Last reviewed
Jun 7, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 ยท 2,362 words
Add missing citations, update stale details, or suggest a clearer explanation.
Between 2025 and 2026 the United States opened two distinct but interlocking trade fronts on advanced AI chips. The first was a Section 232 national-security investigation into semiconductor imports, launched in April 2025, which culminated in a narrow 25 percent tariff on certain advanced computing chips effective January 15, 2026. The second was the export-control saga governing whether Nvidia and AMD could sell AI accelerators to China at all: an April 2025 license requirement that functioned as a ban on the H20, a July 2025 reversal, an unusual August 2025 arrangement under which the two firms agreed to remit 15 percent of their China AI-chip revenue to the US government, and a January 2026 rule conditionally permitting exports of the more capable H200 and AMD MI325X. The two threads run in opposite directions, one taxing chips coming into the country and the other licensing chips going out, and both remain partly contested and fast-moving. This article distinguishes them carefully and flags what is in force versus threatened or proposed.
Section 232 of the Trade Expansion Act of 1962 lets the Secretary of Commerce investigate whether imports of a product threaten US national security and, if so, recommend that the President restrict them, typically through tariffs. On April 1, 2025, Commerce Secretary Howard Lutnick initiated a Section 232 investigation into imports of semiconductors, semiconductor manufacturing equipment, and their derivative products. A public notice followed in the Federal Register on April 16, 2025, with a comment deadline of May 7, 2025; Commerce received 207 comments, including a detailed submission from the Semiconductor Industry Association.
The probe was broad on paper, sweeping in substrates, bare wafers, legacy chips, leading-edge logic, and downstream products that contain chips. The statute gives Commerce up to 270 days to report findings to the President. Throughout mid-2025 Lutnick repeatedly signaled that tariffs were coming, telling reporters they were likely "in a month or two," but no semiconductor tariff was actually imposed during 2025.
The most attention-grabbing moment came on August 6, 2025, when President Trump, at a White House event promoting Apple's pledge to add $100 billion to its US investment, said he would impose a tariff "of approximately 100 percent" on imported chips and semiconductors, with a sweeping carve-out: companies "building in the United States" or that had committed to build would face "no charge." He named Apple as an example. The remark was a verbal announcement, not a proclamation or executive order, and it left central questions unanswered, including how much US manufacturing would qualify a firm for the exemption. No 100 percent blanket tariff was ever formalized. Firms such as TSMC (which had pledged $165 billion of US investment), Nvidia, and GlobalFoundries had already announced domestic manufacturing commitments that the framing was designed to reward.
The formal action came months later and was far narrower than the August threat. On December 22, 2025, Lutnick transmitted his Section 232 report to the President. On January 14, 2026, Trump signed Proclamation 11002, imposing a 25 percent ad valorem tariff on a narrow category of advanced computing chips, effective 12:01 a.m. eastern time on January 15, 2026. Rather than covering all semiconductors, the tariff applies only to chips meeting defined Total Processing Performance (TPP) and DRAM-bandwidth thresholds, the same kind of high-end AI accelerators at the center of the export debate, such as the Nvidia H200 and AMD MI325X, plus certain derivative assemblies.
The proclamation carries extensive exemptions that blunt its practical reach. It does not apply to chips imported for US data centers, for repairs or replacements performed in the US, for US research and development, for US startups, for non-data-center consumer or civil-industrial applications, or for US public-sector use. Because most advanced AI chips entering the country are bound for domestic data centers, analysts characterized the measure as a narrow first phase. The proclamation explicitly frames a second phase: after trade negotiations conclude, the President "may consider imposing significant tariffs" at higher rates, paired with a tariff-offset program rewarding firms that invest in US semiconductor production. That broader, higher-rate tariff remains proposed, not in force.
Running parallel to the import tariff was a separate fight over selling US AI chips to China, governed by export controls administered by the Bureau of Industry and Security (BIS).
Nvidia had designed the H20 specifically for the Chinese market to comply with earlier limits. On April 9, 2025, the government told Nvidia that exporting the H20 (and any chip with comparable memory or interconnect bandwidth) to China would require a license, and that the requirement would stand indefinitely. Officials had pressed for tighter controls partly over reports that H20s were used to train models from DeepSeek. The license rule functioned as a ban. Nvidia initially disclosed a $5.5 billion charge tied to H20 inventory and commitments; it later revised the first-quarter charge to $4.5 billion after reusing some materials. AMD faced parallel restrictions on its MI308.
The policy reversed roughly three months later. In mid-July 2025, following lobbying by Nvidia chief executive Jensen Huang, including meetings at the White House and in Beijing, the administration assured Nvidia that H20 licenses would be granted, and Commerce began issuing them. AMD likewise won clearance to resume MI308 sales.
The reversal came with a price unusual in US trade history. Reported in August 2025, Nvidia and AMD agreed to remit 15 percent of their revenue from China AI-chip sales to the US government in exchange for the export licenses, covering the H20 and the MI308. Trump had initially floated 20 percent before the rate was set at 15 percent following a Huang-Trump meeting. Corporations do not normally pay the federal government a cut of export revenue, and trade lawyers quickly questioned the arrangement's legality, pointing to the Constitution's Export Clause (Article I, Section 9), which bars taxes or duties on exports, and to the Export Control Reform Act of 2018, which prohibits charging exporters for licenses. Skeptics countered that the payments might be treated as a voluntary business decision rather than a tax, and that it was unclear who would have standing to sue, making a successful challenge unlikely.
The next escalation moved up the product stack to more capable silicon. On December 8, 2025, Trump announced he would permit sales of the H200, a substantially more powerful accelerator than the H20, to approved customers in China. BIS formalized the shift in a press release dated January 13, 2026 and a final rule that, by most accounts, took effect around January 15, 2026. The rule moved license review for the Nvidia H200, AMD MI325X, and similar chips from a presumption of denial to case-by-case evaluation. Approval is conditioned on the exporter showing the sale will not reduce semiconductor capacity available to US customers, the Chinese buyer adopting export-compliance and customer-screening procedures, and the product passing independent US-based third-party testing of its performance and security.
The rule also imposes hard ceilings. Eligible chips must fall below a TPP score of about 21,000 and DRAM bandwidth under roughly 6.5 TB/s; reporting put the H200 around a TPP of 15,832 with 4.8 TB/s and the MI325X near 20,800 with about 6 TB/s. A volume cap limits aggregate China-bound shipments of a given product to 50 percent of the volume of that same product shipped into the United States, which effectively bars China-only SKUs. The financial surcharge attached to H200 exports has been reported inconsistently: some accounts describe a continuation of the 15 percent revenue share, while others, including reporting cited by TrendForce, describe a 25 percent revenue remittance to the US government. That figure should be treated as unsettled, and it is distinct from the separate 25 percent Section 232 import tariff. By mid-2026 Commerce had cleared roughly ten Chinese firms, reportedly including Alibaba, Tencent, ByteDance, and JD.com, plus distributors such as Lenovo and Foxconn, to buy the H200, each capped at 75,000 units.
The combined effect by mid-2026 was a system in which the legal pathway existed but commerce did not flow. As of May 2026, despite US approvals, not a single H200 had reportedly shipped to China. Chinese firms pulled back after guidance from Beijing, which discouraged purchases on grounds that heavy reliance on US chips could stall domestic development, that the revenue-remittance arrangement was objectionable, and that imported accelerators raised security concerns. Washington had opened the door, and Beijing largely kept its companies from walking through it. Nvidia's CFO commentary reflected effectively zero H200 China revenue while the company absorbed the earlier H20 charges.
On the import side, the January 2026 tariff's broad exemptions limited the immediate cost to US data-center buildouts, since data-center-bound chips are largely carved out. The larger uncertainty was phase two: the proclamation's signal of significantly higher future tariffs, conditioned on negotiations, alongside an offset program meant to steer investment into domestic semiconductor fabrication. Analysts including the Information Technology and Innovation Foundation warned that broad semiconductor tariffs could raise costs across the electronics supply chain and weigh on US growth.
Reception was sharply divided and, on several specifics, openly skeptical. The revenue-share pact drew the most pointed criticism: legal commentators called it "bizarre" and "ripe for a legal challenge," citing the Export Clause and the Export Control Reform Act, even as they conceded enforcement and standing hurdles made litigation improbable. National-security analysts at CNAS and elsewhere questioned whether loosening H200 access squared with the original rationale for restricting AI chips to China. Chinese state media simultaneously cast doubt on the chips, warning domestic firms about security risks, which dovetailed with Beijing's discouragement of purchases. Supporters in the administration framed the package as leverage: tariffs and licenses together would pull manufacturing onshore while extracting value from a China market that would otherwise be served by smuggling or domestic accelerators. With phase-two tariffs unresolved, the H200 revenue-share figure disputed, and shipments stalled, the policy remained, as of mid-2026, a contested work in progress more than a settled regime. See also AI regulation and broader US AI policy.
| Date | Action | Status |
|---|---|---|
| April 1, 2025 | Commerce initiates Section 232 investigation into semiconductor imports | Completed (report Dec 2025) |
| April 9, 2025 | Nvidia told H20 exports to China require a license (de facto ban); $5.5B charge disclosed, later revised to $4.5B | Superseded by July reversal |
| April 16, 2025 | Section 232 public-comment notice published; comments due May 7, 2025 | Completed |
| July 2025 | Administration reverses H20 ban; Commerce begins issuing H20 and MI308 licenses | In force |
| August 6, 2025 | Trump verbally floats "approximately 100%" tariff on imported chips, exempting US builders | Threatened, never formalized |
| August 2025 | Nvidia and AMD agree to remit 15% of China AI-chip revenue for export licenses | In effect (legality disputed) |
| December 8, 2025 | Trump announces H200 sales to approved China customers will be allowed | Announced |
| December 22, 2025 | Lutnick transmits Section 232 report to the President | Completed |
| January 13, 2026 | BIS revises license review for H200/MI325X to case-by-case | In force |
| January 14, 2026 | Proclamation 11002 imposes 25% Section 232 tariff on certain advanced chips | In force |
| January 15, 2026 | 25% advanced-chip tariff takes effect (12:01 a.m. EST) | In force |
| ~May 2026 | ~10 Chinese firms cleared to buy H200 (75,000-unit cap each); no chips shipped | Stalled |