Sample Quote — 128× L40S Inference Fleet, Air Cooled, Hybrid Committed and On-Demand

1. Quote header

• Section 4 therefore prices the term table in the conventional way, and section 4.2 converts every rate into dollars per million processed tokens against a stated throughput assumption you can test before you sign.

Quote reference

BH-Q-2026-0152

Prepared for

Regional GPU Cloud inference tenant — anonymised

Issue date

14 October 2026

Valid until

15 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, for the committed base. Burst-pool usage is metered and invoiced on the same monthly cycle in arrears. No letter of credit.

• The recommended structure is a hybrid: 96 accelerators on a 1-year committed term to carry the steady-state load, and 32 held as an on-demand burst pool to absorb diurnal and campaign peaks.

2. Headline figures

• The hybrid saves $53,541.12 per month against running all 128 accelerators on the monthly rolling term, a 35.4% reduction, while keeping peak capacity available.
• Its blended effective rate is $1.2318 per delivered GPU-hour, or $0.0535 per million processed tokens on the compute side and $0.0625 once the recurring non-compute schedule is included.

3. Fleet configuration

3.1 Capacity split

• The 128 accelerators are quoted as two pools with different commercial terms and different utilisation assumptions.
• The split is a recommendation, not a constraint: it can be moved at any month boundary on 30 days notice without repricing the committed base.

Committed base

96 accelerators, 1-year committed term, billed at 730 hours per month whether or not they are busy. Carries steady-state request load. 96 × 730 = 70,080 GPU-hours per month.

Burst pool

32 accelerators, on-demand, metered on actual draw, no monthly floor. Quoted at an assumed 40% average utilisation, giving 32 × 730 × 0.40 = 9,344 GPU-hours per month. Reserved for diurnal peaks and campaigns; capacity is held but not billed until drawn.

Delivered total

79,424 GPU-hours per month at the assumed burst utilisation, against 93,440 GPU-hours if the whole fleet ran at 100% for 730 hours.

3.2 Power and cooling envelope

• The 102.4 kW IT load becomes 145.408 kW at the facility boundary at the assumed PUE of 1.42, which is 106,147.84 kWh per 730-hour month.
• At $0.084 per kWh that is $8,916.42 per month of energy, inside the quoted compute rate.
• The PUE is the worst of any sample in this set and the tariff is the highest, because air-cooled fleets are usually placed in older halls where neither the cooling plant nor the power contract is optimised — and yet the rate per GPU-hour is still the lowest, because the accelerator itself is far cheaper per unit of inference throughput.
• Placement and part class pull in opposite directions here, and the part class wins.

4. Pricing by term

• The term table prices the whole 128-accelerator fleet on each term, at full 730-hour utilisation, so that the options are directly comparable.
• Monthly totals are compute only: each is the rate multiplied by 128 GPUs multiplied by 730 hours, giving 93,440 GPU-hours per month.
• Add the $16,480.00 recurring non-compute subtotal from section 6 for an all-in figure.
• The recommended hybrid is priced separately in section 4.1 because it bills two pools on two different bases and cannot be expressed as a single rate against 93,440 hours.

• Against the illustrative on-demand list rate of $1.95 per GPU-hour for this part class and region, the rolling term is 16.9% cheaper, the 1-year term 39.5% and the 3-year term 51.8%.

4.1 Hybrid structure, priced

• It is not the cheapest line in the table — committing all 128 accelerators for three years is — but it is the cheapest structure that still answers a traffic surge without a procurement cycle.

• The blended rate of $1.2318 is the monthly total divided by the GPU-hours actually delivered, not by the fleet's theoretical maximum, and it is quoted to four decimal places because at this rate the second decimal place is worth $794.24 per month across the delivered hours.
• Against running all 128 accelerators on the rolling term the hybrid saves $53,541.12 per month, or 35.4%.
• Against committing all 128 for a year it costs $12,427.52 per month more, and buys 32 accelerators of surge capacity with no minimum term attached.

4.2 Cost per million processed tokens

• On the throughput assumption in section 8 — 6,400 processed tokens per GPU-second, being prompt and completion combined, or 23.04 million processed tokens per GPU-hour — the compute rates above convert as follows.

• At full utilisation the fleet would process 2.15 trillion tokens per 730-hour month; at the hybrid's assumed 79,424 delivered GPU-hours it processes 1.83 trillion.
• The all-in column adds the $16,480.00 recurring non-compute schedule from section 6, spread across the delivered hours, which lifts the hybrid's compute-only $0.0535 to $0.0625 — a 16.8% uplift that is easy to miss if the token cost is quoted on compute alone.

5. Cost composition over the term

• All four bars are monthly figures on the same 730-hour month.

• The hybrid sits between the 1-year and 3-year full-commitment options and is the only one of the four that does not bill 32 accelerators you may not be using.
• The trade is explicit: you give up $9,998.08 per month against a full 3-year commitment, and in exchange those 32 accelerators cost nothing in a month when your traffic does not ask for them.
• If your peak-to-floor ratio is below about 1.3, commit the whole fleet for three years.
• If it is above 1.5, the hybrid is the better structure and the gap widens with every point of peakiness.

6. Non-compute line items

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

• The 180 TB egress allowance is the line to check first on any inference quote, because response payloads are pure egress and a public-facing service can consume the allowance with a fraction of its token volume.
• At the quoted $0.032 per GB overage, 10 TB of unexpected egress costs $320.00 — small in isolation, but inference egress scales with user count rather than with fleet size, and it is the only line in this schedule that can grow without any decision on your part.

7. Volume tiers

• The burst pool counts toward fleet size for tier purposes at its reserved capacity of 32 accelerators, not at its metered utilisation, so this fleet holds the 128-accelerator tier even in a month when the burst pool is barely drawn.

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.
• Throughput basis. 6,400 processed tokens per GPU-second, prompt and completion combined, at the reference model and batch settings named below. That is 23.04 million processed tokens per GPU-hour, and every cost-per-token figure in section 4.2 derives from it. A different model, quantisation or batch policy changes the token cost without changing the hourly rate.
• Reference workload. A 70-billion-parameter dense model at 8-bit weight quantisation, continuous batching, 3:1 prompt-to-completion ratio, median time-to-first-token under 400 ms and p99 under 1.2 s. Latency targets tighter than this reduce achievable throughput and raise the cost per token.
• Burst-pool utilisation. 40% average draw across the month, giving 9,344 GPU-hours. Actual burst usage is metered and billed at $1.62 per GPU-hour with no floor, so a month at 20% draw costs less and a month at 70% costs more.
• Committed utilisation floor. 95% on the 96-accelerator committed base. Under-draw on the base is billed as drawn. The burst pool carries no floor.
• Power cost. $0.084 per kWh at the meter, energy included in the compute rate up to 6.4 kW per node. Sustained metered power above $0.098 per kWh triggers a pass-through surcharge with 60 days notice.
• PUE. 1.42 annualised at the facility boundary for an air-cooled hall, measured over a rolling 12 months, absorbed up to 1.48.
• Region. One named availability region, single site, no cross-region placement, no edge or point-of-presence distribution and no latency guarantee to any end-user geography.
• Rack density. 4 nodes per 42U rack, 4 racks total, 25.6 kW per rack, hot-aisle containment, 24 °C supply air.
• Interconnect ownership. The Ethernet leaf-spine fabric is supplied and supported by us and sits inside the quoted rate. No InfiniBand, no scale-up fabric and no RDMA requirement is assumed; adding one for a large partitioned model is a different quote.
• Serving stack. A supported open-weight serving runtime from the approved list, deployed by you. Request routing, rate limiting, authentication and billing logic on your side of the API boundary.
• 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 8–10 week lead time assumes capacity exists at the named site on the issue date, including 32 accelerators held for the burst pool. Issue of this quote is not an allocation.

9. Exclusions

• Egress above the 180 TB public-internet and 40 TB peered allowances, billed at the section 6 unit rates.
• Cross-region replication, multi-region active-active serving, edge caching and any content-delivery-network charges.
• On-site engineering beyond the 48-hour burn-in window, billed per engineer-day at $1,850.00.
• Software licences of any kind, including serving runtimes with commercial terms, observability platforms, gateways, vector databases and schedulers.
• Model support: quantisation, distillation, fine-tuning of served models, prompt design, evaluation-harness construction, guardrail configuration, latency tuning and any other professional services aimed at your workload rather than at the fleet.
• Fine-tuning or training capacity of any kind. This quote serves models; it does not produce them.
• Request logging storage beyond the 400 TB object allowance, retention processing, redaction and any data-protection tooling.
• Additional accelerators beyond the 128 quoted, and any expansion of the burst pool beyond 32, both of which require 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, VPN termination and dedicated interconnect ports.
• Decommissioning, media sanitisation and asset disposal at end of term.

10. Validity and acceptance

• This quote is valid until 15 December 2026, 23:59 UTC.
• After that date the rates, allowances and the throughput basis all lapse and must be reissued, since inference part pricing and the burst-pool capacity position both move faster than training capacity does.
• Acceptance is by countersignature of the quote reference; the committed base and the burst-pool reservation are both established on acceptance, and the deployment clock starts that day.
• Capacity is not reserved by issue of this quote, only by acceptance of it.