether or not, and to whatever extent, it uses the Cluster. The Charges are reduced only by service credits under section 9, delay credits under clause 3.6 and the abatement for Failed Nodes under clause 7.3.

    3. 4.3

      Provider shall invoice the Reserved Capacity monthly in advance, pro rata for part months, and Additional Services monthly in arrears. Customer shall pay each invoice within [30] days of its date, in the currency in Schedule 2.

    4. 4.4

      [The Charges include the electricity consumed by the Cluster.] / [Electricity consumed by the Cluster is charged separately as set out in Schedule 2, based on Provider’s calibrated meters.] The Charges are fixed for the Initial Term, except for [annual indexation] / [the energy adjustment] in Schedule 2.

    5. 4.5

      Customer shall notify any good-faith dispute about an invoice within [15] days of its date, with reasons, and pay the undisputed part on time. If the dispute is not resolved within [30] days, either Party may refer it under section 19.

    6. 4.6

      Overdue amounts carry interest at [rate] per cent per year above [reference rate] / [the statutory rate for late commercial payments]. The Charges are exclusive of value added tax, which Customer shall pay in addition or account for under the reverse-charge mechanism.

  5. 5Prepayment and security

    1. 5.1

      Within [five] Business Days after signing, Customer shall pay the prepayment in Schedule 2 (the “Prepayment”), which Provider shall apply against the invoices for [the first] / [the last] [number] months of the Initial Term.

    2. 5.2

      Within [ten] Business Days after signing, Customer shall provide the security in Schedule 2, being [a cash deposit (the “Security Deposit”)] / [a bank guarantee] / [a parent company guarantee], and shall maintain it for the Term.

    3. 5.3

      If any amount remains unpaid [ten] Business Days after notice that it is overdue, Provider may apply the security towards it, and Customer shall restore the security to its full amount within [ten] Business Days.

    4. 5.4

      Within [30] days after the end of the Term and payment of all amounts due, Provider shall return the Security Deposit, less any amounts applied, [without interest]. Provider shall hold the Prepayment and Security Deposit [in a separate account] / [as part of its general funds].

  6. 6Customer responsibilities

    1. 6.1

      Customer shall use the Services lawfully and in accordance with Provider’s acceptable use policy [attached as Annex C], and shall ensure that its users do not:

      1. (a)infringe the rights of any person, or store or distribute unlawful content;
      2. (b)attack or gain unauthorised access to any system, distribute malware or send unsolicited bulk messages;
      3. (c)interfere with the Facility, Provider’s network or other customers;
      4. (d)mine cryptocurrency or run similar proof-of-work workloads, unless Schedule 2 permits it; or
      5. (e)modify the firmware or out-of-band management of any Node without Provider’s consent.
    2. 6.2

      Customer represents and undertakes that:

      1. (a)neither it nor any of its directors or controlling owners is a Sanctioned Person;
      2. (b)it will not allow any Sanctioned Person to access the Cluster or its outputs, or otherwise use the Services in breach of Sanctions and Export Control Laws;
      3. (c)it will not use the Services for any prohibited end use, including weapons of mass destruction or restricted military uses; and
      4. (d)it will promptly provide the information on its identity, ownership, users and intended use that Provider reasonably requests, and notify any change that makes this clause untrue.
    3. 6.3

      Customer is responsible for its software, licences, configuration, workloads and backups, for the security of everything it runs on the Cluster, and for anyone who uses its credentials.

    4. 6.4

      Customer shall move workloads off a Node that Provider has identified for repair within [number] hours of request, and its personnel may visit the Facility only by arrangement and under the Facility’s rules.

  7. 7Provider responsibilities

    1. 7.1

      Provider shall provide the Services with reasonable skill and care, in accordance with good industry practice, the Cluster Specification and applicable law. Provider remains responsible for its subcontractors.

    2. 7.2

      Provider shall continuously monitor the health of the Cluster hardware, maintain support arrangements with the manufacturers or their authorised partners, and keep on site the spare Nodes and components listed in Schedule 1.

    3. 7.3

      Provider shall respond to a Failed Node and repair or replace it with a Node that meets the Cluster Specification within the response and replacement times in Schedule 2 (the “Replacement Time”). If it does not, the Charges for that Node abate from the end of the Replacement Time until it is back in service, and no service credit is payable for the same Node and period.

    4. 7.4

      Provider shall keep firmware and management software at manufacturer-supported versions, apply critical security updates within [number] days of release, agree with Customer the timing of updates that may affect running workloads, and make no change that materially alters performance without Customer’s agreement.

    5. 7.5

      Provider shall give at least [five] Business Days’ notice of planned maintenance, carry it out within the windows in Schedule 2 and, unless Customer agrees otherwise, take no more than [number] per cent of the Nodes out of service at once. For emergency maintenance it shall give as much notice as practicable.

    6. 7.6

      Provider shall provide power, cooling and physical security in accordance with the SLA Schedule and maintain the certifications listed in Schedule 1.

  8. 8Customer Data and security

    1. 8.1

      Customer owns all Customer Data. Provider acquires no rights in it other than to host it as needed to provide the Services.

    2. 8.2

      Provider shall not access, copy or use Customer Data except:

      1. (a)as necessary for support that Customer has requested and authorised;
      2. (b)as required by law or binding order, in which case it shall, where permitted, notify Customer promptly and disclose only what is required; or
      3. (c)incidentally and unavoidably when repairing hardware, without reading the data, and subject to clause 8.5.
    3. 8.3

      Provider shall maintain appropriate technical and organisational measures, including logged physical and administrative access, network isolation from other customers, and personnel screening. Customer is responsible for the security of the systems, data and access controls it manages.

    4. 8.4

      Provider shall notify Customer without undue delay, and within [24] hours of becoming aware, of any security incident affecting the Cluster or Customer Data, and take reasonable steps to contain it.

    5. 8.5

      Before any storage media used in the Cluster is reused, disposed of or removed from the Facility, whether after failure, substitution or the end of the Term, Provider shall erase it irrecoverably using a method consistent with [recognised media sanitisation standard], or physically destroy it. Failed drives shall not go back to a manufacturer unless so erased or destroyed. Provider shall keep records and issue a certificate of sanitisation at the end of the Term.

    6. 8.6

      Customer shall retrieve Customer Data before the end of the Term [or within [number] days after it, at the rates in Schedule 2], after which clause 8.5 applies and Provider need retain nothing. Provider does not back up Customer Data unless Schedule 2 includes a backup service.

  9. 9Service levels

    1. 9.1

      Provider shall meet the service levels in the SLA Schedule. Service credits are calculated and claimed under the SLA Schedule, applied against the next invoice, and shall not exceed [number] per cent of the monthly Charges for the Reserved Capacity.

    2. 9.2

      Service credits are a price adjustment and Customer’s sole financial remedy for failing a service level, without affecting Customer’s right to terminate under clause 13.4(b) or its rights for breach of section 8.

    3. 9.3

      The service levels do not apply to failures caused by Customer, by planned maintenance under clause 7.5, by suspension under section 11 or by Force Majeure.

  10. 10Hardware substitution and refresh

    1. 10.1

      Provider may at any time, and shall where a manufacturer ends support for a component, replace a Node or component with one of the same model and specification, giving reasonable notice where a running workload is affected.

    2. 10.2

      Provider may substitute a different GPU model or Node type only with Customer’s consent, which Customer shall not unreasonably withhold where the substitute is Equivalent Hardware and does not require Customer to requalify or materially change its software. “Equivalent Hardware” is hardware at least as capable in GPU memory, compute throughput, memory bandwidth and interconnect bandwidth, which achieves at least the benchmark results in Schedule 3.

    3. 10.3

      Each substitute Node shall pass the relevant Acceptance Tests before going into service. Unless Customer agrees otherwise, Provider shall not substitute more than [number] per cent of the Nodes in any [month].

    4. 10.4

      The Parties may agree to upgrade the Cluster to a newer hardware generation, with revised Charges and any extension of the Term recorded in a signed variation.

  11. 11Suspension

    1. 11.1

      Provider may suspend all or part of the Services by notice if:

      1. (a)an undisputed amount remains unpaid [ten] Business Days after a notice referring to this clause;
      2. (b)Customer breaches clause 6.1 or 6.2, or its use poses a serious and imminent risk to the Facility, Provider’s network or other customers; or
      3. (c)suspension is required by law, a competent authority or Sanctions and Export Control Laws.
    2. 11.2

      Provider shall limit any suspension to what is reasonably necessary, give prior notice where practicable and lawful, not delete Customer Data, and restore the Services promptly once the cause is remedied. Charges continue during a suspension under clause 11.1(a) or 11.1(b).

  12. 12Liability

    1. 12.1

      Nothing in this agreement limits liability for death or personal injury caused by negligence, for fraud, or for anything else that cannot lawfully be limited.

    2. 12.2

      Subject to clause 12.1, neither Party is liable for indirect or consequential loss, or for loss of profit, revenue, business, contracts or goodwill. This does not limit Customer’s obligation to pay the Charges, the Committed Minimum or any amount under section 14.

    3. 12.3

      Subject to clauses 12.1, 12.4 and 12.5, each Party’s total liability in each Contract Year is limited to [100] per cent of the Charges paid and payable for that Contract Year.

    4. 12.4

      The limit in clause 12.3 does not apply to Customer’s payment obligations, to the indemnity in clause 12.6, to Customer’s breach of clause 6.2, or to breach of section 15.

    5. 12.5

      Provider’s total liability for breach of section 8 is limited to [amount] in each Contract Year, in place of the limit in clause 12.3. For loss of Customer Data caused by such a breach, Provider is liable only for the reasonable cost of restoring it from Customer’s latest backup.

    6. 12.6

      Customer shall indemnify Provider against third-party claims relating to Customer Data, Customer’s workloads or its breach of clause 6.1 or 6.2, provided Provider notifies the claim promptly and lets Customer conduct the defence.

  13. 13Term, renewal and termination

    1. 13.1

      This agreement starts on signing. The initial term of the Services is [number] months from the Service Commencement Date (the “Initial Term”).

    2. 13.2

      This agreement then renews for successive periods of [12] months (each a “Renewal Term”) unless either Party gives at least [90] days’ notice before the end of the current term. Renewal Charges are as set out in Schedule 2 or, if none, as agreed at least [60] days before renewal; failing agreement, this agreement expires.

    3. 13.3

      Either Party may terminate by notice if the other commits a material breach that is not remedied within [30] days of notice, or becomes insolvent, enters liquidation, administration or an arrangement with creditors, or anything similar.

    4. 13.4

      Customer may also terminate by notice:

      1. (a)under clause 3.6 after the Long-Stop Date;
      2. (b)if a chronic failure, as defined in the SLA Schedule, occurs; or
      3. (c)under clause 18.1.
    5. 13.5

      Provider may also terminate by notice:

      1. (a)if an undisputed amount remains unpaid [30] days after suspension under clause 11.1(a);
      2. (b)if Customer breaches clause 6.2, or continuing would breach Sanctions and Export Control Laws; or
      3. (c)under clause 18.1.
    6. 13.6

      Customer may terminate for convenience during the Initial Term only on [90] days’ notice and payment of the early termination fee in Schedule 2 (the “Early Termination Fee”).

  14. 14Consequences of termination

    1. 14.1

      If Provider terminates under clause 13.3 or 13.5(a) or 13.5(b), Customer shall pay the Charges accrued and [the Charges for the rest of the Initial Term] / [the Early Termination Fee], [less Provider’s net revenue from re-selling the Reserved Capacity during that period].

    2. 14.2

      If Customer terminates under clause 13.6, it shall pay the Charges accrued and the Early Termination Fee.

    3. 14.3

      If Customer terminates under clause 13.3 or 13.4, or either Party terminates under clause 18.1, no Charges are payable for the period after termination, and Provider shall refund any Charges paid for that period and any unapplied Prepayment and return the Security Deposit.

    4. 14.4

      Clauses 8.5 and 8.6 apply on termination, and Provider shall give reasonable help with exporting Customer Data at the rates in Schedule 2. Accrued rights are unaffected, and clauses 8.5 and 8.6 and sections 12, 14, 15, 17, 18 and 19 survive termination.

  15. 15Confidentiality

    1. 15.1

      Each Party shall keep confidential the terms of this agreement and the other’s non-public information, use it only to perform this agreement, and disclose it only to affiliates, advisers, financiers and subcontractors who need it and are bound by equivalent duties, or as required by law. This section continues for [three] years after this agreement ends. Customer Data is protected under section 8.

  16. 16Introduction through the platform

    1. 16.1

      The Parties acknowledge that they were introduced through AICapacity and that the terms of use of AICapacity that each of them has accepted continue to apply to any agreement between them arising from that introduction. AICapacity is not a party to this agreement and has no rights or obligations under it.

  17. 17Personal data

    1. 17.1

      Each Party processes the business contact details of the other’s personnel as an independent controller, only to manage this agreement and in compliance with data protection law, including the General Data Protection Regulation (EU) 2016/679 (“GDPR”).

    2. 17.2

      To the extent Provider processes personal data within Customer Data on Customer’s behalf, including by hosting the storage it is held on or giving support under clause 8.2(a), Provider acts as processor and the data processing terms in Annex B apply. Provider shall not access the Cluster from outside [the European Economic Area] without Customer’s consent.

  18. 18General

    1. 18.1

      Neither Party is liable for delay or failure caused by events beyond its reasonable control, such as natural disasters, war, epidemics or grid failure beyond the Facility’s backup capacity (“Force Majeure”), if it notifies promptly and mitigates. Force Majeure does not excuse payment and excludes late delivery by Provider’s suppliers. If it prevents a material part of the Services for over [60] consecutive days, either Party may terminate by notice.

    2. 18.2

      This agreement is the entire agreement between the Parties about its subject matter and supersedes any earlier arrangement, including any letter of intent. A variation is effective only if written and signed by both Parties, and a delay in exercising a right is not a waiver.

    3. 18.3

      Neither Party may assign this agreement without the other’s consent, not to be unreasonably withheld, except that Provider may assign its right to the Charges to a financier by way of security on notice.

    4. 18.4

      An invalid provision shall be replaced by the closest valid one, and the rest continues. Notices go by email to the addresses below, with a courier copy for breach or termination. No third party may enforce this agreement. It may be signed in counterparts and electronically.

  19. 19Governing law and disputes

    1. 19.1

      This agreement and any dispute or claim arising out of or in connection with it are governed by the laws of [governing law].

    2. 19.2

      Any dispute arising out of or in connection with this agreement shall be finally resolved by [arbitration under the rules of [arbitral institution], with the seat of arbitration in [city] and the proceedings conducted in English / the courts of [country and city]].

Schedule 1Cluster specification

The Reserved Capacity and Facility standards. Attach network diagrams and bills of materials as annexes where available.

TermTo fill in
Facility[Facility name and address, data hall]
GPU model[GPU model], [memory per GPU] GB
Nodes and GPUs[number] Nodes × [number] GPUs = [number] GPUs
Node configuration[CPU model and count], [system memory] GB, [local NVMe] TB
GPU interconnect within each Node[type and bandwidth]
Cluster interconnect[InfiniBand / RoCE], [speed per port], [topology], [oversubscription ratio]A non-blocking fabric matters for large training jobs. State it if promised.
Shared storage[type], [usable capacity] TB, [read / write] GB/s
External network[bandwidth] Gbps, [number] public IP addresses, [cross-connects]
Software provided by Provider[none, bare metal only] / [operating system, scheduler or orchestration layer, and who maintains it]
Power and cooling[kW per rack], [air / direct liquid], [redundancy]
Spares on site[number] spare Nodes; [spare components]
Certifications[standards and certifications]
Phased delivery[none] / [number of Nodes by date, per batch]
Target Delivery Date[date]
Long-Stop Date[date]

Schedule 2Commercial terms

TermTo fill in
Pricing basis and rate[amount] per [GPU-hour / Node-month]
Monthly Charges for the Reserved Capacity[amount]Write out the monthly amount so the arithmetic is not in doubt.
Initial Term[number] months
Committed Minimum[amount] (monthly Charges × Initial Term)
Prepayment[amount], applied to [the first / the last] [number] months
Security[cash deposit] / [bank guarantee] / [parent company guarantee] of [amount]
Currency and payment terms[EUR / SEK / NOK / DKK / other]; [30] days from invoice
Electricity[included] / [measured consumption × rate or pass-through formula]
Indexation or energy adjustment[none] / [index and adjustment date] / [formula]
Failed Node response and replacement[number] hours to respond; [number] hours to replace
Maintenance windows[day, time, time zone]
Support and remote hands[hours included per month]; [rate] per extra hour
Additional Services[storage], [data transfer], [backup], [data retrieval] at [rates]
Delay credits[amount] per day after [number] days, capped at [amount]
Cryptocurrency mining[not permitted] / [permitted on these conditions]
Renewal Charges[same as Initial Term] / [rate] / [to be agreed]
Early Termination Fee[remaining Committed Minimum] / [percentage] of it

Schedule 3Acceptance tests

Reference figures are agreed before signing. Attach the exact test commands, versions and parameters as an annex so the tests can be repeated.

  1. 1

    Inventory: the GPUs, memory, processors, storage, network adapters, firmware and driver versions of each Node match the Cluster Specification.

  2. 2

    Burn-in: each Node runs a sustained full-load GPU stress test for [number] hours without uncorrectable memory errors, GPU faults, restarts or thermal throttling, within the manufacturer’s limits.

  3. 3

    Single-Node performance: matrix-multiplication throughput, GPU memory bandwidth and GPU-to-GPU bandwidth reach at least [number] per cent of the reference figures.

  4. 4

    Interconnect: all-reduce and all-to-all tests across [all Nodes] reach at least [number] per cent of the reference bandwidth, within [number] per cent of the reference latency, with link errors below [threshold].

  5. 5

    Storage and network: shared storage achieves at least [number] GB/s read and [number] GB/s write from [number] Nodes at once, and external throughput to [endpoint] reaches [number] Gbps.

  6. 6

    Workload: a representative distributed training job, [description], runs across all Nodes for [number] hours without failure at no less than [number] per cent of the reference throughput.

  7. 7

    Pass criteria: a Node passes if it passes paragraphs 1 to 3. The Cluster passes if [all / [number] per cent of] the Nodes pass and paragraphs 4 to 6 are passed. The same tests apply to substitute and replacement Nodes.

Signed for and on behalf of the Provider

Name

Title

Date

Signature

Signed for and on behalf of the Customer

Name

Title

Date

Signature