What the US semiconductor tariff does and does not currently touch

The 25% tariff the United States imposed on advanced semiconductors in January comes with a carve-out that has kept it from touching the AI buildout, exempting covered products destined for American data centres.

That exemption was always provisional, and the Commerce Department report that determines whether it survives was due on 1 July, in a policy environment that has already produced a draft ban on Chinese equipment in data centres.

The report has not been published. Until it is, every operator planning American capacity is working from an assumption rather than a settled rule.

The tariff itself is narrower than the headline rate suggests. It applies to three tariff subheadings covering data processing machines and their parts, and only to logic integrated circuits within specific performance and memory bandwidth bands, which is a deliberately surgical way of hitting AI accelerators without catching laptops.

Seven end-use exceptions sit alongside it, covering repairs, research, startups, public sector use, consumer electronics, civil industrial applications, and data centres. The data centre one is the expensive exception, and it is the one under review.

The proclamation also reserved authority to go considerably further. It explicitly contemplates tariffs on semiconductors, semiconductor manufacturing equipment, and their derivative products, which is the language that would reach servers rather than just the chips inside them.

For an operator, the difference between the two positions is not marginal. A tariff on the accelerators alone is painful, whereas one that reaches assembled servers changes the arithmetic of where a facility gets built.

An analysis published in June by the Computer and Communications Industry Association put a figure on removing the carve-out. Applying the 25% rate to data centres would cost the American economy about $90bn a year and put 243,000 jobs at risk.

The arithmetic behind that is not complicated, which is both its strength and its weakness. Trevor Wagener, the association’s chief economist, multiplies the share of data centre spending that goes on computing equipment, put at 78%, by the share of that compute which is imported, put at 80%, by the tariff rate.

That produces an effective tax of 15.6% on building a data centre in the United States. The study then applies a standard multiplier for knock-on effects and converts the output gap into jobs, which is conventional methodology rather than anything exotic.

The projection that will get quoted is the capacity one. About 20% of planned American AI data centre capacity between 2026 and 2030, roughly $450bn of capital expenditure, would be cancelled, pushed beyond 2030, or relocated abroad.

Whether 80% of American data centre compute is imported is the assumption doing the most work in that chain. It is a defensible figure given where advanced packaging and assembly happen, but it is an estimate rather than a customs return.

The CCIA is a technology industry trade association whose members include the companies that would pay the tariff, so the figure is an advocacy document with a methodology attached rather than a neutral assessment.

The tension the policy is trying to resolve is real enough on its own terms. Tariffs are meant to pull semiconductor manufacturing onshore, and the fabs that would satisfy that goal take years to build, while the data centres are being built now.

Taxing the imports in the meantime raises the cost of the thing the administration also says it wants, which is American AI infrastructure at scale. There is no version of the timeline in which domestic supply arrives before the buildout needs the chips, and Taiwan remains the epicentre of that supply regardless of what any proclamation says.

A tariff offset programme for companies investing in American semiconductor manufacturing capacity was mentioned in the January proclamation, without details. Those details have not appeared either.

Meanwhile, the competitive backdrop has not paused for the review. China has drafted a $295bn data centre plan built specifically to exclude Nvidia, which is a reminder that the cost of American capacity is not being decided in isolation.