Section 232 semiconductor tariff: what it means for servers and GPUs entering the US

The Section 232 semiconductor tariff is a 25 percent US import duty, in force since January 15, 2026, on a narrow band of advanced AI chips and on servers and boards that contain them. Most servers and GPUs are outside that band, and covered hardware headed for large US data centers, R&D and several other uses enters at 0 percent.

That is the short answer, and for many shipments it is good news. The longer answer matters because the tariff works differently from the duties most hardware vendors are used to. It does not depend on the country of origin. It depends on the performance of the chip inside the box and on what the buyer will do with it in the United States. Two identical GPU servers on the same flight can owe very different amounts, depending on the end-use claim made at entry.

It also matters because the rule is not finished. The proclamation that created it describes itself as a first step, the Commerce Department owed the President an update on the data center chip market by July 1, 2026, and senior officials have said publicly that broader semiconductor tariffs are coming. If you sell or ship servers and accelerators into the United States, this guide explains what the Section 232 semiconductor tariff covers today, which exclusions exist, how it interacts with other US duties, who pays it, and how to plan a shipment so that the entry holds up later.

What is the Section 232 semiconductor tariff?

It is an additional import duty created by Proclamation 11002, signed on January 14, 2026, under Section 232 of the Trade Expansion Act of 1962. That law lets the President adjust imports that the Commerce Department finds threaten national security. For semiconductors, the first adjustment is a 25 percent ad valorem duty on a defined set of covered products.

The proclamation, titled "Adjusting Imports of Semiconductors, Semiconductor Manufacturing Equipment, and Their Derivative Products Into the United States", was published in the Federal Register on January 20, 2026, and applies to goods entered for consumption, or withdrawn from a warehouse for consumption, from 12:01 a.m. eastern time on January 15, 2026. You can read the operative text of Proclamation 11002 on the Section 232 semiconductor tariff in the Federal Register.

The design is unusual. Instead of taxing all chips, the measure targets a performance window and then switches the duty off for a long list of uses inside the United States. The White House described the scope as very narrow, and it is. What it creates, in practice, is a duty that applies when advanced accelerators enter the country for a use that does not support the US technology supply chain, while hardware for US data centers, labs, repairs and similar uses keeps flowing at zero.

Section 232 duties are also different from the emergency tariffs that dominated 2025. When the US Supreme Court decided in February 2026 that the International Emergency Economic Powers Act does not authorize tariffs, that ruling did not touch Section 232, which rests on a separate statute and a formal Commerce investigation. For hardware buyers, that makes the Section 232 semiconductor tariff one of the more durable parts of the US tariff picture.

Which servers and GPUs does the 25 percent duty cover?

The duty covers goods classified under three subheadings of the Harmonized Tariff Schedule of the United States, 8471.50, 8471.80 and 8473.30, but only when the article is, or contains, a logic integrated circuit inside one of two performance windows. Everything else in those subheadings is reported under a separate heading at 0 percent.

The three subheadings are the familiar ones for this market. Subheading 8471.50 covers processing units such as servers, 8471.80 covers other units of automatic data processing machines, which is where some accelerator systems and appliances land, and 8473.30 covers parts and accessories of computers, which includes GPU boards and modules shipped on their own. If you need a refresher on how servers, accelerators and network gear are classified, our guide to the HS code for servers walks through the headings.

The performance test comes from CBP's implementing guidance, CSMS 67400472. A chip is inside the first window when its Total Processing Performance, the metric US export controls also use for advanced chips, is greater than 14,000 and less than 17,500, and its total DRAM bandwidth is greater than 4,500 and less than 5,000 GB/s. It is inside the second window when its Total Processing Performance is greater than 20,800 and less than 21,100, and its DRAM bandwidth is greater than 5,800 and less than 6,200 GB/s. Industry summaries of the announcement named the Nvidia H200 and the AMD MI325X as examples of chips that fall within these bands. Chips above or below both windows, including older accelerators and many newer ones, are not covered by this first phase.

Two consequences follow. First, the test is applied chip by chip, but the duty is charged on the covered article. A complete eight-GPU server classified under 8471.50 is covered if its accelerators fall in a window, and the 25 percent is then calculated on the entered value of the server, not only on the value of the chips. Second, the importer needs the chip specifications before the entry is filed. "AI server" on a commercial invoice is not enough to decide between the 25 percent heading and the 0 percent one.

Which end uses are excluded from the Section 232 semiconductor tariff?

Covered hardware enters at 0 percent when it is for one of seven US end uses: large data centers, repairs or replacements, research and development, startups, non-data center consumer uses, non-data center industrial uses and the public sector. The importer claims the exclusion by reporting the matching heading, from 9903.79.03 to 9903.79.09, on the entry.

The data center exclusion is the one most server buyers ask about, and it is narrower than its name. CBP's guidance limits it to US data centers that require more than 100 megawatts of new load for AI inference, training, simulation or synthetic data generation. A hyperscale campus or a large AI build qualifies. A 5 MW enterprise server room, a regional colocation cage or a small inference cluster generally does not, unless another exclusion applies to the buyer, such as research and development or the startup category, which follows the definition of an emerging growth company in US securities law.

The repair and replacement exclusion is useful for vendors who ship spare boards and swap units under support contracts, and the R&D exclusion covers labs and test environments in the United States. The consumer and industrial exclusions mostly protect workstations, gaming hardware, vehicles and factory equipment, which shows the intent of the measure: it is aimed at advanced accelerators that do not end up strengthening the US supply chain.

How do you prove the end use? The proclamation leaves that to the Commerce Secretary, together with the US International Trade Commission and CBP, who are told to decide whether end-use certifications are needed, with any change published in the Federal Register. The CBP guidance does not describe a separate certification form, so the claim is made through the heading itself. That does not make it informal. The importer of record signs the entry under the reasonable care standard, and an exclusion claim that cannot be supported in a later audit means the 25 percent becomes due, with interest and possible penalties. In practice, keep a written end-use statement from the buyer, the site and its power capacity where the data center exclusion is used, and the contract or purchase order that ties the hardware to that use.

How does the tariff stack with other US duties?

When an article is properly entered under 9903.79.01, several other additional duties do not apply on top of it, including the reciprocal tariff headings and the Section 232 duties on steel, aluminum and copper. Free trade agreement preferences do not remove it, drawback is not available, and goods in foreign trade zones are treated as privileged foreign merchandise.

The non-stacking rule is the main relief. CBP's guidance lists the duties that a covered article does not also pay when it is entered under 9903.79.01, and it sets the reporting order for the remaining Chapter 99 numbers on the same line. Duties that are not on that list, such as Section 301 duties on certain origins, can still apply in addition, so the classification exercise has to look at the whole stack, not only at the Section 232 line.

The other rules work against the importer. A free trade agreement or preference program does not reduce the Section 232 duty, which is collected in addition to the preferential rate. Proclamation 11002 states that no drawback is available, so paying the duty on a covered server and re-exporting it later does not bring the money back. And covered goods admitted into a foreign trade zone after January 15, 2026 must take privileged foreign status, which fixes them as subject to the duty when they are entered for consumption. Chapter 98 treatment survives, but for goods returned after repair or alteration abroad under 9802.00.60 the duty is assessed on the full value of the article.

For servers that are not covered, nothing in this paragraph changes the usual analysis. Their duty depends on the ordinary HTSUS rate, the country of origin and whatever other additional duties apply on the date of entry.

Who pays the duty when a foreign vendor ships GPU servers into the US?

The importer of record pays it. That is the party named on the entry, which is responsible for classification, valuation, the end-use claim and the payment of duties to CBP. If the sale is Delivered Duty Paid, that party is the seller or an importer of record acting for it.

This is where the Section 232 semiconductor tariff creates commercial risk. A vendor quoting DDP prices for covered GPU servers to a customer whose site does not meet the 100 MW threshold is quoting with a 25 percent exposure that may not be in the price. The reverse also happens: a vendor adds a tariff surcharge, and the buyer later shows that its R&D use was excluded from the start. In both cases the problem starts with an end use nobody documented before the shipment left.

A foreign company can be importer of record in the United States, but it needs an importer number, a customs bond, a resident agent and a broker relationship, and it carries the audit risk for years after the goods are delivered. Our guide to the Importer of Record USA explains that setup in detail. For companies without a US entity, an Importer of Record for GPU servers in the US takes that role on each entry, files with the broker, pays the duties and keeps the file that supports any exclusion claim. If you are new to the concept, start with what an Importer of Record is and when you need one.

Export controls are a separate layer. The Section 232 semiconductor tariff decides what you pay when advanced chips come into the United States; US export controls decide where those chips can go next. A shipment can satisfy one and fail the other, which is why our guide on how to import AI servers and GPUs without a local company looks at both together.

What could change if a second phase arrives?

Proclamation 11002 calls itself a first step. It allows the President to impose significant tariffs on a broader range of semiconductors and derivative products, with a tariff offset program for companies investing in US chip production. As of late October 2026, no proclamation for that second phase has been published.

The proclamation set two checkpoints. The Commerce Secretary and the US Trade Representative had 90 days to report on negotiations with trading partners, and the Commerce Secretary was told to provide, by July 1, 2026, an update on the market for semiconductors used in US data centers, so the President could decide whether to modify the duty. That update has not been made public. Since late summer, press reports have described a broader measure that could reach data center servers, laptops and other finished products, possibly without the current data center carve-out, and in early September the Commerce Secretary said publicly that new semiconductor tariffs are coming. Rates, product lists and dates have not been announced.

For planning, the useful conclusion is simple. Any second phase will need its own proclamation and its own HTSUS changes, so there will be a published text to read before it applies. What it will not wait for is your shipment schedule. Hardware ordered today for delivery in the first half of 2027 should be priced with a clause that says who absorbs a new duty if one applies on the date of entry, and who keeps any refund if a duty is later reduced.

How should you plan a server or GPU shipment under Section 232?

Plan it in the order CBP will read it: product data, end use, classification, documentation, entry and records. Most problems with the Section 232 semiconductor tariff come from deciding the end use after the goods have arrived.

1. Collect the chip data for every accelerator in the order: model, Total Processing Performance and DRAM bandwidth, from the manufacturer's documentation.
2. Classify each line under the right HTSUS subheading and decide whether it falls under 9903.79.01 or 9903.79.02.
3. If a covered article is involved, identify the US end use and the matching exclusion heading, and obtain a signed end-use statement from the buyer before shipping.
4. Agree in the contract who pays the 25 percent if no exclusion applies, and who carries the risk of a future second phase.
5. Confirm who is importer of record, and that the bond covers the expected duties.
6. File the entry with the Chapter 99 headings in the order CBP requires, and keep the specifications, end-use statement and contract with the entry records.

None of these steps is difficult on its own. The difficulty is that they involve the manufacturer, the seller, the buyer, the broker and the importer of record at the same time, often across three or four countries, and the hardware is usually needed on a fixed installation date.

Getting the entry right the first time

For most server and GPU shipments, the Section 232 semiconductor tariff will be a 0 percent line on the entry. For the ones inside the performance windows, the difference between 0 and 25 percent is a documented end use, prepared before the cargo leaves. That is the kind of detail an importer of record should handle as routine, not discover at the border.

If you are shipping GPU servers, accelerator boards or other high-value hardware into the United States and you do not have a US entity, Comex IOR Services can act as your importer of record, check the Section 232 position of each line, prepare the exclusion documentation with your buyer and keep the entry defensible after delivery. Tell us what you are shipping and where it is going through our Importer of Record services page, and we will tell you plainly what the entry will look like.