We own and finance the servers AI runs on.
Demand for computing capacity does not depend on any single AI application. We finance the NVIDIA GPU servers that AI cloud providers run, through full-stack ownership, where the fund owns the servers and leases them to the operator, and first-lien or second-lien loans. Our results depend on those providers and their customers paying.
Three forces behind the thesis
- Demand is running ahead of supply
The largest buyers of compute are committing capital and signing contracts faster than capacity can be built.
- Power and buildings set the pace
Grid equipment and powered buildings, not chips, now limit how fast new capacity arrives.
- Contracts can repay a server inside its life
Large owners depreciate servers over five to six years. Contracts with customer prepayments can repay a server’s cost in well under half that time.
Demand is running ahead of supply
McKinsey estimates that data centers built for AI will need about $5.2 trillion of investment by 2030 in its base case, within a range of $3.7 trillion to $7.9 trillion. About $3.1 trillion of it, 60%, goes to chips, servers and other computing hardware.1
The largest buyers are spending at that pace. Amazon, Alphabet, Microsoft and Meta plan more than $700 billion of capital spending in 2026, up from about $410 billion in 2025.2 Amazon alone expects to spend about $220 billion on capital projects in 2026, most of it for AWS.3
Microsoft reported a $678 billion commercial backlog at June 30, 2026, up 84% in a year, and said customer demand continues to exceed available capacity.4
At GTC in March 2026, NVIDIA CEO Jensen Huang said he sees at least $1 trillion of demand for Blackwell and Rubin systems through 2027, double the roughly $500 billion he cited a year earlier.5 Outside the largest companies, MarketsandMarkets projects the GPU-as-a-Service market to grow from $8.2 billion in 2025 to $26.6 billion in 2030, about 26.5% a year.6
$5.2trillion1
AI data center investment needed by 2030, McKinsey base case
- Chips, servers and other computing hardware
- About $3.1 trillion, 60% of the base case
- Other investment
- The rest of the base case
- Range across McKinsey’s scenarios
- $3.7 trillion to $7.9 trillion
Demand for AI compute, in three figures
Power and buildings set the pace
Microsoft CEO Satya Nadella said in late 2025 that Microsoft’s constraint had shifted from chips to power and ready data center space, and that it could have chips in inventory it was unable to plug in.7
128weeks8
Delivery time for a large power transformer in mid-2025, about two and a half years
- Grid equipment is slow to arrive
- Large power transformers took about 128 weeks, roughly two and a half years, to deliver in mid-2025, according to Wood Mackenzie. By the first quarter of 2026, generator step-up transformers, which connect power plants to the grid, took more than 160 weeks, a Wood Mackenzie analyst told Reuters.8 Grid equipment, not chips, now sets the pace of new capacity.
- Capacity stays tight
- Goldman Sachs Research expects data center capacity to stay tight through 2026, and global data center power demand to rise about 175% from 2023 levels by 2030.9
- Where we come in
- An operator that has secured power, space and a customer can sell compute under contract. What it often lacks is the capital for the servers. That is what we finance.
Contracts can repay a server inside its life
Large cloud companies depreciate servers over five to six years. Microsoft and Alphabet use six years, Meta 5.5, and Amazon five for part of its fleet since 2025. CoreWeave depreciates its GPU servers over six.10
Customer prepayments shorten the time it takes a server to repay its cost. Nebius expects its Q2 2026 deals to pay back their capex and operating costs in about 22 months, and where those deals include prepayments, they cover 50 to 60% of the capex.11 Nscale’s customer contracts include advance payments averaging 23% of total contract value.12
Rental rates fall as newer chips arrive, and they keep arriving. NVIDIA began production shipments of its next generation, Vera Rubin, in August 2026.13
We underwrite a six-year useful life, with lower rental rates assumed at re-lease. Actual life may be shorter. Each lease or loan we make ends no later than the customer contract that pays for it.
How GPU equipment finance compares
Owning an AI stock is a claim on future earnings, priced every second by the market. A GPU equipment lease or loan is a claim on rental income from specific hardware. Both are exposure to AI. They are different assets, with different risks.
| GPU equipment leases and loans | AI equities | Data center REITs | |
|---|---|---|---|
| Return comes from | Contracted lease and loan payments, and the equipment’s value when the contract ends | Earnings expectations and valuation multiples | Rents, dividends and share price |
| Upside | Limited by the contract and the equipment’s residual value | Uncapped if AI beats expectations | Rent growth and share price |
| Collateral | The equipment, owned or under a lien | None | Buildings and land, not the equipment inside |
| Correlation to the S&P 500 | Not measured. Lessee credit, equipment values and AI demand can be affected by the same conditions that move equity markets. | Measurable from daily prices | Measurable from daily prices |
| Liquidity | Illiquid. Capital is committed for the life of each lease or loan | Daily, on an exchange | Daily, for listed REITs |
| Transparency | Private. Information is limited to what the manager reports to investors | Public filings and daily prices | Public filings and daily prices |
| Minimum investment | Usually high, and open only to eligible investors | One share | One share, for listed REITs |
| Fees | Usually a management fee and carried interest, in addition to fund expenses | Trading costs, and expense ratios for funds | Trading costs, and the REIT’s own operating costs |
| Risk of loss | Lessee or borrower default, and falling equipment values and rental rates. Investors can lose some or all of their capital. | Share prices can fall sharply. Investors can lose some or all of their capital. | Rents, occupancy and share prices can fall. Investors can lose some or all of their capital. |
What industry leaders have said
In their own words, in date order, with the source of each.
November 2025
It’s not a supply issue of chips. It’s actually the fact that I don’t have warm shells to plug into.
Satya NadellaCEO, MicrosoftBG2 podcast7Warm shells are data center buildings with power ready for equipment. November 2025
We’ve entered the virtuous cycle of AI.
Jensen HuangCEO, NVIDIAThird quarter fiscal 2026 results14 February 5, 2026
This isn’t some sort of quixotic top-line grab.
AmazonQ4 2025 earnings call15Answering a question on returns from its capital spending. March 16, 2026
Right here where I stand, I see through 2027, at least $1 trillion.
Jensen HuangCEO, NVIDIAGTC keynote5
Quoted for market context. These companies are not affiliated with GPU Capital and do not endorse it.
The risks we underwrite against
A thesis is only as good as its downside case. These are the four risks that matter most to this one, and what we do about each.
| Risk | What could happen | How we underwrite against it |
|---|---|---|
| Rate compression | Rental rates fall as newer chips ship. | We lock the rate for the contract term and assume lower rates after the contract ends, when the servers are re-leased. |
| Obsolescence | A server loses value faster than planned when a new generation arrives. | We underwrite a six-year useful life, with lower rental rates assumed at re-lease. Actual life may be shorter. Each lease or loan ends no later than its customer contract. |
| Counterparty credit | An operator or its customer stops paying. | We read the customer contract first, take prepayments where available, and hold title or a lien. We can try to recover the servers and re-lease or sell them. That can be slow and costly, and what we recover depends on GPU prices at the time. |
| Power | A contract without power is a promise to deliver capacity that cannot yet run. | Signed power and colocation agreements, at least as long as the deal, are confirmed before capital is committed. |
See how the thesis becomes a deal.
Operators get a straight answer on whether a deployment fits and how we would structure it. Investors can request access to learn more about how we work.