Power, Compute, and the Margin Engine

The 3 September announcement is short on specification and long on intent. The contracts are financially settled against Compute Desk's daily blended GPU price indexes, described as IOSCO-compliant and built on exclusive coverage of private compute transactions, covering Hopper, Blackwell and further architectures, clearing at Nodal Clear, targeted for 2026 subject to regulatory compliance [1]. Paul Cusenza's framing is the operative sentence: power is a significant operating cost for data centers, and offering the leading set of power and compute contracts on the same venue positions Nodal as the exchange for the AI economy. Bloomberg's account adds the detail that Nodal plans to allow electricity and compute positions to offset through portfolio margining, and the competitive context — CME's Silicon Data H100 and B200 contracts targeting 5 October on NYMEX, ICE working with Ornn, Architect building an EFP market on the same Compute Desk benchmarks, and a CFTC review opened in August that could delay every listing [2][6].

Nodal Clear's margin model is the part that is public and precise. Its Disclosure Framework describes an expected-shortfall calculation over "the average of the top portfolio losses simulated with selected historical scenarios over an assumed holding period," with scenarios drawn from "both recent periods and during past periods of market stress," antithetic returns added so that the scenario count exceeds a thousand, a one-day holding period for customer accounts and not less than two days for house accounts, a 99% target for individual portfolios and 99.5% across all participant accounts, and a separate liquidation-cost component driven by the participant's share of contract open interest and the volatilities and correlations of the portfolio [3][4]. The sentence that decides this piece is the one on offsets: Nodal Clear calculates all margins on a portfolio basis. As a result, Nodal Clear does not have explicit margin offset policies. Rather, the price histories incorporated into the margin model determine the amount of offsetting, if any, that will occur.

Two structural facts follow before any data is examined. An expected-shortfall model on joint historical scenarios is sub-additive: two positions whose losses do not coincide always margin at less than the sum of their standalone requirements, whatever their correlation, because the tail of the sum is thinner than the sum of the tails. The offset Nodal describes is therefore automatic for any pair of contracts once both have a history in the scenario set. And that same sentence means the offset is only as large as the joint history makes it, and a contract listed in the fourth quarter of 2026 on an index whose public record is measured in months has a joint history measured in months. The empirical question is what that history will show; the administrative question is what Nodal Clear does about the scenarios it does not have.

Before correlating anything, the pass-through channel can be bounded from the physics. An HGX/DGX H100 server draws about 10.2 kW under load across eight GPUs; at a facility PUE of 1.30 that is 1.66 kWh per GPU-hour delivered. A DGX B200 draws 14.3 kW (2.23 kWh per GPU-hour at PUE 1.25) and a GB200 NVL72 rack roughly 130 kW across 72 GPUs (2.17 kWh at PUE 1.20) [19]. Multiplying by the hub price gives the energy cost embedded in an hour; dividing by the rental index gives the share, and the share is the upper bound on the elasticity of the compute price to the power price if every cent were passed through instantly.

The bound does real work. If the share is 3.5–6.5%, a 100% move in power — Dominion Hub went from $62 in December 2025 to $167 in January 2026, then back to $42 by March — shifts the cost basis of an hour by 3.5–6.5%, and the operator's realised pass-through is slower and smaller than that: neoclouds sell at posted or contracted rates, buy power under multi-year retail or PPA contracts, and price the hour on GPU depreciation and utilisation, not on the day-ahead LMP. In return space the implied daily correlation is the elasticity times the ratio of volatilities; with a prompt-month power proxy near 4.4% a day and the neocloud index near 1.2% a day, even instantaneous full pass-through would put ρ near 0.15, and realistic pass-through puts it at zero to two decimal places. The model will not find in the histories a channel that the physics says is not there.

There is more data on Dominion's data-center load than on any other load pocket in the country, and all of it says the same thing: Northern Virginia is where the power constraint binds. Dominion's January 2026 load-forecast documentation to PJM lists 9.8 GW of connected data-center capacity under executed service agreements, 7.1 GW under construction and 30.1 GW in engineering study — 47 GW of contracted capacity against a 2025 coincident data-center peak near 4 GW and a 2046 demand forecast of 16.6 GW [8]. EIA counts a 30 million MWh rise in Virginia commercial sales between 2019 and 2025, second only to Texas, and a Dominion-zone winter peak of 25,413 MW in 2025–26, 45% above 2019–20 [9]. The capacity market has already repriced it: the 2026/27 base residual auction cleared the RTO at its $329.17/MW-day cap with the Dominion zone separated at $444.26/MW-day [10], and the 2027/28 auction cleared at the $333.44 cap across the footprint, 6.6 GW short of the reliability requirement, with 5.1 GW of the 5.25 GW year-on-year rise in the peak forecast attributed to data centers [11].

Two things are true at once. The Dominion–ERCOT West differential is the most direct price expression of the data-center build-out in the US power market, and it is getting larger: the energy spread more than doubled from 2025 to 2026, the zone-to-hub basis into Northern Virginia nearly doubled with it, and the Dominion capacity premium adds an amount of the same order. And per GPU-hour the whole stack — energy spread, basis, capacity — sums to roughly $0.10 to $0.13, about 4–5% of a neocloud hour and under 2% of a hyperscaler hour.

Against that, the regional structure of compute prices is an order of magnitude larger and not obviously aligned with power. Silicon Data's Q3 2026 regional prints put neocloud H100 near $2.90 in US East and $2.00 in US West, with hyperscaler pricing running the other way (US West $6.10 against US East $5.40). The Nordics — cheap, cool, hydro-fed power — carry the highest neocloud price in the set at $3.30 because capacity there tightened [16]. The compute price is set by where GPUs are and who is bidding for them, and the location of the GPUs is set by where a substation can be energised, which is a quantity constraint, not a price. The Dominion spread is evidence of that constraint; it is not a term in the compute price. A structural link does exist and should be named precisely: sustained Dominion scarcity slows the Virginia build-out, which tightens compute supply, which supports the compute price. That is a positive, slow, quantity-mediated relationship between the power forward curve and the compute level, and it is invisible to a model that looks at one- and two-day returns.

The Compute Desk index is not public, so the compute side of these tables uses the SemiAnalysis H100 spot-contract composite (quarterly to 2Q25, monthly since July 2025) and the Silicon Data neocloud prints its blog discloses [15][16]. The power and gas side is real and daily: PJM day-ahead LMPs for Dominion Hub, the DOM zone and Western Hub, ERCOT day-ahead settlement point prices for the West and North hubs, and Henry Hub spot [12][13][14]. Sample sizes are printed because they are the finding.

Three features of the power-side tables matter for the compute question. Dominion Hub, the DOM zone and Western Hub are one risk at daily horizon (ρ 0.88–0.91, and the hub’s own peak block is 0.96 against its flat), so any compute–PJM offset is a single number regardless of which PJM contract is used. PJM and ERCOT are nearly independent at daily horizon (0.07) and only loosely related monthly (0.5–0.6), so a Dominion-versus-ERCOT West spread position is itself a diversified pair, and the ES engine will find real credit there on genuine histories. And gas-to-power correlation lives at monthly horizon and disappears in daily day-ahead spot — the futures curves co-move more than the spot indices do, but the margin engine sees whichever series Nodal feeds it, and for prompt power that is dominated by weather.

The extension runs through a chain: gas sets the marginal cost of the combined-cycle unit that sets the PJM or ERCOT price in most hours; power sets a few percent of the cost of a GPU-hour; the GPU-hour is priced on capex, utilisation and the supply of accelerators. Each link's correlation multiplies. Henry Hub to Dominion Hub is 0.68 monthly on the 2024–26 window and 0.76 to ERCOT West; Dominion Hub to compute is a number whose sign the sample cannot fix; the product is zero. The direct measurement agrees: compute against Henry Hub is −0.12 monthly on eight observations and +0.41 quarterly on nine, which is the same statement as "no information" made twice.

At the level of the broad curve there is a link, and it is worth stating because it is the kind of thing a stress scenario could encode. The AI build-out raises gas demand — new combined-cycle plants in Virginia, behind-the-meter turbine fleets at the Texas sites, the general repricing of firm capacity — and if the build-out slowed, forward gas demand, forward power in Dominion and the compute price would all soften together. That is a common factor with a positive sign, operating over quarters and years, and it produces exactly the kind of "changes in correlations" that Nodal's sensitivity analysis is designed to probe [4]. It does not produce co-movement inside a one-day or two-day return, which is the only window the initial-margin calculation observes.

So the honest characterisation is orthogonality, and orthogonality is itself an offset in this methodology. A book that is long compute and long Henry Hub in equal risk margins at roughly 65% of the sum of its standalone requirements at ρ = 0, the same as the compute–power pair, and slightly less than the Gaussian √2 rule would predict because the compute index's zero-move days and the gas spot's fat tails rarely land on the same date. The gas leg earns the credit not because it hedges compute but because it does not lose money on the days compute does.

The simulation follows the disclosed method rather than approximating it: historical scenarios over the available window, every scenario mirrored antithetically, expected shortfall as the mean of the top 1% of losses, one-day and overlapping two-day holding periods. The power legs use the real Dominion Hub (Oct 2024 – Sep 2026) and ERCOT West Hub (Jan 2023 – Sep 2026) day-ahead histories, expressed two ways: a 21-day rolling-average proxy for a monthly contract in or near delivery, and the raw daily spot as an upper bound that no listed contract reaches. The compute leg is simulated with the properties Silicon Data discloses for its neocloud index — 18.4% annualised volatility, roughly a fifth of days unchanged — with Student-t tails, and a hyperscaler-tier variant at 3.8% with 70% unchanged days [16]. Correlation is imposed through a common factor so that the credit can be read as a function of ρ, and the relative size of the two legs is set in standalone-ES terms.

The shape of the table is the answer to "how that offset might look across contracts." Against any PJM contract the compute leg is orthogonal and the credit is the ρ = 0 row; the same is true against ERCOT West and against Henry Hub, and the numbers are the same to within simulation noise because the credit depends on the tail geometry of the legs, not on which power series is used. The only power-side pairs that earn correlation credit on real histories are power-against-power: Dominion Hub against Western Hub, the DOM zone against the hub, and — at monthly horizon, less so daily — PJM against Henry Hub. A compute contract dropped into a book of those positions is margined almost exactly as it would be alone, minus a diversification term whose size is set by how large the compute leg is relative to the rest.

Nodal will list a family of contracts — Hopper and Blackwell, several chips, a strip of months on each — and the question of how those contracts margin against one another is separate from the power question and, for a trading book, more consequential. The answer depends on which of three relationships is in play: the same index across months, different chips of the same or different generations, and the same chip on different index constructions. The three behave differently enough that they belong in three rows of a margin policy rather than one.

The evidence base is short but it is real and daily. Ornn publishes three months of settled, transaction-based daily indexes for five GPUs without a key [17]; the gpurentalprices.com open dataset records posted on-demand rates from 22 providers every day from July [18]; and Silicon Data's posts disclose how its posted-rate H100 tiers and its B200 index behaved through the March 2026 repricing [16]. None of this is the Compute Desk history — that caveat stands throughout — but the three constructions bracket what a blended benchmark built from private transactions can look like, and they agree on the shape.

Two things stand out. The common factor is weak at daily horizon — 0.15 to 0.20 between the three chips Nodal will list, with a first principal component of 35% across four GPUs — and it strengthens with horizon, to 0.33–0.53 weekly, which is the signature of a level factor arriving through sticky, provider-by-provider repricing rather than a single tape. And the grade ratios are not stable: B200/H100 ran from 1.50 to 2.90 and closed the quarter at 2.13, H200/H100 from 1.25 to 2.04, with annualised ratio volatility of 75–92% — as large as the outrights themselves. In a mature commodity the quality spread is the low-risk leg; here it is not yet.

A contract that settles on the monthly average of a daily index has an exposure to that index that declines through its delivery month: once d of D publication days have printed, only (D − d)/D of the settlement is still open. The next month is fully exposed. A long-front/short-next calendar spread therefore carries a net index exposure of −d/D — nothing on the first day of delivery, half an outright by mid-month, a full outright on the last day — plus whatever the slope between the two months does on its own. Under historical simulation with the two legs driven by one index history, the engine sees exactly that: a spread charge that starts near zero and grows to an outright as the front month averages out. What it cannot see is the term factor, because a curve that has traded for a quarter has no scenario history of its own. A deflation slope of the size the early strips imply, moving by a few percent a day, is a real risk on a spread the engine would otherwise margin at nearly nothing in the first week of delivery.

So the ranking of intra-compute offsets, in the order the engine would grant them: calendar spreads on one index first and by a wide margin; same-generation chip pairs next, with a credit only modestly better than orthogonal legs would earn; cross-generation and tier pairs as pure diversification; cross-index pairs not at all. Which construction the Compute Desk indexes most resemble decides the absolute level of every number — a transaction-weighted index produces one-day ES near 10–13% on the Ornn record, against 6–17% on the posted-rate medians and 4% on an index simulated to Silicon Data’s disclosed 18.4% volatility — but not the ranking.

A history. Nodal Clear's scenario window is one to three years of recent returns plus stress periods [3]. The Compute Desk indexes have a public history measured in months, and the question of what the CRO does about the rest is the whole offset question. Three answers are conventional and each produces a different offset: truncate the joint window to the overlap (a few hundred scenarios, thin tails, the liquidity component doing the real work); backfill the compute leg with a proxy (Silicon Data or SemiAnalysis history scaled to the Compute Desk level, which imports the proxy's correlation, near zero); or backfill with zero returns through the stress periods, which is the stress-block case above and removes any correlation credit while leaving diversification intact. Expect the first or third in year one.

A curve. Every month of a new strip will be margined off the same index history until the months have settlement histories of their own, which means the engine will treat adjacent months as one risk and calendar spreads as nearly free. That is the right answer for the level factor and the wrong answer for the slope; the disclosed sensitivity analysis is where a term-structure add-on would be justified, and a CRO who does not impose one is under-margining the one trade the early participants are most likely to put on.

A holding period. The customer horizon is one day and the house horizon at least two. Compute's two-day ES is about 40% above its one-day figure (6.0% against 4.4%) because unchanged days do not compound; the power proxies nearly double (21% to 40% for Dominion Hub). The relative size of the legs therefore shifts toward power in house accounts, which moves a data-center book toward the middle of the table where the diversification credit is largest.

An account. Customer initial margin is calculated gross by customer under Rule 3.20.1 [5]. The offset is available only when the compute and power positions sit in the same customer's account at the same FCM at Nodal Clear. A data center that hedges compute at CME under SPAN 2 [7] and power at Nodal earns nothing across the two; a hyperscaler that clears through two FCMs earns nothing across them either. The commercial content of Cusenza's sentence is that the two products sit in one risk engine, and that content is realised only by consolidating both hedges at one FCM's Nodal origin.

A liquidity charge that does not offset. The liquidation-cost component scales with the participant's share of contract open interest [4]. In the first quarters of a listing the early participants are by definition large shares of a small open interest; the component is additive and position-specific and no diversification credit reaches it. For the first natural hedgers the liquidity component is likely to exceed any portfolio credit the price-risk component grants.

A model risk owner. Daily portfolio backtesting, sensitivity analysis over "changes in correlations between contract prices," and the CRO's authority under Rule 3.20.2 to add margin in unstable conditions mean that any correlation the histories do produce is provisional. Given that the plausible structural link between compute and power is a common AI-demand factor whose sign could flip between a build-out and a bust, a risk committee that saw a positive correlation credit emerge in the scenario set would be right to haircut it, and Nodal's disclosed sensitivity practice suggests it would.

Power. PJM day-ahead hourly LMPs for Dominion Hub (pnode 35010337), the DOM zone (34964545) and Western Hub (51288), exported from PJM Data Miner 2 for 1 October 2024 to 7 September 2026 (PJM archives hourly data older than two years under different query rules; the archived Jan–Sep 2024 window is excluded) [12]. ERCOT day-ahead settlement point prices for HB_WEST, HB_NORTH, HB_HOUSTON, HB_HUBAVG and LZ_WEST from the ERCOT "Historical DAM Load Zone and Hub Prices" annual files, 1 January 2023 to 5 September 2026 [13]. Daily series are flat (24-hour) averages; peak (HE 7–22, weekdays) averages behave identically for the purposes here. Henry Hub daily spot from the EIA series to 1 September 2026 [14]; gas-implied power is Henry Hub × 7.0 MMBtu/MWh + $4/MWh.

Compute. Ornn OCPI daily settles for H100 SXM, H200, B200, A100 SXM4 and RTX 5090, 8 June – 7 September 2026, from the public three-month history [17]. Posted on-demand rates from the gpurentalprices.com open dataset (CC BY 4.0), 5 July – 7 September 2026, reduced to the daily provider median per GPU [18]. SemiAnalysis H100 spot-contract composite as published on its index page (quarterly 2H23–2Q25, monthly July 2025 – March 2026) [15]. Silicon Data SDH100RT neocloud and hyperscaler values, B200 values and volatility statistics as disclosed in Silicon Data's own posts and its live index page ($2.53 on 7 September 2026) [16]. The Compute Desk indexes are not public and none of this analysis uses them; the Compute Desk contract size is not yet published and the worked book assumes a one-GPU-month (730 GPU-hour) unit, matching CME's description of its contract as one month of rental cost [6].

Margin engine. Historical simulation with antithetic scenarios; ES at 99% as the mean of the top 1% of scenario losses; one-day and overlapping two-day horizons; 2,662 scenarios for the ERCOT-based runs and 1,384 for the PJM-based runs. The monthly-contract proxy is the log change of a 21-day rolling mean of daily prices (daily volatility 4.4% for Dominion Hub, 6.3% for ERCOT West); the raw day-ahead daily series (26% and 79%) is reported as an upper bound only. Neither is a Nodal settlement history, and a deferred-month contract is materially less volatile than either, which pushes a real data-center book further toward the compute-dominated corner of the offset table. The compute leg is simulated: Student-t(4) innovations scaled to 18.4% annualised volatility with 20% zero-move days, and a 3.8%/70% hyperscaler variant; five replications averaged per cell. Correlation is imposed via a common factor on standardised returns, so ρ is the target, not the realised, correlation. Compute-pair ES ratios on the Ornn record use 182 antithetic scenarios and are reported as indicative.

Cost share. H100 1.66 kWh per GPU-hour (DGX H100 10.2 kW ÷ 8 × PUE 1.30); B200 2.23 (DGX B200 14.3 kW ÷ 8 × 1.25); GB200 NVL72 2.17 (130 kW ÷ 72 × 1.20) [19]. Capacity converted to a flat-load $/MWh equivalent by dividing $/MW-day by 24; a data center's actual capacity cost per MWh depends on its peak-load contribution and can be higher.

Caveats. Eight monthly and nine quarterly compute observations cannot establish a correlation; they can only fail to reject zero, and they do. The proxies for both legs bias the standalone power ES upward and leave the compute ES dependent on disclosed summary statistics rather than prints. Nodal Clear's actual scenario selection, stress periods, scaling and liquidity parameters are not public beyond the Disclosure Framework's descriptions; the engine here reproduces the described structure, not Nodal's numbers. Nothing here is a margin quote.