The spread between the cheapest and most expensive way to rent an H100 GPU was $1.38 to $11.01 per hour in mid-May 2026. That is not a market with functioning price discovery. That is a market before it had one.
ICE and Ornn followed a week later with their own GPU futures based on the Ornn Compute Price Index (OCPI), now live on Bloomberg Terminal.
Three independent pricing benchmarks — Silicon Data, Babbage Index, and CCIR — now track GPU pricing daily across dozens of providers.
This is not a niche financial story. Compute, the literal capacity to run AI models, is following the same arc that oil, wheat, and electricity followed before it. First comes fragmented spot pricing with wide bid-ask spreads. Then a benchmark emerges. Then a futures market. Then the asset class becomes financeable at institutional scale.
Three parallel tracks are laying that arc down at once, in a market that already accounts for an estimated $4 trillion in planned infrastructure investment through 2030.
The benchmark layer
Every commodity market starts with a price people trust. For compute, that layer now has three independent providers.
Silicon Data, backed by trading firm DRW, launched the first daily GPU benchmarks for on-demand rental rates. Its indices cover A100, H100, and B200 pricing and are distributed through Bloomberg and Refinitiv. When CME Group announced compute futures on May 12, Silicon Data's benchmarks were the reference, the first time an exchange of CME's scale had anchored a new commodity contract on GPU rather than energy or agricultural data.
Ornn followed a different path. Its Ornn Compute Price Index (OCPI) is built from printed transactions rather than surveys or rate cards, the first compute benchmark that settles on actual traded prices. OCPI went live on the Bloomberg Terminal in April 2026 and is the reference for ICE's planned GPU futures suite, covering H100, H200, B200, and RTX 5090 classes.
Babbage Index, founded in London in 2026, takes the widest view: five indices covering GPU, CPU, memory, storage, and token inference pricing across roughly 20 providers and 100+ daily price series. Its GPU index (BI-G) tracks H100/H200/B200/A100 pricing across hyperscale, neo-cloud, and marketplace tiers. A single number that captures what the same compute silicon costs depending on where and how you buy it.
CCIR (Compute Credit Index Research) rounds out the set with daily reference rates focused on the credit side: what GPU-backed debt should price at, how utilization affects collateral assumptions, and where the term structure of compute rental commitments sits relative to on-demand spot prices.
Why this matters for capital allocation
Pricing fragmentation is not an abstract problem. A hedge fund modeling AI company revenue needs to know what its portfolio companies pay for compute. A data center operator financing a 100 MW buildout needs to lock in forward rental revenue. A credit fund underwriting GPU-backed debt needs a reference rate that both borrower and lender can agree on.
Before the benchmark layer existed, each party operated from its own private data. The hyperscalers knew their procurement costs. The neoclouds knew their utilization rates. The funds saw only the equity side: GPU companies' stock prices, not the underlying rental market that drives their revenue. The information asymmetry was structural.
The indices change that. Once compute has a daily settlement price, the same financial machinery that works for oil, natural gas, and soybeans can work for GPUs: futures, options, swaps, and eventually ETFs and structured products. As we wrote in July, the compute cost curve is already becoming an institutional asset allocation signal — this is the infrastructure layer forming underneath it.
The futures layer
Two parallel efforts are building the derivatives infrastructure.
CME Group partnered with Silicon Data to launch U.S. dollar-denominated, cash-settled compute futures. Terry Duffy, CME's chairman and CEO, described compute as "the new oil of the 21st century" and called the futures market a response to the fact that "every AI model trained, every transaction cleared, and every byte of data processed runs on compute, which is becoming a fast-emerging asset class in its own right." The contracts are designed for AI builders, cloud providers, financial institutions, and institutional investors — the same participant mix that trades CME's energy and agricultural contracts.
ICE followed a week later with GPU futures based on Ornn's OCPI. Trabue Bland, ICE's SVP of Futures Markets, said the compute market is "in desperate need of a globally accepted pricing mechanism and risk management tool." ICE's contracts will also be dollar-denominated and cash-settled, pending regulatory approval.
| Parameter | CME Group × Silicon Data | ICE × Ornn |
|---|---|---|
| Announced | ✔ May 12, 2026 | ✔ May 19, 2026 |
| Reference index | ✔ Silicon Data GPU benchmarks | ✔ Ornn Compute Price Index (OCPI) |
| Coverage | ✔ A100, H100, B200 | ✔ H100, H200, B200, RTX 5090 |
| Settlement | ✔ Cash-settled, USD | ✔ Cash-settled, USD |
| Status | ◐ Pending regulatory review | ◐ Pending regulatory review |
Don Wilson, founder and CEO of DRW, put the opportunity in perspective: "It has been clear to me for some time that compute will become the largest commodity in the world. The exponential growth in spending on data centers as we move towards that reality has been hampered by the lack of a hedging vehicle."
What a compute futures contract actually hedges
A GPU rental contract is not a barrel of oil. Compute is not fungible in the same way — an H100 hour on AWS is not identical to an H100 hour on a marketplace like Vast.ai or RunPod. The performance depends on neighboring workloads, interconnect topology, storage latency, and power constraints.
The indices handle this by normalising across deployment tiers. Silicon Data's benchmarks track on-demand rental rates across hyperscale, neo-cloud, and marketplace segments. Babbage Index publishes separate indices for each tier and makes the composition transparent — each headline number recomputable from its published components. OCPI is built from printed transactions only, avoiding the self-reported pricing that plagues rate-card-based benchmarks.
A futures contract hedges exposure to the compute price level, not a specific GPU in a specific rack. The same logic applies in wheat: the contract settles on the index, not on a specific bushel from a specific farm. The basis risk between the index and any individual deployment is a separate instrument to manage, just as location and grade differentials are in physical commodity markets.
The institutional adoption signal
The clearest signal that compute is becoming an asset class is not the futures announcements themselves. It is who is paying attention to the benchmarks.
Academic researchers at Johns Hopkins published "AI Compute Asset Pricing" on arXiv in July 2026, the first formal attempt to model the compute futures risk premium. Using data from Silicon Data and Ornn, the paper finds a positive compute risk premium consistent with the idea that compute providers — the natural short hedgers — will dominate the early market, and that institutional investors providing the long side will earn a premium for bearing compute price risk.
The scale of the opportunity draws attention from researchers and analysts alike. A July 2026 BCG study estimated that transparent compute pricing could unlock $140 billion in annual value that exists in the market today but cannot be captured because pricing is opaque, contracts are bilateral, and risk cannot be transferred. Johns Hopkins researchers published a formal asset-pricing model on arXiv in the same month, finding a positive compute risk premium consistent with the structure of early commodity futures markets.
Quant funds are already treating inference pricing data as an alpha signal. Babbage Index explicitly markets to "institutional capital underwriting AI infrastructure — credit funds, infra insurers, treasury desks — alongside quant funds and AI company CFOs." The use case is not speculative. Compute and inference spend is a leading indicator for AI company economics before revenue is reported. A fund that tracks what portfolio companies pay for compute sees margin pressure before it shows up in earnings.
The limits of the parallel
The oil-compute analogy is useful up to a point, and misleading past it. Oil is a physical commodity with grade standards (Brent, WTI), storage costs, transport logistics, and a century of market infrastructure. Compute is a service with no physical delivery, near-zero transport cost, and rapid technological obsolescence — a B200 GPU delivers different economics than an H100, and both will be superseded within two years.
The benchmarks address this through frequent reconstitution. Babbage Index rebases its index cohort every quarter. OCPI adds new GPU types as they reach market significance. The CME contracts reference specific GPU families rather than a single "compute" price, allowing the market to price technological obsolescence through the calendar spread rather than ignoring it.
The harder question is liquidity. Agricultural and energy futures markets took decades to reach the depth that makes them useful for institutional-scale hedging. Compute futures will start with natural hedgers — data center operators who want to lock in forward rental revenue, and AI labs that want to cap their infrastructure costs. Whether speculative liquidity arrives quickly depends on whether the underlying pricing data is trusted. That is the LIBOR question that Babbage Index's latest research addresses directly: can an outsider reproduce the benchmark?
How AI companies are already using the benchmarks
The pricing indices are not hypothetical. Jasper Zhang, co-founder and CEO of Hyperbolic Labs, described the GPU market as one that "increasingly resembles a global commodity market more than a traditional cloud market, yet the financial infrastructure around it is still in its early stages." His company operates one of the largest GPU marketplaces serving AI developers globally and uses benchmark pricing to set its own rates.
Hyperbolic is not alone. Neocloud providers — companies like CoreWeave, Lambda, and Together — operate in a market where input costs (GPU rental rates) can swing by a factor of eight between on-demand and reserved pricing across different providers. Before the benchmarks, each operator priced in isolation. Now an operator can compare its own rate card against the Babbage GPU index, see where it sits relative to the hyperscale, neo-cloud, and marketplace tiers, and adjust pricing or contract structure accordingly.
For AI labs themselves, the benchmarks solve a different problem. A lab training a frontier model commits to millions of GPU-hours months in advance. The difference between locking in H100 capacity at $2.00/hr and paying $4.50/hr at peak demand can amount to tens of millions of dollars in a single training run. Before compute futures, there was no instrument to hedge that exposure. The lab either overpaid for reserved capacity or accepted spot market risk. Futures contracts — even before they clear — create a forward curve that makes the hedging decision visible and quantifiable.
Credit funds are the third leg. GPU-backed debt issuance has already reached investment grade — the first ABS deals referencing AI compute infrastructure closed in mid-2026. But those deals priced without a standardized reference rate for the underlying collateral. CCIR's reference rates fill that gap by tracking what GPU rental revenue actually generates across deployment types, giving lenders a data-backed basis for loan-to-value ratios and margin calls.
What adoption looks like in practice
The timeline from announcement to liquid market is measured in years, not months. CME's energy complex, the deepest commodity futures market in the world, took decades to reach its current depth. Compute futures will follow a faster trajectory — electronic trading, existing participant infrastructure, and a market that already transacts billions in GPU capacity annually — but the early months will be dominated by natural hedgers on both sides.
Data center operators with contracted revenue streams will sell compute futures to lock in forward rental rates. AI labs with committed training schedules will buy them to cap infrastructure costs. The speculative middle — hedge funds, commodity trading advisors, and proprietary trading firms — will enter only once the benchmark has a sufficient track record to backtest against.
The research from Johns Hopkins suggests the early market will carry a positive risk premium: compute providers (the natural short) face more urgent hedging demand than AI labs (the natural long), so the futures curve will trade at a discount to expected spot prices, rewarding the long side for providing liquidity. That is the same structural pattern seen in early agricultural and energy futures markets.
The wildcard is the cross-border dimension. Babbage Index already tracks a US-vs-China frontier inference spread of 14.1× — the same AI model costs 14 times more to serve in the US as in China, driven by GPU export controls and domestic Chinese chip alternatives. That basis alone creates a natural arbitrage for funds that can trade both venues, and it gives the compute futures market something oil never had: a structural geopolitical premium embedded in the price from day one.
Signals to track
CFTC approval timeline for CME and ICE compute futures — whichever clears first sets the contract standard
Babbage Index monthly publication cadence — a sustained record builds the data history needed for institutional adoption
First actual compute futures trade volume and open interest — the gap between announced and traded is where the market proves itself
Credit fund issuance of GPU-backed ABS referencing benchmark pricing — the bridge from derivatives to structured finance
Architect Financial Technologies also joined the race, announcing plans for exchange-traded futures on GPU and DRAM prices using Ornn's benchmark data through its AX exchange. Three venues targeting the same asset class before any of them has cleared a single trade is a strong signal that the market sees compute derivatives as inevitable — the only open question is which contract standard becomes the liquidity magnet.
