Land is available. Fiber is available. Tax incentives are a spreadsheet exercise. Chips ship. The constraint that decides whether a facility exists in 2027 or exists in a press release is power, and specifically the interval between the decision to build and the first electron hitting the bus. We call that interval the energization gap, and it is now the single largest unpriced risk in most capacity plans.
The energization gap, measured
The gap between announced and real is the defining feature of this market. Independent tracking by Sightline Climate counted roughly 12 GW of announced 2026 US data center capacity across 140 projects; only about 5 GW was actually under construction, and a quarter of announced projects had not disclosed a power strategy at all. Hold that ratio: for every megawatt being built, more than two exist only as intention. Against typical build times of 12 to 18 months for the facility itself, the missing ingredient is not construction capability. It is energization.
The supply response is real and slow, which is the worst combination for a buyer. Global Energy Monitor's plant tracking showed US gas capacity under construction topping 29 GW as of January 2026, more than double the prior year, with over 159 GW in pre-construction. Read those two numbers together honestly: the industry knows exactly what it needs to build, and more than eighty percent of it has not broken ground. Capacity in permitting does not train a model.
Why the queue cannot be bought
The mechanism behind the gap has two layers, and most capacity plans only model the first. The visible layer is the interconnection study queue, which runs years in every major market. The structural layer underneath is electrical equipment: high-voltage transformer lead times that ran 24 to 30 months before 2020 now stretch toward five years, with prices up more than half over the same period, and switchgear is no better. Electrical infrastructure is a single-digit percentage of total facility cost and close to one hundred percent of the schedule risk. This is the defining asymmetry of the era: you can buy your way out of nearly every constraint in this industry except a position in a study queue and a slot in a transformer factory.
Money compresses everything but the clock.
The second-order point separates the teams who get it from the teams about to learn it. Data centers are not primarily competing with each other for power. They are competing with utilities' own resource plans, coal-replacement programs, and grid-reliability projects for the same turbines, the same transformers, and the same EPC labor. The equipment market does not care that your load is strategic. As we detail in The Turbine Gap, the manufacturers are sold out to everyone simultaneously, which means the relevant competition set for your 2028 megawatts includes every utility integrated resource plan filed in the last two years.
The phantom load problem, and the market it will create
Now follow the announcement numbers one step further than the coverage does, because this is where the next asset class comes from. The entire supply side of this market is planning against announced load: OEM production ramps, utility resource plans, transformer orders, EPC hiring, all of it calibrated to a pipeline in which the majority of capacity is an intention with no power strategy. A meaningful fraction of that pipeline will not be built. It never is: Lawrence Berkeley National Laboratory's long-running research on interconnection queues, which peaked near 2.6 terawatts and still hold more than two terawatts today, has found that historically only about one in five queued projects ever reaches commercial operation, and only about 13 percent of queued capacity. A 2.4-to-1 announced-to-constructed ratio is not an anomaly of this cycle. It is how queues have always resolved, applied to the largest queue in history.
Here is the part almost nobody has priced: when an announced project dies, its carcass contains exactly the assets this article says cannot be bought. A queue position with study years already behind it. Transformers on order with 2027 delivery dates. Permits, water rights, substation designs, sometimes equipment deposits. Project mortality is about to become the only liquid market in time to power. Between now and 2029, some of the cheapest speed available in this industry will be acquired not from OEMs or utilities but from the balance sheets of projects that ran out of conviction, capital, or tenant. The best 2029 site may well be a 2026 announcement that fails in 2027, and the buyers who build the capability to underwrite distressed projects for their interconnection positions rather than their real estate will be buying years for cents.
Treat competitor announcements accordingly, in both directions. An announcement reserves supply chains, anchors utility negotiations, and signals talent markets whether or not the megawatts materialize; it is a strategic instrument, not a status report. And every announcement is also a future entry on somebody's distressed-asset watch list, including yours.
Price the delay or the delay prices you
The corrective is to put time to power at the top of the underwriting model, above cost per megawatt, and make the cost of delay an explicit line item. The math is not subtle, so run it. A hundred megawatts of AI load supports on the order of 75,000 accelerators once cooling and overhead are counted. At prevailing market rates for that compute, the facility generates revenue in the range of $15 million to $20 million per megawatt per year. Divide by twelve: every de-energized megawatt forgoes roughly $1.2 million to $1.6 million per month. Now hold that against what generation equipment actually costs in today's market, priced precisely in The Turbine Gap, and the conclusion is one sentence: a single month of energization delay costs approximately what the turbine itself costs. The equipment is not the expensive part of this market. The calendar is.
The demand side confirms it at the macro scale. The four largest hyperscalers are guiding to roughly $725 billion in combined 2026 capital expenditure, up 77% from about $410 billion in 2025, and Microsoft's CFO has told investors the company expects to remain capacity-constrained through at least 2026 despite the spend. When demand is provably outrunning a capex program of that scale, a more expensive megawatt in 14 months beats a cheaper one in 60, and the gap between those outcomes is not a procurement variance. It is market share.
Source: Oculus analysis. Inputs: approximately 1.3 kW all-in per accelerator, approximately 75,000 accelerators per 100 MW, priced at prevailing GPU market rates.
The exception worth naming: sites with existing interconnection rights, retired-plant substations, or inherited capacity are genuinely different assets, and they price like it. If you hold one, your problem is different. If you are competing to acquire one, you have already discovered everything this article argues, and the phantom-load section above just told you where the next ones come from.
The regulatory system's response to the gap is covered in Behind the Meter Goes Mainstream. The equipment reality underneath it is covered in The Turbine Gap. Start with the gap itself: how many months sit between your next capacity decision and its first megawatt-hour, and what each of those months is worth.