Power × Compute · Load shape · 28 July 2026
Utility filings, interconnection studies, and most sell-side notes model data-center demand as a flat block: the same megawatts every hour, hedged with a 7×24 strip. Measured facility load profiles say otherwise — they are not flat at any utilization below saturation. Join those profiles against three years of hourly day-ahead settlements across eight ERCOT and PJM locations and the flat-load assumption acquires a price: the load-weighted cost of power runs up to 9.4% above the flat-block cost, the gap peaks at 40% utilization rather than at full load, and it flips sign by season. Every operator hedged with a flat block is carrying that difference, unpriced.
Load shapes are the measured H100 facility simulations from the denominator piece (NLR, arXiv:2604.07345). Prices are 401,379 hourly DA settlement records, Jul 2023 – Jul 2026: four ERCOT hubs, three PJM hubs, and — pulled for this piece — the PJM DOM zone, the settlement zone of data-center alley. The join is reproducible and the interactive module below runs it live on every combination.
Flat is convenient. It is convenient for the utility, for the interconnection study, and for the hedge desk. It is not what the facility does.
The flat-load convention shows up three ways. Interconnection studies size the grid impact of a new facility at its rated draw, every hour, all year. Utility load forecasts add announced campus megawatts as constant blocks. And the standard power hedge for a large consumer is the 7×24 baseload strip — one price for every hour of the delivery period, which is only the right hedge if consumption is the same in every hour.
The convention persists because, until recently, there was no public measurement to check it against. Data-center operators do not publish load traces. What the NLR dataset changed is that a national laboratory measured real H100 training, fine-tuning, and inference workloads at 0.1-second resolution and scaled them through a year-long facility simulation — producing, for the first time, an hour-by-hour load profile for a 10 MW colocation facility and a 1 MW inference facility at four utilization levels, with the workload physics carried through from the device level.
A forecast error in an interconnection study is somebody else's problem for years. A hedge mismatch settles every month. If a facility consumes more in expensive hours than in cheap ones, a 7×24 block leaves it structurally short the peak and long the trough — it pays the load-weighted price but hedged the flat price. The difference is a shape basis, it compounds every hour, and nothing in the standard procurement process quotes it.
Not flat at 20% utilization. Not flat at 80%. Flattest exactly at saturation — which is the one state a growing facility is never in.
The simulated colocation profile has a pronounced daily cycle: demand builds through the afternoon as jobs arrive, peaks in the evening — the summer maximum lands at hour 22 — and troughs before dawn at hour 6. At 80% target node utilization the summer swing from trough to peak is 23% of the trough. At lower utilizations the relative swing is larger still, because idle capacity leaves more room for the arrival pattern to move the total. The inference facility cycles with request rate: heavier through the working day, lighter overnight.
Summer load shape by hour — 10 MW colocation facility, four utilizations
Percent of rated IT load. The shape flattens as the facility fills — the arrival pattern gets absorbed into an ever-fuller schedule — but it never reaches flat, and no real facility operates at the saturated limit.
Two properties of this shape do the work in everything that follows. First, it is evening-loaded: the daily maximum sits in the late-evening hours, after the afternoon job-submission wave has stacked the queue. Second, it is utilization-dependent: the fuller the facility, the flatter the profile. Both are physics of scheduling, not artifacts — a saturated facility has no headroom for the arrival pattern to express itself.
A shape basis is not a property of the load. It is a property of the join between two shapes — the facility's and the grid's. The module runs the join live.
The construction is one line. Take the facility's mean load by season and hour, L(s,h), and the settlement price by season and hour, P(s,h), over the same three years:
The shape-basis engine
Pick a location, facility, and utilization. Bars are the three-year mean DA price by hour; the line is the facility load shape. The readouts run the join across all four seasons.
Positive at every location, every facility type, every utilization. Largest where the grid is scarcest in the evening — and largest, everywhere, at the utilizations a facility ramps through rather than the one it is built for.
Shape premium by location and utilization
Load-weighted vs flat-block DA price, joined across all four seasons. Toggle the facility type — the inference shape runs hotter because daytime serving load overlaps the afternoon price ramp more fully than the evening-peaking colocation shape.
The premium peaks at 40% utilization and falls toward saturation. ERCOT North, colocation: +7.9% at 20% utilization, +9.2% at 40%, +7.3% at 60%, +4.5% at 80%. The ordering is identical at all eight locations. A full facility runs the flattest profile it will ever run; an empty-to-half-full one runs the peakiest. Which means the flat-load assumption is most wrong for facilities in their ramp — the months and years between energization and steady state — and that is precisely the period the interconnection studies and the first hedges are written for. The industry's flattest-load model is applied hardest to the facility state that is least flat.
In dollars, at the 80% colocation case on ERCOT North: flat-block $38.55, load-weighted $40.28 — $1.73/MWh of shape basis. On the facility's own simulated consumption of roughly 51,744 MWh a year, that is about $90k a year per 10 MW facility, before PUE. Scale to a 100 MW campus and the 7×24 hedge is leaving roughly $0.9M a year of shape exposure unpriced — at the flattest operating point. At 40% utilization the same campus carries roughly double that. These are not enormous numbers against a compute P&L; they are enormous numbers against the fee a retailer would charge to shape the hedge properly, which is the actual comparison a procurement desk faces.
The annual number hides the mechanism. ERCOT North colocation at 80% decomposes to +3.6% in summer and −1.3% in winter (negative sign: the load-weighted price runs below flat-block). ERCOT's summer scarcity lives in the late afternoon and evening — squarely under the compute shape's heaviest hours. ERCOT's winter risk lives in the early morning — hours the evening-loaded compute shape barely occupies. Same facility, same grid, opposite alignment. A shape basis is a join between two curves, not an attribute of either one, and it can flip sign twice a year.
The DOM zone — Dominion, Northern Virginia, the densest data-center concentration on any grid — was pulled specifically for this piece: 26,829 hourly DA settlements, Jul 2023 through Jul 2026. Two results. First, DOM's flat-block price is $55.06/MWh, the highest of all eight locations — the zone absorbing the most data-center load already trades above every hub in the sample before any shape effect. Second, its shape premium (colocation +3.32% at 40% utilization, inference +5.20%) runs above the neighbouring PJM hubs but well below ERCOT — because DOM is a winter-morning grid. Its single most expensive average hour is 7 AM in winter ($118/MWh, with a 90th percentile of $231), edging even the summer evening peak ($116). The evening-heavy colocation shape largely misses those hours; the day-heavy inference shape catches more of them, which is why inference carries the larger premium at DOM specifically.
Somebody always owns a basis. Today it is whoever signed the flat hedge — mostly without a price on it.
An operator hedged 7×24 owns the shape basis directly: it settles in their monthly true-up as the gap between the load-weighted cost and the block price. A retailer offering a fixed all-in rate owns it instead, and prices it — the shaping fee inside a data-center retail contract is exactly this number plus margin, which makes the tables above a due-diligence tool: an operator who can compute their own shape basis can audit the fee. And a utility planning on flat blocks owns it socially, in the form of peak capacity built for a coincidence pattern nobody measured.
What is new is that the shape leg is becoming tradable on its own. Nodal's 168 hourly power futures (listing 31 Aug 2026) quote every hour of the week separately; ElectronX's bounded-hour contracts and ICE's TB4 blocks slice the same exposure more coarsely. Against those instruments the correct hedge for the measured compute shape stops being a theoretical shaped OTC contract and becomes a portfolio: overweight the evening hours in proportion to the load shape, underweight the trough — with weights that should decline as the facility fills, per the utilization result above. The shape basis this piece quantifies is, precisely, the mispricing of a 7×24 strip against that portfolio.
Everything here treats the load shape as given. It is not quite — the same NLR queue-time tables that produced these shapes also price what it costs to move them: at 20% utilization, shifting compute is nearly free; at 80%, it costs six-plus hours of queue. A facility that can shift load out of the expensive hours is not hedging the shape basis but selling it back. Valuing that option against the hourly curve is a separate exercise — the curtailment piece in this series.
The caveats are load-bearing. Naming them first is the credibility.
"Data centers are a flat load" is not a stylized fact; it is an unpriced trade. The measured shape is evening-loaded and utilization-dependent; joined against three years of settlements it costs up to +9.4% over flat-block at the hubs where compute is actually being built, peaks precisely in the ramp phase every new facility passes through, and flips sign between an ERCOT summer and an ERCOT winter. Every one of those properties is invisible to a 7×24 hedge and visible to an hourly one — which is, in one sentence, why the power market is going hourly at the same moment compute is becoming a commodity.