Standardizing Compute Trading: Four Key Lessons From Commodities Markets

Executive Summary

Over the past year, compute has started to be talked about, and traded, like a commodity. In August and May 2026 respectively, CME Group (working with index provider Silicon Data) and Intercontinental Exchange, or ICE (working with index provider Ornn), each announced cash-settled GPU futures contracts, explicitly borrowing the language of oil and gas markets to describe what they are building ^1,2,3. Spot marketplaces such as SF Compute have raised institutional capital to build price-discovery infrastructure for GPU-hours that, until recently, traded only through bilateral, opaque hyperscaler contracts ⁵. The market is still young — the CME contracts were not scheduled to begin trading until October 5, 2026, pending regulatory review — but the direction is set: compute capacity is being financialized.

This paper asks a narrower and more useful question than whether compute will become a commodity, since that process is already under way. Instead, it asks which specific mechanics from mature commodity markets — natural gas, electricity, crude oil, gold, wheat, cotton, and coffee — the compute market is likely to adopt over the next one to three years, and which it is likely to reject, given the physical and economic characteristics that make a GPU-hour different from a barrel of oil or a bushel of wheat. Four questions are examined in turn: how compute delivery points are likely to fragment geographically; which contract specifications can be reconciled through simple price adders and which require entirely separate, non-fungible contracts; how the daily and weekly shape of compute demand compares to electricity's peak and off-peak structure, and where it is headed; and who the plausible counterparties are if a delivery-backed compute contract ever emerges.

The central finding is that compute is shaping up to resemble electricity far more than it resembles oil or gas — nodal, non-storable, and priced on local congestion rather than on a single national benchmark — and that the market infrastructure now being built is already consistent with that reading. Both of the exchanges that launched compute futures in 2026 chose independent, transaction-based index providers rather than letting any single hyperscaler set the reference price, and both chose to launch separate contracts for each chip generation rather than one blended "compute" contract with conversion ratios between chip types. Those two design choices, made months apart by competing exchanges, are the strongest available evidence for where this market is heading.

[SECTION 01]

1. Delivery Points: Compute's Coming Nodal Geography

Natural gas and electricity are useful starting points precisely because they are not priced as single national commodities. Natural gas has one dominant national benchmark, Henry Hub in Louisiana, which earns that status because it physically interconnects with thirteen major interstate and intrastate pipelines and sits near deep storage — the combination that produces liquidity ^15,29. But Henry Hub coexists with, and is priced relative to, a set of regional hubs whose prices can diverge from it dramatically (this price differential is defined as “basis” in trading parlance, and that term will be referred to throughout this document). Waha Hub, which serves the Permian Basin in West Texas, trades at a persistent discount because gas production there — much of it "associated gas" produced as a byproduct of oil drilling — has outrun the pipeline takeaway capacity available to move it to market. Waha averaged just $0.17 per MMBtu in 2024 and traded below zero on 36 percent of trading days that year; it’s basis to Henry Hub fell to a record −$10.03 per MMBtu on April 24, 2026, and stayed negative for 47 consecutive days that spring ^16,17. At the opposite extreme, Transcontinental Gas Pipeline's Zone 6 delivery point into New York City spiked to $97.90 per MMBtu on January 17, 2025 — the third-highest nominal price recorded since 1998 — during a cold snap, while Henry Hub itself moved only between $3.54 and $4.23 per MMBtu that same week ¹⁸. Appalachian hubs such as Dominion South and Eastern Gas South trade at their own structural discount for the same reason as Waha — regional production outrunning regional pipeline egress — though that discount has narrowed as new pipeline capacity has come online ^19,20.

Electricity takes this logic further. Because power cannot be stored in bulk and moves instantly across a congested transmission network, wholesale electricity prices are set at the level of individual grid nodes through locational marginal pricing, or LMP — a mechanism the PJM Interconnection itself has described using the analogy of surge pricing for a taxi ride, where the price rises specifically at locations where the grid is nearing its transmission limit ²². ERCOT prices six trading hubs in Texas (North, South, Houston, West, and two aggregate hubs) that are commonly used for hedging, alongside four load zones whose prices reflect the actual, more granular locational congestion used for settlement ²¹. NYISO divides New York State into eleven pricing zones running from Zone A in the west to Zone K on Long Island, with Zone J — New York City — trading at a structural premium to upstate zones because of transmission constraints into the city ²³. PJM similarly divides its region into separate zonal pricing regions rather than a single system price, though the PJM West Hub settlement price acts as the primary liquidity point PJM, and the primary source of trading liquidity / volume for financial power trading in the entire Northeast.

Why compute will sometimes likely follow electricity's nodal model, not gas's single-hub model

Compute’s price differentials split along a line that neither gas’s single-hub model nor electricity’s nodal model quite captures on its own — but the split is not that “a GPU cluster cannot move.” It can’t, but that fact is nearly irrelevant: the thing a buyer actually transacts on — a trained model’s weights, or an inference response — is data, and data moves over fiber at a cost close enough to zero that it isn’t the constraint. The real split is between workloads that have a hard real-time latency budget and workloads that don’t.

Training workloads and inference workloads have different degrees of locational flexibility. A large training run has no inherent need to sit near an end user. If Waymo wants train a new model on 50 million images of red lights, it can send the project to Palo Alto, Nova Scotia, or wherever is cheapest – latency is not a concern. But if a Waymo is driving 45 mph and trying to decide if the object in front of the car is a feather or a four year old, latency becomes a lot more important (note: thankfully Waymo and other autonomous vehicles handle all their inference and calculations on board, and are essentially zero latency mobile data centers). But I use this as an example to show there is a clear split between inference tasks (which need an answer immediately) and training tasks (which just need an answer) Training needs power, cooling, and enough interconnect bandwidth within the cluster, all of which can in principle be sited almost anywhere with available power, much as a gas-fired power plant can be built wherever pipeline capacity and land are cheap. Inference, by contrast, serves live human and application traffic, and network round-trip time creates a real, physics-bound floor on how far inference can be pushed from its users — the same reason NYISO's Zone J commands a premium over rural NY zones is structurally similar to why inference capacity near dense population and business centers should command a premium over inference capacity in remote regions, even if the underlying chips are identical ²³. Training compute, in other words, may behave more like a transportable commodity that can chase the cheapest available delivery node (ie Gold, Silver futures), while some inference compute may behave more like retail power delivered to a specific metro area — a distinction with real consequences for Sections 2 and 3 below. And while there are also different levels of time sensitivity to inference compute jobs (a retail user in California asking Claude to optimize his fantasy football lineup can certainly wait a few extra seconds for the job to be completed in Ohio), studies estimate 72% of enterprise applications require sub 30 millisecond response times, so latency is currently a major factor in most inference calls.

The reason this constraint is likely to intensify rather than fade is that compute and electrical power are becoming contractually inseparable at the infrastructure level, not just economically correlated. The Electric Power Research Institute projects that data centers could consume up to 9 percent of total U.S. electricity generation by 2030 — more than double their current share — with roughly 80 percent of 2023 data center load already concentrated in just fifteen states, led by Virginia and Texas ³². Where the power is available is increasingly where the compute physically has to be. Amazon's expanded agreement with Talen Energy illustrates the trend concretely: a 1,920-megawatt power purchase agreement tied to the Susquehanna nuclear plant, worth roughly $18 billion in revenue to Talen through 2042, with delivery ramping toward that full amount by 2032 ³³. Deals like this are effectively pre-negotiating the location of future GPU capacity years in advance, anchored to specific power assets — which is precisely the kind of fixed, geography-bound infrastructure investment that produces durable regional price hubs in the power market rather than a single fungible national benchmark.

A plausible path for 2026–2029

Extending the gas and power analogy, the most likely outcome over the next one to three years is not a single global compute benchmark, but a small number of liquid regional hubs — probably somewhere between three and eight, echoing how only a handful of gas and power hubs (Henry Hub, Waha, PJM West, ERCOT North, and a few others) can carry meaningful trading volume even though dozens of named pricing points exist. The hubs most likely to become liquid are the ones that already combine four ingredients found at today's liquid gas and power hubs:

1. abundant, contracted electrical power;
2. dense fiber and network interconnection;
3. significant localized demand and, most critically,
4. multiple competing capacity operators physically present, rather than a single company's proprietary campus.

The neocloud sector already shows where such multi-operator density is forming: CoreWeave alone reports 1.5 gigawatts of active capacity and more than 4.2 gigawatts contracted across 51 data centers, with a contracted revenue backlog that reached $99.4 billion in the first quarter of 2026; Crusoe reports roughly 0.2 gigawatts active against 4.9 gigawatts contracted and a pipeline exceeding 40 gigawatts; Nebius is targeting roughly 1 gigawatt of capacity by the end of 2026 with a 5-gigawatt contracted target ^8,9. Regions where several such operators cluster — rather than regions dominated by a single hyperscaler's private campus — are the best candidates for becoming genuine, multi-seller "delivery hubs" in the commodity-market sense, for reasons developed further in Section 4.

[SECTION 02]

2. Fungibility and Contract Specifications: What can be reconciled by an adder, and what can't

Fungibility and commonality are what let a commodity contract work as a shared price. If tradable contracts are interchangeable, one quoted price reflects the whole market, and, most importantly buyers and sellers can offset positions with anyone. That concentrates order flow in a few contracts, which produces the depth and tight spreads that make hedging cheap. Liquidity begets liquidity: The faster a trader can exit position the more comfortable they will be entering a position. More traders means more liquidity, tighter spreads, and smaller volume driven price fluctuations.

When units are not interchangeable, liquidity splits across many narrow contracts, and each one is too thin to trade well. This is why the paper’s comparison matters: CME and ICE launching separate H100 and B200 contracts makes sense for hedging precision, but the cost is fragmented liquidity, which the paper’s closing caveat flags as the central risk to compute market vitality.

Besides price, volume, duration, and counterparty, compute contracts already vary along several secondary dimensions/specifications at once — chip architecture (H100 versus B200, for instance), memory capacity (an 80-gigabyte configuration versus a 140-gigabyte configuration), interconnection, network topology, and location are a few of the more significant variations — each priced differently today in an unstandardized way. The question worth asking is which of these differences other commodity markets have learned to reconcile through a simple price adjustment, and which they have instead handled by creating an entirely separate, non-substitutable contract. The distinction turns out to be consistent across markets, and it offers a fairly precise answer to two illustrative questions worth asking about compute: could three H100s substitute for one B200, and could two 80-gigabyte chips substitute for one 140-gigabyte contract?

How mature commodities draw this line

Wheat is the clearest illustration of a market that separates categorical differences from continuous ones. Within a single wheat class, protein content is treated as a continuous quality attribute and priced with a fixed step schedule. The Kansas City Hard Red Winter wheat contract sets its par delivery grade at 11.0 percent protein; wheat testing at or above 11 percent in the No. 1 grade earns a 1.5-cent-per-bushel premium, wheat between 10.5 and just under 11.0 percent protein is discounted 10 cents per bushel, and wheat below 10.5 percent protein cannot be delivered against the contract at all ³⁸. The Minneapolis Grain Exchange's Hard Red Spring wheat contract works the same way, with par set at 13.5 percent protein and a 10-cent discount applied between 13.0 and 13.4 percent ³⁹. But protein adjustments only operate within a class. Hard Red Winter wheat, Soft Red Winter wheat, and Hard Red Spring wheat are not reconciled against one another through any discount schedule at all — they trade as entirely separate futures contracts, on different exchanges historically, with different delivery grades and even different quality logic (Chicago's Soft Red Winter contract carries no protein specification whatsoever, since it is a lower-protein wheat used for pastry and cake flour rather than bread) ⁴⁰. A trader who wants relative-value exposure between classes uses an inter-commodity spread trade; nobody delivers Soft Red Winter against a Hard Red Winter contract, at any ratio ⁴¹.

Cotton shows the same pattern from a different angle. The ICE Cotton No. 2 contract sets a base grade — 1-1/16 inch staple length, Leaf Grade 4, minimum strength of 25.0 grams per tex, micronaire between 3.5 and 4.9, and a base color grade of Strict Low Middling White — and then applies a schedule of premiums and discounts around that base for every one of those attributes, recalculated weekly from observed cash-market differentials rather than fixed once and left alone. Short staple is penalized asymmetrically: cotton at 1-1/32 inch staple is discounted at 200 percent of the standard average discount rate, reflecting how undesirable that shortfall really is to a spinning mill ⁴². Yet cotton, too, draws a hard non-fungibility line at growth region: only two broad growth categories — Eastern/Memphis/Orleans/Texas Upland, and Far Western Upland grown in Arizona, California, Nevada, New Mexico, and specific Texas valleys — can be commingled in a single certified delivery lot, and mixing bales across that line is treated as a contract default, not a quality adjustment ⁴².

Crude oil follows a related but distinct logic: rather than a continuous adder, NYMEX WTI defines a tolerance band — API gravity between 37 and 42 degrees, sulfur content no higher than 0.42 percent by weight — inside of which all qualifying crude is simply treated as identical for delivery purposes ³⁴. Grades outside that band, or grades buyers want distinguished even within it, trade as entirely separate companion contracts: Mars, WTI Sour, Louisiana Light Sweet, WTI Midland, and WTI Houston all trade as priced differentials to WTI rather than as adjustments within the WTI contract itself ³⁵. Tellingly, even inside the nominally fungible band, the market has re-fragmented on its own: S&P Global has documented three distinct real-world price tiers for crude delivered at Cushing, Oklahoma — generic NYMEX-spec crude at the base price, guaranteed-spec WTI trading roughly 2 cents per barrel higher via an exchange-for-physical trade, and pure, unblended WTI Midland trading closer to a 90-cent premium — because sophisticated buyers still pay for certainty even when the contract formally calls the material identical ³⁶.

Coffee, finally, treats origin as a fixed country-level differential layered on a common cup-quality and defect-count floor: the ICE Coffee C contract most recently priced Colombian coffee at a 4-cent-per-pound premium and Brazilian coffee at a 6-cent discount to par, on top of a shared maximum-defect standard of 15 full imperfections (tightened to 10 for Colombian coffee specifically) ^44,45.

Natural gas offers one further, useful contrast: it has essentially no quality-adjustment schedule inside the futures contract at all. The NYMEX Henry Hub contract does not specify a numeric BTU-per-cubic-foot minimum directly; instead it defines deliverable gas by reference to whatever specification is currently in the FERC-approved tariff of Sabine Pipe Line LLC, the pipeline operator at Henry Hub, delegating quality control entirely to the physical pipeline system rather than encoding it as a contract term ³⁰. Fungibility, in other words, is enforced upstream, at the point of physical interconnection, rather than downstream, in the price.

Table 3. How five mature commodities separate "price adjustment" from "different contract"

Applying the pattern to H100s, B200s, and memory tiers

Put plainly: three H100s are unlikely ever to substitute for one B200 under any published conversion ratio, and the market has already, independently, reached the same conclusion. Both CME/Silicon Data and ICE/Ornn chose to launch separate futures contracts for each chip generation rather than one blended index with built-in conversion factors between chips ^1,3. That decision mirrors exactly how the wheat market handles Hard Red Winter versus Soft Red Winter — different products, priced separately, connected only by a spread trade for those who want relative-value exposure, never by a delivery substitution ratio. The underlying reason is the same one wheat classes and oil grades outside the tolerance band run into: an H100 and a B200 are not scalar multiples of the same thing. Precision support (B200's native FP4 versus H100's FP8 ceiling), memory bandwidth per accelerator, and the size of the NVLink domain each chip participates in are qualitative differences in what the hardware can actually do, not differences of degree along one shared axis — closer to the gap between bread wheat and pastry wheat than to the gap between 11 percent and 10.6 percent protein wheat. An essay specifically addressing this question, titled "A GPU Hour Is Not a Barrel of Oil," makes the point from first principles: a GPU-hour is a resource allocation, not a unit of completed work, and two GPU-hours that are identical on a spec sheet can still differ in cluster topology, achievable bandwidth, and location in ways that make simple unit conversion unreliable. That essay also points to the Deutsche Börse Cloud Exchange, a 2015–2016 attempt to standardize a generic tradable "compute unit" across providers, as a cautionary precedent — it failed specifically because an apparently equivalent virtual machine could differ in ways that made contracted delivery unreliable across sellers ¹³.

Memory capacity looks like a better candidate for a genuine adder — but only within a chip family, not across one. Two 80-gigabyte accelerators are unlikely to be treated as fungible with one 140-gigabyte contract for the same reason two H100s aren't a substitute for a differently architected chip: a single 140-gigabyte memory pool addressable by one accelerator behaves differently, for a large share of AI workloads, than the same total memory split across two separate cards connected by an interconnect with its own latency and bandwidth limits. That is a topology difference, not merely a quantity difference. Within a single chip family, however — say, an 80-gigabyte and a 96-gigabyte variant of the same underlying architecture and interconnect generation — a continuous, per-gigabyte price adder in the style of wheat's per-protein-point schedule looks like a genuinely workable mechanism, since the underlying product is the same and only the quantity of one resource differs. Network and interconnect tier (full InfiniBand or NVLink fabric versus a standard Ethernet-networked instance) and uptime or SLA reliability look similarly well suited to discount schedules along the lines of cotton's strength and leaf-grade deductions, or coffee's defect-count ceiling: the same base unit, priced down for a measurable shortfall in one attribute, rather than reclassified as a different product.

Geography sits in its own category, and it is worth underscoring because it is the one dimension unlikely to be resolved by either mechanism above — not a simple discount schedule, and not a wholesale separate contract for every metro area, but a basis structure. Section 1 already made the underlying case: because locational price differences in compute should track locational price differences in electrical power, and because power basis moves with weather, congestion, and demand rather than settling into one fixed number, geography needs the kind of continuously reset basis-swap mechanism gas markets use to connect Waha or Transco Zone 6 to Henry Hub, not a static per-region premium fixed once and left alone ^16,18,20. A flat geography adder would be quickly overrun by the same congestion dynamics that make Waha trade negative one month and Transco Z6 NY spike to nearly $100 the next.

Compute contract specifications are a significantly deep enough topic that we’ve published a separate report on that topic alone here.

[SECTION 03]

3. Peak and Off-Peak: The Daily Shape of Compute Demand

Electricity markets exist in their current form partly because demand is deeply cyclical and partly predictable: on-peak contracts, under the standard NERC convention followed across most U.S. power markets, cover roughly 7 a.m. to 11 p.m. on non-holiday weekdays, while off-peak contracts cover nights and weekends ²⁵. The New York ISO's tariff is explicit about it, defining off-peak hours as running from 11 p.m. to 7 a.m. ²⁴. Layered on top of that daily on-peak/off-peak split are day-ahead markets, which clear a financially binding hourly schedule the day before delivery, and real-time markets, which settle the actual deviations from that schedule roughly every five minutes as conditions change ²⁶. A third layer, capacity markets, pays generators simply to be available in a future delivery year — a payment for reliability and optionality, separate from any energy actually sold ²⁷. All three layers exist because electricity cannot be stored economically at scale, so the system has to price both the certainty of future availability and the moment-to-moment balancing of supply and demand.

What compute demand looks like today

It is worth being direct about a real gap in the public record: in searching hyperscaler disclosures, neocloud materials, or analyst research, that cleanly plots data-center-wide GPU utilization by hour of day and separates training load from inference load - no dataset was found. That comparison likely exists inside individual hyperscalers and neoclouds as internal capacity-planning data, but what is available, and cited below, are several credible proxies that point in a consistent direction.

SemiAnalysis has published detailed analysis of power draw during large training runs that is a reasonable stand-in for a load-shape description, even though it is not framed in daily peak/off-peak terms. Training clusters exhibit extreme, sub-second power fluctuations rather than a smooth daily curve: a Meta Llama 3 training run on roughly 24,000 H100 GPUs, drawing about 30 megawatts, showed instantaneous swings of tens of megawatts within fractions of a second, and Google Cloud data cited in the same analysis found roughly a fifteen-fold difference in load fluctuation magnitude between a traditional cloud data center and an AI training data center — swings between 1.5 and 15 megawatts rather than a gentle ramp ¹¹. The mechanism is technical, not human-driven: GPUs alternate between synchronized computation and synchronized communication (an operation called AllReduce), pause almost entirely during periodic checkpointing, and oscillate with the batch-processing cycle — a load signature closer to a chemical plant's batch-cycle demand spikes than to a city's evening electricity ramp. Inference looks structurally different, and closer to a genuinely diurnal, usage-driven pattern: the same SemiAnalysis analysis notes that inference workloads hit maximum power during the initial "prefill" phase of processing a request but often run under 50 percent utilization during the subsequent "decode" phase, and it flags that the largest current reasoning models, which require many nodes working together to serve a single inference request, are starting to "resemble more training" in their volatility as they scale — meaning the clean training/inference distinction on load shape may already be blurring at the frontier even as it holds for smaller, simpler models ¹¹.

Epoch AI's research on how AI labs allocate compute between training and inference is not a time-of-day study either, but it is a genuinely useful data point on aggregate scale: Epoch estimates that OpenAI generates on the order of 36 trillion inference tokens annually against training runs consuming roughly 10 trillion tokens, with each training token costing roughly three times as much compute as an inference token — implying that OpenAI's annual inference compute spend is now on a similar order of magnitude to its annual training compute spend, even though a single training run remains far more schedulable and geographically flexible than an always-on inference fleet serving live user traffic ¹². If that balance is now roughly even at a leading lab, and if inference's share continues to grow as more AI usage moves into consumer and enterprise applications running continuously rather than in the research lab, then the demand-driven, diurnal component of total compute load should mechanically grow relative to the batch-driven, schedulable component.

Where this is likely headed over five years

Extending that trajectory, and treating this explicitly as a forecast rather than an established fact, compute demand should develop more daily peak-and-off-peak variability over the next several years, not less — the opposite of what a market maturing toward smooth, continuous utilization might otherwise suggest. The reasoning follows directly from the training/inference distinction developed above and in Section 1. As inference's share of total compute grows, and as inference is inherently usage-driven and tied to human and business activity patterns across time zones, its aggregate demand curve should increasingly resemble electricity's — rising through the business day in whichever regions are active, falling at night and on weekends in those same regions. Training, by contrast, retains its scheduling flexibility precisely because it does not need to track live users, which creates a natural complementary role: training capacity can be preferentially scheduled into the overnight and weekend troughs that inference leaves behind, in much the same way flexible industrial gas and power customers today accept interruptible-service contracts, at a price discount, in exchange for being curtailed whenever the system needs the capacity for firm, time-critical demand ²⁷. Expect, over the next one to three years, a pricing structure to emerge in which reserved or training-oriented capacity is sold at an off-peak-style discount and on-demand or inference-oriented capacity at peak windows carries a premium — a structure that is already visible in embryonic form. Tomasz Tunguz has documented B200 spot prices moving from $2.31 per hour in early March 2026 to $4.95 per hour by late April 2026 — a 114 percent move in six weeks — coinciding with major model releases (OpenAI's GPT-5.3-Codex in February 2026 and GPT-5.5 in April 2026) that presumably drove sudden inference demand spikes ⁷. These are demand-event-driven price moves rather than clean daily cycles so far, but they are the same underlying phenomenon peak electricity pricing is built to capture, just not yet organized into a standard contract structure.

One further wrinkle is worth flagging, because it cuts against a pure electricity analogy: compute's "transmission" — meaning the ability to route a workload to wherever capacity is currently idle — is far more flexible than the electric grid's physical transmission network, at least for workloads that tolerate some added latency. A training job, and even a fair amount of asynchronous inference traffic, can in principle be shifted across time zones far more easily than power can be wheeled across a congested grid interconnection. That flexibility could partially flatten the aggregate global demand curve even as any single regional node continues to show a sharp daily peak-and-off-peak pattern — much as a global cloud operator can shift batch workloads to whichever of its regions is currently in its overnight trough. If that kind of cross-region load-balancing scales the way this paper expects, the more durable trading product may end up being a peak/off-peak contract defined per hub, alongside a separate, more constant-load contract for large operators sophisticated enough to arbitrage the daily cycle across their own global footprint — a distinction electricity markets do not need, because power genuinely cannot be moved between interconnected grids the way a training checkpoint can be moved between data centers.

[SECTION 04]

4. Physical Settlement: Who Would Need to Sign On

The most direct evidence available on this question comes from what the two exchanges that actually launched compute futures in 2026 chose to do, rather than from theory. Both CME Group, working with Silicon Data, and ICE, working with Ornn, built cash-settled contracts referencing an independent index, not physically delivered contracts, and not a price set by any single hyperscaler ^1,3. That is a real design choice with real alternatives available, and it is worth taking seriously as the market's own answer to the question this section asks.

Why cash settlement against an independent index, not physical delivery

As mentioned earlier, a GPU-hour's economic value depends on a number of characteristics: location, networking quality, uptime, security posture, software stack, minimum reservation windows, queue time, and data-movement costs, all varying by provider — making a GPU-hour economically inconsistent across providers in a way that makes physical-delivery verification genuinely hard to standardize. Jason Hoffman's essay makes a related point using a specific historical precedent: the Deutsche Börse Cloud Exchange, launched around 2015 and shut down within roughly a year, attempted to create exactly the kind of neutral, multi-provider tradable compute-unit market a physically settled contract would require, and it failed because an apparently equivalent virtual machine could differ enough in processor generation, contention, storage behavior, network performance, geography, and support to make delivered units unreliable substitutes for one another. Hoffman's broader argument is that AWS Spot succeeded where Deutsche Börse's exchange failed not because Amazon solved the standardization problem, but because Amazon sidestepped it — Spot works because Amazon controls both the inventory and the settlement within one single administrative domain, not because it created a neutral market across competing sellers ¹³. That is a genuinely important distinction for anyone imagining a physically settled compute contract: the only physical-delivery model that has actually worked at scale so far is a single company delivering its own capacity to its own customers, not a multi-party exchange delivering interchangeable capacity from competing sellers.

Would a single large hyperscaler need to set the index price?

The market's actual behavior argues against it. CME chose Silicon Data, an independent index provider backed by the trading firm DRW rather than by any cloud operator, to publish the reference price for its H100 and B200 contracts ^1,2. ICE chose Ornn, whose Ornn Compute Price Index is explicitly marketed as built "only from printed transactions" — meaning actual executed trades rather than indicative quotes — precisely to avoid the manipulation risk of a price based on unexecuted or self-reported figures ³. This mirrors, almost exactly, the role S&P Global Commodity Insights (formerly Platts) and ICE itself already play in gas and power markets, where neutral third-party assessors publish the settlement indices used across the industry rather than any single utility or producer publishing its own price as the market standard ²⁸. A single hyperscaler setting the reference price for a compute benchmark would face the same conflict-of-interest problem a single oil major would face if it alone published the WTI settlement price.

Should the physical counterparty be a neocloud?

The index-setting question and the physical-counterparty question are separable, and the evidence points differently on each. If a compute contract does eventually move toward genuine physical delivery — most likely, per Section 4's next point, through privately negotiated bilateral arrangements rather than as the default settlement path — neoclouds look like more natural physical counterparties than hyperscalers, for a few concrete reasons. Neocloud contracts are already structured more like the kind of negotiable, shorter-commitment arrangement a trading relationship needs, in contrast to hyperscalers' standardized consumption-commitment and minimum-spend terms, which are built for enterprise procurement rather than spot-style trading ¹⁰. Neoclouds have also built real scale: CoreWeave alone reports 1.5 gigawatts of active capacity, more than 4.2 gigawatts contracted, 51 data centers, and a contracted backlog that reached $99.4 billion in the first quarter of 2026, growing roughly fourfold year over year ^8,9. And neoclouds generally lack the broader platform moat a hyperscaler is protecting — AWS, Azure, and Google Cloud each have strong incentives to keep customers inside a proprietary ecosystem of adjacent services, while a neocloud's business model is closer to pure capacity provision, making participation in a neutral exchange more likely as it is not a competitive threat.

There is a close historical parallel worth naming directly: Cushing, Oklahoma became the WTI delivery point not because a single major oil company offered its own terminal as the benchmark location, but because it is a neutral interconnection hub, served by multiple independent pipeline systems and many separate storage operators, none of which controls the point outright ^34,35. A vertically integrated hyperscaler offering its own data center as the physical delivery point for a compute benchmark would concentrate exactly the single-party counterparty risk every other commodity market with a genuine delivery-based benchmark has structurally avoided. A neutral, multi-tenant hub — whether a large neocloud's facility, or, plausibly, a carrier-neutral colocation or interconnection site analogous to the role companies like Equinix or Digital Realty already play in networking, hosting capacity from more than one seller under standardized interconnection terms — is the closer analog to Cushing than any single hyperscaler's proprietary campus would be.

One universal delivery counterparty, or different companies for different geographies?

Section 1's argument that compute delivery is fundamentally nodal — power- and latency-constrained like electricity, not a single national point like Henry Hub — points fairly directly to the answer here. A single universal physical counterparty across every geography looks unlikely for the same reason no single company operates the entire U.S. electric grid; PJM, ERCOT, NYISO, and MISO each function as the regional authority for their own footprint rather than one operator running the whole system ^21,22,23. The more probable structure over the next one to three years is close to the pattern already embedded in the framing of this question: a small number of large, capacity-rich operators — likely a mix of neoclouds and, opportunistically, hyperscalers selling genuinely excess capacity — become the de facto physical anchors in the specific regions where they already have the deepest infrastructure, echoing something like a CoreWeave-anchored hub in one region, a Nebius-anchored hub in another, an xAI-anchored hub in a third, etc. Layered on top of that regional physical structure, a single independent index provider — playing the Platts or ICE role rather than the Exxon or utility role — aggregates transaction data across all of those regional anchors into one or a small number of tradable, cash-settled national or global benchmark contracts, the same way Henry Hub functions as the national gas price reference even though Waha, Transco Zone 6, Dominion South, and SoCal Citygate remain economically distinct regional physical markets connected to it only through continuously reset basis differentials ¹⁵–^20,29–³¹.

In that structure, actual physical delivery — as opposed to cash settlement against the index — would most plausibly occur the way it already occurs in oil markets: through privately negotiated exchange-for-physical trades between large, sophisticated counterparties, settled off the public order book, while the exchange-traded contract itself stays overwhelmingly cash-settled ³⁷. That is not a hypothetical caveat; it is close to how Henry Hub itself already works. Henry Hub futures are nominally physically deliverable, obligating delivery via the Sabine Pipe Line network for anyone who holds a contract to expiration, but the overwhelming majority of contracts are closed out financially before expiry rather than delivered ^29,30. A compute contract following the same path — legally capable of physical delivery, used that way rarely and mostly by the largest, most sophisticated buyers and sellers, but functioning day to day as a cash-settled financial instrument — is the most likely outcome of the counterparty structure described above.

[SECTION 05]

5. Synthesis: What to Watch For Over the Next One to Three Years

Pulling the four analyses together produces a reasonably specific set of predictions, each grounded in a market design choice mature commodities have already made and, in several cases, one the nascent compute-futures market has already echoed.

Delivery points will fragment along power and connectivity lines, not settle into one national benchmark. Expect a handful of liquid regional hubs — plausibly three to eight over this window — to emerge specifically where abundant contracted power, dense fiber connectivity, and multiple competing capacity operators coincide, mirroring why only a few gas and power hubs actually carry deep trading volume despite dozens of named pricing points existing on paper ^15,21,22.

Contract specifications will move towards more standard/common GPU configurations when possible, and the nascent futures market has already drawn the line in the same place this paper's commodity-market comparison predicts. Chip generation and precise geography are unlikely to ever be resolved through a simple adder — they define contract identity, the way wheat class and delivery hub already do — while memory tier within a chip family, network/interconnect tier, and uptime or SLA reliability are good candidates for continuous premium/discount schedules, the way wheat protein and cotton staple length already work ^38,42. CME and ICE's decision to launch separate contracts per chip generation, rather than one blended index, is the clearest confirming signal available today ^1,3.

Demand shape should grow more peaked and more time-zone-differentiated, not smoother, as inference's share of total compute continues to grow relative to training's. Training is likely to increasingly play the role flexible, interruptible industrial demand already plays in gas and power markets — filling the troughs that usage-driven inference leaves behind — while compute's greater workload portability, relative to electricity's physically fixed transmission grid, may partially flatten the picture at the level of a single large multi-region operator even as any individual hub continues to show a sharp daily cycle ^11,12,27.

Settlement infrastructure is unlikely to converge on one hyperscaler setting the price and one hyperscaler serving as the universal delivery point. The pattern already visible — independent, transaction-based index providers (Silicon Data, Ornn) playing the Platts/ICE role, paired with a small number of regionally dominant, more neutral capacity operators serving as physical anchors — looks far more durable, with genuine physical delivery staying rare and privately negotiated even as cash-settled trading volume grows, closely tracking how Henry Hub itself actually trades today ^1,3,28,29.

A closing caveat

None of this should be read with more confidence than the underlying market deserves. Compute differs from every commodity examined here in one respect that could upend parts of this forecast: the pace of the underlying hardware cycle. Wheat has a crop year; oil grades change slowly, over decades; but GPU architectures turn over roughly every one to two years, and each new architecture can change what a "unit" of compute is actually worth in ways slower-moving commodities never have to price around. Friedman's primer names this directly as a tension between standardization and hedge precision, and warns that fast hardware-generation turnover risks splitting liquidity across too many hardware-specific contracts before any one of them reaches the trading depth gas and power hubs took years to build ⁴. The compute market's next one to three years will likely be shaped as much by whether Nvidia's, AMD's, and any future entrants' release cadence outruns the market's ability to build liquidity in each new contract, as by any of the structural analogies to gas, power, wheat, cotton, coffee, and oil developed in this paper.