Sample quote — illustrative data. The figures below are structural
samples published to demonstrate format and analytical depth. They were not collected from live
market sources. For decisions, commission a quote built from contracted rate cards and
observed transaction data.
Sample quote. This is a structural example published to show the format and completeness of a Bramcott GPU fleet price quote. Every figure is illustrative and was not derived from a live rate card or a real supplier agreement. A commissioned quote is built from contracted rates for your specific configuration.
1. Quote header
- This quote prices a small, short-horizon fleet: 64 accelerators across 8 nodes, intended for fine-tuning, evaluation and research rather than for pre-training.
- The 1-year and 3-year terms are priced alongside it so the value of committing is visible as a number rather than as an argument.
- Quote reference
- BH-Q-2026-0131
- Prepared for
- Applied AI research group, Tier 3 — anonymised
- Issue date
- 5 October 2026
- Valid until
- 30 November 2026, 23:59 UTC
- Currency
- USD, exclusive of sales tax, VAT, goods and services tax and withholding
- Payment terms
- Monthly in advance, net 15, by ACH or direct debit. No letter of credit and no minimum term on the rolling option.
- A 30-day notice period only means something if the rate behind it is current, so this document expires after eight weeks rather than sitting in a procurement system for a year.
2. Headline figures
$3.05Quoted compute rate per GPU-hour on the monthly rolling term
$158,936.00All-in monthly cost, compute plus recurring non-compute
30 daysNotice to exit, either side, with no termination charge
6–8 weeksDeployment lead time from countersignature to acceptance
- The all-in monthly figure of $158,936.00 is $142,496.00 of compute plus $16,440.00 of recurring non-compute charges from section 6.
- Divided across the 46,720 GPU-hours a 64-accelerator fleet delivers in a 730-hour month, the all-in effective rate is $3.40 per GPU-hour — a 11.5% uplift on the compute-only rate, which is the number worth remembering when you compare this fleet against a larger one where storage and support are spread further.
3. Fleet configuration
| Line item | Specification | Per node | Fleet total |
|---|---|---|---|
| Compute nodes | 8-accelerator SXM GPU baseboard, 8U, single contiguous block | 1 | 8 |
| Accelerators | H200, 141 GB HBM3e, 4.8 TB/s, 700 W envelope | 8 | 64 |
| Host CPUs | 48-core x86, 2.4 GHz base, PCIe 5.0 | 2 | 16 |
| System memory | DDR5-4800 ECC RDIMM, 1,024 GB per node | 1,024 GB | 8,192 GB |
| Local NVMe | 3.84 TB U.2, mixed-use, per-node scratch and dataset cache | 4 × 3.84 TB | 122.88 TB |
| Compute NIC | 400 Gb/s OSFP, one per accelerator | 8 | 64 |
| Fabric switches | InfiniBand NDR 400 Gb/s, 64-port: 2 leaf plus 2 spine, non-blocking at 64 ports | — | 4 |
| Storage NIC | 100 Gb/s Ethernet, single port, shared filesystem path | 1 | 8 |
| Rack power | 10.2 kW per node, single busway feed plus UPS-backed control | 10.2 kW | 81.6 kW |
| Cooling | Rear-door heat exchanger, chilled water, 20 °C supply | 1 | 8 |
| Racks | 48U, 2 nodes per rack, hot-aisle containment | — | 4 |
| Out-of-band management | Baseboard controller per node plus 1 management switch, isolated VLAN | 1 | 8 + 1 |
3.1 Memory-to-accelerator ratio
- Fine-tuning workloads are governed by what fits, not by what is fast.
- At 141 GB per accelerator a 64-accelerator fleet presents 8,960 GB of aggregate high-bandwidth memory, which is what lets a 70-billion-parameter model be fine-tuned with full-rank adapters and a comfortable batch, or a 200-billion-parameter model be handled under tensor and pipeline parallelism with a modest activation budget.
- The rate here is therefore higher per GPU-hour than a comparable H100 fleet and the fleet is smaller: you are buying memory bandwidth and capacity per accelerator, and you are buying fewer of them.
- The fabric is sized for a 64-port non-blocking domain.
- That is sufficient for full-bandwidth collectives across the whole fleet, so no experiment on this cluster is fabric-limited — a property worth paying for at research scale, where an unexplained slowdown costs a week of investigator time and the week costs more than the switches.
3.2 Power and cooling envelope
- The 81.6 kW IT load becomes 106.896 kW at the facility boundary at the assumed PUE of 1.31, which is 78,034.08 kWh per 730-hour month.
- At $0.076 per kWh that is $5,930.59 per month of energy, already inside the quoted compute rate.
- Smaller fleets carry a worse PUE than large ones because the fixed hall load is spread across fewer racks; that is a real cost and it is one reason the per-GPU-hour rate does not scale linearly downward as fleets shrink.
4. Pricing by term
- Monthly totals are compute only: each is the rate multiplied by 64 GPUs multiplied by 730 hours, giving 46,720 GPU-hours per month.
- Add the $16,440.00 recurring non-compute subtotal from section 6 for an all-in figure.
| Term option | Rate per GPU-hour | Monthly total | Contract months | Total contract value |
|---|---|---|---|---|
| Monthly rolling — primary offer, 30 days notice either way | $3.05 | $142,496.00 | 0 minimum | $1,709,952.00 12-month annualised run-rate, no commitment |
| 1-year committed, capacity guaranteed | $2.68 | $125,209.60 | 12 | $1,502,515.20 |
| 3-year committed, capacity guaranteed | $2.34 | $109,324.80 | 36 | $3,935,692.80 |
| Totals, fleet of 64 GPUs, 46,720 GPU-hours per month | $2.34 – $3.05 | $109,324.80 – $142,496.00 | — | $1,502,515.20 – $3,935,692.80 |
- Against the illustrative on-demand list rate of $3.60 per GPU-hour for this part class and region, the rolling term is 15.3% cheaper, the 1-year term 25.6% and the 3-year term 35.0%.
- The rolling discount is shallow because the supplier keeps the risk: at 30 days notice it must hold this capacity saleable to someone else, and it charges you for that option.
On-demand list (reference)
$3.60
Monthly rolling
$3.05
1-year committed
$2.68
3-year committed
$2.34
4.1 Short-commitment comparison
- Committing for 12 months saves $17,286.40 per month against the rolling rate, or $207,436.80 over the year.
- That is a genuine saving, and it is still frequently the wrong decision on a research fleet, because the saving is only realised if the fleet is both wanted and wanted at that size for all twelve months.
- Three things commonly break that assumption: the experiment concludes early, the model grows and needs a wider domain, or the funding line moves.
- On a rolling term each of those costs 30 days of notice.
- On a 1-year term each of them costs the remainder of the commitment, which is why we quote the rolling rate first here even though it is the higher number.
5. Cost composition over the term
- On this 64-accelerator fleet the fixed elements — support, spares, and the standing capacity of the filesystem — are a material share of the bill.
On-demand list, compute only
$3.60
Monthly rolling, all-in
$3.40
1-year committed, all-in
$3.03
3-year committed, all-in
$2.69
- The three all-in monthly figures are $158,936.00 rolling, $141,649.60 on 1 year and $125,764.80 on 3 years.
- Each is the corresponding compute monthly total from section 4 plus the same $16,440.00 of recurring non-compute charges, which is why the gap between the bars narrows as the compute rate falls: the fixed component becomes a larger proportion of a smaller total.
6. Non-compute line items
- Storage allowances are in decimal terabytes, where 1 TB is 1,000 GB.
| Line item | Basis | Included allowance | Unit rate | Monthly (USD) |
|---|---|---|---|---|
| Egress to the public internet | Per GB above allowance | 60 TB | $0.040 | included |
| Parallel filesystem, performance tier | Per GB-month, checkpoint and adapter store | 120 TB | $0.062 | $7,440.00 |
| Object storage, warm tier | Per GB-month, dataset and weight archive | 250 TB | $0.019 | $4,750.00 |
| Object storage, cold and archive tiers | Per GB-month | not provisioned | $0.006 | $0.00 |
| Remote management and support, Tier 1 | Flat monthly, business hours plus P1 cover, P1 response in 4 hours | 1 contract | flat | $2,400.00 |
| On-site engineering beyond the burn-in window | Per engineer-day, 8 hours | 0 days | $1,850.00 | on use |
| Spares pool held on site | Flat monthly, 1 partial node and 4 accelerators | 1 pool | flat | $1,850.00 |
| One-time: rack and stack, cabling, firmware baseline | One-time, 8 nodes | — | flat | $18,500.00 once |
| One-time: 48-hour burn-in and acceptance test | One-time, 8 nodes, includes fabric validation | — | flat | $8,000.00 once |
| Recurring non-compute subtotal | Monthly, added to the compute monthly total in section 4 | — | — | $16,440.00 |
| One-time subtotal | Invoiced at acceptance, not recurring | — | — | $26,500.00 |
- On a rolling term the one-time subtotal deserves scrutiny, because $26,500.00 amortised over a 3-month experiment is $8,833.33 per month against a $142,496.00 compute bill — a 6.2% uplift that disappears over a 12-month horizon.
7. Volume tiers
- The quoted 64 accelerators land in the third band, and the table extends downward as well as upward because research groups frequently start at 8 or 16 and grow.
| Fleet size | Monthly rolling | 1-year committed | 3-year committed |
|---|---|---|---|
| 8 – 31 GPUs | $3.28 | $2.89 | $2.58 |
| 32 – 63 GPUs | $3.15 | $2.78 | $2.44 |
| 64 – 127 GPUs — quoted tier | $3.05 | $2.68 | $2.34 |
| 128 – 255 GPUs | $2.96 | $2.59 | $2.26 |
| 256 GPUs and above | $2.88 | $2.51 | $2.18 |
- Below 32 accelerators a fleet still occupies whole racks, still needs a fabric, still needs a spares pool and a support contract, and all of that fixed cost is divided across fewer GPUs.
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.
- Committed utilisation floor. None on the monthly rolling term. The 1-year and 3-year rates assume a floor of 90% of the fleet's GPU-hours drawn each month, with under-draw billed as drawn.
- Power cost. $0.076 per kWh at the meter, energy included in the compute rate up to 10.2 kW per node. Sustained metered power above $0.091 per kWh triggers a pass-through surcharge with 60 days notice.
- PUE. 1.31 annualised at the facility boundary, measured over a rolling 12 months, absorbed up to 1.37.
- Region. One named availability region, single site, no data-residency constraint beyond that site and no cross-region placement.
- Rack density. 2 nodes per 48U rack, 4 racks total, 20.4 kW per rack. The hall's chilled-water capacity is assumed sufficient for rear-door heat exchangers at this density.
- Interconnect ownership. The InfiniBand NDR fabric — switches, cables and optics — is supplied and supported by us and sits inside the quoted rate. Customer-supplied fabric reduces the rolling rate by $0.11 per GPU-hour and moves fabric support to the customer.
- Workload shape. Intermittent, bursty fine-tuning and evaluation runs, not a continuously saturated training loop. The spares pool and support tier are sized for that shape; a continuously saturated fleet needs a larger pool and is re-priced.
- Operating stack. A supported Linux distribution, a supported container runtime and a scheduler from the approved list. Driver and firmware updates fall inside the maintenance window of the support tier.
- 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.
- Capacity availability. The 6–8 week lead time assumes capacity exists at the named site on the issue date. Issue of this quote is not an allocation.
9. Exclusions
- Egress above the 60 TB public-internet allowance, billed at $0.040 per GB.
- Cross-region replication, remote backup and any second-site or disaster-recovery capacity.
- On-site engineering beyond the 48-hour burn-in window, billed per engineer-day at $1,850.00.
- Software licences of any kind, including schedulers, experiment trackers, cluster managers, monitoring agents and commercial model weights or datasets.
- Model support: hyperparameter selection, recipe design, evaluation-set construction, adapter management, checkpoint recovery assistance and any other professional services aimed at your workload rather than at the fleet.
- Data preparation, labelling, deduplication, decontamination and transformation of training or evaluation corpora.
- Additional accelerators beyond the 64 quoted, including temporary burst capacity, which requires a side letter at the applicable tier rate.
- Physical security beyond the site baseline, custom badge regimes and compliance attestation work.
- Network circuits from your premises to the site, including transit, wavelength services and VPN termination.
- Decommissioning, media sanitisation and asset disposal at end of term.
10. Validity and acceptance
- This quote is valid until 30 November 2026, 23:59 UTC.
- On expiry the rates and allowances lapse and must be reissued.
- Acceptance is by countersignature of the quote reference; on the rolling term no payment instrument is required in advance of the first monthly invoice, and the deployment clock starts on countersignature.
- Capacity is reserved for 10 working days from issue as a courtesy and is not held beyond that without acceptance.