◆
Kinetic Alpha
Compute × Power Trading

Compute × Power Trading Dashboard

D. Kaufman | Built June 2026 | Power (electricity) options & trading strategies pertaining to compute. Yellow-background inputs are indicative placeholders — re-mark to live ICE/Nodal/Silicon Data/Ornn quotes. Backtesting modules deferred pending data access.
Market State
Compute Factors
Peak / Off-Peak / 24x7
Transmission & FTR
Options Lab
Trade Builder
Margin (ICE / Nodal)
NG Seasonality (UNG)
Index Dispersion
Compute Heat Map
Peak/Off-Peak Convergence
Take-or-Pay vs Spot
Data & Method
$136.53/MWh
PJM wholesale avg Q1-26 (+76% y/y; 63% data-center attributed)PRINT
$329.17/MW-d
PJM capacity 26/27 (FERC cap; was $28.92 in 24/25)PRINT
$3.26/MMBtu
Henry Hub — fell ~9% post-Hormuz (LNG capacity maxed)PRINT
5.48
SDB200RT Mar-30 (CME settlement index family)PRINT
$95 / $108
PJM-W RT Peak Jul26 / Aug26 fwdINDICATIVE
16.8 GW
NoVa DC pipeline vs 4.0 GW inventory (4.2x tightness)PRINT

GPU Rental Indices — Q1 2026 (Silicon Data published series)

B200 +24% in March while both H100 indices barely moved: hardware grades decouple (low implied correlation). H100 Hyperscaler holds a 3.0x premium over H100 Neocloud for identical silicon — the SLA/firmness tier is the largest differential in the complex. CV: B200 11.4% vs H100-Neo 2.6% vs H100-Hyper 0.5% (the maturity-vol curve).

Why power is the compute trade today

ChannelMechanismInstrument
Load growthDC pipeline = power demand 12–36mo fwd; queues are publicZone basis, deferred cals
DecouplingPower vol now load-driven, not gas-driven (HH fell, PJM +76%)Power vol structures (P1–P3)
SparkEnergy is 4–6% of GPU-hr price; availability binds, not priceCompute fut vs power fut (B3, at listing)
SettlementICE/Ornn compute futures settle Asian-style like powerCross-index basis (B1)

Futures status

CME × Silicon Data (ann. May 12-26): cash-settled vs quote-based daily indices; specs unpublished — watch CFTC filing. ICE × Ornn (May-26): cash-settled vs transaction-VWAP OCPI, Asian settlement, regional weighting; H100/H200/B200/RTX-5090 series. Neither has a firm launch date.

Conversion Stack: Chips → MW

Anchors: GB200 NVL72 rack = 120 kW / 72 GPUs (NVIDIA spec). Grid MW = GPUs × all-in kW × PUE / 1000. MWh per 1,000 GPU-hr = all-in kW × PUE — this is the B3 hedge ratio.

Compute Spark Spread

Regional Pipeline → Power Demand (the basis grid)

RegionGrid/ZoneTightnessPower instrument
N. VirginiaPJM / DOM4.2xDOM-West Hub basis, FTRs, capacity
Texas (DFW+Austin)ERCOT North1.2xERCOT-N fwds, N-Houston spread
Central OhioPJM / AEP3.0xAEP-zone basis, deferred cals
PhoenixWECC / PV2.4xPalo Verde fwds
AtlantaSOCO (non-RTO)2.1xmonitor only (bilateral)

Lease-rate term structure (C1 solver)

Peak / Off-Peak / ATC Decision Tool

Logic: peak window = 5x16 (HE7–22 Mon–Fri, ~47% of hours). Flat high-LF load splits MWh ~46/54 peak/off-peak → ATC (7x24) is the clean hedge where listed; liquidity usually concentrates in on-peak. Inference share ↑ → daytime-weighted → overweight on-peak. Curtailable training is a physical short option vs peak scarcity — hedge less peak, monetize curtailment.

Load Shape: DC vs System

Stylized PJM summer day. Data-center load (~flat) lifts off-peak LMPs relatively more than peak over time, compressing the peak/off-peak ratio — but scarcity convexity means the marginal peak MWh prices on a much steeper supply curve. Trade implication: long off-peak/ATC in compute zones for the structural lift; keep peak upside via options rather than linear length.

Transmission Constraints → Basis (compute corridors)

CorridorConstraint storyIndicative basis $/MWh
DOM zone vs PJM-W HubNoVa import limits; 16.8 GW queue vs lagging transmission build+4 to +12 IND
AEP zone vs PJM-W HubColumbus DC growth; AEP-DOM transfer interface+1 to +5 IND
ERCOT North vs HoustonDFW load growth vs coastal generation; N-H interface+2 to +8 IND
ERCOT West vs NorthWest gen-rich (wind/solar) feeding DFW; congestion rents-5 to -15 IND
FTRs (PJM) / CRRs (ERCOT) monetize expected congestion directly: an FTR from hub to load zone pays the hourly DA LMP difference. Compute signal: interconnection queue + substation filings → future congestion BEFORE the FTR auction reprices. Margining: FTR positions are collateralized at auction (mark-to-auction + credit add-on), not SPAN-margined — capital treatment differs from futures.

FTR / Basis Position Calculator

Queue-to-basis signal model (A1)

1. Scrape queues (PJM, ERCOT) + utility filings (Dominion, Oncor, AEP)
2. Convert announced DC capacity → MW via Conversion Stack
3. Rank zones by tightness = pipeline / inventory
4. Compare implied load growth vs what zone forwards/FTR auctions price
5. Express: buy zone basis / FTRs / deferred cals where gap largest (DOM 4.2x is the standout)
Status: signal layer is buildable now from public data; this tab holds the framework and the sizing math until the scraper feeds it. ~60% of the 35 GW NA pipeline is pre-leased — announcements are dated, regional, and tradeable.

1. Black-76 Pricer (monthlies — building block)

2. Bullet vs Monthlies — Implied Correlation (Trade P1)

Identity: σ²_bullet = w&sub1;²σ&sub1;² + w&sub2;²σ&sub2;² + 2w&sub1;w&sub2;σ&sub1;σ&sub2;ρ, w = F·MWh weights. Basket of monthlies ≥ bullet always (Jensen). Sell bullet / buy monthlies when implied ρ > 0.90. Remember the expiry mismatch: the Aug monthly outlives the bullet by ~5 weeks — the solver is conservative to the buy-monthlies side.

3. Daily vs Monthly (Asian) — Trade P2

Monthly future settles to the average of N daily prices → the monthly option is an Asian option on dailies. Factor = √((N+1)(2N+1)/6N²) → 1/√3 ≈ 0.577 large-N. Sell dailies AS CALL SPREADS only (cap at scarcity adders) vs buy monthly straddle. Same Jensen logic prices the ICE/Ornn Asian settlement vs any point-settled CME window (B1 convexity leg).

4. Vega-neutral package sizer

A1 — Long DOM-zone basis (compute info → power)

Buy DOM zone / sell PJM-W Hub, deferred months; or DOM-sink FTRs at auction
Signal: 4.2x tightness, 16.8 GW queue. Entry: basis below model-implied congestion. Risk: transmission builds faster than expected; DC cancellations (half of 2026 US DC projects delayed/cancelled — size for it).

A2 — Deferred steepener, compute zones

Buy Cal-28/29 DOM/AEP/ERCOT-N vs sell front or low-growth zone equivalents
Thesis: deferred curves under-price confirmed load growth (thin liquidity past 24mo). Proof: Q1-26 PJM +76% was the mechanism realizing.

P1 — Buy monthlies / sell summer bullet (dispersion)

Long Jul+Aug PJM-W straddles (own-strike, vega-weighted) / short Jul-Aug strip bullet
Trigger: implied ρ > 0.90 (Options Lab #2 prints 0.912 on defaults). Keep the post-bullet Aug leg.

P2 — Sell daily call spreads / buy monthly straddle

Short Aug26 daily call spreads / long Aug26 monthly straddle, ~1.7–1.9:1 vega ratio
Trigger: daily/monthly IV ratio > fair (Lab #3). Short tail by construction — spreads not naked, halve size on reserve-margin warnings.

P3 — Prompt vs shoulder vol calendar

Buy Jul26 straddle / sell Oct26 straddle, vega-weighted (~2.5:1 Oct:Jul lots)
Extreme Samuelson in power: vol concentrates in the delivery-proximate weather window; shoulder months over-marked by smooth surface fits.

B3 — Compute spark spread (at listing)

Long compute futures / short host-hub power futures at ~2.1 MWh per 1,000 GPU-hr (Conversion Stack)
Structural one-way hedger flow: DC operators/neoclouds short compute to lock revenue (ICE's stated use case) — the trader takes the other side at a premium, jet-crack style.

Scenario tester

Stylized scenario P&L on default sizing — a design tool for trade construction, not a backtest. Wire to historical ISO data later.

Contract-Level IM (indicative) RE-MARK VIA FCM

Contract (1 MW monthly)Nodal IM $ICE IM $~% notional
PJM-W RT Peak future5,0005,50015%
ERCOT North RT future7,5008,00020–25%
PJM-W monthly option (short ATM)6,5007,1501.3x fut
Daily option strip (short, per MW-mo)8,0009,0001.6x fut
Long optionspremium paid, no IM—
FTR (PJM)auction collateral + credit add-on—

Portfolio offset assumptions (sliders)

Nodal Clear margins the whole portfolio on expected shortfall (options included since Jul-2023; cross-commodity power-gas recognized). ICE applies inter-month/intra-hub spread credits. Cross-CCP offset = 0 always. These offsets are indicative calibration knobs — request actual IM runs from your FCM / venue margin calculators before sizing.

Strategy Margin Calculator

Gross = Σ short-leg IM + futures IM (long options consume premium, not IM). Net = gross × (1 − offset). The cross-CCP row margins gross at BOTH houses — consolidate venues where possible; the margin delta often exceeds dealer price improvement. Hold cash ≥ 2x net IM: vol-spike margin calls arrive the same morning as the P&L event.

Trade Setup — UNG vs NG Jan27 Options

Scenario

Sizing is delta-equivalent at entry: 1 NG option (10,000 MMBtu) vs N UNG options (100 sh). UNG path simulated through 6 monthly rolls with 1.17% expense drag; entry curve shape 3.26/3.30/3.33/3.40/3.65/3.95/4.10 (indicative — re-mark). The wedge: UNG strikes at spot; the NG leg strikes at the winter premium, $0.84 'ahead'. Pricing honesty: roll drag is NOT in UNG option prices (risk-neutral drift = r) — this is a position on the curve's seasonal risk premium, not an arbitrage.

Terminal Payoff vs Front Price (both curve modes)

P&L Through the Life of the Trade (selected scenario)

Path matters: Variant A's trough is weeks 8–16 (theta paid before the winter window); Variant B's mark-to-market low is early December (UNG drag accrued, Jan put time value not yet intrinsic). Size to survive the trough, not the endpoint. Margin: UNG legs at OCC, NG legs at CME SPAN — NO cross-margin between them.
$1.80–1.90
OCPI-H100 SXM (Ornn live, Jun-26)SCRAPED
$2.43–2.63
SDH100RT neocloud (Silicon Data, Mar-26)SCRAPED
~$0.55–0.75
Raw SD−OCPI H100 gap (~part is Mar-vs-Jun timing)DERIVED
7.43–7.52
SD H100 Hyperscaler (Mar-26) — 3.0x neoSCRAPED

Factor Decomposition — SD − OCPI = Σ factors

Contribution Waterfall

Factors C (non-NVIDIA processor scope), H (config normalization), J (thin-SKU breadth) are held at zero by default — watch/noise terms, not tradeable today. The two structural drivers are D (tier) and B (SKU): both become knowable the moment CME publishes which sub-index and SKU basket it settles on. Residual = observed − modeled is your missing factor or your edge. Full 10-factor detail and signs in the memo §2 and the workbook Dispersion_Factors tab.

Physical Compute × Power — Lease Clusters, Power State, Trading Nodes

WEST TEXAS MIDWEST MID-ATLANTIC SOUTHEAST Northern Virginia — ~4,040 MW (+16.8 GW pipeline) — PJM/DOM — largest US load shock NoVa Dallas–Fort Worth — ~2,900 MW (+3.1 GW) — ERCOT North — scarcity-priced DFW Columbus OH — ~1,600 MW (+2.4 GW) — PJM/AEP — cheap, congesting CMH Atlanta — ~1,280 MW (+1,892 u/c) — Southern Co (non-RTO) — bilateral ATL Chicago — ~692 MW (+244) — PJM/ComEd — liquid CHI Phoenix — ~600 MW (+1.5 GW) — WECC/Palo Verde — heat-constrained PHX Santa Clara / Silicon Valley — ~700 MW — CAISO/NP15 — expensive, capped SCL PJM Western Hub — benchmark trading node PJM DOM zone — NoVa pricing node PJM AEP zone — Columbus pricing node PJM ComEd — Chicago node ERCOT North Hub — DFW node ERCOT Houston Hub Palo Verde — WECC node CAISO NP15 — N. California node HOU
Bubble size = operational MW (the OCPI regional-weight vector). Color = host-grid power state: red largest load shock, orange scarcity/constrained, amber cheap-but-congesting, green liquid, grey bilateral/opaque. Blue diamonds = the actual power trading nodes that price each cluster. Hover any marker for detail. Positions are projected from lat/lon (schematic, not a survey map).
The visual is the trade: a power shock in a heavily-weighted cluster (NoVa, DFW) moves OCPI (regionally weighted) while SD (composite) lags, and the SAME shock prices that cluster's blue-diamond power node. One event → OCPI basis + power-node basis + operator spark, all correlated. MW from JLL/CBRE/C&W Q1-26.

Flat DC Load → Peak/Off-Peak Price Convergence

Is it priced in?

Spread vs DC Block (the threshold)

Flat load lifts peak & off-peak equally, so the absolute DEMAND spread is unchanged — but the quantity RATIO compresses, and on a 3-segment supply stack (flat → steep → scarcity cap) the PRICE spread first WIDENS (peak climbs the steep mid while off-peak stays on baseload) then CONVERGES once peak hits the cap and off-peak climbs the steep region. The turning point is the threshold. Whether a zone is left or right of it is empirical: DOM and ERCOT-North are at/near it; lighter zones are still in the widening regime. The model is stylized — calibrate the stack to the zone's actual heat-rate/scarcity curve.

Take-or-Pay vs Spot — Capacity Decision Optimizer

Cancellation clauses are puts on power rates; expansion/extension rights are calls; take-or-pay floors with usage flexibility are swing options (gas-trading technology, directly portable). Systematically given away free in current capacity negotiations because neither side prices them. A desk with the option toolkit can extract real value structuring around counterparty optionality.

Term Cost Path

Blue line = ToP cumulative fixed cost net of option. Green = spot expected cumulative cost. Shaded band = spot ±1σ. Where the bands overlap, ToP is "within band" — consider the operational benefits (price certainty, billing simplicity) before sizing.

How This Dashboard Works

TabWhat it doesEngine
Index DispersionDecomposes the SD−OCPI basis into 10 isolated factors that sum to the observed spread; isolates the two structural drivers (tier, SKU)Linear factor model; residual = edge/missing factor
Compute Heat MapMaps lease MW (= OCPI weight vector), host-grid power state, and the actual power trading nodes — the geography of the compute×power tradelat/lon projection; MW-scaled bubbles
Peak/Off-Peak ConvergenceModels whether flat DC load compresses the peak/off-peak price spread, finds the threshold, and tests whether the forward already prices it3-segment supply stack with scarcity cap
Options Lab / NG / Margin(existing) Black-76, bullet-vs-monthly correlation, Asian factor, UNG seasonality MC, ICE/Nodal marginverified vs Excel workbooks

Data Provenance — Real vs Proxy (fully disclosed)

Data pointStatusValue / sourceIf proxy: why & how assigned
OCPI-H100 SXM levelREAL~$1.80–1.90, Ornn homepage (Jun-26)—
SDH100RT neocloud / Hyperscaler / SDB200RTREAL$2.43–2.63 / $7.43–7.52 / 4.40→6.11→5.48, Silicon Data (Q1-26)—
OCPI hardware coverage; transaction-only; Asian settle; regional weightingREALOrnn / ICE filings (H100/H200/B200/B300/RTX5090)—
SD methodology (80%+ H100 mkt, DRW/Jump, normalize specs/interconnect/geo, Bloomberg SDH100RT)REALSilicon Data product page—
Regional lease MW by clusterREALJLL/CBRE/C&W/Avison Young Q1-26—
DC load shape (90–95% LF, inference 80–90%), PJM/DOM growth 5.4%/yrREALPJM 2026 Load Forecast, EIA, Deloitte—
Live same-date SD−OCPI spread per SKUPROXY~$0.55–0.75 H100 derivedThe spread is not published; computed from two scraped levels at DIFFERENT dates (SD Mar vs OCPI Jun). Part is timing, not methodology — treated as factor A′. Re-mark same-date from Bloomberg/Ornn API.
Factor contribution magnitudes (A–J)PROXYSliders, default sum ≈ observedIndicative; calibrate by regressing realized SD−OCPI on each driver once both daily series are licensed. Signs are defensible; magnitudes are priors.
OCPI regional weight vectorPROXYApprox = operational MW shareOrnn doesn't publish weights; MW share is the best public proxy for where transactions print. Refine with Ornn regional sub-indices.
Convergence stack (knee, slopes, cap)PROXYStylized 3-segmentReal heat-rate/scarcity curves are zone-specific and licensed (ISO). Shape is qualitatively correct; calibrate to the zone's bid stack.
Forward-implied peak/off-peak spreadPROXYInput default $72Placeholder; pull live on-peak vs ATC forwards from ICE/Nodal for the target zone.
Power state per cluster; CME contract SKU/sub-indexPROXYQualitative / TBDCME settlement SKU & tier not yet published (the single biggest unknown for factors B & D) — watch the CFTC filing.
Principle: every number is either a dated, sourced observation or a clearly-labeled proxy with its assignment rationale. Nothing is presented as live market truth that isn't. The ISO feed toolkit (delivered separately) replaces the power-side proxies with real PJM/ERCOT data; licensing Silicon Data + Ornn APIs replaces the compute-side proxies.
Sources: Monitoring Analytics/Bloomberg (PJM Q1-26), PJM BRA, EIA (HH), Silicon Data published Q1 series, ICE/CME announcements May-26, JLL/C&W/Avison Young regional reports, NVIDIA GB200 specs. All forwards, IVs, basis ranges and IM levels marked indicative are placeholders for desk re-marking. Not investment advice.
Disclaimer. This tool is provided by Kinetic Alpha for research and educational purposes only. It is not investment advice, an offer, or a solicitation to buy or sell any security, derivative, or strategy. All forward prices, implied volatilities, basis levels, margin figures, index spreads, and convergence parameters marked indicative are illustrative placeholders or clearly-labeled proxies (see the Data & Method tab) and must be re-marked to live data before any real-world use. Scraped provider levels are dated and may be stale. Markets involve risk including loss of principal. Consult a qualified professional before making any financial decision. © 2026 Kinetic Alpha.