Contact us

For AI cloud operators and investors

Asset-backed financing for AI compute.

GPU Capital owns NVIDIA GPU servers and leases them to AI cloud providers under multi-year contracts, or lends against them with a first or second lien.

Apply for financingRequest access

GPU Capital, Seattle, Washington. Led by Marek Omilian, CFA.

Firm facts

As of September 2026

Deployment size
$1M to $5M1Loans from about $1M. Full-stack ownership up to about $5M.
Structures
3Full-stack ownership, first-lien loan and second-lien loan.
Useful life we underwrite
6 years2With lower rental rates assumed at re-lease. Actual life may be shorter.
Cloud providers in discussion
70+3AI cloud providers in financing discussions with us.
GPU equipment projects
104Our Managing Partner has invested in them personally.

For operators

Financing for GPU servers, sized for neoclouds

  • $1M to $5M per deployment, as full-stack ownership or a first-lien or second-lien loan.
  • Terms sized to your customer contract and to an appraisal of each server.
  • You work directly with our Managing Partner, who makes credit decisions.
Apply for financingHow operator financing works

Financing is subject to our due diligence, credit approval, available capital and signed final agreements. Nothing on this website is a commitment to lend or lease.

For investors

Contracted payments, secured by hardware

  • The fund we manage holds title to the servers it leases, or a first or second lien on the servers it finances.
  • Its cash comes from contracted monthly payments for compute, which depend on lessees and borrowers paying as agreed.
  • Access is by request. We review every request ourselves before sharing any investor materials.
Request accessHow we invest

We own the servers, or we lend against them

We own and finance the NVIDIA GPU servers that AI cloud providers run on. In both structures a signed customer contract is the source of repayment, and the servers are the collateral.

Key to the diagrams

  • Cash
  • Title
  • Lien
  • Runs the servers

Full-stack ownership

Who holds title
The fund. It buys the servers, and the operator runs them for its customer.
How cash comes back
After data center costs, the customer’s prepayment and monthly fees go to the fund until it has recovered what it paid for the servers. Profits after that are shared with the operator.
Full-stack ownership: title and cashThe fund holds title to the servers and places them with the operator, who runs them. The end customer pays a prepayment and monthly fees to the operator. After data center costs, that cash goes to the fund until it has recovered what it paid for the servers, and then to a profit share with the operator.End customerPays for computeOperatorAI cloud providerThe fundOwner and lessorPrepayment and monthly feesData center costsCapital recovery, then profit shareRuns themHolds titleGPU serversFull-stack ownership: title and cashThe fund holds title to the servers and places them with the operator, who runs them. The end customer pays a prepayment and monthly fees to the operator. After data center costs, that cash goes to the fund until it has recovered what it paid for the servers, and then to a profit share with the operator.End customerPays for computeOperatorAI cloud providerThe fundOwner and lessorPrepayment andmonthly feesAfter data center costs:capital recovery,then profit shareRuns themHolds titleGPU servers

First-lien or second-lien loan

Who holds title
The operator. The fund holds a first or second lien on the servers and on the assigned customer contract.
How cash comes back
The operator pays interest and principal monthly, and the loan is repaid inside the customer contract term.
First-lien or second-lien loan: title and cashThe operator holds title to the servers and the fund holds a first or second lien on them. The fund advances the loan at closing. The end customer pays monthly fees to the operator, and the operator pays the fund interest and principal each month.End customerPays for computeOperatorAI cloud providerThe fundSecured lenderLoan at closingFees for computeInterest and principal, monthlyHolds titleFirst orsecond lienGPU serversFirst-lien or second-lien loan: title and cashThe operator holds title to the servers and the fund holds a first or second lien on them. The fund advances the loan at closing. The end customer pays monthly fees to the operator, and the operator pays the fund interest and principal each month.End customerPays for computeOperatorAI cloud providerThe fundSecured lenderFees for computeInterest andprincipal, monthlyLoan at closingHolds titleFirst or second lienGPU servers

If an operator defaults, we can take back the servers, by repossession or by enforcing our lien, and re-lease or sell them. That can be slow and costly, and it may recover much less than is owed, because it depends on GPU prices at the time.

How we structure each deployment

Hypothetical illustration

How capital comes back

One 8-GPU HGX B300 server on a four-year compute contract, built from public market inputs. The line shows how much of the server’s cost has been recovered. Recovery depends on the customer paying as agreed, and it is not investor liquidity, which a fund’s terms govern.

  • Hypothetical B300 server, until recovered
  • After recovery, cash shared with the operator
Capital recovered from contract cash, hypothetical B300 serverLine chart. Hypothetical B300 server: the share of the server's cost recovered starts at 48% at signing and reaches 100% at about month 22 of the 48-month contract. After recovery the line holds at 100% for the months in which cash is shared with the operator. Resale value at the end of the contract is not counted.0%25%50%75%100%Capital recovered, % of server cost012243648Months from signingServer cost recovered in full48% at signingHypothetical B300 serverCapital recovered, about month 22After recovery, cash is shared with the operatorContract ends at month 48Resale value not countedCapital recovered from contract cash, hypothetical B300 serverLine chart. Hypothetical B300 server: the share of the server's cost recovered starts at 48% at signing and reaches 100% at about month 22 of the 48-month contract. After recovery the line holds at 100% for the months in which cash is shared with the operator. Resale value at the end of the contract is not counted.0%25%50%75%100%Capital recovered, % of server cost01224364848% at signingRecovered, aboutmonth 22Then shared with theoperatorContract endsResale not counted
Hypothetical industry illustration from public market inputs. Not a fund result, deal or projection.Inputs: server $640,000; 4 years at $4.40 per GPU-hour on 8 GPUs; customer prepayment of 25% of contract value (about $308,000) credited evenly against billing; about 11 kW at $350 per kW-month (about $3,850 a month). If the server draws about 14 kW, the figure NVIDIA lists for a DGX B300 system, data center costs rise to about $4,900 a month and recovery moves to about month 23. Source: GPU Capital calculation from Rack Rate (July 2026), Petronella (as cited October 2, 2026) and NVIDIA DGX B300 (viewed October 30, 2026) data.
Show data
MonthShare of server cost recovered
048%
663%
1277%
1892%
21.5 (recovered)100%
48 (contract ends)100%

  1. The customer prepays at signing

    The customer pays 25% of the contract value up front, about $308,000. That returns about 48% of the server’s $640,000 cost on day one.

  2. Monthly fees return the rest

    The customer then pays for compute every month. After data center costs of about $3,850, about $15,400 a month goes toward the server’s cost.

  3. Capital is recovered at about month 22

    Cumulative cash reaches 100% of the server’s cost about 22 months into the 48-month contract.

  4. Then cash is shared with the operator

    For the rest of the contract, the operator that runs the servers shares in the cash. This illustration tracks recovery only, so the line stays at 100%.

  5. The hardware still has resale value

    When the contract ends, the servers can be leased again or sold. Their value is appraised separately from the contract and is not counted in this curve.

Underwriting in five steps

Each deployment goes through the same five steps, from the first conversation to monthly monitoring.

  1. Origination

    Direct relationships with operators and server manufacturers, and referrals from investors and industry contacts.

  2. Appraisal

    Each server is valued on market comparables and on income, assuming a six-year useful life. The valuations draw on GPU Appraisals, an affiliated appraisal practice, so its valuations of equipment we own or finance are not independent. Conflicts of interest

  3. Structuring

    Tenor matched to the customer contract and inside the hardware’s useful life. Prepayment and reserves where available. Title or lien perfected.

  4. Approval

    Our Managing Partner, who makes credit decisions, approves each deployment against the fund’s investment criteria.

  5. Monitoring

    Monthly payments, utilization and hardware condition are tracked, with a remarketing plan ready if a lessee fails.

What the fund finances

Four NVIDIA server platforms appear in our deal flow. We underwrite each one at the node, because the 8-GPU server is what gets bought, racked, rented and financed.

Read the payback comparison of B300, H200 and RTX PRO 6000

PlatformConfigurationWhat drives payback
NVIDIA HGX B3008 Blackwell Ultra GPUs, 288 GB of HBM3e each. Air-cooled and liquid-cooled servers.5The contract rate locked at signing. It needs the most capital per node, and its successor, Vera Rubin, began production shipments in August 2026.6
NVIDIA HGX B2008 Blackwell GPUs.The term rate, not the spot rate. On September 7, 2026, the B200 index stood at $5.69 per GPU-hour spot and $5.39 for a 12-month term. Lenders size on the lower number.7
NVIDIA HGX H2008 Hopper GPUs, 141 GB of HBM3e each.5What it earns after the contract ends, when it re-leases two generations behind the newest chips.
NVIDIA 8-GPU RTX PRO 6000 Server Edition8 passively cooled PCIe cards, 96 GB of GDDR7 each.5The purchase price. The node pays back well when it is bought well below list.

  • NVIDIA HGX B300

    Configuration
    8 Blackwell Ultra GPUs, 288 GB of HBM3e each. Air-cooled and liquid-cooled servers.5
    What drives payback
    The contract rate locked at signing. It needs the most capital per node, and its successor, Vera Rubin, began production shipments in August 2026.6

  • NVIDIA HGX B200

    Configuration
    8 Blackwell GPUs.
    What drives payback
    The term rate, not the spot rate. On September 7, 2026, the B200 index stood at $5.69 per GPU-hour spot and $5.39 for a 12-month term. Lenders size on the lower number.7

  • NVIDIA HGX H200

    Configuration
    8 Hopper GPUs, 141 GB of HBM3e each.5
    What drives payback
    What it earns after the contract ends, when it re-leases two generations behind the newest chips.

  • NVIDIA 8-GPU RTX PRO 6000 Server Edition

    Configuration
    8 passively cooled PCIe cards, 96 GB of GDDR7 each.5
    What drives payback
    The purchase price. The node pays back well when it is bought well below list.

Demand, power and financing, in three figures

  1. $678 billion8

    Microsoft’s commercial backlog at June 30, 2026, up 84% in a year. Microsoft said customer demand continues to exceed available capacity.

  2. 175%9

    Approximate rise in global data center power demand from 2023 levels by 2030 expected by Goldman Sachs Research, which also expects data center capacity to stay tight through 2026.

  3. $8.5 billion10

    GPU-backed loan closed by CoreWeave in March 2026 and rated A3 by Moody’s, the first investment-grade financing of its kind. The rating applies to that facility, not to the company. Those financings differ greatly in size, structure and credit quality from ours.

Quoted for market context. These companies are not affiliated with GPU Capital and do not endorse it.

Capital spending by Amazon, Alphabet, Microsoft and Meta

The four companies plan more than $700 billion of capital spending in 2026, up from about $410 billion in 2025.

Capital spending by Amazon, Alphabet, Microsoft and MetaBar chart. 2025: $410B; 2026, planned: $725B.$410B2025$725B2026, plannedCapital spending by Amazon, Alphabet, Microsoft and MetaBar chart. 2025: $410B; 2026, planned: $725B.$410B2025$725B2026, planned

Source: Tom’s Hardware, April 30, 2026. Company guidance as of late October 2026 adds up to about $725 billion to $745 billion.

Show data
Value
2025$410B
2026, planned$725B

Demand alone does not repay a loan. Customers paying under signed contracts do. We look for deployments whose contracts are scheduled to repay their cost inside the servers’ useful life.

Read our thesis

Market notes from the Managing Partner

Sourced notes on GPU infrastructure credit: contracts, rental rates, hardware values and how we underwrite them.

All insights

Market

Fiber in 1999, GPUs in 2026

GPU servers are not the new fiber. A fiber glut could wait decades for demand to catch up; a GPU glut expires with the hardware. A GPU loan has to be repaid by today’s customers before that happens.

By Marek Omilian, CFA

Start the conversation.

Operators can find out whether a deployment may fit and how it could be structured. Investors can request access, and we review each request before discussing anything further.

Apply for financingRequest access