Your Data Center's Chips Have Been Tariff-Free Since January. That's About to Cost You a Factory.
A 25% US tariff on advanced AI chips has exempted anything headed for a data center since mid-January. Washington is now reportedly weighing whether to kill that exemption outright — and replace it with a quota sized to how much you've committed to building in an American fab.
Play with the investment-tied tariff quota
Illustrative model, not the actual proposal: assumes a hyperscaler needs to import $30B of tariff-covered chips a year, and that under the reported "Phase 2" formula, every $1B of committed US fab investment buys $1.5B of duty-free import allowance. The real ratio, thresholds, and demand figures haven't been made public. Drag the slider, then compare the two rules.
Under Phase 2, this buys a $15B duty-free allowance — half your chip imports still get taxed.
The plain version
Picture a customs checkpoint that opened back in January: any advanced AI chip entering the US pays a 25% toll — Nvidia's top data-center GPUs, AMD's, all of it — unless the truck is headed straight for a data center, a research lab, a startup, or a laptop factory. Those trucks get waved through free. That free lane is a big reason building AI data centers in the US hasn't gotten more expensive this year.
Now Washington is reportedly rethinking the free lane, and reportedly wants to widen the toll booth's reach too — not just the raw chip, but the laptop, server, or game console built around it. Multiple outlets, first Politico and then CNBC and others, report the administration is drafting a "Phase 2" that would scrap the free-lane exemptions altogether.
The replacement idea, championed by Commerce Secretary Howard Lutnick, isn't "pay the toll" or "don't." It's a quota: the amount of chips you can bring in tariff-free gets sized to how big a factory you've promised to build on US soil. Pledge a lot of fab investment, and you get a bigger free lane. Pledge nothing, and even your data-center chips start getting the same 25% haircut everyone else pays.
The catch reporters keep flagging: the free lane this formula implies is reportedly much narrower than what the biggest cloud companies actually need to import during the current AI building spree. So even a company making a real factory pledge could still eat a large tariff bill on everything above its quota. Nothing here is locked in — no rate, no exemption list, no start date has been finalized, and officials say it could still change substantially over the next few months.
Why it matters for tech + supply chain: the exemption that's kept AI data-center construction insulated from tariffs for eight months might not survive the fall — and the fix on the table doesn't remove the cost, it just prices it in factories instead of dollars.
The expert version
Presidential Proclamation 11002, issued January 14, 2026 under a Section 232 national-security investigation, imposed a 25% ad valorem tariff (HTSUS 9903.79.01/.02) on semiconductor articles meeting defined Tensor Processing Performance and DRAM-bandwidth thresholds, effective January 15 — a narrowly scoped category that captures parts like Nvidia's H200 and AMD's MI325X. The proclamation simultaneously created a carve-out (HTS 9903.79.03) exempting imports that support US data-center buildout, R&D, startups, and consumer end-uses from the duty.
Reporting first attributed to Politico on August 27, and corroborated by CNBC, Tom's Hardware, and Tech Times, describes a "Phase 2" package under internal review that would (1) extend Section 232 coverage from the bare semiconductor article to downstream finished goods incorporating covered chips — servers, laptops, gaming consoles — and (2) eliminate the 9903.79.03 exemption categories rather than merely narrow them.
The mechanism reportedly favored by Commerce Secretary Howard Lutnick replaces the blanket exemption with a tariff-rate-quota-like structure: duty-free import volume allocated per company in proportion to its committed US semiconductor-manufacturing capital investment, rather than to the import's end use. Above that allowance, imports would be dutiable at the standard rate. Structurally this converts a use-based exemption into an investment-indexed quota — a policy instrument more familiar from agricultural tariff-rate quotas than from tech trade policy, now aimed at reshoring fab capacity rather than just raising revenue or curbing volume.
Reporting on the same story notes industry concern that the quota volumes implied by Lutnick's formula are structurally short of current hyperscaler chip-procurement volumes during the ongoing AI capital-expenditure cycle, meaning even firms with substantial announced fab commitments could face material tariff exposure on the balance. No rate schedule, exemption taxonomy, or effective date has been finalized; officials describe a possible staggered rollout, and the framework remains subject to change over the coming months.
Why it matters for tech + supply chain: it would convert Section 232 chip policy from a use-based carve-out into a capital-investment-indexed allocation mechanism — forcing hyperscalers, OEMs, and chipmakers to model tariff exposure as a function of their own reshoring commitments rather than end-use classification alone.