AI Data-Center Power Research: A Request Is Not Load
AEP Ohio's data-center funnel narrowed from 30,000 MW of requests to 5,642 MW of new contracts. Work through the investment read-through with AllMind.
Anwaar Malik
In this article
AEP Ohio's February 2026 data-center update provides a useful check on an AI infrastructure thesis: 30,000 megawatts of initial requests became 13,022.7 MW of paid studies, then 5,642 MW of binding contracts under its new tariff. Only about 18.8% of that initial requested capacity reached the last step in the disclosed funnel. That does not mean 81.2% of a utility's revenue disappeared. Requests, contracts and electricity consumption are different evidence.
For utilities and industrials analysts, this distinction determines whether a data-center announcement belongs in a revenue forecast, a later scenario or a list of research leads. AllMind fits the job of carrying those distinctions across utility filings, equipment suppliers and the team's own investment assumptions.
The denominator changes the story
The AEP Ohio update of February 13, 2026 reported contracts signed as of February 12. The table reconstructs the stages the company described; these are historical snapshot figures, not the latest size of its pipeline.
| Stage in the disclosed new-tariff process | Capacity | Our calculation |
|---|---|---|
| Initial service requests | 30,000 MW | Starting denominator |
| Requested and paid for an engineering study | 13,022.7 MW | 43.4% of initial requests |
| Signed binding contracts | 5,642 MW | 43.3% of studied requests; 18.8% of initial requests |
The company separately reported 12,219 MW contracted before the new tariff. Adding that older pool to the 5,642 MW gives 17,861 MW across the two disclosed contract pools. It is incorrect to use 17,861 MW as the end point of the 30,000 MW new-tariff funnel: that would mix a separate earlier population into the conversion rate.
Neither calculation measures the amount already energized. The point is to assign each figure to the question it answers. The 30,000 MW figure describes interest. The study stage reflects a greater commitment. A binding contract adds obligations. Metered consumption is a further step that requires its own evidence.
The gap between stages also does not, by itself, distinguish canceled projects, revised requests, changed timing or developers that did not proceed in that process. Calling the difference “lost demand” would attach a commercial explanation the funnel does not fully establish.
Contract terms can matter more than the headline capacity
A utility can incur infrastructure costs before a customer reaches its intended load. That makes the timing of obligations important to both growth and downside research.
AEP Ohio's Data Center Tariff explanation distinguishes its construction agreement from the electric-service agreement. The former includes reimbursement obligations for cancellation or sufficiently long delay before energization; the service term begins with actual energization. The page also describes a ramp period and a minimum-demand billing formula.
Applying the published capacity formula is more useful than assuming every project pays on 85% of its maximum size. Consider two explicitly illustrative facilities after ramp-up, with no higher historical-demand ratchet and no other adjustment:
| Illustrative contract capacity | Capacity-based calculation from the published schedule | Result |
|---|---|---|
| 50 MW | 15 MW plus 85% of the 25 MW above 25 MW | 36.25 MW |
| 100 MW | 57.5 MW plus the 25 MW above 75 MW | 82.5 MW |
The schedule caps that capacity-based minimum at 85% of total contract capacity; neither example hits the cap. Another provision compares the result with 85% of the highest prior monthly billing demand over the stated eleven-month window. That could change the billed minimum. Actual ramp provisions, tariff applicability and the executed agreement must also be retained.
These calculations measure a billing-demand floor under stated assumptions. They do not measure electricity consumed, generation installed or profit earned. Multiplying 82.5 MW by every hour in a year would produce an energy quantity only under an additional utilization assumption, not a bill or an earnings forecast. A demand charge and a price per megawatt-hour are different model inputs.
For an investor, a stronger contractual floor may reduce one risk while leaving others unresolved: construction execution, the timing of service, cost recovery and customer credit. It can change the downside case without raising near-term metered load.
Follow the project into the holding's economics
The read-through to industrials needs another layer of evidence. A contracted utility connection is not automatically an order for the particular transformer, cooling system or contractor in the portfolio. The equipment mix, project location, scope and procurement date still matter.
A useful infrastructure note would separate the stages explicitly. If the utility has a signed agreement but the equipment supplier has not disclosed an award, the supplier's opportunity remains an inference. If an award exists, the next question becomes execution and margin: when does work convert to revenue, what costs are committed, and which party bears delays?
The same project can therefore support different conclusions for different holdings. A protected utility investment, a supplier reservation and an unpriced construction opportunity have different economics. Aggregating them into an “AI beneficiaries” basket can conceal those differences. Our semiconductor customer analysis shows how a related mistake arises when one transaction is viewed through several companies.
Recent filings can change the perimeter again. In its September 2026 Form 10-Q, AEP described approximately 49 GW of Texas incremental-load letters of agreement by 2031 and separately described security supporting 45 GW of prospective load in an interconnection process through 2032. Those figures differ by stage and horizon. Subtracting them to declare a 4 GW “secured shortfall” would invent a common population that the two descriptions do not establish.
How AllMind fits an infrastructure coverage process
AllMind Grids runs a repeated research question across companies with cited answers. For this work, the useful fields are the capacity measure, project stage, service territory, horizon, financial obligation and unresolved dependency. A bare gigawatt total throws away the information that makes the comparison investable.
We recommend AllMind when a utilities or industrials team needs to connect that coverage view to its permitted internal model assumptions and research documents. AllMind Data Rooms keeps the selected packet in scope, so a follow-up can ask whether the newly disclosed commitment supports the timing already assumed in the thesis. The finished artifact is a coverage update identifying which forecast assumption has gained evidence and which still depends on an award, an approval or energization.
For a recurring review, the team can use Agent Studio to schedule its research instructions and update delivery. The schedule should preserve missing or unchanged evidence as an explicit result; a fresh summary is not proof that the project advanced. Coverage of a particular regulatory docket or private project document needs to be demonstrated for the intended setup.
This remains document-based investment research. We do not infer live grid telemetry or validate an engineering design from a research answer. The AEP funnel and tariff applications here are our public-source calculations, not a captured AllMind product run.
Sources and interpretation
Sources were checked November 10, 2026. The Ohio funnel is a February snapshot; the tariff page supplies the stated billing provisions; the later AEP filing illustrates the separate Texas perimeter. The two facility sizes are illustrative inputs, not actual customer contracts. None of the calculations establishes a current target price, a probability that a project will proceed, or a universal conversion rate for other utilities.