Sample Quote — 256× H100 SXM5 Training Cluster, InfiniBand NDR, 1-Year and 3-Year Terms

1. Quote header

• This document prices one deliverable: a 256-accelerator training cluster, delivered as 32 nodes of 8, on a non-blocking rail-optimised InfiniBand NDR fabric, at three commitment terms.

Quote reference

BH-Q-2026-0148

Prepared for

Neocloud Tier 2 — anonymised

Issue date

12 October 2026

Valid until

31 December 2026, 23:59 UTC

Currency

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

Payment terms

Monthly in arrears, net 30. The 3-year term moves to monthly in advance from month 13 and requires a standby letter of credit equal to two months of compute charges.

2. Headline figures

• Including the recurring non-compute line items in section 6, the all-in effective rate on the 3-year term is $2.07 per GPU-hour, which is the number to carry into a cost-per-token or cost-per-training-run model.
• The 42.6% discount is measured against an illustrative on-demand list rate for the same part class in the same region, not against any published price of any named supplier.

3. Fleet configuration

3.1 Fabric and topology

• The fabric is rail-optimised rather than fully switched at the node level.
• Each of the 8 accelerator positions in a node lands on its own leaf switch, so a collective that walks the rails stays inside one switch hop for the majority of its traffic and only reaches the spine at the reduction boundary.
• For dense all-reduce at 256 GPUs this is the difference between a fabric that is merely non-blocking on paper and one that is non-blocking under the actual traffic pattern of your training loop.
• The spine is sized at 4 switches to keep the oversubscription ratio at 1:1 for the worst-case permutation, not the average case.
• NVLink is intra-node only, giving a 8-accelerator scale-up domain.
• Cross-node scale-up — a single memory-coherent domain larger than one node — is not part of this quote.
• If your model needs a domain wider than 8, see the rack-scale sample, which prices that capability explicitly.

3.2 Power and cooling envelope

• The 336 kW IT load becomes 430.08 kW at the facility boundary at the assumed PUE of 1.28, which is 313,958.4 kWh per 730-hour month.
• At the assumed $0.072 per kWh that is $22,605.00 per month of energy, already inside the quoted compute rate up to the 10.5 kW per node envelope.
• Energy drawn above the envelope is recharged at cost plus 8%.

4. Pricing by term

• Three commitment levels are offered against the same 256-accelerator fleet.
• The monthly totals below are compute only: multiply the rate by 256 GPUs and by 730 hours to reproduce every figure in the table, and add the recurring non-compute subtotal from section 6 for an all-in number.

• The on-demand list rate used as the discount reference throughout this quote is $3.10 per GPU-hour for the same part class, region and fabric.
• Against it the monthly rolling term is 21.0% cheaper, the 1-year term 31.6% and the 3-year term 42.6%.

• It is the transfer of idle-capacity risk from the supplier to you, plus the supplier's ability to finance against a signed 36-month receivable.
• That is why the 3-year rate carries a utilisation floor, a letter of credit and a phased acceptance schedule, and why the monthly rolling rate is barely below list: at 30 days notice the supplier retains almost all of the risk and prices accordingly.

5. Cost composition over the term

• The comparison below normalises everything to a 36-month horizon so the options are commensurable: the rolling and 1-year cases are renewed at their stated rates for the full 36 months, and the 3-year case carries its one-time setup and burn-in charge.

• Committing for 3 years rather than rolling month to month is worth $4,359,545.60 over the horizon, a 23.7% reduction, and that is before any volume-tier improvement from growing the fleet.
• The all-in 3-year figure of $14,060,070.40 is $13,912,070.40 of recurring charges across 36 months plus $148,000.00 of one-time setup and burn-in.
• The recurring component divides to $386,446.40 per month, or $2.07 per GPU-hour across the 186,880 GPU-hours the fleet delivers.

6. Non-compute line items

• Storage allowances are quoted in decimal terabytes, where 1 TB is 1,000 GB.

7. Volume tiers

• Above 2,047 GPUs the tier table stops being a schedule and becomes an opening position: at that size the price is set by allocation, by the supplier's own cost of capital and by whether you are the anchor tenant in a hall.

8. Assumptions

• Standard month. Every monthly figure assumes 730 hours, which is 8,760 annual hours divided by 12. Billing is on the calendar month at 730 hours regardless of whether that month has 28 or 31 days.
• Committed utilisation floor. The 1-year and 3-year rates assume you draw at least 95% of the fleet's GPU-hours every month. Under-draw is billed as if it had been drawn; there is no credit for idle capacity you reserved and did not use.
• Power cost. $0.072 per kWh at the meter, energy included in the compute rate up to 10.5 kW per node. Sustained metered power above $0.086 per kWh triggers a pass-through surcharge with 60 days notice.
• PUE. 1.28 annualised at the facility boundary, measured over a rolling 12 months. A worse annual PUE is absorbed up to 1.34; beyond that the delta is recharged.
• Region. One named availability region, single site, no cross-region placement and no data-residency constraint beyond that site. A second region is a separate quote.
• Rack density. 4 nodes per 48U rack, 8 racks total, 42 kW per rack. Higher density needs a different cooling design and is re-priced.
• Interconnect ownership. The InfiniBand NDR fabric, including switches, cables and optics, is supplied and supported by us and is inside the quoted rate. If you supply your own fabric the compute rate falls by $0.09 per GPU-hour and fabric support moves to your side of the line.
• Operating stack. A supported Linux distribution, a supported container runtime, and a scheduler from the approved list. Driver and firmware updates are applied in the maintenance window defined in the support tier.
• Currency and tax. All figures are USD and exclude every form of sales, value-added, goods-and-services and withholding tax, all of which are charged additionally where legally applicable.
• Capacity availability. The quoted lead time assumes capacity exists at the named site on the issue date. It is not an allocation until the quote is countersigned.

9. Exclusions

• Egress above the 250 TB public-internet and 50 TB peered allowances, billed at the section 6 unit rates.
• Cross-region replication, cross-region backup and any second-site disaster-recovery capacity.
• On-site engineering beyond the 72-hour burn-in window, billed per engineer-day at $1,850.00.
• Software licences of any kind, including schedulers, cluster managers, monitoring agents, compilers and frameworks, whether per-node or per-core.
• Model support, training-run support, hyperparameter work, checkpoint recovery assistance and any other professional services aimed at your workload rather than at the fleet.
• Data ingestion, dataset preparation, labelling, transformation and egress of training corpora into the cluster.
• Physical security beyond the site baseline, custom badge regimes, escorted-access arrangements and compliance attestation work.
• Accelerator generations other than the H100 SXM5 80 GB part specified in section 3, and any mid-term substitution.
• Network circuits from your premises to the site, including dark fibre, wavelength services and transit.
• Decommissioning, asset disposal and media sanitisation at end of term, quoted separately on request.

10. Validity and acceptance

• This quote is valid until 31 December 2026, 23:59 UTC.
• After that date the rates, allowances and lead time lapse and must be reissued, because the underlying capacity position and the power assumptions are both time-sensitive.
• Acceptance is by countersignature of the quote reference and receipt of the initial payment instrument where the term requires one; the deployment clock starts on the later of the two.
• Capacity is not reserved by issue of this quote, only by acceptance of it.