Article — "Power and supply chains, not chips, are delaying nearly half of planned US data centers": In the red dirt of Abilene, Texas, more than 6,000 workers move around on electric buggies, day and night, building eight sprawling buildings that will draw 1.2 gigawatts — roughly the electricity of a million American homes.
Crusoe won that contract on a promise of speed: a portion of the campus energized in under a year from the start of construction. The way it bought that speed was by ordering switchgear and transformers early, securing supply before export barriers went up.
That is the tell. The edge in the AI buildout now belongs to whoever locked down the electrical equipment first, not whoever signed the biggest GPU order.
The edge in the AI buildout now belongs to whoever locked down the electrical equipment first, not whoever signed the biggest GPU order.
Last week this column argued that power, not chips, had become the ceiling on the US AI buildout. This week the number behind that claim is in: Sightline Climate estimates that between 30% and 50% of the 16 gigawatts of US data center capacity slated to come online before the end of 2026 will be delayed or canceled.
Across 140 construction projects, only about 5 GW is actually under construction. Typical build times run 12 to 18 months. Another 16 GW sits in "announced" with no visible progress, and of the more than 25 GW scheduled for 2027, less than 10 GW is being built.
Those are not financing failures. The money is there. The big four hyperscalers paid $165.1bn for property and equipment in Q2 2026 alone, against $88.2bn a year earlier — Amazon at $54.2bn, Alphabet at $44.9bn, Microsoft at $35.8bn, Meta at $30.1bn.
The delay engine is industrial. Generator step-up transformer lead times have passed 160 weeks, up from an average of 143 weeks two years ago. High-voltage circuit breakers have climbed to 125 weeks against a historical baseline of 77.
Then there is the queue. In Northern Virginia, Phoenix, and Dallas, clearing an interconnection queue for a firm power commitment now routinely takes four to seven years. Utilities cannot build substations and string transmission fast enough for a schedule built by software companies.
And a shortage of roughly 439,000 skilled construction workers, per Allianz Research, means even fully equipped sites lack the crews to energize them.
The equipment, meanwhile, is largely imported — much of it from the country the AI race is framed against. Wood Mackenzie's Benjamin Boucher puts it plainly: "There's not enough domestic capacity to go around, so people are pretty much forced to go to the export market."
Crusoe's energy and infrastructure lead, Andrew Likens, describes the procurement problem as "a pretty wild puzzle." His colleague Cully Cavness, the chief strategy officer, says the company pledged energization in under a year and bought equipment early to make that promise real.
Electrical infrastructure is under 10% of a data center's total cost, and without it nothing runs. "If one piece of your supply chain is delayed, then your whole project can't deliver," Likens says.
Reshoring rhetoric does not substitute for megawatts. Tariffs on Chinese exporters do not shorten a 160-week transformer lead time, and a domestic factory that does not exist yet cannot fill a 2026 order book.
Here is where the slip rate becomes a financing problem rather than a scheduling one. JPMorgan calculates that each gigawatt of new capacity needs roughly $35bn in upfront capex; stabilized, it throws off $10bn to $15bn in annual revenue and about $9.4bn in annual cash EBITDA.
That math works if the capacity is live. CoreWeave had 0.85GW of active power in early 2026 and added 1.0 GW during it, leaving an active-to-new ratio of 0.9 times. JPMorgan says it must add active power 4 to 5 times faster — annual additions rising to 1.5 GW in 2027, 1.6 GW in 2028, 1.6 GW in 2029 — to reach a positive free cash flow inflection in 2030, reach 4.5 times when 1.5GW is added in 2030.
The bank's own framing is that the deficit comes from the absolute pace of expansion, not from bad unit economics at any single facility. That distinction matters, because it means the fix is speed. Speed is exactly what the grid will not supply.
CoreWeave cannot self-finance the climb from 1.85 GW at the end of 2026 to 8 GW in 2030. It is not alone in that structure. CoreWeave reports a 2.5-year capital payback against typical five-year GPU contracts — two and a half years of margin, assuming chips do not obsolete early and customers do not default.
Vertically integrated operators move faster; xAI and SpaceX recoup cluster costs in roughly one year. When a debt-financed operator's payback stretches to 30 months and spot rental pricing softens, the equity buffer erodes before the debt amortizes, and bondholders absorb the difference.
Gartner's October warning sizes the aggregate exposure: more than $3.5tn of AI infrastructure spending faces possible impairment by 2030 if current spending continues. The firm expects nearly $1tn in AI infrastructure spending in 2026 and another $6tn between 2027 and 2030 — against AI data center revenue it forecasts at no more than $3.6tn.
Gartner analyst Will Sommer reportedly said the firm has not been able to find a way to make this math work, and the reason is structural: for every dollar spent on compute, only a fraction reaches the owners of compute.
Gartner expects more than 80% of neoclouds to stop operating independently by 2030. Hyperscalers have diversified revenue to absorb years of AI spending; specialist GPU clouds must earn returns on AI-optimized infrastructure to keep running.
The connection between a transformer factory in China and a neocloud bond is direct. Energization dates set revenue start dates. Revenue start dates set debt-service coverage. Every quarter a 160-week lead time slips, a facility that was priced to generate cash in 2027 generates it in 2028 — or does not get built.
Andy Jassy's remark, as recounted by Ben Horowitz, is the cleanest statement of the demand side: "It's beyond 2028. Like, I don't think we get stuff back."
That is not a soft market. It is a market where the buyer cannot take delivery, which is precisely the condition that produces stranded assets rather than a demand collapse.
The capital cycle is priced for a buildout the grid cannot energize. The 2026 cohort is the first hard measurement: 16 GW scheduled, about 5 GW under construction, and 30% to 50% expected to slip. The 2027 cohort looks worse on the same measure.
The exposure runs through whoever borrowed against an energization date. Hyperscalers can re-phase a campus, eat a year, and book the depreciation. A neocloud that financed a $35bn-per-gigawatt build against a 2.5-year payback and a five-year contract has no such room.
Watch two things from here. Whether the transformer order books start clearing faster than the interconnection queues. Until they do, the constraint is not the chip. It is the wire, the breaker, and the substation.