Sample Quote — 1,024× B200 Rack-Scale NVLink Domain Deployment, Liquid Cooled, 3-Year Committed

1. Quote header

• At 1,024 accelerators organised into 16 rack-scale NVLink domains, with direct-to-chip liquid cooling and a 3-year committed term, this is a capacity programme: it has phases, acceptance gates, a cooling plant, a spares strategy and a financing shape.

Quote reference

BH-Q-2026-0117

Prepared for

Hyperscaler A, sovereign capacity programme — anonymised

Issue date

1 October 2026

Valid until

31 January 2027, 23:59 UTC

Currency

USD, exclusive of sales tax, VAT, goods and services tax and withholding

Payment terms

Quarterly in advance, net 45, against milestone acceptance for each delivery phase. A standby letter of credit equal to one quarter of compute charges is required at countersignature. Compute charges on a phase begin on that phase's acceptance date, not on signature.

2. Headline figures

• All-in across compute and the recurring non-compute schedule in section 6, the 3-year term costs $2,918,365.60 per month, or $3.90 per GPU-hour across the 747,520 GPU-hours the fleet delivers in a 730-hour month.
• Adding the $1,240,000.00 of one-time integration and burn-in brings the 36-month total cost of ownership to $106,301,161.60.

3. Fleet configuration

• Each domain is 64 accelerators in a single memory-coherent scale-up domain, housed in two liquid-cooled compute racks plus one NVLink switch rack, sharing a cooling distribution unit.
• Sixteen domains give 1,024 accelerators.

3.1 Phased delivery and acceptance

• A phase is accepted when it passes a 168-hour burn-in, a fabric validation across its own scale-up domain and the scale-out fabric, and a reference-workload throughput gate within 5% of the agreed figure.

• The monthly totals in section 4 are stated for the fully deployed fleet of 1,024 accelerators.

3.2 Scale-up and scale-out pricing

• Inside a domain, 64 accelerators share a coherent memory fabric at 1.8 TB/s per accelerator, which is roughly an order of magnitude beyond what the scale-out network can offer and is what makes expert parallelism and long-context tensor parallelism viable without stalling on communication.
• Between domains, traffic crosses the 800 Gb/s scale-out fabric, rail-optimised and non-blocking at 1:1.
• Both fabrics are inside the quoted rate.
• A quote that prices only the accelerators and leaves the fabric as an add-on is not a quote for a cluster; it is a quote for a pile of cards.

3.3 Power and cooling plant

• The 2,355.2 kW IT load comprises 1,830.4 kW across 128 compute trays at 14.3 kW each, 352 kW across 64 NVLink switch trays at 5.5 kW each, and 172.8 kW across 96 scale-out switches at 1.8 kW each.
• At the assumed PUE of 1.09 — achievable only because the load is liquid-cooled at the chip — the facility boundary draw is 2,567.168 kW, or 1,874,032.64 kWh per 730-hour month.
• At $0.061 per kWh that is $114,315.99 per month of energy, inside the quoted rate up to the stated envelope.

4. Pricing by term

• The 3-year committed term is the offer; the other two are shown for comparison and, at this fleet size, are not generally available without an allocation agreement.
• Monthly totals are compute only and are stated for the fully deployed fleet: each is the rate multiplied by 1,024 GPUs multiplied by 730 hours, giving 747,520 GPU-hours per month.
• Add the $466,500.00 recurring non-compute subtotal from section 6 for an all-in figure.

• Against the illustrative on-demand list rate of $5.20 per GPU-hour for this part class and region, the rolling term is 11.5% cheaper, the 1-year term 26.0% and the 3-year term 36.9%.
• The rolling discount is the shallowest of any published sample because at 1,024 accelerators a supplier will not hold saleable capacity on 30 days notice; the term exists mainly to cover a gap between phases, and it is priced as an exception rather than as a product.

5. Cost composition over the term

• Everything below is normalised to 36 months: the rolling and 1-year cases are renewed at their stated rates for the full period, and the 3-year case carries its $1,240,000.00 of one-time integration and burn-in.

• Committing for three years rather than rolling is worth $34,282,150.40 over the horizon, a 24.4% reduction, before any volume-tier improvement.
• The all-in monthly figure on the 3-year term is $2,918,365.60, being $2,451,865.60 of compute plus $466,500.00 of recurring non-compute charges; over 36 months that is $105,061,161.60, and with the one-time integration and burn-in of $1,240,000.00 the total reaches $106,301,161.60.
• Note the proportion: non-compute recurring charges are 16.0% of the all-in monthly bill.

6. Non-compute line items

• Storage allowances are in decimal terabytes, where 1 TB is 1,000 GB, and decimal petabytes where 1 PB is 1,000 TB.

• Storage dominates the recurring non-compute schedule at $348,000.00 of the $466,500.00 total, or 74.6%.
• During the ramp, storage is provisioned per phase and billed at the same unit rates on the capacity actually stood up, so the $466,500.00 figure is the full-deployment number and not the phase-1 number.

7. Volume tiers

• Above 4,096 accelerators the schedule stops applying and pricing becomes a negotiation against allocation, because at that size the binding constraint is supplier capacity and financing rather than discount policy.

• That retroactive step is a material term at this scale: crossing from 1,023 to 1,024 accelerators reprices the entire fleet from $3.44 to $3.28 per GPU-hour, worth $119,603.20 per month at 747,520 GPU-hours, so the 1,024th accelerator arrives carrying a discount on the other 1,023 with it.

8. Assumptions

• Standard month. Every monthly figure assumes 730 hours, being 8,760 annual hours divided by 12. Billing is on the calendar month at 730 hours regardless of the month's actual length.
• Full-deployment basis. Section 4 totals are for all 1,024 accelerators. During the phased ramp the charge is proportional to the accelerators accepted, per section 3.1.
• Committed utilisation floor. 97% of the fleet's GPU-hours drawn each month on the 3-year term. Under-draw is billed as drawn. A maintenance allowance of 3% is included and is not counted as under-draw.
• Power cost. $0.061 per kWh at the meter, energy included in the compute rate up to the 2,355.2 kW IT envelope. Sustained metered power above $0.074 per kWh triggers a pass-through surcharge with 90 days notice.
• PUE. 1.09 annualised at the facility boundary, achievable only with direct-to-chip liquid cooling as specified. Absorbed up to 1.13; beyond that the delta is recharged. An air-cooled variant of this fleet is a different quote at a different rate.
• Region and site. One named site, one hall pair, single sovereign jurisdiction, with the cooling plant, busway and fibre entry assumed already permitted. Planning consent, grid connection upgrade and substation work are not in scope.
• Rack density. 56 racks as itemised in section 3, 57.2 kW per compute rack and 22 kW per switch rack. Any change to rack density changes the cooling plant and is re-priced.
• Interconnect ownership. Both fabrics — the NVLink scale-up switching inside each domain and the 800 Gb/s scale-out network between domains — are supplied, cabled and supported by us and sit inside the quoted rate, including all optics and copper cartridges.
• Domain size. 64 accelerators per NVLink domain. A 72-accelerator domain changes the tray and switch counts, the rack geometry and the rate; it is not a substitution within this quote.
• Spares and repair. The on-site pool in section 6 supports a 4-hour hardware-replacement objective inside the maintenance window. Repair turnaround beyond that pool assumes supplier return-material-authorisation lead times of 15 working days.
• Operating stack. A supported Linux distribution, a supported container runtime and a scheduler from the approved list, with driver and firmware updates applied in the defined maintenance window.
• Currency and tax. All figures are USD and exclude every form of sales, value-added, goods-and-services and withholding tax, charged additionally where legally applicable. Import duty on hardware brought into the jurisdiction is excluded.
• Capacity availability. The phased lead times assume supplier allocation is confirmed at countersignature. Issue of this quote is not an allocation.

9. Exclusions

• Egress above the 1 PB public-internet and 2 PB peered allowances, billed at the section 6 unit rates.
• Cross-region replication, geo-redundant checkpointing, remote backup and any second-site disaster-recovery capacity.
• On-site engineering beyond the 168-hour burn-in window per phase, billed per engineer-day at $2,400.00.
• Software licences of any kind, including schedulers, cluster managers, observability platforms, compilers and frameworks, whether per-node, per-core or per-accelerator.
• Model support, training-run support, parallelism-strategy design, throughput tuning, checkpoint recovery assistance and any other professional services aimed at your workload rather than at the fleet.
• Facility capital work: grid connection, substation, generator and UPS plant, chilled-water plant, cooling towers, structural floor loading work and fibre entry.
• Planning consent, environmental permits, water-use permits for the cooling plant and any regulatory approval process.
• Data ingestion and dataset preparation at petabyte scale, including transformation, deduplication and staging into the parallel filesystem.
• Physical security beyond the site baseline, custom badge regimes, escorted-access arrangements and sovereign-personnel requirements.
• Compliance attestation, audit support and certification work of any kind.
• Accelerator generations other than the B200-class part specified in section 3, and any mid-term substitution or technology refresh.
• Network circuits from your premises to the site, including dark fibre, wavelength services and transit.
• Decommissioning, media sanitisation, coolant disposal and asset disposal at end of term, quoted separately on request.

10. Validity and acceptance

• This quote is valid until 31 January 2027, 23:59 UTC.
• At this fleet size the validity window matters more than usual: the rates depend on a supplier allocation and on a power tariff that are both time-sensitive, and a quote held past its expiry is not a quote.
• Acceptance is by countersignature of the quote reference together with establishment of the standby letter of credit.
• The phase-1 deployment clock starts on the later of the two.