Seven months after imposing a deliberately narrow 25% tariff on a thin slice of advanced AI chips, the Trump administration is weighing a second, far broader round that would extend duties to the finished machines those chips go into — data-center servers, laptops and gaming consoles — and strip away the exemptions that have kept the American AI buildout effectively tariff-free since January.
Politico reported the deliberations on August 27, citing eight people familiar with the talks. The Guardian and others followed. Nothing is signed. Rates, country treatment, exemptions and timing remain unsettled, and officials have discussed a phase-in period. But the direction of travel has alarmed an industry that spent the first half of 2026 assuming its chip supply was insulated.
What Phase One Actually Did
The current regime rests on a presidential proclamation signed January 14, 2026, under Section 232 of the Trade Expansion Act of 1962, effective for goods entered on or after 12:01 a.m. EST on January 15. It imposed a 25% ad valorem duty on imports under three Harmonized Tariff Schedule subheadings — 8471.50, 8471.80 and 8473.30 — but only where the article contains a logic integrated circuit above specific performance thresholds, such as total processing performance over 20,800 paired with DRAM bandwidth above 5,800 GB/s. Trade lawyers read those parameters as targeting Nvidia H200 and AMD MI325X parts.
It also created seven end-use carve-outs at HTSUS headings 9903.79.03 through 9903.79.09, covering chips for US data centers, repairs, domestic research, startups, non-data-center consumer electronics, civil industrial applications and public-sector use. Those exceptions, it said, would spare goods that “contribute to the buildout of the United States technology supply chain and the strengthening of domestic manufacturing capacity for derivatives of semiconductors.” Between them, they cover the overwhelming majority of what American technology companies actually buy.
Lutnick’s Offset Formula
The proclamation promised any future expansion would come with an offset program “to enable companies investing in United States semiconductor production and certain parts of the United States semiconductor supply chain to obtain preferential tariff treatment.” Commerce Secretary Howard Lutnick favors tying duty-free import allowances proportionally to a company’s committed US fabrication investment: build more here, import more without paying. A version already exists in the Taiwan trade agreement, under which TSMC can import 2.5 times its current US capacity duty-free while plants are under construction, tightening to 1.5 times once operational. Officials have said Korean and Taiwanese firms declining to invest could face tariffs as high as 100%, a stated negotiating position.
Industry representatives argue the arithmetic fails. TSMC holds roughly 73% of the global foundry market and essentially all leading-edge logic production, including the custom silicon Amazon, Google and Microsoft design for their own fleets. Its $265 billion Arizona commitment is the largest foreign direct investment in US history, yet only about 30% of the company’s most advanced capacity is projected to sit there at full build-out. Hyperscalers own no fabs and would generate no quota of their own. One person involved in the talks told Politico that meaningful domestic capacity is more than five years away.
Jonathan McHale, digital policy chief at the Computer and Communications Industry Association, whose members include Amazon, Google and Meta, compared the buildout to the transcontinental railroad. “Anytime you add to the cost and decrease predictability, you make it more difficult to invest, and you are putting that in jeopardy,” he told Politico.
The White House pushed back. “Reshoring semiconductor manufacturing is a top priority for President Trump, whose policies have already secured hundreds of billions of dollars of investments in this key sector,” spokesperson Kush Desai said, adding that unannounced tariff plans amount to speculation. In a statement to Benzinga, the White House was blunter: “Unless officially announced by the Administration, any reporting about tariffs should be regarded as baseless speculation.”
Analysis
The exposure is not the accelerator die — it is the rack. A modern AI server is a chassis full of tariffable content: GPUs, high-bandwidth memory, networking silicon, power delivery. A duty assessed at the finished-server level captures far more value than taxing the chip alone, because it lands on the assembled unit’s full import price. For hyperscalers running capital programs in the tens of billions of dollars per quarter, even a mid-single-digit effective rate on imported compute means billions in incremental capex.
That reaches cloud economics with a lag. GPU capacity is priced off five- to six-year depreciation schedules, so higher landed hardware cost raises the basis for every instance-hour sold off that fleet. Providers can absorb it, pass it to enterprise customers, or slow deployment — and the third option worries the industry most, because compute scarcity, not compute price, is the binding constraint on AI training today.
There is little room to substitute. SK Hynix holds roughly half the HBM market and Samsung about 30%, leaving Micron as the only substantial US-based source. DRAM contract prices rose 90 to 95% quarter over quarter in the first quarter of 2026, and Chinese semiconductors already face combined tariff exposure near 70%. Nintendo, Sony and Microsoft have raised gaming hardware prices on component costs alone; a derivative tariff would stack on top.
The administration’s counterargument holds on its own terms: exemptions shielding the largest buyers remove the pressure that drives reshoring. Phase One tried to hold both goals at once. Phase Two, if the carve-outs go, would choose.
What to Watch
Any expansion requires a new proclamation and HTSUS amendments, so it will surface in the Federal Register first. The Commerce Department’s July 1 report on the data-center semiconductor market, mandated by the January proclamation, is complete but unreleased; it, plus the April assessment of talks with Taiwan, South Korea and Japan, will shape the framework. Watch the length of any phase-in, whether the offset formula counts committed or operational capacity, and whether the data-center exception at 9903.79.03 survives. That single subheading, more than any headline rate, determines what the AI buildout pays.
“Anytime you add to the cost and decrease predictability, you make it more difficult to invest, and you are putting that in jeopardy.”— Jonathan McHale, Digital policy chief, CCIA