The chip supply chain did not broaden. It changed address.

The decision this is for

A hardware procurement or supply-chain lead deciding whether to pay the premium a supplier charges for diversified or China-free sourcing now, or wait because the diversification is nominal. Operations or finance approves the premium.

If the flows genuinely broadened, the premium buys real optionality and paying early is cheap insurance. If concentration simply moved to one substitute, the premium buys a relabel and the exposure being retired is still there under a different flag.

On this evidence it is one of each, depending on which of the three lines you buy. That is the usable part, and it is the opposite of a category answer.

Before any of these numbers: Taiwan is in here under another name

One more thing that would corrupt any reading of this data: asking the source for partner detail returns the individual countries and a World aggregate in the same list. Summing that list gives $80.77bn against a $40.39bn world total for integrated circuits in 2024 — a clean double count. The World row is excluded from the country rows here and kept as the denominator, and after that filter the partner sums reconcile to the world total to within 0.00% on all 17 commodity-years. Of 310 entries in the source’s partner reference, exactly one is flagged as a group.

Integrated circuits: concentration did not fall, it moved

Malaysia’s 23.3 points went to two places. The top three partners hold 62.3% of US integrated-circuit imports against 64.8% five years earlier — a move of two and a half points, while the identity of the leader changed completely.

Three lines, three different answers

Discrete devices dispersed into four places at once — Viet Nam +9.6, Thailand +9.3, India +6.0, Cambodia +5.9 — while Malaysia lost 13.3 and China lost 5.8. Yet the top three hold more than they did in 2019. Wider participation, same concentration.

Manufacturing machines genuinely diversified. Japan and the Netherlands each lost 14.7 points and the gains spread across four partners with no single winner above 6.3. It is the only one of the three that looks like the diversification story describes.

What we said beforehand

Four explanations with probabilities, written down before any data was pulled. Every one of them is true of at least one line and none is true of all three.

The third study to return the same shape

Three unrelated categories, three unrelated measures, the same answer. The spread inside a category is larger than the movement of the category, so a question asked at category level returns a number describing almost nobody. That is no longer a finding about chips or about food. It is a finding about category-level questions.

The filings: margins rose, and the company that onshored hardest lost the most

The AI confound was tested rather than assumed. Removing the four largest AI-exposed names moves the semiconductor median from +2.22 to +1.83 points, so the result is not an artifact of one or two companies.

For the procurement seat this is the relevant line: supplier gross margins expanded through the period. A premium charged for diversified sourcing is not funding a margin squeeze at the supplier, because there was not one.

The hardware panel, published in full

Drawn by the same SIC rule at the same revenue floor, held separately so the semiconductor move has something to be compared against.

What was excluded, and the guard’s own false positives

Eleven filers of 75 produced no usable FY2019-to-latest pair, including one that tags no usable revenue-and-cost pair in any year at all.

Fourteen filer-years were rejected by the plausibility guard, and inspection showed the guard was wrong about two of them.

A licensing business legitimately runs a gross margin in the high nineties, and a distressed producer legitimately runs a negative one. The band is a blunt instrument tuned for goods producers and it misfires in both directions. It is kept because it catches the class of defect that turned a $100bn filer’s gross margin into −240% in an earlier study, and that trade is declared here rather than hidden.

What this study cannot tell you

• Trade data records a border crossing, not a fab. A chip packaged in one country from a wafer made in another is attributed entirely to the packager. This is the largest limit on the study, and it cuts directly against reading these shares as manufacturing location.
• Nothing here reads a filing’s prose. No management explanation of any margin move; the document host is unreachable from the workspace that ran this.
• Margin movement is not attributed to sourcing. The window spans the AI build-out, the memory cycle and the solar subsidy regime. The AI-exposed names were tested out of the median; the memory and solar cycles were not.
• Five years is two data points. The endpoints are 2019 and 2024 and the path between them is not characterised here.
• Share is not volume and not capacity. A partner’s share can move because it grew or because another shrank.
• One of 18 trade requests failed with a rate limit — manufacturing machines, 2020 — named rather than reported as absence. It affects no endpoint year and no figure here.
• No qualitative half, so this read is not finished by the standard this estate holds itself to. No procurement lead was asked anything.

How this page changes

A standing study, not an article. Corrections are published in place and dated, never applied quietly.

Related reading, free and in full

• All standing studies — the index and the method each one follows
• Costs did not fall, they split — the same “it depends which one you are” result in food and household
• Your category’s advertising benchmark is noise — 1,633 filers
• Can I calculate gross margin from SEC XBRL data?