Something is clearly going wrong with the American artificial intelligence buildout. In the first three months of 2026, according to a tally by the advocacy group Data Center Watch, 75 projects worth 130 billion dollars were blocked or delayed — roughly as many as in the whole of 2025. In March, OpenAI and Oracle abandoned plans to take their flagship Abilene site in Texas from 1.2 gigawatts to 2. In July, New York became the first state to stop large data centres outright, when Governor Kathy Hochul signed an executive order pausing permits for anything at 50 megawatts or above.
The conclusion drawn from this, in a great deal of commentary, is that the buildout is collapsing and the bubble is about to go. The most repeated version holds that nearly half of the capacity planned for 2026 has been cancelled.
That figure does not survive checking. What does survive is stranger and more consequential: the industry is not being stopped by a loss of appetite for compute. It is being stopped by transformers, by neighbours, and by an electricity grid that cannot be persuaded to arrive faster.
What actually stopped
The Abilene cancellation is the clearest single data point, and its causes are instructive. Reporting attributed it to power grid delays exceeding a year, a breakdown in financing negotiations and changing capacity projections — with the additional embarrassment that winter weather outages had already forced buildings offline by disrupting liquid cooling equipment. Meta was reported to be in talks to take the capacity from operator Crusoe. Notably, Oracle's separate July 2025 commitment to 4.5 gigawatts with OpenAI across multiple sites was reported to remain on track. One expansion died; the programme did not.
The regulatory turn is real and fast. Lawmakers introduced more than 300 data centre bills in the first six weeks of 2026 and 14 states proposed moratoriums. Tracking of local restrictions counted 225 moratoriums across 30 states, of which 151 were in force. The pause in New York runs for a year and could affect more than a dozen projects.
The number that does not survive checking
The claim that half of 2026 United States capacity has been delayed or cancelled traces to a Bloomberg article of 1 April 2026 citing Sightline Climate, which estimated that of roughly 12 gigawatts expected in 2026 only about 5 gigawatts was under construction. SemiAnalysis, which tracks construction directly, rejected it. Its objection is arithmetical rather than rhetorical: the denominator, it said, is off by multiples, and satellite imagery of the top two hyperscalers alone yields a higher under-construction figure than 5 gigawatts. Over six months its own year-end 2026 forecasts moved about 1 percent for hyperscaler self-builds and less than 5 percent for North American colocation.
The reason the two accounts diverge so wildly is that "cancelled" has been doing three different jobs at once. SemiAnalysis separates them. Most of what gets counted is early-stage announcements — projects with no financing, no equipment orders and no interconnection agreement, which realistic models never placed in 2026 in the first place and would have landed in 2028 or later. A second, much smaller group is genuinely delayed: Nebius in New Jersey, Core Scientific at Denton, an Oracle and STACK site in New Mexico pushed to 2029 by pipeline permitting. A third group was killed by local opposition, but existed only on paper. Cancelling a press release is not the same as stopping a building.
The bottleneck is a transformer, not a graphics processor
The most useful correction to the whole debate is a customs one. Most graphics processors enter the United States duty-free. The import duty on an artificial intelligence buildout falls instead on the power and cooling equipment — transformers, switchgear, busways — a significant share of which is manufactured in China, and for which domestic manufacturing capacity remains insufficient. The expensive object is cheap to import; the boring object carries the tariff.
The volumes involved explain why this now governs schedules. Combined United States imports of transformers, switchgear and lithium-ion batteries grew from 33.2 billion dollars in 2020 to 77.1 billion dollars in 2025. Electrical equipment procurement has become a primary development constraint, with transformer lead times extending project timelines by as much as 72 months, and shortages of transformers, switchgear and batteries cited among the main causes of delay. Contractors have responded by writing escalation clauses into contracts to cover tariff-driven increases in steel, aluminium and electrical equipment.
The reported outer bound for power infrastructure procurement — six years for the equipment that connects a data centre to the grid. No amount of capital shortens it quickly.
The nuclear detour arrives too late
The industry's answer to the power problem has been to buy its own generation, and the commitments are substantial: 9.8 gigawatts of nuclear capacity contracted for data centres. The timelines are the problem. Meta holds the largest commitment at 5.2 gigawatts, but not before 2032 to 2035. Amazon agreed a 50 billion dollar partnership with X-energy for 960 megawatts of small modular reactors. Google signed a 615 megawatt power purchase agreement that enabled a decommissioned plant to restart. Microsoft will get power first — from the 835 megawatt Crane Clean Energy Center in 2027 — precisely because it chose an existing reactor restart rather than a new build.
That ordering is the whole lesson. Restarts deliver; new reactors promise. Every small modular reactor project has already slipped at least once, and none has yet demonstrated the factory-floor learning curve on which its economics depend, even as plans for up to 25 gigawatts globally are counted with data centre demand as the largest single driver. Nuclear power is a credible answer to the next decade. It is not an answer to a 2026 interconnection queue, which is when the power is needed.
The money that goes around
If the physical shortfall is smaller than advertised, the financial structure is more exposed than the headline cancellations suggest. In September 2025 Nvidia and OpenAI announced a partnership to deploy at least 10 gigawatts of Nvidia systems, with Nvidia intending to invest up to 100 billion dollars as those systems were deployed. Jensen Huang put the hardware equivalent at 4 million to 5 million graphics processors. The full 100 billion never materialised: Nvidia contributed 30 billion dollars to the funding round OpenAI closed in 2026.
That gap matters because of the shape of the arrangement rather than its size. Bloomberg mapped a web of interlocking deals in which money leaves a supplier's balance sheet labelled investment and returns to its income statement labelled revenue, having passed through a customer and a cloud provider. The reported commitments are enormous: a 300 billion dollar OpenAI contract with Oracle, roughly 1.4 trillion dollars in total OpenAI compute commitments, and a wider loop estimated at more than 800 billion dollars by 2026. OpenAI was reported to be on course to lose about 14 billion dollars in 2026, nearly triple its 2025 losses, against a projection of 100 billion dollars of revenue by 2029.
These concerns are no longer confined to sceptics. Stacy Rasgon of Bernstein Research wrote that the arrangements would fuel "circular" concerns. A venture capitalist described an Oracle statement expressing high confidence in OpenAI's ability to meet its commitments as "bank-run language"; Oracle shares fell about 2.8 percent.
The concrete is mostly still being poured. It is the commitments that are stacked on top of each other.
Everyone is quietly building an exit
The clearest sign that buyers expect this dependency to become uncomfortable is that all of them are engineering around it. OpenAI unveiled its first custom inference processor, designed with Broadcom, under a collaboration announced in October 2025 that targets 10 gigawatts of compute by the end of 2029, with deployment beginning in the second half of 2026. Broadcom simultaneously serves as the implementation layer for Google's tensor processing unit programme, alongside other custom accelerators including Meta's MTIA. The risks to OpenAI's schedule are reported as capacity constraints at TSMC's N3 node — competing with Apple, Nvidia and other custom programmes — first-generation ASIC yields, whether Samsung can supply committed HBM4 memory at volume and quality, and the software stack. Samsung supplies the memory; it does not design the processor.
The counter-example is instructive rather than contradictory. On 4 August 2026 Elon Musk said SpaceX would build exclusively on Nvidia, calling the Vera Rubin architecture the best, with 2 gigawatts of compute planned by the end of 2026 and 10 gigawatts by the end of 2027. That is a SpaceX commitment, not an xAI one — a distinction that matters, because Musk had confirmed in March a semiconductor plant in Austin, Texas as a Tesla and SpaceX joint venture, drawing in xAI and aimed at more than a terawatt of compute annually. Asked in April whether the plant existed to fight Nvidia, he said it did not, describing it instead as survival infrastructure for his own compute shortage. Buying exclusively from a supplier while building a fab is not a contradiction; it is a hedge with a delivery date.
China's mirror-image problem
The same constraint appears in inverted form on the other side of the export controls. Beijing required in August 2025 that data centres source at least half their chips locally, then in November barred state-funded projects from foreign accelerators altogether, with builds less than 30 percent complete instructed to strip out Nvidia, AMD and Intel hardware already installed. A drafted 295 billion dollar plan would run a national grid of artificial intelligence data centres on 80 percent domestic silicon by 2028, a timeline analysts said could collide with the limits of local chip production.
The cost of that policy is measurable. Analysts describe a roughly 50 percent premium on training workloads run on domestic Chinese accelerators — a structural disadvantage rather than a rounding error. Chinese executives nonetheless said they would put about 46 percent of artificial intelligence budgets into domestic products within a year, and Huawei's Ascend line was reported to hold roughly half the Chinese market. Both superpowers, in other words, are paying a tariff: one on transformers, the other on silicon.
So is it a bubble?
The comparison most often reached for is 2008, and it is a poor fit in one specific way. The mortgage crisis was a failure of instruments written on top of an asset whose value nobody could independently verify. Compute is not like that: a gigawatt either exists and draws power or it does not, and satellite imagery settles the argument. That is precisely why the SemiAnalysis correction is possible at all, and why the physical panic is overstated.
The dot-com comparison lands closer. What is genuinely fragile here is not the buildings but the promises stacked against them — commitments measured in hundreds of billions, resting on revenue that does not yet exist, held up by a small number of counterparties who are also each other's investors, suppliers and customers. A cancelled expansion in Abilene costs somebody a construction schedule. A repricing of the loop would cost considerably more.
For now the honest reading is unglamorous. The buildout has slowed, but less than reported and for duller reasons: neighbours who object, permits that expire, and a six-year queue for the equipment that connects a building to a grid designed for a different century. The financial risk is real, but it lives in the contracts, not in the concrete.
