Core argument: The US government is attempting to build a "compute empire" through chip export controls and the repatriation of capital — yet nearly forty percent of its AI data-center projects face the risk of stalling out. The root cause is not technological competition, but the systemic degradation of American manufacturing capacity and national infrastructure. The essence of technological competition has never been merely a race over code in the laboratory; it is, above all, a competition over physical industrial systems and infrastructure capability.
I. Nearly Forty Percent of AI Data Centers Face Stalled Construction
According to an April 18, 2026 report by the UK's Financial Times (cited by the Global Times, Chang'anjie Zhishi — a commentary account affiliated with Beijing Daily — and others), close to forty percent of the data-center projects originally planned in the United States face the risk of delay. Key projects from Microsoft, OpenAI and others have been pushed back by three months or more across the board.
Satellite imagery from the geospatial analytics firm SynMax paints an even more vivid picture: of the projects originally slated to come online in 2027, more than sixty percent have not even had their foundations laid — some have yet to break ground at all.
Of the AI data-center projects scheduled to come online next year, more than sixty percent have not even had their foundations laid — this is not a lag in progress; the entire project has been put on pause.
II. The Four Mountains: Structural Root Causes
The essence of AI compute is converting electricity into intelligence. When the ideal blueprint collides with America's industrial reality, four systemic structural obstacles surface in sequence.
The American grid — dilapidated, privatized, and fragmented — is simply unable to carry new loads of hundreds of megawatts, or even above one gigawatt. A single AI data center's power demand rivals that of a small city, yet in many parts of the United States the power infrastructure has gone decades without a major upgrade. Meanwhile, critical equipment such as transformers is in severe shortage, with delivery lead times stretched from a few months to more than two years.
From electricians to plumbers, the United States faces a severe shortage of skilled construction workers. An executive associated with an OpenAI project conceded helplessly that the country simply does not have enough skilled workers to complete these builds. This is not any single company's problem — it is the accumulated result of three decades of deindustrialization across American industry as a whole: as factories moved offshore in large numbers, what vanished with them was the trained industrial workforce — and the soil that cultivates it.
Federal directives often fail to travel beyond Washington. The states' attitudes toward data centers have "fractured completely" — environmental lawsuits and state-level approval barriers obstruct projects layer upon layer. A single data center can bring thousands of jobs, yet a single environmental organization can shut it down. Policy fragmentation means that no national-level compute strategy can bypass the veto power of local politics.
The United States imposes strict export controls on advanced GPUs, demanding that American customers be served first. Yet cutting off other countries' chip supply does not automatically translate into compute staying in the United States — because compute needs physical infrastructure to host it. When you cannot build enough data centers, ever-tighter chip controls only force other countries to accelerate indigenous substitution, rather than locking demand inside American borders.
III. The Real Foundations of a Compute Empire
Faced with the hard evidence of satellite imagery, America's tech giants have talked tough: an OpenAI spokesperson insisted projects are "proceeding as planned," while Oracle said construction is "proceeding in an orderly fashion."
What is genuinely missing, however, are the four things required to build large industrial projects:
- Stable power — a dilapidated grid cannot support gigawatt-scale loads
- Skilled workers — a blue-collar generation gap after three decades of deindustrialization
- A unified market — policy fragmentation between the federal government and the states
- An effective government — the driving force to cut through environmental lawsuits, state-level barriers, and administrative review
These four are precisely America's systemic weaknesses.
The confidence behind US chip sanctions against China rests on the assumption that "America can build faster." But when forty percent of data centers stall and sixty percent of projects have yet to break ground, that assumption itself becomes the largest exposure of risk.
IV. Key Evidence at a Glance
| Category | Content |
|---|---|
| Data point 1 | In 2026, nearly forty percent of the US's originally planned AI data-center projects face delay; satellite data shows sixty percent have not started construction |
| Data point 2 | Key projects from Microsoft, OpenAI and others have been pushed back by three months or more across the board |
| Labor data | The United States faces a severe shortage of construction workers; an OpenAI executive acknowledged "there are not enough skilled workers to complete the projects" |
| Infrastructure | Major grid upgrades have lagged for decades; transformer delivery lead times have stretched to more than two years |
| Policy environment | State attitudes toward data centers have "fractured completely"; environmental lawsuits and approval barriers obstruct projects layer upon layer |
V. Extended Reflection: The Three Layers of Compute Competition
On the surface, AI competition is competition over chips; the middle layer is competition over compute (electricity plus data centers); and the deepest layer is competition over national infrastructure capability. The United States holds the advantage at the chip layer, but structural fragility is exposed at the middle and bottom layers.
The essential difference in US–China technological competition is not "who is more advanced," but "whose system is more complete." When one side cannot convert its advantage into productive capacity because of infrastructure bottlenecks, the other side — as long as it is not completely cut off — has already won the window of time.
This is not a problem only the United States has. Building AI infrastructure requires the cooperation of manufacturing and engineering. China's advantages in ultra-high-voltage transmission and construction speed are real, but it too faces constraints in high-power chip design, advanced process nodes, and other segments — the dimensions of constraint are simply different.