Where GPU infrastructure fits in asset allocation
GPU server financing and AI stocks are both exposure to AI, but they are different assets. Where each return comes from, how cash comes back, and what can go wrong.
Marek Omilian, CFAManaging Partner, GPU Capital
Key takeaways
- GPU server financing is a claim on contracted rental income from specific hardware. AI stocks are a claim on future earnings, priced every second. Both are exposure to AI, and they behave differently.
- In a hypothetical B300 example, a 25% prepayment returns about 48% of the server’s cost at signing, and billing recovers the rest by about month 22 if the customer pays as agreed. If it stops after a year, about 77% is back.
- The risks are real: falling rental rates, obsolescence, customer default, operator execution, power and illiquidity. A deep AI downturn could hit rates, customer credit and hardware values at the same time as equities.
- Its correlation with equities has not been measured. Institutional investors generally classify it as private credit or real assets, not growth equity.
Contents
Institutional investors generally classify GPU server financing as private credit or real assets, not growth equity. It gives exposure to the AI build-out through contracted rental cash flows, backed by hardware that loses value as newer chips arrive. The return comes from rent, month by month, rather than from selling a stock at some future price. That is a different claim on AI than owning Nvidia, with different risks, and it behaves differently in a portfolio.
The gold rush analogy and its limit
In a gold rush, the steadiest fortunes went to the people supplying the miners. In 1848, San Francisco merchant Sam Brannan bought up picks, shovels and pans and sold them to arriving prospectors at steep markups; he became the city’s first millionaire, largely from selling supplies. Levi Strauss built his fortune selling canvas pants to the same crowd.
Most miners went home with little. The suppliers got paid whether or not any single claim struck gold.
The AI boom has the same shape. Frontier labs and AI applications are the miners, betting on which models and products will win. Compute is the shovel. Every one of them needs GPUs to train and serve models, and they pay for that compute by the hour, whoever ends up winning. The analogy has a limit: a shovel seller was paid at the counter, while a GPU owner is paid over the life of a contract, and only while its customer keeps paying.
The allocator’s dilemma: most AI exposure is one bet
The AI build-out is the largest capital program in the market today. Microsoft, Alphabet, Amazon and Meta plan about $725 billion of capex in 2026, up 77% from $410 billion in 2025, most of it on data centers and chips.
Most investors already own this theme, often without choosing to. According to RBC Wealth Management, the 10 largest S&P 500 companies reached a record 40.7% of the index at the end of 2025, and Nvidia alone was nearly 8%. RBC’s point is blunt: the index has become a directional bet on AI adoption, with fewer offsetting exposures if expectations slip.
So the question for an allocator is not whether to own AI. It is which part of AI to own, and through which return driver. Public equity gives you valuation-driven exposure to the miners and to the shovel maker. GPU server financing gives you the rental income from the shovels themselves.
GPU server financing compared with AI stocks
Nvidia stock is a claim on future earnings, priced every second by the market. GPU server financing is a claim on today’s rental income from specific hardware. Both are AI exposure; they are not the same asset.
| Feature | Nvidia and AI public equities | GPU server financing |
|---|---|---|
| What you own | A fractional share of a company’s future profits | Specific servers, plus the cash flows from their compute contracts |
| How return arrives | Price appreciation, realized when you sell; minimal dividend | Monthly cash from compute billing, starting in the first months |
| What drives value | Earnings expectations, valuation multiples, rates, index flows | Contracted $/GPU-hour, utilization, contract term, residual value |
| Pricing and volatility | Marked to market daily; sharp single-day moves | Not traded daily; cash flows follow the contract |
| Downside protection | None beyond the business itself | Hardware collateral and customer prepayments, which may cover only part of what is owed |
| Upside | Uncapped if AI beats expectations | Capped by contract economics and residual value |
| Liquidity | Daily | Illiquid for the life of the deal |
| Transparency | Public filings, audited accounts and a market price every day | Private reporting to investors; no market price |
| Minimum investment | The price of one share | Usually high, and open only to eligible investors |
| Fees | Low in index funds and ETFs | Usually higher in private vehicles: management fees and a share of profits |
| Risk of loss | Prices can fall sharply, and a single stock can lose most of its value | Defaults, falling hardware values and illiquidity can cause losses, up to the whole investment |
| Correlation with equities | Nvidia alone is nearly 8% of the S&P 500 | Not measured. Lessee credit, equipment values and AI demand can be affected by the same conditions that move equity markets |
A general comparison of two kinds of exposure. It does not describe any fund, security or investment.
Source: GPU Capital.
The volatility row is not theoretical. On January 27, 2025, a single headline about a cheaper Chinese AI model sent Nvidia shares down nearly 17% in one day, erasing about $593 billion of market value, a record for any US company. A GPU under a multi-year contract kept billing the same hourly rate that day, as long as its customer kept paying.
The tradeoff runs the other way on upside. If AI outruns every forecast, equity holders capture it and a financing position does not. That is why the two are complements, not substitutes: equities for growth, GPU financing for contracted income. Investors also give up daily liquidity, and part of the return is compensation for that.
What full-stack ownership is
In full-stack ownership, a capital provider buys GPU servers, owns them, and places them with a neocloud operator that runs them in a data center and sells the compute to end customers on multi-year contracts. The operator contributes no capital. The capital provider funds the whole asset, not a slice of a loan.
Capital provider
Buys the GPU servers and owns them. Puts in all the capital.
Neocloud operator
Runs the servers in a data center and sells the compute. Puts in no capital.
End customer
Signs a multi-year contract for the compute.
Capital provider: places the servers with the operator. Neocloud operator to End customer: compute, under a multi-year contract. End customer to Neocloud operator: prepayment at close, then monthly fees.
How the customer’s payments are applied, in order
- Data center costsPower, space and cooling are paid first.
- The owner’s capitalThe rest goes to the capital provider until it has recovered what it paid.
- Shared profitsOnly then are profits shared with the operator.
Source: GPU Capital.
The return path is the point. The end customer typically prepays part of the contract at close and then pays monthly for compute. After data center costs, that cash goes to the owner until it has recovered its capital. Only then are profits shared with the operator.
Why time is the main risk
In a fast-moving technology cycle, the main risk is time. Every month capital stays outstanding is another month of exposure to new chip generations and falling rental rates. A structure that aims to recover its capital early shrinks that window, but only if the customer pays as agreed.
Capital recovered from contract cash: a hypothetical B300 server
One 8-GPU node on a 4-year contract, capital recovered as a percent of its cost. If the customer pays as agreed, the prepayment returns about 48% at signing and billing recovers the rest by about month 22. If the customer stops paying after month 12, about 77% is back, and the rest depends on re-leasing or selling the node.
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. The dashed green segment marks the months after recovery, when cash is shared with the operator. The illustration counts recovery only, up to 100% of the server’s cost, and leaves out any residual value of the hardware. The downside line assumes the customer pays nothing after month 12 and nothing is yet recovered from re-leasing or selling the node.
Source: GPU Capital calculation from Rack Rate (July 2026), Petronella (as cited October 2, 2026) and NVIDIA DGX B300 (viewed October 30, 2026) data. Hypothetical industry illustration from public market inputs. Not a fund result, deal or projection.
Show data
| Months from signing | Hypothetical B300 server | Customer stops paying after month 12 (hypothetical) |
|---|---|---|
| Month 0 | 48% | |
| Month 6 | 63% | |
| Month 12 | 77% | 77% |
| Month 18 | 92% | 77% |
| Month 21.5 | 100% | 77% |
| Month 24 | 100% | 77% |
| Month 30 | 100% | 77% |
| Month 36 | 100% | 77% |
| Month 42 | 100% | 77% |
| Month 48 | 100% | 77% |
The example is a hypothetical industry illustration built from public market inputs. It does not describe any specific fund, deal or investment:
| Input | Assumption |
|---|---|
| Server cost | $640,000 per 8-GPU HGX B300 node, the low end of Rack Rate’s indicative allocation range |
| Contract | 4 years at $4.40 per GPU-hour, billed on all 8 GPUs |
| Customer prepayment | 25% of contract value, about $308,000, paid at close |
| Data center cost | About 11 kW per node at $350 per kW-month, about $3,850 per month. H100 and H200 class HGX nodes typically draw 10 to 12 kW; NVIDIA lists about 14 kW for a DGX B300 system, which would raise the cost to about $4,900 a month |
Two things drive the curve. The prepayment returns almost half the capital on day one. Monthly billing, after data center costs, returns the rest by about month 22 of a 48-month contract, if the customer pays on time. If it stops paying after a year, about 77% of the cost is back, and the rest depends on re-leasing or selling the node at whatever GPUs fetch then.
The dashed green segment marks the months after payback, when cash is shared with the operator. The curve stops counting at 100% of the server’s cost and leaves out any residual value of the hardware at contract end.
Where it fits in an asset allocation
Institutional investors usually classify GPU server financing as an alternative, alongside private credit and equipment finance. It is income-producing, backed by equipment whose value falls as newer chips arrive, and illiquid for the life of the deal.
- Price appreciation; Daily liquidity.Public growth equityNvidia and AI stocks. Marked to market daily.
- Price appreciation; Locked for the life of the deal liquidity.Private equity and ventureGains realized when a company is sold or listed.
- Contracted income; Daily liquidity.Public bondsCoupons, with principal at maturity.
- Contracted income; Locked for the life of the deal liquidity.Private credit and real assetsPrivate loans, equipment finance such as aircraft and railcars, and real assets.GPU server financing
Qualitative placement by how each return arrives and how liquid it is. Not a model portfolio or a recommendation.
Source: GPU Capital.
The closest mental model for most allocators is equipment leasing: aircraft, railcars, shipping containers. In each case an owner finances a productive asset, an operator runs it, and contracted payments repay the capital while the asset still has resale value. GPUs fit that pattern, but with much shorter economic lives and faster price declines, so their contracts have to repay the capital sooner. We underwrite a six-year useful life, with lower rental rates assumed at re-lease. Actual life may be shorter.
It also fits where private credit is heading. Preqin projects private credit AUM will more than double to $4.5 trillion by 2030. Macquarie expects allocations to asset-backed finance to rise as investors look for differentiated exposure. GPU financing is asset-backed finance with a specific, fast-growing collateral type.
Sizing is a decision for each investor and its advisers, and this note does not suggest one. Institutions that hold asset-backed finance usually count it within their private-credit or real-asset allocations.
The risks, stated plainly
GPU financing has real risks, and the biggest one is well documented. The H100 rented for $7 to $10 per GPU-hour at launch and $2 to $4 by late 2025, according to Silicon Data. Anyone underwriting GPUs has to assume newer chips will push older ones down the price curve.
| Risk | What can go wrong | How the structure responds | What it does not fix |
|---|---|---|---|
| Rental rate compression | Market rates fall as new chip generations ship | The contracted rate is fixed for the term; underwriting assumes lower rates after the contract ends | Re-leasing after the term happens at the lower market rate |
| Obsolescence | Hardware loses value faster than planned | A payback target well inside the hardware’s life, and residual value set by appraisal | Residual values are uncertain and can fall quickly; an appraiser affiliated with the owner is not independent |
| Customer default | The end customer stops paying | Prepayment collected at close; the owner holds title and can redeploy the servers | Recovery can be slow and costly, and a resale may bring much less than is owed |
| Operator execution | Delays in installation, weak uptime | Operator diligence; the operator is paid from profits after capital recovery, which aligns incentives | An operator failure can stop billing while the servers are moved |
| Power and data center | Capacity or cooling is not ready on time | Power and site confirmed before capital is committed | Outages or a site dispute can still interrupt billing |
| Illiquidity | Capital is locked up for the deal’s life | No structural fix | An investor may not be able to exit early, or only at a discount |
Source: GPU Capital.
Time is the thread that ties the mitigants together. Most of these risks do the most damage late in an asset’s life, so the sooner capital comes back, the smaller the window in which they can bite. None of the mitigants removes a risk, and several can arrive together.
Diversification: a different return driver, not zero correlation
GPU financing may diversify an equity-heavy portfolio because its return comes from a different place. Contract cash flows depend on compute demand, contract terms and operating execution, not on daily sentiment or valuation multiples.
It would be wrong to call it uncorrelated. Its correlation with equities has not been measured, and lessee credit, equipment values and AI demand can be affected by the same conditions that move equity markets. A deep AI downturn would pressure rental rates, customer credit and hardware values at the same time equity prices fall. The asset class is also young, with only a few years of rental-rate history and no long track record through a full credit cycle.
The market is maturing quickly, though. Daily rental indices now exist. Silicon Data describes hedging H100 rental costs with newly listed compute futures. Better price data makes the asset easier to underwrite, value and, eventually, trade.
Eight due diligence questions for GPU equipment finance
Due diligence on any GPU equipment financing comes down to questions like these. The answers separate disciplined GPU underwriting from AI enthusiasm.
- How many months to principal recovery, under the base case and a downside rate case?
- Is the hourly rate contracted for a fixed term, or exposed to the spot market?
- What rental rate is assumed after the contract ends, and what benchmark is it based on?
- How much of the capital is covered by customer prepayment at close?
- Who owns the hardware, and what happens to it if the customer or operator fails?
- How is residual value set, and is it independently appraised?
- Are power and data center capacity confirmed before capital is committed?
- How is the operator paid, and are its incentives aligned with capital recovery?
The bottom line
In a gold rush, you can bet on the miners or own the shovels. Public AI equities are a bet on who wins and at what valuation. GPU server financing is a claim on the rent every winner has to pay, with the hardware as collateral. It carries its own risks: the renter can stop paying, and the hardware can lose value faster than planned.
That is why institutional investors generally classify it with private credit and real assets, not with AI equities.
Neocloud operators: if you have customer demand and data center capacity but not the capital for the servers, talk to GPU Capital about full-stack ownership. Lenders and lessors: for valuation of GPU collateral, see GPU Appraisals, a practice affiliated with GPU Capital; its valuations of equipment GPU Capital owns or finances are not independent (conflicts of interest).
Sources
- Quote Investigator: the origin of the pick-and-shovel adage quoteinvestigator.com
- Yahoo Finance: hyperscaler capex plans for 2026 finance.yahoo.com
- RBC Wealth Management: The Great Narrowing, S&P 500 concentration rbcwealthmanagement.com
- Korea JoongAng Daily (Reuters): record Nvidia one-day loss, January 2025 koreajoongangdaily.com
- Rack Rate: HGX B300 allocation pricing rackrate.ai
- Petronella: HGX server power requirements petronellatech.com
- NVIDIA: DGX B300 system power nvidia.com
- IQ-EQ: private credit market trends for 2026 iqeq.com
- Macquarie Asset Management: private credit outlook 2026 macquarie.com
- Silicon Data: H100 rental price over time silicondata.com
- Silicon Data: hedging with compute futures silicondata.com
Quoted for market context. These companies are not affiliated with GPU Capital and do not endorse it.
Updated October 30, 2026. We revised this note as general market commentary: it no longer refers to any fund, it gives the risks the same weight as the cash flows, and it adds a downside case to the hypothetical example. We also corrected the description of GPU Appraisals, which is affiliated with GPU Capital and not independent, and noted the effect of a higher B300 power draw on the example. Exhibit 3 now shows capital recovery only: the comparison with a bond and the cash figures beyond full recovery were removed.
Disclosure
The payback example is a hypothetical industry illustration built from public market inputs. It does not represent any GPU Capital investment or predict future results.
GPU Capital is affiliated with GPU Appraisals. See conflicts of interest.
This post is general market commentary for information only. It reflects the author’s views on the date shown, which may change. It is not investment advice, and it is not an offer to sell or a solicitation of an offer to buy any security. GPU Capital, its funds or its affiliates may have relationships with, or interests in, companies or equipment mentioned.
GPU Capital finances NVIDIA GPU servers for AI cloud providers.
Through full-stack ownership and first-lien or second-lien loans, at $1M to $5M per deployment.
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.