Tariffs on AI Chips Aren't Really About Trade — Here's What's Actually at Stake

On January 15, 2026, a new 25% tariff on a narrow category of advanced AI semiconductors took effect in the United States — covering chips like Nvidia's H200 and AMD's MI325X. Read the fine print, though, and it becomes clear the tariff isn't primarily a trade measure at all.

The legal basis

The tariff was imposed under Section 232 of the Trade Expansion Act of 1962 — the same national-security authority that lets the president adjust duties on goods deemed to threaten US security interests when imported under certain conditions. Framing an AI-chip tariff as a national-security measure, rather than ordinary trade policy, already signals the administration's real concern: dependency, not just trade balance.

The exemptions tell the real story

The tariff comes with broad end-use exemptions — for data-center, R&D, startup, consumer-electronics, industrial, and public-sector uses. But the more revealing mechanism is the US-Taiwan semiconductor deal: companies that commit capital to TSMC Arizona's construction can import AI chips tariff-free at a multiple of their invested capacity — 2.5x during construction, dropping to 1.5x once operational. Amazon, Google, and Microsoft have already committed capital at the scale needed to cover their own GPU procurement volumes under this formula.

What that structure actually incentivises

Put together, the tariff isn't designed to raise revenue or simply restrict imports — it's designed to make domestic manufacturing investment the cheapest way to avoid the tariff altogether. Hyperscalers that fund US chip-fabrication capacity buy their way out of the tariff; those that don't, pay it. That's industrial policy wearing a trade-policy label.

The dependency problem underneath it all

The US currently manufactures only around 10% of the chips it requires domestically — the structural vulnerability this entire policy is aimed at. A two-phase approach is reportedly planned: Phase 1 is this narrow, immediate 25% tariff; Phase 2 involves potentially broader tariffs on semiconductors and derivative products, paired with a tariff-offset program to further incentivise domestic manufacturing.

What it means for AI hyperscalers and chip buyers

If your company is a major AI-compute buyer, the TSMC Arizona investment-multiplier mechanism is the detail worth modelling now — it's a genuine tariff-avoidance path, but only for companies that can commit capital at hyperscaler scale. Everyone else is likely paying the 25% as a straightforward cost of doing business in advanced AI compute.

---

Sources: Thomson Reuters · Thompson Hine · Mayer Brown