Owning AI GPUs: Economics, Risks and the Asset-Class Test

Review period: August 29, 2023–October 3, 2026 Scope: U.S.-oriented GPU ownership, leasing and cloud-operation economics, compared with data-center infrastructure, equipment leasing and other technology assets. Source policy: Prioritize primary filings, transaction announcements, official pricing and statistics; use original institutional and academic research for context, and label issuer claims and modeled results.

Analysis horizon: Four operating years in the central illustration, with three-to-six-year remaining-life sensitivities. This is separate from the evidence review period.

Executive summary

AI GPUs can provide a distinct, investable exposure—but the strongest case is currently for a specialized segment of equipment finance and AI infrastructure, not a proven, low-correlation standalone asset class. The evidence establishes rental markets, substantial equipment purchases and financeable customer contracts. It does not yet establish representative, through-cycle hardware-owner returns after capital consumption, operating costs, defaults and resale losses. GPU-backed investment-grade financing is meaningful progress, but its security includes customer contracts and infrastructure rights, not merely chips. (s205.q4cdn.com)

• Separate the investment structures. Merchant GPU rental, a fixed equipment lease, a managed AI cloud and a powered data-center shell expose investors to different risks. Their revenue, costs and returns cannot be compared interchangeably. (s205.q4cdn.com)
• A high cash yield can conceal weak investment returns. In the Revoqra-derived illustration below, a $35,000 installed GPU allocation generates $8,935 of annual unlevered cash flow—a 25.5% cash yield—but only approximately 4.5% IRR over four years, including 10% net resale value. These are hypothetical calculations, not observed market returns; the acquisition and power evidence used to frame the assumptions is identified below. (irisenergy.gcs-web.com)
• Entry price, paid utilization and realized rent dominate. At a 12% illustrative hurdle, the central case has approximately −$5,600 NPV per GPU. Holding other assumptions constant, 85% paid utilization at $3.50/hour produces approximately +$27,000 NPV; a 20% annual rental-price decline produces approximately −$15,100 NPV. No scenario probabilities are assigned.
• Useful life is uncertain, not necessarily short. CoreWeave reports selected A100 and H200 recontracting examples that support continuing value for older architectures. Conversely, Amazon shortened the accounting life of a subset of servers and networking equipment because of faster AI-related technological development. Neither observation establishes a universal GPU depreciation curve. (sec.gov)

Investment implication: A compelling transaction should recover capital primarily from credible contracted cash flows, with conservative financing and limited dependence on an optimistic exit valuation. Merchant GPU ownership can be profitable, but it requires underwriting an operating business and technology cycle—not merely owning scarce hardware.

1. What exactly would the investor own?

The relevant economic unit is usually a deployable system plus contractual rights, rather than an isolated accelerator. Provider offerings bundle GPUs with CPU capacity, RAM, storage and a software environment; large deployments also require an appropriate interconnect and operating platform. NVIDIA's own asset-class argument concerns a full-stack AI factory, not just a chip. (lambda.ai)

The table is a Revoqra classification, informed by disclosed cloud offerings, equipment-finance accounting and GPU-backed financing structures—not a claim that these exposures have identical historical return behavior. (lambda.ai)

This distinction matters particularly for financing. CoreWeave's March 2026 facility takes security over infrastructure, the underlying cloud-services customer contract and corresponding data-center leases. Its investment-grade rating should therefore not be interpreted as a rating of the resale stability of uncontracted GPUs. (s205.q4cdn.com)

2. What the available evidence establishes

2.1 Acquisition costs: use complete-system quotes, not headline chip prices

Two disclosed purchases provide useful, but non-comparable, reference points:

Method: Divide the company-reported aggregate purchase amount by its reported GPU count. These are nominal purchase/order-level benchmarks, not verified current standalone-card prices. Package scope, volume terms, support and architecture differences prevent a clean price-change comparison.

The difference between a chip and an installed deployment is substantial enough to require a detailed bill of materials. IREN's November 2025 Microsoft transaction, for example, paired a five-year, approximately $9.7 billion cloud-services contract with approximately $5.8 billion of GPU and ancillary-equipment purchases from Dell. The announcement separately describes liquid-cooled data-center delivery. It does not provide a clean bare-GPU unit price. (sec.gov)

Underwriting requirement: Obtain executable quotes covering accelerators, hosts, memory, storage, networking, racks, cooling interfaces, support, delivery, applicable taxes and commissioning. Allocate shared infrastructure consistently, and identify reusable components separately from generation-specific equipment. The reviewed evidence does not supply a sufficiently robust current purchase-price index to replace those quotes.

2.2 Rental pricing: customer prices are not owner receipts

The following are posted USD prices reviewed for this report, not verified transaction prices or assurances of available capacity:

Sources: Runpod's page labels its update September 27, 2026; Lambda publishes configuration-specific rates. N = two providers and six quoted configurations, selected for illustration—not a market average. CPU, RAM, storage, deployment size, support and networking differ, so these are not perfectly matched offers. (runpod.io)

The investor must then bridge:

Posted customer rate → negotiated and collected rate → owner receipts → cash flow after operating costs.

Distribution can materially change that bridge. Vast's FAQ states that hosts receive 75% of successful-job revenue, while its earnings article describes public prices as typically about 25% above host earnings. Those descriptions are not arithmetically identical: a 25% markup on host receipts is a 20% share of customer spending. Executed terms and settlement records should govern underwriting; the model below uses a 25% revenue-share sensitivity rather than assuming a small distribution charge is universally available. (console.vast.ai)

2.3 Financing: the customer contract can matter more than the GPU

Disclosed financing demonstrates substantial differentiation by structure and customer credit:

These are dated, source-reported terms—not current October 3 borrowing quotes or expected lender returns after fees and credit losses.

Coupon or spread is only part of financing cost. CoreWeave's filing describes undrawn fees, substantial interest-rate hedging requirements and restricted-cash requirements in its project financing. Such protections improve lender resilience but can increase the equity capital tied up or the effective funding cost. (sec.gov)

Customer prepayments also need correct treatment. IREN's Microsoft announcement includes a 20% prepayment. That improves liquidity, but it is payment for future services—not additional lifetime revenue or automatically free economic capital. (sec.gov)

2.4 Useful life and resale: distinguish three different concepts

1. Physical life: How long equipment operates reliably.
2. Accounting life: The depreciation estimate used in financial statements.
3. Economic life: How long continued use or resale generates sufficient value, considering alternative equipment and scarce site capacity.

The available evidence argues against both simplistic extremes—“obsolete in two years” and “six years of unchanged economics”:

• Evidence for durability: CoreWeave's investor presentation describes an A100 renewal extending into 2029 and an H200 three-year renewal at a premium to the original contract. These are two selected, company-reported examples, without fleet denominators, original installation dates or complete cash-cost histories. They support continuing demand for older architectures, but not a portfolio-wide survival rate or a claim that particular devices have already operated profitably for nine years. (sec.gov)
• Evidence for shortening assumptions: Amazon's 2025 annual filing says it reduced the useful life of a subset of servers and networking equipment from six to five years, effective January 1, 2025, because of faster technology development, particularly AI and machine learning. This is an accounting estimate for a subset—not a universal GPU retirement date. (sec.gov)
• The competitive frontier is moving: NVIDIA reported on August 26, 2026 that it had commenced Vera Rubin production shipments earlier that month. New-generation competition is therefore a current consideration, not only a distant forecast. (investor.nvidia.com)

Revoqra synthesis: Software improvements and established workloads may extend economic life, while better competing systems may reduce rent, occupancy or value per unit of scarce power. A still-profitable older GPU can nevertheless have a lower opportunity-adjusted value if replacing it produces more site-level cash flow. NVIDIA's argument that software and redeployability extend installed-system value is relevant, but remains an issuer interpretation rather than a demonstrated universal residual-value floor. (blogs.nvidia.com)

No representative, multi-cycle series of net realized GPU resale recoveries was verified in the reviewed evidence. Consequently, resale assumptions below are sensitivities, not market forecasts.

3. Transparent unit-economics model

3.1 Assumptions and cost boundary

This is a Revoqra-derived hypothetical model, expressed per GPU as an allocation within an eight-GPU H100-class deployment. It is not an executable October 2026 purchase budget, a market-average rental forecast or a model of every accelerator generation.

The illustration assumes an established operator's shared-cost allocation. A new business operating only one node may face substantially higher staffing, compliance and customer-acquisition costs.

Evidence anchors—not validations of every input: The disclosed equipment orders frame acquisition scale, and posted provider rates frame the rental market. A Brookhaven study measured a maximum of approximately 8.4 kW on one eight-GPU H100 node, compared with a 10.2 kW manufacturer rating; that supports modeling complete-system power rather than accelerator TDP alone, but is not a fleet-average draw. (irisenergy.gcs-web.com)

Site pricing must be checked separately. CBRE reported Chicago asking rents of $200–230/kW-month for 250–500 kW requirements in Q1 2026; those are market/size-specific capacity quotes, not prices for a small eight-GPU installation. EIA sector averages likewise do not replace the actual electricity tariff. (cbre.com)

Utilization warning: Paid reserved time, hardware busy time and model FLOP utilization are different measurements. A take-or-pay customer can pay while equipment is idle; a computationally busy machine can still produce no collected revenue. Performance-utilization metrics should not be substituted for paid occupancy. (sec.gov)

3.2 Formulas

Let u be paid utilization, a physical active-load fraction, p realized hourly customer pricing, f the distribution charge, F fixed annual cash costs, C acquisition cost and S net resale.

Revenue = 8,760 × u × p

Energy cost = 8,760 × PUE × electricity price
              × [idle kW + (active kW − idle kW) × a]

Unlevered cash flow = Revenue × (1 − f) − F − Energy cost

NPV = −C + Σ[annual cash flow / (1 + hurdle)^year]
         + S / (1 + hurdle)^remaining life

All model outputs below are Revoqra-derived deterministic calculations using these disclosed assumptions. They are nominal USD, annual end-of-year cash flows, before tax, and are not source-published investment returns.

3.3 Central result

The central lesson is capital consumption. Straight-line recovery of $35,000 less $3,500 resale over four years is $7,875 annually. The $8,935 cash surplus therefore leaves only about $1,060 annually above that simple capital-consumption allowance. Time value then further reduces the investment's attractiveness.

This explains why corporate EBITDA is insufficient evidence of hardware-owner returns. CoreWeave reported Q2 2026 revenue of $2.575 billion and adjusted EBITDA of $1.510 billion, a 59% margin, alongside $640 million of net interest expense and a $626 million net loss. Its approximately $104 billion June 30 backlog is subject to delivery and service-availability requirements. These demonstrate business scale and contracted demand—not individual-GPU investment IRRs. Nor do corporate losses, by themselves, prove that every deployed GPU cohort is uneconomic. (investors.coreweave.com)

4. Sensitivity analysis

4.1 Rental price and paid utilization

NPV per GPU at a 12% hurdle, $ thousands. Four-year life, 10% net resale and all other central assumptions unchanged; physical activity equals paid utilization in this table.

Revoqra-derived scenarios, not observed outcomes. Price and occupancy are not assumed to have independent probabilities; the table does not estimate price elasticity or the likelihood of its favorable cells.

At the central acquisition cost, a 12% return requires approximately:

• $2.85/hour at 65% paid utilization; or
• 74.5% paid utilization at $2.50/hour.

Alternatively, maintaining 10% proportional resale and the central cash flows, the acquisition allocation needs to be approximately $29,000 or less, rather than $35,000.

A contracted illustration also shows the upside: at the same $2.50 price and 95% paid utilization, the model produces approximately +$12,200 NPV. But that is valuable only if payments are collected and service obligations can be fulfilled. CoreWeave itself identifies the possible loss of its take-or-pay model as a risk to forecastability and margins. (sec.gov)

4.2 Other assumptions that can change the conclusion

Each row changes the stated input while otherwise retaining the central case.

All figures are Revoqra-derived, rounded to $100. Life sensitivities retain the same 10% terminal percentage to isolate remaining service life; that is not a claim that six-year-old equipment has the same resale percentage as three-year-old equipment. The 20% annual rent decline is a stress, not an inferred depreciation forecast; the selected older-generation renewals discussed above are credible counterevidence to inevitable rental collapse. (sec.gov)

The delay stress is economically important. Amazon describes infrastructure outlays occurring typically 6–24 months before billing, depending on the component. That is not a GPU-specific lead-time estimate, but it reinforces the need to model commissioning and power availability rather than assume purchased equipment immediately earns revenue. (aboutamazon.com)

4.3 Financing does not repair weak unlevered economics

Assume 60% acquisition financing, or $21,000 debt and $14,000 initial equity, with full four-year amortization and annual payments.

Revoqra-derived alternative financing assumptions at inception—not available loan offers or a forecast of floating-rate payments. Origination fees and required liquidity reserves are excluded.

At 10%, the leveraged equity IRR is approximately −3.2%, despite positive annual cash distributions. Equity receives approximately $12,740 in total cash, including resale, against $14,000 invested. The debt costs more than the underlying asset earns.

Under the 20% annual rental-price decline, annual unlevered cash flow falls to approximately $8,935, $6,316, $4,220 and $2,544. The modeled 10% loan becomes undercovered beginning in year two. High initial utilization or an acceptable first-year coverage ratio therefore does not establish a safe lifetime financing structure.

Revoqra underwriting preference: Amortize against credible contracted cash flows; avoid relying on refinancing or optimistic collateral sale proceeds after the contract expires. A balloon loan against short-lived, merchant equipment combines operating risk with a potentially adverse exit-market requirement.

5. Exposure to demand, technology and interest-rate cycles

The following are plausible mechanisms and underwriting risks, not estimated causal coefficients or cycle probabilities.

The mechanisms are consistent with disclosed take-or-pay risks, financing protections and delivery conditions. They should be stressed jointly rather than treated as independent favorable inputs. (sec.gov)

Three portfolio implications follow:

1. Short hardware life is not the same as low risk. Earlier capital recovery can reduce unlevered interest-rate duration relative to long-lived infrastructure, but technological and operating leverage can remain severe.
2. No automatic inflation hedge exists. Costs can rise without an equivalent increase in rent. A contract's escalation and power pass-through provisions must be examined rather than inferred from the asset's physical nature.
3. Manufacturer support changes the exposure. NVIDIA disclosed approximately $36 billion of typically six-year cloud-service commitments as of July 26, 2026, under arrangements in which capacity can be redirected to third-party customers. It also says it may selectively provide limited residual-value support. Returns on supported transactions partly reflect the contract/support provider's credit and terms, not unsupported GPU-market economics. (investor.nvidia.com)

6. Risk-adjusted comparison with adjacent investments

6.1 Qualitative comparison

A mature infrastructure maintenance metric also differs from whole-hardware capital recovery. Equinix's July 2026 guidance put recurring capital expenditure at approximately 3% of annual revenue, while separately anticipating large expansion spending. Neither that metric nor a GPU operator's EBITDA margin is an investment IRR, and their maintenance/expansion definitions are not directly comparable. (investor.equinix.com)

6.2 A standardized capital-recovery comparison

To isolate finite life, use the same hypothetical 12% hurdle for all assets. The annual cash flow required is:

Required annual net cash flow / acquisition cost
= [1 − residual fraction / (1 + hurdle)^life]
  ÷ annuity factor(hurdle, life)

Revoqra-derived hypothetical patterns—not empirical industry averages or forecasts. Cash flow is after operating costs and sustaining maintenance, before financing and tax. The common hurdle deliberately isolates capital life; it does not assert equal risk or equal required returns across assets.

This is why a GPU can show a much higher annual cash yield than infrastructure yet deliver a lower lifetime IRR. Conversely, a competitively purchased, highly occupied GPU can generate substantial returns over a shorter horizon.

For a pure net equipment lease, where the tenant pays operating costs, the same $35,000/four-year/10%-resale example requires approximately $10,791 annual net owner cash flow, before additional owner expenses, to meet the 12% hurdle. Do not confuse that amount with the retail cloud price, or charge the lessor twice for costs contractually borne by the operator.

Risk-adjusted conclusion: Without representative total-return and loss data, there is no defensible empirical Sharpe-ratio ranking among these exposures. The useful comparison today is capital recovery, downside coverage, credit quality, operational complexity and liquidity—not heterogeneous corporate margins or unmatched headline yields.

7. What would substantiate an asset-class case?

Progress is real, but announcements are not a return history

NVIDIA's August 10, 2026 announcement proposed financing platforms with six major capital providers designed to mobilize more than $500 billion over time. Its subsequent filing cautions that the memorandums of understanding and other preliminary arrangements may not lead to definitive agreements. The headline is therefore evidence of financing ambition and market development—not $500 billion of capital already deployed or evidence of net investor performance. (investor.nvidia.com)

Rental benchmarks and hedging infrastructure are also emerging. Silicon Data describes standardized GPU rental observations adjusted for specifications, term, performance and geography. CME's proposed H100 and B200 contracts are financially settled rental-index exposures, not delivery of physical GPUs. As of this report's October 3, 2026 cutoff, the announced listing is still future: the October 4 session, with October 5 trade date, pending completion of regulatory review periods. Established trading liquidity should not be assumed. (silicondata.com)

An early July 2026 academic working paper constructs synthetic compute-futures returns from existing term rentals. That is relevant to market design, but is not a record of realized hardware-owner returns. Compute hours also cannot be stored: owning a GPU preserves future production capability, not yesterday's unsold service hours. (arxiv.org)

Evidence needed for a stronger case

The following are Revoqra's proposed evidentiary tests, not claims that all are currently satisfied:

1. Audited acquisition and disposal cohorts. Track complete installed cost and net binding-sale recoveries by architecture, form factor, installation date, condition, location and support status. Report both device counts and independent project/transaction counts.
2. Collected utilization and pricing. Distinguish installed, commissioned, available, reserved, billed, collected and physically busy hours. Include discounts, refunds, credits, defaults and distribution charges—not only advertised prices.
3. Net lifetime owner cash flows. Report returns after site costs, energy, networking, operations, repair/refresh, financing fees, reserves, fund expenses and terminal disposal. Separate maintenance from expansion and disclose reinvestment needed to continue the business.
4. Multiple vintages and adverse conditions. Include failed operators, unoccupied fleets and liquidations, not only successful renewals. Cover several architectural transitions and a meaningful demand/credit downturn.
5. Credit and operational recoveries. Show whether lenders or owners can continue operating after default, enforce/assign contracts, retain site access and sell equipment promptly. Distinguish hardware recovery from guarantees or customer-contract recoveries.
6. A genuine total-return benchmark. Combine income with credible changes in market value. Estimate exposure to technology equities, AI capital spending, rates, credit spreads and infrastructure using comparable periods and definitions. Avoid treating stable accounting marks as evidence of low volatility.
7. Usable hedges. Demonstrate actual volume, open interest, bid–ask depth and basis performance. A rental-rate hedge does not automatically hedge paid occupancy, customer default, power cost or physical resale.
8. Standalone economics without hidden support. Identify supplier guarantees, prepayments, subsidized financing and related-party demand separately. Establish whether returns remain adequate without those protections—or classify the supported credit exposure explicitly.

Practical underwriting priorities

• Secure suitable power, site access and operating capability before committing to equipment.
• Prefer capital recovery inside a credible contractual period; treat extensions as upside until evidenced.
• Match debt amortization to conservative cash flows and remaining life, not optimistic accounting depreciation.
• Stress lower rent, lower occupancy and weaker resale together.
• Obtain actual owner settlement records and executable exit bids, not just customer price pages.
• Evaluate investment liquidity separately from the physical transferability of hardware.

Bottom line

Owning AI GPUs can be investable and economically distinctive. The potential return driver is near-term, monetizable compute capacity combined with procurement, deployment and operating advantages. But the return profile is highly sensitive to capital basis, collected occupancy, contract quality and the pace at which competitive economics change.

The strongest current asset-class argument is for repeatable, contract-backed AI-compute financing with explicit technology and residual-risk controls. The weaker argument is that scarce GPUs inherently deliver infrastructure-like durability or diversification. The model shows that attractive annual cash receipts can coexist with mediocre or negative lifetime equity returns.

A dedicated allocation would be substantiated by independently verifiable, net, multi-vintage owner returns and recoveries—not by AI demand growth, financing announcements, EBITDA margins or two successful renewal examples alone.

Sources and evidence limitations

Key original evidence used in this report:

• Equipment purchases: IREN's August 2023 H100 announcement and filing describing its B200/B300 order. (irisenergy.gcs-web.com)
• Customer contract and financing: IREN's Microsoft announcement and August 2026 results. (sec.gov)
• GPU-backed project structure: CoreWeave's March 2026 financing presentation, June-quarter SEC filing and Morningstar DBRS's original rating announcement. (s205.q4cdn.com)
• Corporate results and older-generation renewals: CoreWeave's Q2 2026 results and SEC-filed investor presentation. (investors.coreweave.com)
• Rental offers and owner-price bridge: Official Lambda, Runpod and Vast publications. (lambda.ai)
• Life and infrastructure comparisons: Amazon's annual filing/shareholder letter, Equinix's results, and Dell, United Rentals and American Tower disclosures. (sec.gov)
• Power and site costs: Original H100-node measurement research, EIA statistics and CBRE's market research. (arxiv.org)
• Asset-class and market-development claims: NVIDIA's original announcement/filing, Silicon Data's methodology description, CME's product/listing notices and original compute-pricing research. (investor.nvidia.com)

Evidence limit: These sources substantiate purchases, posted rental prices, selected renewals and financing structures. They do not provide a representative, multi-cycle panel of net GPU-owner returns, paid fleet utilization and realized resale recoveries. Accordingly, this report does not estimate an empirical expected return, correlation, Sharpe ratio or statistical significance for GPU ownership.