Demand is not a contract.
AI needs electricity. Capital still needs a paying customer. Separate expressed demand, paid investigation and a commitment to absorb the cost of failure.
Contents & sources
Finding a shortage and earning a return from it are different jobs. Infrastructure must be built, customers must arrive and capital must be recovered. A bottleneck owner can also become the owner of expensive idle assets.
Read the unit of the forecast
The IEA’s 2026 outlook projects global data-centre electricity consumption rising from 485 TWh in 2025 to around 950 TWh in 2030. The latter is a projection, not observed consumption. All data centres and AI-focused data centres are different aggregates.
Our inference: a global forecast cannot be copied into one company’s revenue. Location, required equipment and financing determine the economic result. Global growth does not replace a local contract.
When interest becomes a commitment
AEP Ohio’s February 2026 update separated 30,000 MW of initial interest, 13,022.7 MW entering paid studies and 5,642 MW in new tariff contracts. Earlier contracts of 12,219 MW brought the total to 17,861 MW. Contracted capacity is not energised load.
We do not infer that every unsigned project was fictitious. Service dates, collateral and obligations can change a project’s decision. Follow the stages between an expression of interest and capital at risk.
Interest. Investigation. Commitment.
Distinct stages of the new process. Earlier contracts are excluded. This is not energised load.
Original report & definitions ↗Keep the scope and date fixed
In its 30 July 2026 release, AEP reported 69 GW of contracted load growth through 2030 across its system. It added six GW of signed load agreements in the quarter, mainly in Texas. Customers included data centres and industrials. This differs in scope and time horizon from the Ohio February process.
Choosing the smaller number does not make an argument more rigorous. February’s 5.642 GW and July’s 69 GW are not before-and-after values for the same population. Align the company, territory, customer types, observation date and target energisation period before testing the commitments and their execution.
Who pays when the project fails?
AEP Ohio’s published process requires paid studies for new facilities or expansions of at least 25 MW. Its LOA requires reimbursement of buildout costs if, before the target energisation date, a customer cancels the project or delays it by more than twelve months. The ESA’s initial term adds eight years to a ramp period of up to four years. Read the published conditions for application.
Our interpretation: allocating failure costs is part of the productive system. A clause turning interest into a payment obligation determines who absorbs overbuilding. Put minimum billing, collateral, cancellation, delay and the supplier’s obligations beside the demand forecast.
From a bottleneck to recovered capital
A strong contract does not determine shareholder returns by itself. Capital expenditure, financing costs, regulated recovery and returns, customer credit and utilisation still matter. Inspect what the market price already assumes.
The countertests are concrete: studied and contracted projects can fail to become load; construction and financing costs can outgrow recoverable cash; new supply can reduce scarcity. Needed by AI is the beginning of an investigation, not an investment conclusion.
Find what is needed. Identify its owner. Then read who has committed to pay.
Primary sources checked
An analysis extending Kamui’s physical-bottleneck thesis with primary sources, checked on 4 October 2026. The queue figures describe the initial process reported by AEP Ohio in February, not today’s total queue or energised load. The company’s account of the tariff’s effects is distinct from our interpretation. This is not a company valuation or a buy/sell conclusion.
Correction, 4 October 2026: the initial Korean edition reversed the reimbursement direction. It now correctly states that the customer reimburses the utility.