Somebody is filling in a landed-cost cell this week for accelerators that clear customs late next year, and there are two defensible numbers to put in it. If the data-centre exception holds, the headline rate never reaches the order [1]. If the review lands the other way, the same order arrives materially dearer, and the person who signed it gets to explain the variance.
The exception is not the only open variable in that cell. The January proclamation reserved authority over semiconductor manufacturing equipment and derivative products, which is the drafting that would reach an assembled server rather than only the logic chip inside it [5]. A tariff on bare accelerators is a procurement problem you can argue about with a vendor. One that catches finished systems moves the question from what you buy to where you put it [6].
Here is the arithmetic behind the number that will get quoted at you. The Computer and Communications Industry Association's chief economist, Trevor Wagener, multiplies the 78% of data centre spending that goes on compute equipment by the 80% of that compute which is imported by the 25% rate, arriving at an effective 15.6% tax on building in the United States [8]. On a $1bn facility that is $156m [1]. Divide the projected output loss by the projected job losses and you get about $370,000 of annual output per job at risk [2], which is a sign the multiplier is doing ordinary textbook work rather than anything clever [9]. Read the capacity projection backwards and the implied pipeline is roughly $2.25tn of planned American AI capacity across those five years [3], a denominator nobody has audited either.
The 80% import share is the input carrying the most weight, and the source describes it as defensible but an estimate rather than a customs return [11]. The association's members include the firms that would write the tariff cheques, so the figure is advocacy with a methodology attached [12]. That does not make the direction wrong. It does mean the magnitude is a submission, not a finding.
The underlying squeeze is not rhetorical. Fabs take years and the data centres are going up now, and by the source's account there is no version of the timeline where domestic supply arrives before the buildout needs the chips, with Taiwan still the epicentre [13]. The same policy environment has already produced a draft ban on Chinese equipment in data centres [16], and China has drafted a $295bn data centre plan built to exclude Nvidia [15], so the cost of American capacity is being set against a moving comparison.
For Monday, the useful cut is a two-by-two. One axis: does the tariff reach only chips, or finished systems too. The other: does your purchase order put reclassification risk on you or on the OEM. Three of those four boxes are survivable with a spreadsheet revision. The expensive box is finished systems covered and the risk sitting with you, and that is the one worth raising with the vendor's contracts team while the report is still unpublished, because the answer is cheaper to get in writing before it has a price attached. Anyone describing the carve-out as stable is describing a report they have not read.
