Simple sketch of inference cloud unit economics

Example GPU-rental model for a single rack of 64 H200 GPUs
Downside · $0.90mDistressed value
Base · $1.50mBase case resale market
Upside · $2.20mStrong continued utility
Super bull · $2.70mScarcity value retained
Equity IRR incl. resale, net of tax
Equity IRR excluding resale
Exit value required for 15% IRR
Upfront equity required
EBITDA asset yield
EBIT asset yield
Simple EBITDA payback
Three-year debt interest
Three-year equity return

Free cash flow to equity and cumulative equity cash position

Operating and cash bridge

USDToday202720282029
UtilisationBillable utilisation, not technical uptime. 2027 is set by the slider; the model adds five percentage points in each following year, capped at 95%. Idle capacity, outages and non-billable maintenance should all be reflected here.—
Billable GPU-hours—
Price per GPU-hourRealized GPU-rental price. 2027 is set by the slider; 2028 and 2029 compound the selected annual change. Public on-demand list prices are reference points, not evidence that a 64-GPU rack can sustain the same blended realization at the modeled utilisation.—
Revenue—
Power & colocation$241k base. Approximately $74k of metered electricity on 70kW average IT draw plus $167k of colocation rent on 80kW nameplate capacity. This fixed run-rate does not fall with billable utilisation. PUE is total facility power divided by IT-equipment power; 1.20 adds 20% for cooling and facility overhead.—
Network & maintenance$140k base. Bandwidth, IP transit, cross-connects, operating repairs, spares and roughly 3% hardware support. No separate maintenance capex is modelled. The main swing factors are committed bandwidth, egress and the support contract.—
Software$100k base. Orchestration, monitoring, security, scheduler and licence reserve. This assumes an off-the-shelf stack rather than a large proprietary control-plane build.—
Rack operations allocation$180k base. Fractional site operations and support coverage shared across several racks. Sales, platform development and corporate overhead are captured separately below.—
Total direct operating cost3% annual inflation. Each direct rack operating-cost category increases by 3% per year.—
Rack contribution profit—
Rack contribution margin—
Rack contribution yieldAnnual unlevered rack yield. Each year's rack contribution profit divided by all-in deployment cost: rack capex plus the one-off setup cost. Central overhead, interest and tax are excluded.—
Central platform, sales & G&A allocationSet by the slider. Illustrative per-rack allocation for sales and marketing, central platform engineering, finance, legal, insurance and management. It increases by 3% annually and should decline per rack as the platform scales.—
EBITDA—
EBITDA margin—
Cash tax on operations25% tax rate after depreciation. Applied annually to EBITDA less interest and five-year straight-line rack depreciation. Tax is zero whenever that calculation is negative; no tax benefit or loss carry-forward is modelled.—
Interest paidOpening-balance interest. Interest is charged on debt outstanding at the start of each year, so it declines as principal is repaid.—
Scheduled principal repayment$250k per year. Scheduled principal is flat at $250k annually, capped at the remaining debt balance.—
Rack capex64-H200 rack hardware. Eight 8-GPU systems plus rack-level networking, cabling and integration. Facility-side setup is captured separately. No maintenance capex is modeled after deployment.———
Debt fundingApplied to rack capex only. The one-off setup cost is equity-funded. No arrangement fee, original issue discount, commitment fee or refinancing cost is modeled.———
One-off setup cost$250k equity-funded at deployment. Facility-side fit-out, commissioning and initial launch costs. No tax deduction is modeled. The rack is assumed ready for a full operating year in 2027, with no construction or ramp delay beyond the utilisation assumption.———
Rack market value at exit2029 hardware value. The selected rack market value is treated as a cash inflow at the end of 2029. Use a 4–6 year accounting life as a modelling anchor, then cross-check the assumption against per-GPU and per-system resale evidence.———
Tax on equipment saleTax recapture on disposal. The model taxes sale proceeds above the remaining tax book value at 25%. Tax book value is 40% of rack capex after three years of five-year straight-line depreciation. No tax benefit is modelled when sale value is below book value.———
Debt repaid at exitFull repayment in 2029. Any debt remaining after scheduled principal repayments is repaid from exit proceeds.———
Free cash flow to equity
Cumulative equity cash positionInvestment recovered to date. Cumulative net equity cash flow from initial funding through each period. A negative figure is unrecovered equity, not an accounting equity value.
Credit metrics
Gross debtFunded debt. Each forecast year reflects the $250k scheduled principal repayment. The 2029 figure is shown before the remaining balance is repaid at exit.
Less: cash balanceCash is retained. Annual free cash flow to equity accumulates on the balance sheet rather than being swept to shareholders. A negative balance indicates an additional funding requirement.
Net debt
Net debt / exit equipment valueResidual-value coverage. Net debt at each period end divided by the selected 2029 rack market value. Below 100% indicates the assumed exit equipment value covers net debt; a negative result means the rack is in a net cash position.
Net leveragen.m.
EBITDA / interest coveragen.m.

Per MW economics

Annualized on 0.08 MW of nameplate rack power capacity · same arithmetic basis as annual value / MW
USD per MWToday202720282029
Annual revenue / MWThe comparable price-per-MW measure. Annual rack revenue divided by 0.08 MW, matching the published maximum power of a 64-H200 liquid-cooled rack. Compare it with annual contract value per MW, while allowing for different GPU generations, service scope and contract terms.—
Rack contribution profit / MW—
EBITDA / MW—
Rack capex / MW———
All-in deployment cost / MW———
Rack market value at exit / MW———

Current H200 price benchmark

Public on-demand list prices per GPU-hour, checked 3 September 2026 · median uses the first four like-for-like rentals
Provider / productOn-demand priceProvider vs your price
Your rack · 8× H200 nodeYour starting point
Hyperstack · H200 SXMLive price ↗$3.99 / GPU-hour
Crusoe · H200 HGXLive price ↗$4.29 / GPU-hour
Runpod · H200 SXM clusterLive price ↗$4.31 / GPU-hour
Nebius · HGX H200Live price ↗$4.50 / GPU-hour
Together AI · HGX H200 (published for 256+ GPUs)Live price ↗$5.99 / GPU-hour
Google Cloud · A3 Ultra H200 (8-GPU VM incl. host resources)Live price ↗$10.60 / GPU-hour
Assumptions back-up

Power & colocation build

Illustrative US basis · electricity and capacity rent modelled separately
DriverCurrent modelCalculationExternal benchmark and read-through
Average IT power draw70kW1.09kW per H200, including its share of host and network equipmentGIGABYTE's purpose-built 42U liquid-cooled rack contains 64 H200 GPUs and lists up to 80kW rack power. Average modeled draw is 88% of that maximum.Source ↗
Nameplate critical IT capacity80kWPublished maximum for the 64-H200 liquid-cooled rackUsed for colocation capacity rent and per-MW normalization. This keeps average energy consumption separate from capacity that must be provisioned.Source ↗
Facility efficiency (PUE)1.20×PUE = total facility power ÷ IT power. 70kW IT × 1.20 = 84kW facility draw.A 1.20 PUE means cooling and other facility overhead add 20% on top of IT load. Uptime Institute's 2025 global average was 1.54, so this assumes an efficient, modern facility.Source ↗
Metered electricity$0.10/kWh
$74k per year		84kW × 8,760 hours × $0.10 = $73.6k2025 US averages were $0.0862/kWh for industrial users and $0.1341/kWh for commercial users. The model sits between them.Source ↗
Colocation capacity$174/kW/month
$167k per year		80kW × $174 × 12 months = $167.0kCBRE's Q1 2026 ranges were $200–$230 in Chicago and $190–$235 in Northern Virginia for 250–500kW requirements. The model is below those wholesale rents, before a possible small-scale or high-density premium; CBRE pricing excludes metered electricity.Source ↗
Total$241k per year$74k electricity + $167k capacity rent$251 per provisioned IT kW/month including electricity, or $0.43 per available GPU-hour. Replacing only the capacity rate with CBRE's range gives approximately $266k–$294k per year.

H200 hardware value anchors

Public system-level evidence · not fully like-for-like · no clean secondary-market index
DateReference pointImplied value / GPUWhat the number contains
Nov 2023NVIDIA H200 announcementSource ↗n.a.Launch milestone, not a price point. Commercial systems were expected to begin shipping in Q2 2024.
Sep 2024IREN purchase · $43.9m / 1,080 H200sSource ↗$40.6kObserved purchase including servers, InfiniBand, racks, cabling and labour. Strongest disclosed all-in transaction anchor.
Feb 2025WhiteFiber S-1 estimateSource ↗~$30.0kManagement estimate including networking, setup and shipping; useful directionally, but not an observed transaction.
Feb 2025Texas A&M DGX SuperPOD · $45m / 760 GPUsSource ↗$59.2kPremium integrated SuperPOD with Quantum-2 InfiniBand; not comparable with a bare GPU or standard node purchase.
Jun 2026Lenovo ThinkSystem · 8× H200Source ↗$49.7kLenovo's usual customer sale price of $397.8k for a configured 8-GPU system; a current replacement-cost reference, not resale evidence.
ModelYour all-in deployment cost
2029Your selected rack market value at exit