Semiconductor subsidy should be judged on demand met at home, not on fabs opened

Two measures are being used to judge the American semiconductor programme and neither is the one that matters. I would judge the money on the share of domestic demand met from domestic plant, and on nothing else. That figure is published nowhere in the data this site holds, which is the difficulty and not a reason to reach for a substitute.

Take the last full year in these files. Construction spending on manufacturing plant ran at an annual rate of 244 billion dollars in 2024 and 215 billion in 2025, a fall of 29 billion. U.S. semiconductor production went from 144 to 165 on a 2017 base of 100 across the same two years, a step of 21 index points and the largest of the four steps in this window. On the building yardstick 2025 was a retreat. On the output yardstick it was the strongest year of the five. Both cannot be the score.

The count that exists becomes the standard

A programme is judged on the numbers that are to hand. Plant spending is published by a federal agency on a monthly cycle. Openings are counted by anyone who can read a press release. The share of American demand met from American plant is counted by nobody in these files, and a standard that nobody computes does not compete with one that arrives every month. The fab count won this argument by turning up, not by being defended.

Convenience is a poor reason to adopt a test, and a worse one when the convenient number is an input. Construction spending records what was bought. Its rise from 75 billion dollars at an annual rate in 2020 to 244 billion in 2024 is a record of purchases across manufacturing plant of every kind, and plant of that sort is what the subsidy set out to buy rather than what it set out to achieve. Scoring a programme on a measure of what it meant to pay for makes the score a restatement of the intention.

What the money was bought to change

The case for the subsidy was never that the country should own buildings. It was that when supply is interrupted, what the economy uses should be reachable. That is a claim about a ratio between domestic supply and domestic use. Plant is the means. The ratio is the end. Confusing the two is the standing error of industrial policy, and it survives because the means can be photographed.

A fab count can rise while the ratio falls. If use grows faster than domestic capacity, every new building is compatible with a deeper dependency than the one it was meant to close. A construction series cannot detect that. Nor can a production index, and for the same reason: each counts one side of a fraction and leaves the other side unwritten.

An index has 2017 in its denominator

The semiconductor series is an index on a 2017 base. A reading of 165 means output is 65 percent above its 2017 level. It does not mean 65 percent of anything the economy consumes. The denominator is a year. Computer and electronic product output behaves identically, moving from 118.3 in 2024 to 126.8 in 2025 on the same base, a step of 8.5 points that is again a comparison with 2017.

This is not a complaint about the Federal Reserve Board, which labels the base plainly and claims nothing more for it. It is a complaint about the use the number is put to. An output index answers how much more is made than was made before. Substituting that for how much of what is used is made here changes the question without telling the reader that it has changed.

The missing half of the ratio

A share of demand met at home needs the other half of the fraction, which is what arrives from abroad at the level of the product. These files do not hold it. The nearest measure is U.S. goods imports from China, a monthly average of 26 billion dollars in 2025 against 45 billion in 2022. Each annual point there is a monthly average rather than a year's total, and it counts every good from a single partner. No device-level figure can be lifted out of it.

Imports of goods and services, at 4.25 trillion dollars in 2025, carries the same defect on a larger scale. Neither series separates a wafer from a washing machine. Anyone who has claimed that dependency narrowed has claimed it from a number that cannot show dependency, and that is the position I am objecting to whichever direction the claim happens to run in.

What I would commission

The recommendation is dull. Before the next tranche, build the denominator. Domestic shipments of semiconductors, imports of the same products at the same level of product detail, exports netted out, published annually by a named agency with a stated revision policy. That is a measurement task rather than a research programme, and it is the only thing that would let anyone say whether the money did what it was voted for.

Until that exists, I would rather the programme were reported as unscored than scored on the wrong quantity. Plant spending states what was bought. Output states what was made. Only a share states what was depended on, and dependency was the thing the policy was for.

Against my own argument

Three cautions, because the case above is only as good as the series under it. The first is the construction figure. It covers manufacturing plant of every kind, so it is a proxy for fabs rather than a count of them, and it is stated in dollars without adjustment for building costs, so the fall from 244 to 215 can be less plant or cheaper plant. Its monthly figures are seasonally adjusted annual rates, so 215 billion dollars describes a year of factory construction; treating it as one month's outlay would put the annual total at roughly twelve times what was actually spent.

The second caution is that a production index cannot say which devices were made, at which generation, or with which imported inputs, so even a rising index is compatible with the exposure the policy names. The third is that none of these series attributes anything to any programme. Those limits are the condition under which the choice of yardstick matters most. Where there is no measure, the available number becomes the verdict by default, and the available number here counts spending on manufacturing plant of every kind, which is an input and not the thing the policy was for.