What Makes Compute an Asset Class? Price Indexes, Futures and the Credit Layer
Summary
- It has a price. Compute is a scarce, capital-intensive input with a rental market, daily price indexes and futures announced by two exchange operators.
- It can't be stored. An unused GPU-hour is gone for good, so compute prices like electricity and futures hedge only the rental rate.
- Credit makes it holdable. The GPU and its contracted cash flow are the asset, and standardized GPU-backed loans are how institutions own it.
Introduction
Compute moved from a cost line to a priced market in 2026. Index providers now publish daily GPU rental benchmarks, and CME Group and ICE have each announced compute futures. BlackRock's Larry Fink told the Milken Institute Global Conference in May 2026: "I actually believe a new asset class will be buying futures of compute."
The financialization of compute leaves institutions with a practical problem. A fund that wants exposure sees a rental price, a chip and a contract, and each carries a different risk. A credit committee has to decide which of the three it is actually underwriting.
We cover what qualifies compute as an asset class, how indexes and futures price it, and what the rental price leaves unpriced. We then set out how institutions hold compute today, with a checklist a credit committee can use.
What Makes Compute an Asset Class?
Compute is an asset class because it is a scarce, capital-intensive input with a measurable price and a large, growing capital base. Daily indexes track that price, and announced futures would let investors take or hedge exposure to it. The price is quoted in dollars per GPU-hour, the rental of one GPU for one hour.
Johns Hopkins economists Federico Bandi and Yinan Su write that "Compute capital expenditure and service flow already exceed 1% of U.S. GDP." Their October 2026 working paper has not been peer-reviewed. They put the rental value of the GPUs installed by the fourth quarter of 2025 at about $430 billion to $1.3 trillion a year.
That range is a rental-equivalent estimate that assumes full utilization, and the authors caution against reading it as provider revenue. The paper also notes that the four largest hyperscalers plan more than $700 billion of capital expenditure for 2026, about 2.2% of U.S. GDP. That capex figure includes spending beyond AI.
NVIDIA's results show the demand behind that spending: Data Center revenue was $89.0 billion in the quarter ended July 2026. That was up 117% from a year earlier.
An investor can own compute in two layers. The first is the service, the hourly rental rate that indexes track and futures would settle against. The second is the capital good: the GPU itself and the cash flow it earns under contract, which is the layer a lender finances.
How Is Compute Priced Today?
The rental market is the underlying for every compute price, and it is fragmented by GPU generation, provider type and contract form. Operators sell processing power by the hour through their own clouds and through compute marketplaces. The rate a buyer pays depends on all three variables.
Rental rates, indexes and price dispersion
Three contract forms set the rate. On-demand rentals, billed by the hour with no commitment, are the closest thing to a financial spot price. Cloud "spot" capacity is a different product: interruptible excess capacity the provider can reclaim. Reserved or term rentals lock in capacity for one to three years.
Provider type moves the price further. Hyperscalers are the largest public clouds, and neoclouds are GPU cloud operators built to rent out GPU capacity. The Bandi and Su data table shows the spread. Silicon Data's neocloud H100 index was $2.50 per GPU-hour in April 2026, against $7.43 for its hyperscaler H100 series.
Silicon Data's H100 index stood at $2.96 per GPU-hour in November 2026. Silicon Data is also CME's index partner for compute futures, an interest to weigh when using its series. The company builds its series from observed rates across neocloud providers and GPU marketplaces, and Ornn publishes a competing index.
Price also climbs with each generation. In April 2026, Silicon Data's neocloud series priced the A100 at $1.43 and the H100 at $2.50 per GPU-hour. The B200 rented at $5.10 per GPU-hour on the same Silicon Data neocloud series in the Bandi and Su data. The B200 ships as an eight-GPU HGX Blackwell system, covered on our page for NVIDIA B200 systems.
| Index series | Segment | Price ($ per GPU-hour) | Date |
|---|---|---|---|
| Silicon Data A100 | Neocloud | 1.43 | April 2026 |
| Silicon Data H100 | Neocloud | 2.50 | April 2026 |
| Silicon Data H100 | Hyperscaler | 7.43 | April 2026 |
| Silicon Data B200 | Neocloud | 5.10 | April 2026 |
| Silicon Data H100 | Neocloud | 2.96 | November 2026 |
In our view, for a lender the rate in the offtake contract matters more than the index level, because the contract repays the loan. The index marks where the market sits on a given day, while the contract fixes the rate the operator's customer pays.
Compute futures and where they stand
Compute futures are scheduled, and their listing still depends on regulators. CME Group and Silicon Data said on May 12, 2026 that CME "will launch a first-in-class compute futures market later this year, pending regulatory review."
The contracts are scheduled to list on NYMEX, pending CFTC review: the Commodity Futures Trading Commission extended its review to November 9, 2026. Per CME's product page, each contract covers 730 GPU-hours, with financial settlement and 36 monthly listings. Both contracts track neocloud indexes only, so hyperscaler pricing sits outside them.
ICE and Ornn announced plans on May 19, 2026 to launch cash-settled GPU compute futures on Ornn's Compute Price Index. Their announcement states, "The contracts will be launched subject to regulatory approval," and gives no date.
The idea has a precedent in Deutsche Börse Cloud Exchange, which in 2014 described its aim as "turning compute, memory and storage into tradable commodities." In our view, if today's contracts list, three groups would use them. GPU cloud operators would sell futures to hedge rental revenue, and AI developers would buy them to hedge compute cost. Speculators and arbitrageurs would supply liquidity.
Why Does Compute Price Like Electricity Instead of Oil?
Compute prices like electricity because a GPU-hour can't be stored. The paper's authors write that "A GPU-hour not used today cannot be carried forward and delivered in the future." That breaks cash-and-carry pricing, the trade that buys a commodity today, stores it and sells it forward. For storable goods, that trade ties futures to spot prices.
For that reason the authors place compute "closer to electricity" in how it prices. Three consequences follow for an investor:
- Forward curve. The futures price reflects expected future rental rates less a risk premium, the compensation paid to whoever takes on the price risk.
- Term rentals. A one-to-three-year reservation works as a synthetic future, and its price is likely an upper bound on a financial futures price.
- Risk premium. Early evidence is "consistent with a positive compute risk premium," which the authors attribute to compute providers hedging their revenue.
Term rentals sit above financial futures because a physical reservation also locks in access to capacity, which a cash-settled contract can't deliver. The risk premium evidence comes from a short sample of synthetic futures, so we treat it as direction only.
Non-storability also explains where the risk sits. An idle GPU-hour earns nothing and can never be sold later, so the hardware owner carries utilization risk every hour of the asset's life. That is why the contract behind the GPU matters to credit.
What Does a Rental Price Leave Unpriced?
A futures contract hedges the hourly rate on a standard index. An institution that holds compute through hardware or credit carries two further risks the index doesn't price. They are what the GPU is worth when the loan ends, and who pays for its time.
Residual value
Residual value, the GPU's worth at the end of a loan, depends on what the GPU can earn at that point. SemiAnalysis's Daniel Nishball said in October 2026 that banks "have to know residual value, which is very, very hard with GPUs." In his framing, residual value "is what you can earn on tokens, or what you can reasonably rent it at."
Nishball added that operators "underwrite it to make money on a five to six-year period," without banking on residual value. On the same episode, Jordan Nanos said one-year H100 deals are approaching first-half-2023 prices, "like three bucks an hour." That is close, in our observation, to the $2.96 level of Silicon Data's on-demand neocloud H100 index in November 2026. The two measure different contract forms, though.
A liquid forward curve would give lenders a market reference for residual value, and that curve doesn't exist yet. Until it does, amortization does the work: each payment shrinks the balance that depends on resale. In our view, paying the loan down within the contracted term matters more than forecasting resale prices.
Counterparty and contract
The same H100-hour is worth different amounts to a lender depending on who pays and on what terms. An offtake contract is the agreement under which a customer commits to buy an operator's compute. Under a 36-month take-or-pay contract, a creditworthy customer pays whether or not it runs workloads. That revenue supports more debt than on-demand revenue that can stop any hour.
The rated market already reflects this. CoreWeave said its $8.5 billion facility, closed in March 2026, is "secured by HPC infrastructure and an associated customer contract." The company said the facility is rated A3 by Moody's and A (low) by DBRS.
Compute IQ, USD.AI's agentic underwriting system, is built on more than $25 billion of GPU financing data points as shown on its page in November 2026. It scores every financing originated through USD.AI on one 100-point scale, using three inputs: the operator, the customer and the contract. The customer's credit and the offtake contract terms carry most of the review, as the Compute IQ scoring rubric shows.
On-demand revenue doesn't disqualify a deal, and it is priced to reflect its shorter commitment. The contract turns a volatile rental price into a creditworthy cash flow. That conversion is the credit layer.
How Can Institutions Hold Compute Today?
The holdable form of compute today is a GPU-backed loan, a senior claim on the GPUs, the offtake contracts and the revenue. Futures, once listed, would hedge the rental rate around that claim. Standardization is what makes the loan investable, because bespoke GPU credit costs millions to issue and that cost barely changes with deal size.
Comparable, repeatable loans spread that work across many deals and let a credit committee assess the structure once. Through October 2026, we reviewed nearly 700 GPU financing deals representing about $25 billion of loan demand. That pipeline includes a fast-growing group of own-use compute buyers.
Every financing through USD.AI uses the same core structure:
- Ring-fenced collateral. A bankruptcy-remote special-purpose vehicle (SPV) holds the GPUs, offtake contracts and revenue accounts, and lenders hold a first-priority claim on them.
- Funding on verification. Capital is held in escrow until the servers are installed and independently verified.
- Reserve at closing. Part of the operator's equity funds a debt-service reserve at closing, a cash buffer that covers loan payments if collections are delayed.
- Hardware coverage. Hardware insurance and monitoring stay in place for the term of the loan.
- Full-term servicing. A dedicated servicer manages each loan for its lenders for the full term.
Where a deal includes a residual-value warranty, it may cover part of a shortfall on qualifying server sales, subject to its terms.
Loans cover up to 80% of the deployment cost, with the operator contributing at least 20% as equity. Terms run 3–5 years with straight-line monthly amortization.
The GLX institutional marketplace connects institutional capital with these financings. Institutions can take a position at origination, join a syndicate, or buy all or part of a seasoned loan. Seasoned financings may be sold through GLX. Each opportunity carries the same credit package and structure, whoever holds it.
Private credit funds use that package to support their own independent review, as set out in GPU credit for private credit funds. Banks examine the documentation, structure and servicing against their own credit policies, covered in GPU credit for banks. The checklist below maps each risk to the tool that covers it.
| Risk | What to check | Covered by futures | Covered by credit diligence |
|---|---|---|---|
| Rental rate | Index level and forward curve for the GPU generation | Partly, for the neocloud H100 and B200 rate once listed | Contracted rate in the offtake contract against the index |
| Residual value | What the GPU can earn at loan maturity | No | Amortization profile against the contract term |
| Counterparty | Credit quality of the customer buying the compute | No | Customer credit, scored by Compute IQ |
| Contract term | Length and commitment of the offtake contract | No | Term, take-or-pay provisions and on-demand share |
| Structure | Ring-fence, escrow, reserve and servicing | No | SPV, first-priority claim, escrow until verified install, debt-service reserve |
Futures reach one row of five. Compute is becoming an asset class that needs a credit market, and the other four rows are where that market does its work.
Conclusion
Indexes and scheduled futures give compute a price, and non-storability makes that price behave like electricity's. Residual value and counterparty risk sit outside the index, so institutions price them through credit. In our view, that makes the GPU-backed loan the practical way to hold compute as an asset class today.
GPU financing through USD.AI standardizes every loan, so each one is scored the same way and can be compared side by side. That consistency is what lets financings be placed with institutional capital through GLX, on one credit package and one structure.
Operators can see how a deployment is financed on USD.AI's GPU financing page. Institutions reviewing GPU credit can ask about positions at origination, syndicates and seasoned loans.
FAQ
Is compute an asset class?
Yes, in our view. Compute has macro-scale capital spending, a daily indexed price, futures scheduled pending regulatory review, and hardware that is financed against contracted revenue.
What are compute futures?
Compute futures are cash-settled contracts on a GPU rental index. CME's H100 and B200 contracts are scheduled to list on NYMEX, pending CFTC review. ICE and Ornn plan to launch theirs, subject to regulatory approval.
What is a compute price index?
A compute price index is a daily benchmark of GPU rental rates, published by GPU generation and market segment. Silicon Data and Ornn publish the main series, and each underlies announced futures.
How do investors get exposure to compute?
Investors can hold equity in companies that own GPUs, use futures once they list, or hold GPU-backed credit. Credit gives a claim on the hardware and the contracted revenue that repays the loan.
