The biggest problem facing American AI is not technology — it is that there is not enough electricity and not enough physical equipment. The power bottleneck, industrial hollowing-out, and dependence on Chinese equipment overlap threefold, forming the ceiling on the scaling of America's AI industry. By contrast, China not only enjoys ample power supply (the capacity it added in five years equals the entire stock the United States accumulated over more than two hundred years), but also holds more than 70 percent of the world's transformer and critical-mineral supply in its hands.
US Electricity Prices Surge: Data Centers Eat Into Household Power
In September 2025, the NBER (National Bureau of Economic Research) analyzed 25,000 grid nodes across the United States and found:
- In data-center-dense regions, peak single-month wholesale electricity prices rose by as much as 267 percent above 2020 levels
- Consumers on the PJM grid (covering 13 states) overpaid by roughly $14 billion in 2025–2026
- Of that, about $9.3 billion was burned by data centers — not by households' own consumption
- A single 1-gigawatt AI data center consumes as much electricity per year as 700,000 American households — roughly 2.2 million Chinese households
Per-capita household electricity consumption in the United States is 3.3 times that of China. The gap in awareness shows up in every kind of waste — Chinese people switch off the lights and the air conditioning when they leave a room; Americans wonder, "Wait, you're supposed to turn this thing off? What do you mean, saving energy?"
The Institutional Deadlock of Power Supply
Level One: Grid Companies Refuse to Invest
Once grid equipment is built, it becomes a dead asset — it cannot be dismantled, and it cannot be moved.
The decision-making logic of American grid executives: rather invest too little than ever build a stranded asset or ever shoulder the blame.
The reasons:
- The grid-connection demand that AI companies file is seriously inflated — duplicate applications for the same project, vague plans, placeholder filings to squat on capacity
- Grid investment must be proven prudent before it can be recovered through electricity rates
- If an AI project falls through, the losses from idle equipment can only be borne by the company's shareholders
- For the decision-making committee, corporate interests > national development > technological competition
Level Two: The Government Is Unwilling to Intervene
Trump's "electricity-bill protection pledge" is a verbal commitment — demanding that AI data centers solve their own power problems and not ride on the residential grid. It looks like protecting ordinary households; in practice, it amounts to doing nothing.
Level Three: Its Own Solutions Are Infeasible
| Option | Problem |
|---|---|
| Nuclear power | Construction cycles too long |
| Natural-gas turbines | Construction costs persistently high; new-order prices up 50 percent year on year — order one today and the delivery queue extends past 2030 |
| Ancillary equipment | Global capacity constraints |
Installed Capacity, China vs. the US: Five Years > Two Hundred Years
| Indicator | United States | China |
|---|---|---|
| Total installed capacity | 1.37 terawatts (accumulated over more than 200 years since the nation's founding) | 3.89 terawatts |
| Newly added, 2021–2026 | — | 1.76 terawatts |
| Increment vs. stock | China's five-year additions ≈ 128% of America's entire stock accumulated over more than 200 years | |
"The United States accumulated this 1.37 terawatts over the more than two hundred years since its founding, while China alone installed 1.76 terawatts of new capacity between 2021 and 2026 — our five-year increment is larger than America's entire stock built up over more than two hundred years."
— Wushi Caijing
The Most Awkward Layer: Serious Dependence on Chinese Equipment
A Bloomberg survey shows that the construction of US AI data centers depends heavily on imports from China:
| Equipment / Material | Degree of Dependence |
|---|---|
| High-power transformers | More than 8,000 units imported in the first ten months of 2025 (fewer than 1,500 for all of 2022); exports to the US surged 182 percent year on year in January–February 2026 |
| Switchgear | Imports account for roughly 30 percent of certain types |
| Energy-storage batteries | China's share has consistently stayed above 40 percent |
| Optical fiber and optical modules | Highly dependent |
| Critical minerals (rare earths, etc.) | US import dependence for rare-earth permanent magnets exceeds 90 percent; graphite is 100 percent dependent on China |
One Very Concrete Awkwardness
In January 2026, a group of American utility executives visited a transformer factory in China — only to discover that half of the transformers rolling off the line at that moment were bound for the United States.
Here lies the awkwardness: the United States shouts about decoupling in politics and brandishes the tariff stick — while quietly ordering frantically from China on the side.
America's Real Trump Card — and Its Limits
In Wushi Caijing's view, Nvidia GPUs + the CUDA ecosystem + the frontier models form the three peaks that are hardest for China to climb:
- Frontier models (GPT-5, Gemini) — the technological lead objectively exists
- Nvidia GPUs — domestic chips lag by more than one or two generations
- The CUDA ecosystem — more than 5 million developers locked in, more than a decade of accumulation (the ecosystem moat > the chip moat)
Yet the ceiling imposed by stacking these three layers is precisely what the power bottleneck and industrial hollowing-out lock down. A technological lead that fails to convert into the capacity for deployment at scale cannot produce an overwhelming advantage in genuine industrial competition.
The United States leads in AI algorithms and chips, yet runs into a structural bottleneck at the infrastructure layer of deploying AI at scale. Once AI competition moves from the laboratory to industrial adoption, grid capacity, transformer production capacity, and critical-mineral supply constitute constraints harder than chip process nodes. This is not a question of technology — it is a competition of entire industrial systems and infrastructure capabilities.