Underwriting that starts with an appraisal.
We value the servers before we set terms. Our Managing Partner also leads GPU Appraisals, an affiliated appraisal practice. Because it is affiliated, its valuations of equipment we own or finance are not independent. See conflicts of interest.
- Eight GPUs: the collateral
- 800 Gb/s networking
- NVLink switch fabric
- Rack power and cooling
Three values, and the one that sets the advance rate
We value a server two ways: against market comparables, and on the income it can still earn over its useful life. The valuation gives three values, and each answers a different question.
Going-concern value
What is it worth at work?
Its value in continued operation: installed, connected and earning. It supports the credit case.
Fair market value
What is it worth today?
What the hardware trades for today, between a willing buyer and a willing seller.
Orderly liquidation value
What would we recover?
What the servers would realistically bring in an orderly sale. It sizes the downside, so it sets the advance rate.
The loan amount
Set from orderly liquidation value
The advance rate is the share of the node’s cost we lend. It is set from the appraised orderly liquidation value, and it moves with the quality of the contract and the age of the node. In full-stack ownership the fund buys the whole node, so there is no advance rate.
The usual order, not to scale. Orderly liquidation value is usually the lowest of the three, because it assumes a sale within a limited time.
The five things we underwrite
In the order we work through them. The contract comes first, because GPUs lose value every year and the contract is what repays the capital.
The customer contract
Can the customer pay for the whole term?
Who the customer is, how long the contract runs, whether it is take-or-pay, and how much has been prepaid. We size on the contract rate, not the spot rate.
The operator
Has this team run GPUs for a customer before?
How many nodes, for how long and for whom, with uptime history from your monitoring system and your customers’ SLA reports, not from a slide.
The hardware
Will the node still earn at the end of the financing?
The whole node, not a per-card price: generation, age and exact spec, including memory, networking, CPU, and HGX or PCIe. An older node earns less, so it supports a shorter term.
The site and power
Is the power signed, sized and as long as the financing?
Without power there is no cash flow. We read the power and colocation agreements, and check that the servers could be repossessed, re-leased and maintained where they sit.
The structure
Does the financing repay inside the contract?
Tenor matched to the customer contract and inside the hardware’s useful life. An advance rate set from orderly liquidation value. Prepayment and reserves where available. Title or lien perfected.
See the full checklist for operatorsRead the lender’s checklist
Residual value: what a server is worth when its contract ends
Rental rates fall as newer chips arrive. We underwrite a six-year useful life, with lower rental rates assumed at re-lease. Actual life may be shorter.1
Loans are structured to be repaid inside the customer contract, so repayment does not count on a renewal or a sale. If a lessee or borrower defaults, we may repossess the servers and try to re-lease or sell them. That can be slow and costly, and what we recover depends on GPU prices at the time. That is why the advance rate comes from orderly liquidation value, not from the purchase price.
Illustrative
The value of one server over a six-year life
On these assumptions a server is worth about a quarter of its cost when a three-year contract ends, against half on a six-year straight line.
Source: GPU Capital illustration, not fund data, an appraisal or a forecast. Here a server is worth the rent it can still earn over the rest of a six-year life, with market rates for its generation falling 30% a year, the steepest decline in our payback analysis, and nothing left at year six. The straight line is a six-year depreciation schedule.
Show data
| Years from purchase | Illustrative case | Straight line |
|---|---|---|
| Year 0 | 100% | 100% |
| Year 0 | 90% | 96% |
| Year 1 | 82% | 92% |
| Year 1 | 73% | 88% |
| Year 1 | 66% | 83% |
| Year 1 | 59% | 79% |
| Year 2 | 53% | 75% |
| Year 2 | 47% | 71% |
| Year 2 | 42% | 67% |
| Year 2 | 38% | 63% |
| Year 3 | 33% | 58% |
| Year 3 | 29% | 54% |
| Year 3 | 26% | 50% |
| Year 3 | 22% | 46% |
| Year 4 | 19% | 42% |
| Year 4 | 16% | 38% |
| Year 4 | 14% | 33% |
| Year 4 | 12% | 29% |
| Year 5 | 9% | 25% |
| Year 5 | 8% | 21% |
| Year 5 | 6% | 17% |
| Year 5 | 4% | 13% |
| Year 6 | 3% | 8% |
| Year 6 | 1% | 4% |
| Year 6 | 0% | 0% |
Payback calculator
With the base-case inputs it reproduces our payback analysis: 27 months for HGX B300 and 22 for HGX H200 and RTX PRO 6000, on three-year take-or-pay contracts with no prepayment.2
Hypothetical industry illustration from public market inputs. Not a fund result, deal or projection, and not an offer or a commitment to lend.
Illustrative calculator
Estimate payback for a GPU deployment
Pick a server platform, set the contract and your costs, and see when the node’s cost comes back from contract cash. Presets follow GPU Capital’s September 29, 2026 analysis of B300, H200 and RTX PRO 6000 payback.
Payback
26.5 months
Capital recovered in month 27 of a 36-month contract.
- Monthly net cash
- $21,594 a month
- Revenue
- $26,747 a month
- Less colocation
- $3,816 a month
- Less operations reserve
- $1,337 a month
- Less prepayment credit
- $0 a month
- Contract value
- $962,899
- Customer prepayment
- $0 at signing
- Capital at risk node cost less prepayment
- $572,364
- Cash returned by end of term
- 136% of node cost
- Rate needed to pay back within the term
- $3.55 per GPU-hour
With the loan
- Loan amount
- $400,655
- Monthly debt service
- $13,307 a month for 36 months
- Debt service coverage (DSCR) monthly net cash / monthly debt service
- 1.62x
- Operator equity node cost less loan and prepayment
- $171,709
- Repaid inside the contract term
- Yes
The loan amortizes over 36 months, inside the 36-month contract, and monthly net cash covers the payment.
Lenders commonly look for a DSCR of about 1.15x to 1.3x.
An estimate from your inputs and public market rates. It is not an offer of financing or a result of any GPU Capital fund.
Methodology and sources
- Revenue a month = 8 GPUs × nodes × contract rate × 730 hours × billed utilization.
- Colocation = nameplate kW ÷ 0.8 (provisioned kW) × $ per kW-month × nodes.
- Operations reserve = a share of revenue, for remote hands, replacement parts and monitoring.
- Contract value = revenue a month × term. The prepayment is a share of contract value, paid at signing and credited back evenly against billing.
- Monthly net cash = revenue less colocation, the operations reserve and the prepayment credit (prepayment ÷ term).
- Capital at risk = node cost less prepayment. Payback = capital at risk ÷ monthly net cash, in months.
- Cash returned by end of term = (prepayment + monthly net cash × term) ÷ node cost.
- Rate needed to pay back within the term = (node cost ÷ term + colocation per node) ÷ (1 less the operations reserve) ÷ (8 × 730 × utilization).
- Loan: a level monthly payment = loan × r ÷ (1 less (1 + r) to the power of minus n), where r is the yearly rate ÷ 12 and n is the amortization in months, capped at the contract term. DSCR = monthly net cash ÷ monthly debt service. Operator equity = node cost less the loan and the prepayment, not below zero. The loan counts as repaid inside the term when it amortizes within the term and monthly net cash covers the payment.
- Left out: networking fabric, storage, installation, taxes, operator fees and the hardware’s residual value. Market rates move weekly; check current inputs before relying on any figure.
- With the B300, H200 and RTX PRO 6000 presets, a 36-month term and no prepayment, payback is 27, 22 and 22 months, as in the September 29, 2026 analysis.
Preset sources
- B300: node price and nameplate power from the UpStation catalog (September 9, 2026); rate is the low end of 36-month 8-GPU listings on GetDeploying (September 24 to 29, 2026).
- B200: midpoint of the UpStation catalog range of $390,000 to $410,000 for an Exxact build, and its 9.6 kW nameplate (prices as of October 13, 2026); rate is the low end of 36-month 8-GPU listings on GetDeploying (October 30, 2026).
- H200: node price and nameplate power from the UpStation catalog (September 9, 2026); rate is the Silicon Data H200 index (late September 2026).
- RTX PRO 6000: a recent quote in GPU Capital deal flow; rate is the longest published 8-GPU listing, 12 months, on GetDeploying (September 2026), held for the term.
- Colocation: $225 per kW-month on provisioned kW, UpStation power and colocation pricing, three-year term (September 2026).
Illustrative arithmetic, not an offer of financing or a projection of fund returns.
Contract rates, not on-demand rates
Headlines quote on-demand prices. Lenders get paid on contract, and the gap is wide.
On October 30, 2026, the rates in our base case sat at 48% to 64% of the on-demand median for the same GPU. We size financing on the contract rate.3
Rental rates for one GPU, October 30, 2026
The base case sits well below the on-demand median for every platform.
- On-demand median
- Reserved median, all terms
- Base case we used
Source: GetDeploying medians as of October 30, 2026; base case from our September 29, 2026 payback analysis, with the Silicon Data H200 index for the H200 base case. GPU Capital calculation.
Show data
| On-demand median | Reserved median, all terms | Base case we used | |
|---|---|---|---|
| HGX B300 | $9.50 | $5.16 | $4.58 |
| HGX H200 | $4.88 | $3.62 | $3.10 |
| RTX PRO 6000 | $2.36 | $1.90 | $1.39 |
Payback is one test, not the only one
Residual value, network fabric costs and the customer’s credit all move the result. So alongside payback we screen deployments on an internal measure we call the GPU Capital Yield Score. It is a screening tool for our own credit decisions, not a published score or a rating.
Quoted for market context. These companies are not affiliated with GPU Capital and do not endorse it.
Bring us the contract and the quote.
Operators can find out whether a deployment may fit and how it could be structured. Investors can request access to learn more about how we work.