Industrial policy usually fails in public and slowly: money is appropriated, an agency cannot spend it, and the shortfall shows up years later as a project that never opened. What is happening in semiconductors is a faster and less familiar failure, and it is being produced by two policies that each work exactly as designed.
The first is a deadline. The Section 48D Advanced Manufacturing Investment Credit returns 35 percent of qualifying capital spending, and it is available only to projects that begin construction before 31 December 2026. That date is the entire instrument. It exists to convert intention into concrete, which is a reasonable thing for a government to want and a reasonable way to get it.
The second is an uncertainty. A proclamation signed in January imposed a 25 percent tariff on certain advanced computing chips, with a Commerce report on the data-centre semiconductor market due at the start of July as the trigger for a broader second phase. The scope of that phase — which chips, which tools, which countries — is the operative fact, and it is precisely the fact a buyer cannot yet look up.
The rational response to both is to wait
Put the two together from the position of somebody who has to sign for a fab. Committing now captures 35 percent of the capital cost, and exposes you to a tariff schedule whose contents are not published, on equipment you must import because it is not made anywhere else. Waiting protects you from buying dear and costs you a third of the project.
Neither policy is irrational on its own. Together they price delay as the safe option, and delay is the one outcome both were written to prevent.
The industry has said as much in the ordinary way. In May the Semiconductor Industry Association and seventeen allied groups wrote to Congress asking for the credit to be extended, describing it as the anchor for some $640bn of announced domestic investment. That letter is usually read as lobbying, and it is, but the more useful reading is as a statement about sequencing: the announcements are real and the construction starts are the part that has to clear a date.
This is the same failure mode this publication keeps describing from other directions, and it is worth saying plainly that it is one failure rather than several. A federal order can bar foreign-made grid equipment on the day it is signed, but you cannot ban the equipment you cannot find, so the instrument that governs the next purchase does nothing about the installed base. Buyers who cannot inspect what they are acquiring have moved the question upstream, so that when you cannot audit the code, you audit the vendor. In each case the policy that can be executed quickly is a procurement rule, and the thing that actually determines the outcome runs on a physical clock nobody can compress.
Fabs run on that slower clock. Construction is measured in years, the tools have lead times of their own, and the people who install them are a small and fully booked population — the same shape as an interconnection queue that has become the binding constraint on new power, where the money arrived long before the capacity could.
There is an unglamorous fix and it does not require choosing between the two policies. Publish the second-phase scope, or extend the credit, or both; what the decision cannot survive is being left open past the point where the deadline forces a commitment. A government can offer certainty about cost or certainty about time. Offering neither, in the same quarter, to the same buyer, is how a programme with $640bn behind it produces a year of nothing being built.
The terms of the Section 48D credit, its 31 December 2026 construction deadline, the May letter from the Semiconductor Industry Association and seventeen allied groups, the $640bn figure, and the January proclamation and its 25 percent rate are as reported by TechTimes and TrendForce and as set out in the proclamation itself. The analysis is our own.
