Semiconductor Tariff: The 25% Rule Is Narrower Than the Headlines

Through most of 2025 the trade press carried a 100% chip tariff as a near certainty. What actually arrived on 15 January 2026 was 25%, and it reaches a far narrower set of goods than almost anyone planned for. Proclamation 11002 covers advanced AI-accelerator-class hardware and the assemblies built around it. It does not cover memory, discrete devices or general-purpose integrated circuits.

The scope surprise has a classification twist attached. The duty attaches in Chapter 84, under headings for automatic data processing machines and their parts, not in Chapter 85 where most people go looking for semiconductors. An importer checking 8541 and 8542 for exposure will find nothing and conclude the tariff does not apply, which is right for now and wrong for the reason they think. This guide sets out what the semiconductor tariff actually covers as of 26 August 2026, and what a second phase would change.

What Proclamation 11002 Actually Reaches

Proclamation 11002 was signed on 14 January 2026, published at 91 FR 2443 and took effect at 12:01 a.m. EST on 15 January 2026. The rate is 25% ad valorem.

Coverage runs to advanced logic integrated circuits of the class used for AI acceleration, and to the derivative assemblies they are built into, classified in headings 8471.50, 8471.80 and 8473.30. An article only becomes dutiable if it falls inside defined bands for Tensor Processing Performance and DRAM bandwidth. Those bands are narrow windows rather than simple floors, so hardware can sit above the range and fall outside the duty just as hardware below it does. Anything that classifies in those headings without meeting a band is reported under a separate Chapter 99 subheading and is not dutiable.

The White House named the NVIDIA H200 and the AMD MI325X as covered examples, which gives a useful sense of the tier involved. This is data-centre accelerator hardware, not the silicon inside a laptop or a car.

Chapter 84, Not Chapter 85, and Why That Trips People Up

The instinct is to look for a semiconductor tariff among the semiconductor headings. Chapter 85 is where diodes, transistors, memory and integrated circuits live, under 8541 and 8542, and that is where importers and their brokers naturally check first.

This action does not sit there. It attaches to automatic data processing machines and parts thereof in Chapter 84, because the covered goods are accelerator cards and modules rather than bare die. A GPU board entering as a unit of an automatic data processing machine is caught; the same silicon entering as an integrated circuit under 8542 is not, at least in the current phase.

That gap is a real planning consideration and also a real risk. Importers who bring in accelerator hardware at the board or module level should be checking their Chapter 84 classifications against the thresholds rather than assuming their Chapter 85 review answered the question. Where the classification between a part of a machine and a component is genuinely arguable, the reasoning has to run through the General Rules of Interpretation in order rather than being chosen to fit the preferred outcome.

The Exemptions Are Use-Based, Which Is Unusual

Most Section 232 relief is product-based: an article is on a list or it is not. The semiconductor action instead exempts on the basis of what the hardware will be used for in the United States, which puts an evidentiary burden on the importer that product-based programmes do not.

Hardware destined for a qualifying US data centre enters under 9903.79.03, where the qualifying facility is defined as one requiring more than 100 megawatts of new load dedicated to AI inference, training, simulation or synthetic data generation. Repairs and replacements, US research and development, and US startups and emerging growth companies each have their own subheading, and three further subheadings cover non-data-centre consumer and gaming hardware, non-data-centre civil industrial uses such as factory robotics, and US public-sector applications. The full range runs from 9903.79.01 to 9903.79.09.

The compliance implication is that the entry has to be supported by facts about the end use, and those facts sit with the buyer rather than the supplier. An importer claiming the data-centre exemption is making a representation about a facility, and that representation has to be documented at the time of entry rather than reconstructed during an audit.

Two mechanical points matter for anyone structuring around this. Duty drawback is not available on these goods, so the usual re-export recovery route is closed. And goods entering a foreign-trade zone must be admitted in privileged foreign status, which fixes their tariff treatment at admission and removes the flexibility that a zone normally provides. That materially changes how a foreign-trade zone works for this category.

Phase 2 Is Announced, Not In Force

Proclamation 11002 required the Secretary of Commerce to report by 1 July 2026 on the US data-centre semiconductor market, so the President could decide whether to modify the rate or widen the scope. A broader phase covering essentially all semiconductors and semiconductor manufacturing equipment has been described as the intent.

As of 26 August 2026 no public action following that review has been announced. The correct position for planning is that Phase 1 is in force at 25% on a narrow list, Phase 2 is pending, and the outcome of the July review has not been published.

The same caution applies to the tariff offset programme for companies investing in US semiconductor production. Commerce recommended it and it has been reported as though it were available. It is a Phase 2 item and it has not been implemented. An importer building a duty model around an offset that does not yet exist is planning against a press release.

Where the Duty Lands in a Chip Supply Chain

Semiconductor supply chains separate design, fabrication, packaging and test across different countries by design, which makes origin a harder question here than in almost any other sector. A device designed in the United States, fabricated in Taiwan, packaged and tested in Malaysia and assembled onto a board in Mexico has four plausible origin stories and only one correct answer for customs purposes.

Origin follows the last substantial transformation, and for semiconductors that determination has historically pointed to the fabrication step rather than to assembly and test. Getting it wrong changes not just Section 232 exposure but Section 301 exposure and preference eligibility at the same time, which is why a documented country of origin determination belongs in the file before the first entry rather than after the first CBP question.

The packaging and test economies carry the commercial consequence even when they do not carry the origin. Importers sourcing through Taiwan and Malaysia should be reading their exposure at the level of the finished assembly, because the board entering the United States is the article being classified, not the die inside it.

There is a related action worth tracking alongside this one. Proclamation 11052, signed on 6 August 2026 and published at 91 FR 51975, imposes Section 232 duties on polysilicon and its derivatives effective 4 December 2026, together with minimum import prices. It is a separate programme, but it touches the same upstream material base and the same set of importers.

What to Check Before the Next Entry

Pull every line classified in 8471.50, 8471.80 and 8473.30 and test each against the Tensor Processing Performance and DRAM bandwidth thresholds. That test decides between 9903.79.01 at 25% and 9903.79.02 at zero, and it is a technical specification question rather than a customs one, so it needs engineering input.

For anything claiming a use-based exemption, build the supporting file at entry. A qualifying facility claim needs the facility identified and its capacity substantiated; an R&D or startup claim needs the same discipline. These are representations, and the time to document them is before they are questioned.

Then confirm nothing in the model still assumes IEEPA duties on top. Those were invalidated in February 2026 and are no longer collected, so a semiconductor duty model carrying a reciprocal line is overstating cost. Where the classification or the threshold test is genuinely close, the analysis belongs with a trade advisory services specialist before the goods ship rather than after the entry is filed.

Frequently Asked Questions