The most advanced technology contest in human history has stalled in a queue outside a machine shop.
Nvidia can ship the world's best AI processors in weeks. The transformer needed to power them carries a delivery quote in 2030. That single contrast, drawn from TAFI's September commentary, is the whole story of this issue compressed into two sentences.
Last week we tracked Asian heavy industry planting US beachheads. This week the beachheads have line items: a $528 million transformer plant in Mississippi, a semiconductor map splitting along export-control fault lines, and a critical-minerals chain running through Brazil with political risk baked into the acquisition.
Start with the machine shop. Hitachi Energy will build a $528 million factory in Gallman, Mississippi — its single largest US manufacturing investment — expected to start operations in 2029.
It will produce the mid-range transformers used in data center projects and semiconductor plants. Roughly the size of a detached garage, not the two-story-house scale of the largest grid units.
Read the timeline again. Announced September 2026. Producing 2029. Dozens of units annually.
"We're going to build big things in our electrical infrastructure, but we need the components," Energy Secretary Chris Wright said at the Houston conference where Hitachi's CEO announced the plant. "We need the transformers to build that big expansion."
That is the gap between the AI narrative and the AI build. Software iterates in weeks. Chips refresh every eighteen months. A transformer plant takes years to permit and build, and its workforce takes longer still.
Software iterates in weeks. Chips refresh every eighteen months. A transformer plant takes years to permit and build, and its workforce takes longer still.
The order books confirm it. The average wait for a large power transformer in the United States has stretched to 128 weeks, up from roughly a year before the pandemic. Specialty units run to half a decade. Prices are up 77 percent.
GE Vernova's gas turbine backlog hit 116 gigawatts in the June quarter, with worldwide orders exceeding annual manufacturing capacity by nearly half. Siemens Energy carries a record €162 billion backlog. Mitsubishi warns buyers that a turbine commissioned today arrives at the decade's end.
The arithmetic behind the queue is not subtle. The IEA projects global data centers will consume roughly 945 terawatt hours by 2030 — more than Japan's total annual electricity demand — with the US and China accounting for nearly four-fifths of the expansion.
Sightline Climate tracked sixteen gigawatts of data center capacity slated worldwide for 2026 and found barely five under active construction. The gating item was heavy electrical plant, not silicon, not capital.
Now the counterintuition. If you follow the AI competition debate, you would think the concentration problem sits at the model layer — a handful of frontier labs consolidating into an oligopoly.
Measured on the Herfindahl-Hirschman Index, the index US antitrust agencies actually use, that layer scores near 1,150 for frontier-model usage and 1,410 for cloud. Both sit below the 1,800 threshold at which a market is treated as highly concentrated.
Go upstream and the numbers detonate. Advanced packaging scores 8,100, with a single firm performing close to ninety percent of the work current AI accelerators require. Leading-edge lithography hits the ceiling of 10,000 — one company. Three firms supply essentially all high-bandwidth memory. Gallium approaches the maximum.
The competition the debate is searching for is real. It is simply located three layers below where the debate is focused.
There is a second insight buried in that measurement, and it matters more than the ranking. A rise in the price of an upstream input barely changes the cost of the product built from it, because the input is a small share of that cost. A chokepoint's threat is therefore the loss of the input itself, not a higher price for it.
That is why export controls are the instrument, not tariffs. Since October 2022, the Bureau of Industry and Security has issued seven major rule updates. The February 2026 expansion added semiconductor equipment and EDA software to restricted lists, extended the Foreign Direct Product Rule to more than 40 countries, and added 180-plus entities.
The controls are working on capacity, unevenly. CHIPS Act grants of $52 billion catalyzed $450 billion in private investment and lifted the US share of advanced chip capacity from 12 percent to roughly 22 percent. TSMC's Arizona fab ships 4nm wafers; Samsung's Taylor site runs 3nm gate-all-around; Intel's Ohio module produces 18A-class wafers.
The bill arrives as fragmentation. Duplicate qualification, logistics, and compliance overhead add 25 to 35 percent to landed costs for advanced chips in controlled markets. TSMC has raised advanced-node wafer prices 3 to 10 percent, with 2nm wafers approaching $30,000. Texas Instruments and Analog Devices pushed through 10 to 30 percent hikes.
China is not standing still, but the constraint is physical. Huawei targets up to 1.6 million Ascend dies in 2026, including roughly 600,000 Ascend 910C chips, and launched the Ascend 950PR at 1.56 petaflops FP4. ByteDance has committed $5.6 billion in orders. DeepSeek v4 reportedly runs entirely on Huawei silicon.
Yet SMIC lacks EUV access, Cambricon yields hover around 20 percent, and HBM plus advanced packaging remain constrained. Progress below 5nm is capped by equipment restrictions. The reported consequence for restricted markets: next-generation model training delayed 6 to 12 months.
Then there is the layer below the chips, where the Payne Institute's authors argue the real constraint sits. Data centers will reshape critical-mineral demand not primarily through AI chips but through power infrastructure, electrical equipment, and processing capacity.
India's experience shows what that dependency looks like at the chemistry level. Industry estimates put specialty chemical import dependence for its new fabs above 85 percent. These inputs are consumed continuously and cannot be swapped mid-process without full requalification — weeks of qualification downtime after a supply shock. The binding constraint on localizing them is qualification capacity, not manufacturing capacity.
At the ownership layer, the Peterson Institute flagged two deals that landed at a pivotal moment for Latin America's relationship with Washington: the Venezuela oil deal and the less-reported USA Rare Earths acquisition of Brazil's Serra Verde, a critical minerals producer. Political risk is baked into both.
Carnegie Mellon's CMIST researchers offer a way to triage rather than panic. Their "electrotech moneyball" framework sorts technologies by urgency, whether risk can realistically be reduced, and whether the US can become a strong domestic producer. Not every imported component carries equal risk.
Their sharpest distinction is between basic physical components and the smart systems that control them. A battery or solar panel may present less of a security concern than the software managing how it charges, discharges, and connects to the cloud. Blocking all Chinese-made parts immediately could delay construction, raise costs, and threaten grid reliability.
Here is the tension worth holding. By every conventional measure, the US is still winning the AI race — better models, more compute, far more revenue. Even the most alarming Chinese benchmark result this year, Z.ai's GLM-5.3 edging Anthropic's Mythos on one cybersecurity test, came with Mythos comfortably ahead on converting discovered flaws into working attacks, 78.0 percent to 54.4 percent.
But the lead rests on a stack whose upstream layers are single points of failure. The scarce resource in the AI race is no longer compute. It is the infrastructure that sustains it, and the queue outside the machine shop is the hierarchy.
The countries that build, power, and protect the underlying electrical technologies gain the advantage. Everyone else waits until 2030 for a transformer.