The United States pays chip factories back 35 cents on every qualifying dollar, in cash if they ask. The offer closes to new projects on December 31, 2026. It does not require anyone to finish a fab by then, only to have started one, in a sense Treasury’s regulations define and the IRS can police for years. On every $100 of qualifying buildings and equipment, a project that misses the date pays $100 instead of $65: about 54 percent more, by my arithmetic, before tax and timing effects.

The credit and its two dates

The credit is “an amount equal to 35 percent of the qualified investment for such taxable year with respect to any advanced manufacturing facility of an eligible taxpayer.” The facility’s primary purpose must be “the manufacturing of semiconductors or semiconductor manufacturing equipment.”

The credit was enacted at 25 percent in 2022, for property placed in service from 2023. The 2025 budget law raised it to 35, and that increase “shall apply to property placed in service after December 31, 2025.” The deadline is a different test: “The credit allowed under this section shall not apply to property the construction of which begins after December 31, 2026.”

One date sets the rate by when property is placed in service, which in tax terms means a “state of readiness and availability for a specifically assigned function”. The other sets eligibility by when construction began.

It can be taken as cash: an electing taxpayer “shall be treated as making a payment against the tax” equal to the credit, so a sponsor with no tax bill still collects. It covers buildings and tools that are “integral to the operation of the advanced manufacturing facility”, with office and administrative space carved out.

What counts as starting

Treasury’s 2024 regulations define “begins,” with two routes. Under the physical work test, construction “begins when physical work of a significant nature begins, provided thereafter that the taxpayer maintains continuous construction or continuous efforts.” Examples: “the excavation for the foundation and the pouring of the concrete pads of the foundation”, or a toolmaker building equipment under a binding contract, if the equipment “is not held in the manufacturer’s inventory.”

Excluded: “planning or designing, securing financing, exploring, researching, obtaining permits, licensing, conducting surveys, environmental and engineering studies, or clearing a site”. A ceremonial groundbreaking is not, in my reading, a start.

The second route is a safe harbor, a bright-line test: pay or incur five percent of the property’s total cost, then maintain continuous construction or efforts. It is measured against final cost, so an overrun can shrink the share after the fact. For a single indivisible property the taxpayer then “will not satisfy the five percent safe harbor with respect to any portion of the single property in such earlier year.” A start resting only on a thin 2026 spend can vanish years later, when there is no 2026 left. For facilities treated as one project the loss can be partial: the start holds for those whose combined final cost is within twenty times what was paid or incurred.

One start, several fabs, ten years

For capacity, what matters most, in my view, is how large one start can be. Several facilities count as one project if one taxpayer owns them at some point during construction and “any two or more of the following factors are present”: seven are listed, such as contiguous land. In the regulation’s example, work on one building of a three-building campus meant “the taxpayer performed physical work of a significant nature that constitutes the beginning of construction of Project C in 2023.”

Continuity has a safe harbor: it is met if the property is placed in service “no more than 10 calendar years after the calendar year during which construction of the property began”. By the rule’s own example, a 2026 start runs to the end of 2036.

There are two catches. Whether several fabs are one project “is determined in the taxable year during which the last of the multiple properties or facilities is placed in service.” And continuity outside the safe harbor is a question of facts, on which “The Commissioner will closely scrutinize a property”. A start is therefore a claim that may be examined years later, not a fact anyone can read off on January 1. If a sponsor took cash and the claim fails, it repays the excess, plus 20 percent unless it shows reasonable cause.

What the credit costs

Treasury’s tax expenditure tables, its estimates of revenue given up through tax breaks, put the credit at $23.1 billion of forgone receipts over fiscal 2026 to 2035, plus $17.7 billion booked separately as spending. The claimants are few: in its 2024 elective payment rules Treasury expected “50 taxpayers to be impacted by the reporting requirements”.

What this changes in a model

For a US fab or tool factory announced but not visibly under way, the question is not when it opens but whether it began construction under Treasury’s test by December 31. If it did not, and Congress does not act, its qualifying spend costs up to about 54 percent more than a rival’s that did. On a whole project the gap is smaller (see below). A model of announced capacity should say which side of the line each project sits on, and whether its return assumed the credit.

A groundbreaking press release does not show that. Look instead for a statement that construction began for purposes of section 48D, the credit’s code section, though even that can be examined years later. For a sponsor that did start a campus, the start works, in my judgment, like an option on later phases at 35 percent, good only while continuity holds and the later phases still count as the same project.

Not starting may also be a bet on Congress. A House bill would move the date to 2031, and its extension “shall apply to property, the construction of which begins after December 31, 2026.” If enacted with that wording, a law passed in 2027 could reach projects that missed. As of early October no extension had been enacted.

Where I could be wrong

Congress may move the date. If an extension like the House bill passes, the 2026 line gives way to 2031. As of mid-October, that bill had not left committee.

The 54 percent is a ratio on qualifying property only. Land, offices and anything not integral to production are outside it. The cash generally comes with the tax return for the year property enters service, though an election for progress expenditures can pull some into the construction years. On a whole project, counting non-qualifying spend and the wait for the cash, the difference is smaller than 54 percent.

A start is not the end of the risk. The credit can be recaptured on a sliding scale if property is sold or stops qualifying within five years of entering service. And if within ten years a claimant makes a significant transaction “involving the material expansion of semiconductor manufacturing capacity” in China or another country of concern, all of it can be recaptured, unless the deal mainly involves older-generation chips or is abandoned within 45 days of Treasury’s notice.

I cannot see who is on which side. No public source counts how much planned US capacity has started under this test. Projects already begun have cleared the deadline, though their starts can be challenged; the date matters most for what comes next.

The credit was written to pay for fabs. From January, it pays only for those that can defend a start made by December 31: 35 cents back on each qualifying dollar for them, full price for the rest, unless Congress moves the date.