Why Power Is the Real Constraint

Megawatts, not land or chips, decide which AI data centers get built. How interconnection actually works, what a non-binding power proposal is worth, and the questions that separate a real site from a hopeful one.

If you remember one thing from this site, make it this: in AI data centers, the scarce input is electricity delivered to a specific place at a specific time. Land is available. Buildings can be built. Equipment can be ordered. Utility capacity at the scale and on the schedule these projects need is the bottleneck, and it is the thing most likely to turn a good-looking projection into a late, expensive one.

The demand side, in national numbers

US data centers used 58 TWh of electricity in 2014 and 176 TWh in 2023 — about 4.4% of total US electricity consumption. The Department of Energy's 2024 report, prepared by Lawrence Berkeley National Laboratory, projects 325 to 580 TWh by 2028, equivalent to roughly 6.7% to 12% of national consumption (DOE, December 2024).

A range that wide — nearly a doubling from the low case to the high — is itself informative. The people with the best data do not know how much power this industry will draw in a few years' time. Any offering document that implies certainty about demand five years out is overstating what is knowable.

What is not in doubt is the direction, and that the increase has to be physically delivered: generation, transmission, substations, transformers.

What "10 MW site" can mean

The phrase appears constantly in marketing and covers at least five quite different situations, in ascending order of how real they are:

1. A site where 10 MW might be obtainable — someone looked at a map and a substation is nearby.
2. An interconnection request submitted — the utility has been asked. There is a queue, and being in it is not an allocation.
3. A written proposal from the utility — the utility has described terms on which it could serve the load. Normally non-binding: it can be withdrawn, repriced or re-timed, and it commits nobody.
4. A signed electric service agreement — a contract to deliver, with a date, a tariff and obligations on both sides. This is the first point at which "10 MW site" is a fact about the world.
5. Energised capacity — power is physically at the property.

The distance between 3 and 4 is where a great many retail data center offerings sit, and it is the single most common place where the headline overstates the asset. A non-binding proposal is a genuine milestone and worth something; it is not contracted power.

Why interconnection takes so long

Connecting a large new load is not a matter of running a thicker cable. The utility has to study what the new load does to the local network under normal and fault conditions, identify upgrades — a new substation, a transformer bank, reconductoring, sometimes transmission work — and then schedule and build them.

Several things make this slow in ways no sponsor controls:

• Studies are sequential and queued. Requests are generally studied in order, and a change upstream can require restudy downstream.
• Long-lead equipment. Large power transformers and high-voltage switchgear are built to order with lead times measured in quarters to years, and the order book is full because every data center, utility and renewable project wants the same parts.
• Upgrade cost allocation is negotiated. Who pays for the substation is a commercial discussion, and the answer can move a project's capital cost materially.
• Public process. Rezoning, environmental review and local objection are now routine for large loads, and data centers have become politically contested in several of the busiest markets.

The market effect is visible in CBRE's data: "power capacity constraints are forcing aggressive preleasing and extending new construction timelines to 2027 and beyond," with Northern Virginia preleasing facilities scheduled for delivery "as far out as 2028" (CBRE, Global Data Center Trends 2025).

Why this is good news for owners and bad news for builders

Scarce power is the reason the economics look attractive at all. It is why North American vacancy hit an all-time low of 1.6% and why 74.3% of capacity under construction was preleased (CBRE, H1 2025). If power were easy, supply would arrive, rents would normalise and there would be no unusual return to capture.

So the same constraint that makes owning energised capacity valuable makes getting to energised capacity risky. As an investor you need to know which side of that line your money is on:

• Buying a listed REIT buys mostly existing, energised, leased capacity — the constraint works in your favour, and it is priced in.
• Buying into a development project puts you on the wrong side of the constraint until the power is contracted and delivered. The higher advertised return is payment for that.

Neither is wrong. Confusing them is.

Behind the meter, and the other workarounds

Because the queue is slow, sponsors increasingly propose alternatives, each with its own catch:

• On-site generation (gas turbines, fuel cells) — faster than a transmission upgrade but needs fuel supply, air permits and local consent, and shifts fuel-price risk onto the project.
• Co-location with existing generation — siting next to a power plant. Elegant, and contested: regulators have had to decide how much of the grid's shared cost such arrangements should bear, and the rules are not settled.
• Bridge power — temporary generation until the utility arrives. Expensive per MWh, and a cost the model may not carry.
• Demand flexibility — agreeing to curtail during system peaks in exchange for faster or cheaper connection. Reasonable for some workloads, awkward for training runs that cannot pause.

If an offering relies on one of these, the question is whether its cost is in the model and whether the permits exist.

The questions to ask about power

1. Is there a signed electric service agreement? If not, what exactly exists, and is it binding?
2. What is the contracted energisation date, and what is it conditional on?
3. Who pays for network upgrades, and is that cost in the budget?
4. What is the substation and transmission position — one feed or two, existing or to be built?
5. What happens to the schedule if the utility slips, and who bears the carrying cost while rent is not being earned?
6. Is there bridge or on-site generation in the plan, and is its cost in the model?
7. What is the assumed power price, and who bears an increase? See how data center leases work.

Next: the risks, in rough order of how often they actually bite.