The Chips Changed Address. They did not change how concentrated they are.
A de-risking program moved US integrated-circuit sourcing off its largest supplier and onto a different one. Across three semiconductor trade lines the same five years give three different answers, and only one of them looks like diversification.
You know the slide. There is a map on it, and the map has fewer pins in one country than it had three years ago, and the heading says something about resilience. Somebody earned that slide. It took two years, a qualification program, a second source for parts that had one, and a travel budget nobody enjoyed spending.
The question underneath it is simpler than the map, and the map is not built to answer it. Did the exposure go away, or did it move?
The top three held almost exactly what they held before
US imports of integrated circuits ran $33.09bn in 2019 and $40.39bn in 2024, across 114 trading partners in both years. That is the line most people mean when they say chips, and it is the line most de-risking programs were pointed at.
Over those five years the share held by the three largest partners went from 64.8% to 62.3%.
Share of US integrated-circuit imports held by the top three trading partners, 2019 against 2024. Five years of supply-chain diversification moved the top of the table by two and a half points.
Read that number alone and the conclusion is that nothing happened. That conclusion is wrong too, and the way it is wrong is the whole point of the exercise.
Malaysia left and Taiwan arrived
Malaysia × Taiwan × Israel
2019–2024In 2019 Malaysia was the largest single source of US integrated circuits at 47.1% — close to half the line, from one country. By 2024 the largest source was Taiwan at 28.1%, and Malaysia had given up 23.3 points.
Those points did not spread out. Taiwan took 16.9 of them and Israel took 10.0, and the next largest gains — Costa Rica at 3.0, Ireland at 2.2, Germany at 1.1 — are rounding beside those two. One concentrated position was replaced by another concentrated position in a different place.
The single-partner risk did fall, and that is real: 47.1% in one country is a worse position than 28.1%. What did not happen is the broadening the slide claims, because the top three ended the period holding 62.3% against 64.8%.
A buyer who moved off Malaysia to retire concentration risk is now holding Taiwan concentration instead.
The same question returns a different verdict per trade line
The same supply chain, the same five years, three different answers. Manufacturing machines are the only line where the top three gave up real ground.
Integrated circuits relabeled. Discrete devices did something else entirely: Vietnam gained 9.6 points, Thailand 9.3, India 6.0 and Cambodia 5.9, while Malaysia lost 13.3 and China lost 5.8. More countries participating, by a wide margin — and the top three finished holding more than they started with, 55.1% against 52.3%. Wider participation, same concentration, and those two facts are not in conflict once you look at who the three are.
Manufacturing machines are the only line that behaved the way the diversification story describes. Japan and the Netherlands each gave up 14.7 points, the gains spread across Malaysia, Singapore, Korea and Germany with no single winner above 6.3, and the top three fell from 77.9% to 54.2%.
So the answer to “did the semiconductor supply chain diversify” is yes for the machines that make chips, no for the chips themselves, and the opposite of yes for discrete devices. Four explanations were written down before the data was pulled, and each one turned out to be true of at least one line and none was true of all three.
Taiwan is filed under a name you probably excluded
UN Comtrade does not report Taiwan as Taiwan. It reports it as Other Asia, nes — not elsewhere specified.
An analyst who filters that dataset by country name, as anyone would, gets a clean-looking table in which Taiwan does not appear in US semiconductor imports at all. The correct reading is that Taiwan is the largest single source of US integrated circuits in 2024 and the biggest share mover of the five-year window. The row is sitting there. It is just not wearing the name anybody searches for.
This is the kind of defect that never announces itself. The table parses, the percentages sum, the chart renders, and the single most consequential line in the dataset is missing from it.
Supplier margins went up, and Intel's went down
Across 41 semiconductor filers, median gross margin rose 2.22 points between FY2019 and the latest filed year, with 28 of the 41 expanding. Taking the four largest AI-exposed names out of the panel moves that median to 1.83 points, so the rise is not one company carrying the field. The 23 hardware and systems filers moved 1.41 points on the same measure.
For a procurement seat that line matters more than it looks. A premium being charged for diversified sourcing is not funding a margin squeeze at the supplier, because over this window there was not one.
Intel
FY2019–latest filedIntel’s gross margin went from 58.6% to 34.8% — a fall of 23.79 points, the worst in the 41-filer panel, against a panel median that rose by two. The next worst move was Methode Electronics at 6.73 points down.
Intel is also the filer in that panel that pursued domestic capacity hardest. Those two facts sit next to each other and it is tempting to join them, so here is the join and here is why it does not hold: onshoring is capital-intensive and depreciation lands in cost of goods, which would push gross margin down. Product mix, foundry losses and server share loss would also push it down, and none of them are controlled for here.
The claim that survives is narrower. The one company that built domestic capacity hardest did not get paid for it in gross margin over this window. Why it did not is a different question and this data cannot answer it.
A supplier’s margin tells you what happened to the supplier, not what caused it.
The honest limits
Trade data records a border crossing, not a fab. A chip packaged in one country from a wafer made in another is credited entirely to the packager, which cuts directly against reading any of these shares as manufacturing location. That is the largest limit on the whole study and it does not have a workaround in public data.
Share is not volume and not capacity — a partner’s share can move because it grew or because somebody else shrank. The endpoints are 2019 and 2024, two data points, and the path between them is not described. Nothing here reads a line of management commentary from any filing, because the document host is unreachable from the workspace that ran the pull.
One registered prediction died outright: that the period would show real diversification. It is dead for integrated circuits, where the falsifier fired, and it holds for manufacturing machines. Fourteen filer-years were rejected by a plausibility band tuned for goods producers, which misfires on IP licensors above its ceiling and on loss-making producers below its floor. All fourteen are published by name.
No procurement lead was interviewed for any of this. It is an outside read of public filings and public trade records, and it is missing the half that would come from asking the people who signed the contracts.
- Your de-risking program targeted a country rather than a line — moving off one supplier retires single-partner risk and leaves top-three concentration roughly where it was.
- You buy from more than one of the three trade lines — circuits, discretes and the machines that make them moved in three different directions over the same five years.
- Your exposure analysis filters by country name — Taiwan is filed as “Other Asia, nes” and a name filter drops the largest mover in the dataset without erroring.
- You are paying a premium for diversified sourcing — supplier gross margins expanded over this window, so the premium is not covering a squeeze at the other end.
- It is not yours if you need to know where a chip was made — trade data credits the last border crossed, and no public source fixes that.
Split the sourcing review by trade line, and re-run the country filter with “Other Asia, nes” in it.
Take your highest-spend commodity line, pull the partner shares for it alone from UN Comtrade for 2019 and the latest year, and compute the top-three share at both ends. Do it per line rather than for semiconductors as a category, because the category average is three different stories laid on top of each other. Then check whether the country list your own exposure model uses has a row for Other Asia, nes; if it does not, the model has been running without the largest single source of US integrated circuits in it.
- What it costs
- About a day of an analyst’s time against free public data — Comtrade charges nothing. The cost that bites is that a de-risking program already reported as complete may come back open.
- How you’ll know
- You will know inside a week. If top-three concentration on your highest-spend line has moved less than about five points since the program started, it moved the name and not the risk.
Your supplier map, read line by line instead of country by country.
Where your own sourcing concentration actually sits once the trade lines are separated, so the de-risking claim has a number under it rather than a map.
Get a free positioning review Or read the frameworks →