Sizing for a Load That Has Not Arrived Yet

Every energy decision rests on a load. When the load already exists, it can be measured, and the argument moves on to technology and cost. When the load has not arrived yet, the most consequential number in the entire study is a forecast, and the study's honesty depends almost entirely on how that forecast is handled.

Section 01The forecast is the weakest input in the model

Energy models present all their inputs in the same typeface. Fuel price, equipment cost, maintenance intervals, and projected load appear as columns of equal weight. They are not of equal weight. Most of the inputs describe things that already exist and can be quoted. The load forecast describes a decision that has not been made yet, by people who may not have made it, about a business that may look different when the time comes.

The scale of that uncertainty is not a matter of opinion. It is visible in the most scrutinized load forecast on the continent. In its 2025 Long-Term Reliability Assessment, released January 29, 2026, the North American Electric Reliability Corporation projected ten-year summer peak demand growth of 224 gigawatts, more than a 69 percent increase over the forecast published in the prior year's assessment for substantially the same horizon. Winter peak growth was projected at 246 gigawatts. NERC attributes most of the increase to new data centers and the digital economy.¹

The second figure makes the same point from the other direction. The Department of Energy's December 2024 report, prepared by Lawrence Berkeley National Laboratory, put data center consumption at roughly 4.4 percent of U.S. electricity in 2023, or about 176 terawatt-hours, and projected a 2028 range of approximately 6.7 to 12 percent, or 325 to 580 terawatt-hours.² The authors did not publish a point estimate. They published a spread of roughly 255 terawatt-hours because that spread is the honest answer.

Institutions with far more data than any single site, and no commercial interest in the outcome, decline to give a single number four years out. An owner who has been handed one for their own site, without a range and without named drivers, is holding a weaker document than they think.

This is not an argument for paralysis. Capital decisions get made under uncertainty every day in every industry. It is an argument for treating the forecast as what it is: the input most likely to be wrong, and therefore the input the entire design should be stress-tested against first.

A forecast is not a fact about the future. It is a statement about what someone intends today, and intentions are revised.

Section 02Requested capacity and arriving capacity are different quantities

The gap between what is asked for and what gets built is now visible in utility disclosures. Pacific Gas and Electric Company reported in its second-quarter 2026 results, released July 23, 2026, an overall data center pipeline in its service area of more than 12 gigawatts.³ Three months earlier, in first-quarter results released April 23, 2026, the company reported approximately 4.6 gigawatts of customer data center projects in final engineering.⁴

Those two figures come from different quarters and describe different stages, so they do not produce a clean attrition rate and should not be presented as one. What they do establish is that the utility itself distinguishes between a pipeline and a project in final engineering, and that the two numbers are of different magnitudes. Any capacity figure quoted in a study should carry the same distinction: which stage is this, and who has committed what.

The practical lesson for an individual owner is narrower and more useful. If the aggregate pipeline in your territory is substantially larger than the volume of work that reaches engineering, then two things follow at once. Your own request is competing against a queue that is partly speculative, which argues for moving early. And your own forecast is subject to the same discount that everyone else's is, which argues against sizing to the optimistic case.

Section 03Reserved capacity is no longer free

For most of the last half century, asking a utility for more capacity than you ended up using was close to costless. That is changing, and the change is being written into tariffs.

The clearest published example is the data center tariff approved by the Public Utilities Commission of Ohio in Case No. 24-508-EL-ATA. The commission adopted a stipulation on July 9, 2025, and the compliance tariff took effect July 23, 2025. Its terms are worth reading closely by anyone in any state who is about to request capacity for a load that does not yet exist. Under the published schedule, facilities requesting 25,000 kilowatts or more enter a mandatory load study process, with load study fees stated at $10,000 to $100,000 depending on capacity requested. A load ramp period of up to four years applies, with graduated minimum capacity obligations of 50 percent in year one, 65 percent in year two, 80 percent in year three, and 90 percent in year four. After the ramp, monthly billing demand is set at no less than the greater of 85 percent of the customer's highest previously established monthly billing demand during the past eleven months or tiered percentages based on contract capacity. The initial term equals the ramp period plus eight years, up to twelve years in total. A customer may exit after the fifth year following the ramp by paying minimum charges for thirty-six months after notice of termination. Customers without an A-minus or A3 credit rating and substantial cash reserves must post a guarantee or collateral equal to 50 percent of the total minimum charges for the full term.⁵

Read that as a price list rather than a rule book. It is one utility's published answer to the question this paper asks. The answer is that a request for capacity you might not use will be converted into a multi-year minimum payment obligation, secured by collateral, with a defined cost to walk away.

California is moving on the same axis by a different route. On July 24, 2025, the California Public Utilities Commission approved interim implementation of PG&E's Electric Rule 30 for transmission-level retail service, conditioned on applicants paying for necessary transmission infrastructure work up front, with the terms and cost allocation of any refunds left to a future decision. The commission described the framework as protecting ratepayers from potential risks associated with serving large new loads.⁶ The mechanism differs from Ohio's. The direction of travel is the same: the party requesting capacity carries more of the risk that the capacity is never used.

An owner sizing a service request in 2026 should therefore assume that the number written on the application is closer to a commitment than it used to be, and should size it the way a commitment is sized.

Section 04The two failure modes, priced against each other

Sizing errors are not symmetrical, and the asymmetry runs in different directions at different sites. The purpose of the table below is to force both columns into the same document, because most sizing conversations only argue one of them.

The table is not a recommendation. It is the shape of the trade. Which column dominates depends on one question that a study can actually answer: is the deadline the binding constraint, or is the capital the binding constraint? Sites where the deadline binds should accept some oversizing risk. Sites where capital binds, and where the growth is contingent on a decision that has not been made, should not.

Section 05Part load is where the technology ranking changes

If a site's load will sit well below design capacity for the first several years, part-load behavior stops being a specification footnote and becomes a first-order selection criterion. It is one of the few places where the honest ranking of technologies at partial load differs materially from the ranking at full load.

The U.S. Environmental Protection Agency's Catalog of CHP Technologies provides manufacturer-neutral figures. Gas turbines typically show efficiency decreases on the order of 15 to 25 percent at half load. Spark-ignition gas engines compare favorably, with decreases on the order of 8 to 10 percent at 50 percent load, and compression-ignition engines show still better part-load characteristics.⁸ Fuel cells have good part-load performance; at 50 percent of full load the efficiency of a fuel cell typically declines less than 2 percent relative to its full-load value.⁹

Those numbers do not decide anything by themselves, and a study that stops there is doing the same selective work it was hired to prevent. The honest version of each case:

• Reciprocating engines. Comparatively low capital cost per unit of capacity, short delivery, deep service ecosystem, and a moderate part-load penalty that makes them workable under load-following duty. Against: they are combustion machines, and in stricter California air districts new combustion capacity faces demanding control-technology requirements and, in some districts, offset obligations. Sustained very low-load operation carries maintenance and emissions consequences that belong in the operating model rather than the appendix.
• Gas turbines and microturbines. Compact, with high-grade exhaust heat that rewards a site with genuine, coincident thermal demand. Against: the steepest part-load efficiency penalty of the three, which makes them a poor fit for a machine expected to run at a fraction of design output for years. They remain permitted combustion equipment.
• Fuel cells. The flattest part-load efficiency curve of the three, non-combustion operation with near-zero criteria pollutants that in several California districts translates into materially lighter air permitting, quiet operation, and good pairing with heat recovery. Against: higher capital cost than combustion peers, economics exposed to delivered gas price, and stack replacement as a real lifecycle line item. A flat efficiency curve does not rescue a project whose capital cost was set by capacity that never gets used.
• Solar with storage. Buys no fuel and burns nothing, and the incremental unit is genuinely modular, which suits phased growth. Against: for continuous demand the capacity factor means it rarely stands alone, and land or roof area becomes the binding constraint earlier than most first models expect.
• Storage alone. The most naturally staged asset of the group and the easiest to add later. Against: it shifts energy rather than creating it, so it cannot cover sustained growth in total consumption, and its usefulness depends on the service you actually have.
• Doing nothing yet. Waiting has no part-load penalty and no stranded capital. Against: it consumes the one input that cannot be bought back, which is time, and unpriced delay remains the most common hidden cost in this field.

The point is not that one row wins. It is that a load which will spend years at partial output ranks these rows differently than a load that arrives at full size on day one, and a study that never states which case it is analyzing has not analyzed either.

Section 06Modularity is an option, and options have a price

The standard answer to an uncertain load is modularity: install less now, add units as the load appears. It is usually the right instinct and it is never free.

What modularity buys is real. Capital is staged against demonstrated load rather than forecast load. Turndown improves, because a plant of several units can shut units off rather than run one large machine at low output. Redundancy arrives as a side effect. And the decision to expand gets made later, with better information.

What modularity costs is equally real and is routinely left out of the comparison. Capital cost per unit of capacity is generally higher for several small units than for one large one. Maintenance events multiply with unit count. Footprint grows, which matters where land or roof area is already tight. Each permitted unit is a unit the air district evaluates, and in a constrained district the count itself can become a permitting variable. Balance-of-plant, controls, and interconnection work must be designed once for the full build-out or redone at each phase, and the first of those choices is itself a sizing decision made under the same uncertainty.

The disciplined treatment is to price modularity as what it is: an option on the load forecast. An option is worth buying when the underlying uncertainty is large and the cost of being wrong is high. It is not worth buying when the load is contractually committed and dated, because then the owner is paying an uncertainty premium on a certainty.

Section 07What to fix before capital moves

The following can be worked through with a finance lead and a facilities lead in an afternoon. Every item is a question about the load rather than about equipment, which is the correct order.

1. State the load as stages with dates, not as an endpoint."We will reach X by 2030" is not a sizing input. A dated schedule of what is expected to be running in each year is. The design is then tested against the schedule, not against the endpoint.
2. Separate contracted load from expected load from aspirational load.Three different confidence levels routinely arrive in one spreadsheet cell. Split them, label each, and note who would have to decide what for the second and third tiers to become real.
3. Get the ramp and minimum-demand terms in writing before you size to them.Ramp schedules, ratchets, minimum billing demand, term length, exit charges, and collateral are the actual cost of reserved capacity. A verbal indication of flexibility is not a term.
4. Model year two, not just year ten.Most designs are validated at full build-out. The years that determine whether the project survives its own business case are the early ones, at partial load, paying minimums.
5. Ask what the recommendation would be at half the forecast.If the answer is the same design, the forecast was not load-bearing and the study can proceed. If the answer is a different design, the forecast just became the most important number in the room, and it deserves a range.
6. Price the part-load case explicitly for each candidate.Full-load efficiency comparisons are the wrong test for a load that has not arrived. Ask for efficiency and maintenance treatment at the output the site will actually see in its early years.
7. Write down the trigger that reopens the decision.A named, observable event: a signed lease, a permit issued, an executed interconnection agreement, a production line ordered. Without a trigger, a phased plan quietly becomes a single-phase plan.

Section 08Sizing is a decision about uncertainty

The failure this paper describes is rarely arithmetic. Nobody multiplies wrong. The failure is that a forecast enters the model with the same authority as a quoted equipment price, travels through the analysis unlabeled, and emerges at the other end as a capacity number on a signed application with a twelve-year term behind it.

The correction is not more precision in the forecast. Precision is not available, which is why the institutions with the most data publish ranges. The correction is to build the decision so that it survives being wrong: staged where staging is cheap, committed where the deadline binds, tested at half the forecast, and reopened on a named trigger rather than on a feeling.

Our conviction is that green energy is only sustainable if it is profitable, and profitability under uncertainty comes from designs that do not require the forecast to be right. All the paths, sourced numbers, dated facts, one decision, tested against the load that actually shows up.