A commodity isn't a single price; it's a stack of differentials. Brent doesn't trade alone — it sits on top of WTI-Brent grade spreads, locational basis (Mars, LLS, MEH), quality tiers (sweet vs sour, sulfur content), tenor (prompt vs cal-strip), credit (cleared vs bilateral). Power doesn't trade alone — it sits on top of zone-vs-hub spreads, peak vs off-peak, firm vs non-firm, FTRs.

For compute to commoditize the same way, it needs the same skeleton. Six levels, each of which has a known commodity-market analog and a known pricing technology. The hierarchy below is the spine of everything that follows.

The point of the hierarchy is not that compute will be traded exactly like crude or power. It's that every layer above the benchmark already exists in a market where pricing technology is mature. If a desk can price an FTR, it can price an L2 compute basis. If it can price a 3:2:1 crack, it can price an L1 grade spread between H100 and B200. The work isn't inventing new pricing math; it's reading the compute tape through the lens of the math that already exists.

The richest layer is L4 (tenor). $/FLOP deflates ~30-40% per year through generational succession, so the curve should sit in structural backwardation in grade-adjusted terms — but scarcity (2026: +48% in two months) puts spot premium on top. Disentangling decay from scarcity is the compute version of separating roll yield from convenience yield. L3 × L2 interaction is the second richest: training demand is geographically mobile (chases cheap power + capacity); inference is sticky (chases users). Over time this bifurcates L2 into training basins (ERCOT, frontier markets) and inference premia (NoVa, metro-adjacent) — the compute equivalent of off-peak vs on-peak and zone-vs-hub that power traders price daily.

Two compute-futures listings filed in May 2026 differ in every dimension that matters. The contrast is the central fact for the next twelve months of this market.

• CME × Silicon Data is a quote-based assessment. Silicon Data computes daily benchmarks (SDH100RT for H100 Neocloud, SDB200RT for B200, plus a separate H100 Hyperscaler index) from quoted and listed rental rates sampled across roughly 50 chipsets and providers. The CME contract settles cash vs the on-demand indices. Settlement methodology — single window or monthly average — was not in the May announcement. Read this as the Platts-style assessment of compute: quote-derived, composite, settlement-window sensitive.
• ICE × Ornn is a transaction VWAP. Ornn's Compute Price Index (OCPI) is built from printed and cleared transactions on live GPU marketplaces, volume-weighted, normalized across hardware/provider/deployment, and explicitly regionally weighted. The ICE contract is Asian-style — averaging volume-weighted executed transactions over the contract period, deliberately modeled on power settlement. Read this as the power-market analog: transaction-anchored, period-averaged, regionally weighted.

The methodology gap is not a complication; it's the cleanest visible evidence that the asset class is still figuring out what counts as a "price." Two exchanges, two philosophies, one underlying. Either the two indices converge — in which case compute commoditizes in the Brent-vs-WTI sense, with a stable structural spread — or they don't, in which case the asset class has further to go before it earns the label.

The dashboard sits here, at the center of the piece. The three sections that follow each lean on a specific view to test one component of the commoditization thesis: the SD vs OCPI dispersion view tests whether the indices will converge; the Grade × Tier × Region matrix tests whether the dispersions across hardware, SLA, and geography are stable and tradeable; the Compute → Power conversion view tests whether the supply curve is already sitting in power-market basis that desks can trade today.

A commodity has, eventually, one price. Brent and WTI quote separately, but the spread between them is bounded, well-modeled, and tradeable. If the SD-OCPI spread behaves the same way — mean-reverting around a stable band that's a function of sample mix, settlement convention, and regional weighting — compute looks like a commodity. If the spread drifts or regime-shifts based on whatever the market is arguing about that week, the two indices are measuring different things and compute hasn't yet earned the label.

The dashboard's first view decomposes the spread into four observable legs. Each leg is calibratable from the six months of OCPI history that's been on Bloomberg since April 2026, plus the published Silicon Data Q1 tape.

• Quotes vs prints. Listed and cleared rates diverge most when the market fragments. In March 2026 the listing ceiling spiked past $8 while the floor sat near $2.85. In a squeeze, aspirational listings drag a quote-based index up faster than volume clears; in a glut, stale listings make it sticky on the way down.
• Sample / tier mix. The 3.0× hyperscaler/neocloud premium on identical silicon means a 10-point difference in tier composition between the two samples moves relative index levels materially. Marketplace volume drifting toward enterprise SLAs trends the basis without a price move at all.
• Settlement shape. OCPI is confirmed Asian (period average); CME's methodology is unpublished. If CME settles on anything other than the identical window, the spread embeds average-vs-point convexity — the dailies-vs-monthly Jensen inequality from the power book, hiding inside what others will price as flat basis.
• Regional weighting. OCPI is regionally weighted; SD benchmarks are composites. A region-specific shock moves one settlement and not the other.

What the dashboard shows, scenario by scenario: each of the four legs has a stable functional form. None of them is mysterious. If a sponsor publishes the SD methodology in detail (sample composition, trim rules, tier weights), the band collapses to a modelable function of observables. That's the Brent-vs-WTI case for compute: there will be two prints, the spread will be tradeable, and commoditization holds.

What would falsify the case: an SD methodology that's deliberately opaque, or a sample whose composition shifts every quarter without warning. Both would prevent the band from converging. That outcome is possible — and is the most important thing to watch in the next two CFTC product filings.

A commodity needs differentials that are persistent enough to support hedging and speculative structures. WTI-Brent is a tradeable grade spread because it's been stable for decades and is driven by known fundamentals (gravity, sulfur, logistics cost). For compute to commoditize beyond the flat price, the analogous differentials — across hardware generations, SLA tiers, and regions — need to behave the same way.

The dashboard's matrix view crosses five grades (H100, H200, B200, GB200, RTX 5090) with three firmness tiers (interruptible spot, on-demand neocloud, on-demand hyperscaler) and five regions. Pairwise dispersion in any direction is ranked by magnitude. Drag the scarcity shock or the regional shock and watch which pairs blow out and which absorb.

The Q1 2026 tape already shows three durable signals:

• Grades decouple. B200 +24% in March while both H100 indices barely moved. Hardware grades carry idiosyncratic shocks driven by supply-chain bottlenecks (HBM3e contract prices +20%) and demand-side substitution. Implied correlation between grade contracts should be priced LOW.
• SLA tier is the largest single differential in the complex. H100 Hyperscaler index prints ~3.0× the H100 Neocloud index on identical silicon. This is bigger than grade dispersion, bigger than regional dispersion, and structurally permanent — a hyperscaler-tier basis market that any neocloud-sampled index leaves unhedged by construction.
• Vol scales with grade age. B200 coefficient of variation 11.4% (year one) vs H100 Neocloud 2.6% (year two) vs H100 Hyperscaler 0.5%. That's a vol term structure across the grade ladder. Young grades trade like prompt power, mature grades like deferred. Whoever lists options on the first listed contracts will mark new-generation vol off equity-vol intuition and misprice the commodity-style term structure.

The matrix view is what makes those signals actionable — you can isolate "hold tier and region fixed, vary grade" or any other combination, and see what the dispersion ranks look like. If those pairwise spreads stay stable across regimes, compute has a grade ladder, a tier ladder, and a basis grid worth trading. That's a commodity.

Compute has every structural property of a tradeable commodity, but its single largest physical input is electricity — and the electricity grid is one of the longest-lead, most-public supply curves in any commodity market. The dashboard's third view makes the link explicit.

Every compute quantity converts mechanically into a power quantity. Chip TDP scales to rack density, racks scale to facility load via PUE, and facility load grosses up to a grid number that shows up as basis exposure in PJM, ERCOT, and the rest of the host-region power markets. A 100,000-GPU GB200 deployment is roughly 209 MW at the grid (1.67 kW × 1.25 PUE × 100k); over a year at 85% utilization, that's 1.55 TWh of consumption — meaningful at the zone level, especially in tight zones like Dominion or ERCOT North.

Two consequences matter for the commoditization argument.

First, the compute spark spread is fat. Energy is 4-6% of revenue for a Blackwell-class platform at $65/MWh against $4-6 GPU-hr rentals. Variance comes from the compute leg, not from electricity. Compute is capex- and scarcity-driven, not fuel-driven — closer to crude grades than to gas-fired generation. The cross-product hedge math works, but the dominant leg is the compute side.

Second, and more important: power's influence runs through availability, not price. Grid interconnection is the binding constraint on supply growth. Northern Virginia alone has 16.8 GW under construction or planned against roughly 4.0 GW of existing inventory; the North American pipeline is about 35 GW, 64% of it in emerging markets, and around 60% pre-leased. That's the compute supply curve, and it's sitting in interconnection queues that anyone with access to PJM filings can read. The lead time on a 175 MW data center is 24-36 months from queue position to commercial operation. Whoever reads the queue prices the supply curve before it shows up as anything called "compute basis."

This is the strongest single argument for the commoditization case: the supply curve is already public, dated, and indirectly tradeable through power-market basis. The asset class doesn't need to wait for a futures listing to be priced — it's being priced, in DOM-zone and AEP-zone forwards, right now.

Five structural reasons the process might take longer than the futures filings suggest. None of them is a reason it can't happen — only reasons it might not happen on the timeline that May 2026 implies.

• Benchmark integrity. GPU-hours are heterogeneous in ways crude and power are not — networking, SLA, software stack, cluster scale all affect the economic value of an hour. Index methodologies are young and gameable; expect settlement-window squeezes and methodology revisions that gap the basis. The Q1 2026 sample fragmentation (B200 ceiling past $8 while floor sat near $2.85) is the warning sign.
• Technological discontinuity. An architecture break — a new chip generation that reshapes FLOP economics, an algorithmic-efficiency jump that compresses demand per unit of model quality — can obsolete the grade ladder itself. There is no analog in oil, where a barrel stays a barrel. The L1 hierarchy is the most exposed to this risk.
• Demand reflexivity. Compute demand is one capex cycle, highly reflexive, correlated to equity markets in a way oil and power are not. An AI-capex pause is a correlated drawdown across every long-compute expression including the power-basis proxies. The supply curve is durable (already-interconnected load shows up regardless); the demand curve is not.
• Regulatory re-segmentation. Export controls, data-sovereignty rules, and energy-allocation policy can re-segment L2 overnight. A regulatory basis-blowout has no weather analog and no mean-reverting structure — it's a one-way regime shift. Most acute on non-US legs but increasingly relevant domestically as states pass datacenter-specific rules.
• Liquidity asymmetry. Day-one listed markets will be thin; the natural hedgers (neoclouds, AI labs) are not yet futures users. Without designated market-maker programs to bootstrap quoting, the bid-ask story collapses and the cost case for the dashboard's arbitrage-triangle view doesn't survive contact with real execution.

Each of these is a place to watch. None of them is fatal. The Brent complex needed two decades to settle into the shape it has today, and that was an asset class with a single clearly-defined unit of measure. Compute has more dimensions to organize and a faster clock. Whether the commoditization completes in three years or ten depends on which of the five constraints binds first.

The hierarchy is recognizable. Six levels, every one with a mature commodity-market analog and a known pricing technology. The dashboards already show the dispersion forming — SD vs OCPI under different fragmentation states, grade × tier × region spreads under scarcity and regional shocks, compute load converting cleanly into grid MW. The supply curve is already sitting in PJM and ERCOT basis, dated 24-36 months ahead of when the load actually materializes.

None of that proves compute will commoditize on the futures-listing timeline. It does prove that the structures required for commoditization are visible right now, in published data, and that the dispersions trade structures rely on are already present in the underlying market — not waiting on a CFTC product filing to start existing.

The question that's not "will compute be a commodity?" is which version of "commodity" it will be. A Brent-style benchmark with a clean assessment? A power-style transaction-anchored VWAP with explicit locational weighting? A natural-gas-style hub-and-spoke with persistent basis structures? The dashboard is built to let you watch that question resolve in real time, using the data that already exists.