Nvidia's Chips Carry a 25% Tariff; the Server Around Them Does Not

U.S. markets closed higher on Thursday, August 27, carried by AI names. The Nasdaq Composite rose 1.32% to 26,474.45 and the S&P 500 gained 0.62% to 7,723.43. Nvidia climbed close to 10%, according to Forbes, pushing its market value back above $5.5 trillion.

The same day, Politico reported, citing eight people familiar with the discussions, that the administration is considering extending tariffs to the servers that fill AI data centers. Reuters carried the report and said it could not verify it; the White House and the Commerce Department did not respond to requests for comment.

Buyers and sellers of servers both rose that day. That does not mean the market missed the story. To see why, you have to read the text of the tariff schedule rather than the headline.

The gap between the first two figures is the subject of this piece. The third shows how large that gap is in dollars, and the fourth shows how much legal weight the news currently carries.

What Actually Took Effect in January

Proclamation 11002, signed January 14, 2026, imposed a 25% duty on certain advanced computing chips effective January 15. The proclamation labeled the step "Phase 1."

The scope is narrow. The duty applies only to logic integrated circuits under tariff headings 8471.50, 8471.80 and 8473.30 that meet one of two technical parameter sets: total processing performance of 14,000-17,500 paired with DRAM bandwidth of 4,500-5,000 GB/s, or 20,800-21,100 paired with 5,800-6,200 GB/s. Customs advisories cite Nvidia's H200 and AMD's MI325X as the examples.

The duty is therefore tied to a measurement, not to a product category. A good can sit under the same tariff heading and still owe nothing if it fails the test.

The shape of those parameters is worth noting. Two independent customs and legal publications describe the thresholds as closed bands rather than floors: 14,000-17,500 and 20,800-21,100. A floor would describe a class, and every future chip above it would be pulled in automatically. A band describes specific products. The proclamation calling itself "Phase 1" and ordering a negotiation status report within 90 days is consistent with that reading: what took effect in January is a starting point, not a regime.

U.S. Customs and Border Protection built the distinction into the schedule with separate lines. 9903.79.01 carries the 25% for covered chips. 9903.79.02 is a line created for goods under the same headings that do not meet the technical parameters. End-use exemptions follow: 9903.79.03 for chips destined for U.S. data centers of 100 MW or more, 9903.79.04 for repairs and replacements performed in the United States, 9903.79.05 for research and development, and 9903.79.06 for early-stage companies.

The Exemption Is Not an Oversight. It Is Its Own Line.

The structure has three layers, and each opens a separate door.

The first layer is the product test. A server rack is a finished computer, not a logic integrated circuit; it cannot pass the technical threshold and lands in 9903.79.02. Customs advisories state that servers and laptops incorporating covered chips are not directly subject to the Phase 1 duty.

The second layer is the end-use test. Even when a chip is imported on its own, it is exempt under 9903.79.03 if it is bound for a U.S. data center above 100 MW. Nearly every AI data center clears that threshold.

The third layer is the carve-out of electronics from country-based tariffs. On an analysis built from customs data, Apricitas Economics puts the electronics exemption at roughly $34B of imports per month, the single largest such carve-out, and estimates that about half of all U.S. imports currently enter duty-free through exemptions of this kind. The same analysis calculates that without the carve-out importers would have paid roughly $8.9B so far this year at a 10% baseline rate, or $19.2B at current country-specific rates.

Most of that increase is large server computers. Specialized parts from Taiwan run at roughly $30B a year. Over the same period electronics coming from China declined, because the 20% blanket duty applied to China carries no electronics exemption.

The Mechanism: Where the Duty Attaches Decides Who Pays It

A tariff rate on its own tells you nothing. What tells you something is the link in the chain where the duty is triggered.

The version of Phase 2 under discussion changes that arithmetic in two places at once. Moving the duty to the finished server widens the dutiable base from the chip's value to the whole rack. Removing the data center exemption closes the escape hatch. The two steps can be taken separately, but they push in the same direction.

There is a second-order consequence. If the duty attaches to the finished rack, assembling the rack inside the United States becomes the way around it. The Nvidia and Foxconn server plant in Houston sits on that side of the line. But assembling a rack in the United States does not change where the wafer was made; what moves is assembly, not fabrication.

The chart shows where the share price stood while this was being debated. It is not evidence of the report's effect, since earnings landed in the same week.

The Rule Is Already Written Into the Taiwan Agreement

According to Politico, Commerce Secretary Howard Lutnick favors a structure that ties foreign companies' relief from the tariff to their investment in U.S. chip manufacturing. The idea is not new. It already sits in a signed document.

The U.S.-Taiwan trade agreement signed on January 15, 2026 grants graduated relief in exchange for Taiwanese commitments of at least $250B in direct investment and $250B in credit guarantees for U.S. semiconductor and AI production. Under its terms, companies building new chip capacity in the United States may import duty-free up to 2.5 times their planned capacity while construction is under way, and up to 1.5 times their new U.S. production capacity once the plant is complete.

In that structure the rate stays fixed while the bill moves. The same chip, cleared at the same port on the same day, generates two different duties for two importers. What separates them is not the product but the importer's U.S. plant. Lutnick has previously said the rate could reach 100% for countries that do not commit to U.S. manufacturing investment.

The first three lines rest on documents. The fourth rests on unnamed sources.

A Statement or an Instrument

This is where regulatory coverage most often blurs. Three questions have to be asked separately.

What exactly was said? No official said anything on the record. Politico, citing eight unnamed sources, reported that the administration is weighing an option. White House spokesperson Kush Desai offered a general line rather than addressing the substance: "reshoring chip manufacturing is a top priority for the president."

Is there an instrument that makes the preference binding? There is a proclamation in force, but for something else: 11002 covers only chips that clear the technical threshold. No Federal Register notice has been published for servers, laptops or consoles. Politico reported that the framework could be substantially revised in the coming weeks or months, and that a phase-in period is also under discussion.

Does the existing instrument's scope reach this case? No. Line 9903.79.02 explicitly separates goods that sit under the same headings but fail the technical test. A finished server entering duty-free today is not a loophole. It is a written rule.

Jonathan McHale of the Computer and Communications Industry Association framed the objection around uncertainty rather than cost: "Anytime you add to the cost and decrease predictability, you make it more difficult to invest."

Server and chip names rising on the day the report landed says the market does not regard it as binding. That need not reflect optimism; there is simply no instrument to price. Nvidia's earnings were released the same week and were the actual catalyst.

A story that does not move prices is not therefore unimportant. Markets price the rule in force, not the intention behind it. When the rule changes, the repricing happens at once rather than in stages.

What Is Left

An AI server rack entering the United States today passes through three separate doors duty-free: it fails the product test, its destination data center is above the 100 MW threshold, and electronics are held apart from country-based tariffs. All three doors can be closed independently of one another.

Who ends up paying is also open. Duties are paid by the importer, and for AI servers the importer is usually a system builder such as Dell or Super Micro, or the data center operator itself. Server assembly is the lowest-margin link in this chain; a duty worth 25% of a rack's sale price has to be passed through as long as it exceeds what the assembler earns on that rack. In that case the bill lands on the capital expenditure budgets of the large cloud operators.

What to watch is not a headline but a tariff line. A new subheading added to the 9903.79 series, or an amendment to the .03 line, would say far more than an interview. For an investor who can read a set of financials, the question surfaces in gross margin: a 25% increase in server cost feeds directly into the growth assumptions that carry the multiple in a valuation.

This piece draws on the text of Proclamation 11002, customs advisory publications describing the U.S. Customs and Border Protection tariff lines, Reuters and secondary outlets carrying Politico's August 27 report, the published terms of the U.S.-Taiwan trade agreement, and independent analysis of U.S. customs import data. Server rack prices are institutional estimates rather than verified transaction prices and are given as ranges. The Politico report rests on unnamed sources and has not been independently verified by Reuters.