Washington announced it would buy fewer Korean transformers. Three days later our ledger said Korea was selling more of them.
Ask what is scarce in 2026 and you will be told: memory. It is a good answer. DRAM supply is growing sixteen per cent against a historical norm of twenty to thirty. IDC expects smartphone shipments to fall 13.9 per cent this year, the steepest annual decline in the history of the product. Tim Cook, on Apple’s July earnings call, called it a hundred-year flood on memory pricing.
It is also the shortage with the shortest tail. Fabs are being built. The supply arrives late and all at once, as it always does — Intel’s chief executive put the turn at 2028, SK Group’s chairman at 2030, and both are describing a queue that eventually clears.
There is a second shortage inside the same boom, and it does not clear by building a fab. In July the Financial Times reported that forty per cent of American data centres scheduled to come online this year will miss the date. Not for want of chips. For want of the equipment that connects a building to a grid.
On 13 August the US Department of Energy put up to $375m behind fixing that at home — standardising specifications, substituting American materials, shortening lead times. The stated aim was to reduce reliance on foreign suppliers.
So which way is this trade actually going — and how would you know before the annual figures arrive?
The policy logic is sound, and the market agrees with it. American power capex is a domestic story now: reshoring, grid hardening, tariffs on the horizon. If Washington intends to make its own transformers, the sensible expectation is that imports flatten and then fall.
We can confirm the first half of that expectation from our own book. For most of this year the Korean shipping line was flat — no spike, no story, month after month at roughly the same level. Flat is what the policy expects. Flat is what we recorded.
Then it stopped being flat.
Annualised, Korean transformer shipments to the United States now run at 1.108× last year’s full-year figure. On 16 August that multiple crossed a threshold we had published in advance, and our ledger reading for the power axis moved from flat to diverging up.
It is the first state change in that log since we began writing these judgments down on 10 August. Nothing in it is backfilled. And it landed three days after Washington said it would be buying less of exactly this.
The order books side with the ledger rather than the announcement. Hyosung Heavy’s backlog stood at ₩10.4trn in the first quarter, roughly double the year before, with the United States estimated at more than thirty per cent of it. On 6 August HD Hyundai Electric raised its order target for the year by 22.8 per cent, to $5.19bn, citing North American data-centre demand. The three Korean makers together carry about ₩16.6trn of work. Their plants — Changwon, Busan, Ulsan — are running flat out.
The market has priced two of them
If the ledger is turning, the shares should know. Mostly they do not.
We track five pairs linking American power capex to Korean power-equipment names. Their median coupling ratio — how closely the Korean side follows the American side — sits at 1.155. Our control group, the same computation run against unrelated Korean indices, sits at 1.616.
Subtract one from the other and the power axis runs 0.46 below the market’s own drift. In plain terms: Korean power-equipment shares are following American power spending less closely than the broad Korean market is, not more.
Inside the five, the split is clean. Two have re-rated against American capex. Three have not moved at all.
The ledger is rising for the country. The market has picked two companies in it.
The Department of Energy’s programme is about specifications and materials more than about factories. It is a statement of intent before it is a statement of capacity.
Intent is what got announced this week. Capacity takes about as long as a fab — and the data centres are being poured now.
What we cannot see
A customs book records dollars, not units.
That is not a footnote here. It is the centre of the question. A transformer that costs more this year and ships in the same quantity looks identical, in this data, to one that ships more often at last year’s price. We cannot tell you which happened. Anyone claiming to, from this source, is claiming something the source does not contain.
Three further limits, plainly. Seven months is not a year: the 2026 figure is annualised from a partial book and will move. Our coupling ratios run on a 250-day window, which makes them slow by construction — they will not register a ten-day rally, and that is what they are for. And customs records where a shipment cleared, not who eventually installed it.
We also do not know whether $375m changes anything. It is a small number next to ₩16.6trn of backlog.
What would prove this wrong
Two conditions, both a single number in a monthly release.
If the annualised multiple falls back below 1.10, the state change of 16 August reverses and the power axis returns to flat. That is the reading this piece rests on.
If the coupling excess climbs above zero — Korean power names following American capex more closely than the market does — the divergence has closed and the market has caught up with the book.
We will record whichever arrives, on the day it arrives, in the same log that produced this piece.
Figures marked as ours — the annualised multiple, the coupling ratios, the control comparison, and the 16 August state change — are computed daily from government customs releases and exchange data, and logged without backfill from 10 August 2026. Monthly series, per-name coefficients and the method behind them are not published here; they ship in /chain/power.
Data: Korea Customs Service, Korea Exchange, US Department of Energy, Financial Times, IDC, and company disclosures reported in the Korean press. Findings may be cited with attribution to Asia Pulse (asia-pulse.com).