Section 232 Semiconductor Tariff Report: What AI Infrastructure Buyers Must Model Now

7 min read

The Section 232 semiconductor tariff process reached a critical deadline on 14 April 2026. The US Trade Representative (USTR) and the Department of Commerce had a statutory deadline on 14 April 2026 to deliver a joint report to President Trump on the outcome of negotiations triggered by January’s Section 232 semiconductor proclamation. That report will either recommend Phase 2 tariff escalation against Taiwan, South Korea, and Japan — the three countries that fabricate nearly all of the chips in scope — or signal a negotiated path that moderates the existing 25% rate. For every organisation buying data centre GPUs, high-bandwidth memory (HBM), or AI networking silicon at scale, this is a direct cost variable, not a background policy event.

The report’s conclusions were not public at time of writing. What is public is the tariff structure already in force, the technical parameters that define it, and the second-wave review due 1 July 2026. Procurement teams waiting for the report before building scenarios are already behind.

What Section 232 Is — and Why This Application Is Different

Section 232 of the Trade Expansion Act of 1962 authorises the President to restrict imports that the Secretary of Commerce has determined threaten national security. Previous Section 232 actions targeted steel and aluminium (2018). The January 2026 proclamation is the first application to advanced semiconductors and represents a structural departure from prior chip-trade policy.

The distinction matters. Export controls administered by the Bureau of Industry and Security (BIS) restrict what US entities can sell to designated foreign parties. Section 232 tariffs operate on the import side: they raise the cost of buying foreign-manufactured chips in the US, irrespective of who is buying. A hyperscaler, an enterprise AI team, and a government contractor all pay the same tariff premium. Unlike export controls, which primarily affect chip vendors’ revenues, Section 232 directly raises procurement costs for chip customers.

The Specific Chips in Scope

The January proclamation is technically narrow in a way that has significant commercial implications. Two performance thresholds define the tariff band:

• Processing performance: chips falling between 14,000 and 17,500 in the applicable performance metric defined in the proclamation
• Memory bandwidth: DRAM configurations with bandwidth between 4,500 and 5,000 GB/s

Both thresholds are drawn tightly around current-generation data centre accelerator products from Nvidia and AMD — the H-series and MI-series GPU families used in large-scale AI training and inference. HBM stacks at the density used in those products sit inside the bandwidth range. The proclamation text also extends coverage to “derivative products” — without defining that term with precision — a deliberate ambiguity that importers are testing through exclusion petitions. Whether fully assembled servers containing in-scope chips qualify as derivative products is one of the active questions in those filings, per trade counsel analysis from Cassidy Levy Kent and Torres Trade Law.

Networking ASICs — Broadcom’s Tomahawk and Jericho series, used in spine-leaf AI fabric designs — based on publicly reported performance specifications, appear to fall outside the January proclamation’s parameter bands. However, importers should verify their specific SKU specifications against the proclamation’s precise threshold definitions before treating any product as definitively excluded. The 1 July Commerce review is specifically examining the data centre semiconductor market, and networking silicon has been identified as a potential expansion target.

What 25% Means for GPU and HBM Procurement Costs Today

The 25% tariff is applied to the customs value of imported chips or derivative products at the US border. For enterprise buyers, the effective landed-cost increase depends on where in the supply chain the tariff is absorbed versus passed through.

For direct chip importers — primarily the large system integrators and hyperscalers who purchase wafers or packaged chips directly — the 25% is a hard cost applied to the import value. On a data centre GPU with a pre-tariff value of $30,000 to $40,000 (current list pricing for high-end H-series accelerators, per multiple technology publications in 2024–2025, though actual contract pricing and customs import value may differ), a 25% tariff adds $7,500 to $10,000 per unit at the import stage. At scale — a 10,000-GPU cluster — that is $75 million to $100 million in additional cost before any other line item. Note: the tariff applies to customs import value, not retail list price. Import value for complex electronics is typically declared at manufacturer transfer price, which may be materially lower than the list figures used here. Procurement teams should verify their actual customs valuation basis before modelling specific dollar impact.

HBM is subject to the same tariff band via the memory bandwidth threshold. SK Hynix and Samsung produce the overwhelming majority of HBM used in AI accelerators; both are South Korean companies, making South Korea one of the primary counterparties in the negotiations whose outcome today’s report assesses. A Phase 2 escalation targeting South Korea specifically would raise HBM import costs beyond the current 25% floor, compounding the GPU cost impact for buyers who are separately sourcing memory. For organisations buying through value-added resellers or OEM system vendors, the tariff is typically embedded in the vendor’s cost of goods and passed through as a list price increase, making it less visible but no less real.

Scenario Matrix: Three Paths from the Report

The report’s conclusions will not resolve all uncertainty immediately, but they will significantly narrow the distribution of outcomes. Three scenarios are worth modelling now.

Scenario A: Phase 2 Escalation

If the report recommends Phase 2 action — higher tariff rates, expanded product scope, or both — implementation may involve a further rulemaking process. If Phase 2 follows a notice-and-comment rulemaking process, buyers could have 60–90 days of advance notice before new rates take effect. However, the January 2026 proclamation itself took effect without a prior public comment period — as did the original 2018 Section 232 steel and aluminium actions — meaning there is no procedural guarantee of advance notice. Procurement teams should model both an immediate-effect scenario and a 60-day notice scenario.

The relevant modelling questions are: Does Phase 2 raise the rate above 25%? Does it extend scope to networking ASICs or to a broader definition of derivative products? Does it impose country-specific differentials (e.g., a higher rate on South Korean HBM than on Taiwanese logic chips)? In a Phase 2 escalation scenario, procurement teams with multi-year hardware refresh cycles should treat the current 25% as a floor and build upside cost exposure into their models. ML Strategies’ 2026 AI policy outlook flagged this timing dynamic as a key procurement consideration earlier in the year.

To illustrate the magnitude: if Phase 2 raises the effective rate on South Korean HBM to 35%, and a standard AI training cluster uses 1,000 HBM stacks at an import value of $3,000 each, the incremental tariff exposure per cluster is approximately $90,000 — a meaningful but manageable figure at enterprise scale, and a potentially significant one for mid-market deployments.

Scenario B: Negotiated Reduction or Tariff Pause

If negotiations with Taiwan, South Korea, and/or Japan have produced commitments sufficient to satisfy the report’s criteria — investment pledges, supply chain disclosures, or other concessions — the report could recommend a rate reduction or a tariff pause pending compliance milestones. This is a materially different procurement environment: teams that have been deferring purchases or building in 25% cost contingencies could have an opportunity to revise models downward.

A negotiated outcome would not necessarily eliminate the tariff; partial reductions or phase-down schedules tied to third-country investment commitments are plausible structures. The exclusion petition pathway is also still live — Cassidy Levy Kent has tracked the administrative process since January, and approved exclusions would reduce effective tariff exposure for individual importers on a product-by-product basis.

Scenario C: Status Quo Continuation

If the report is inconclusive, deferred, or recommends continuation of the current structure pending the 1 July review, procurement planning effectively remains in the current environment: 25% tariff on in-scope chips, exclusion petitions in process, Phase 2 risk unresolved. This is the least useful planning scenario because it sustains uncertainty rather than resolving it. Teams in this scenario should prioritise getting visibility into their vendors’ exclusion petition status and their own contracted pricing structures to determine how much of the tariff they are actually bearing.

What Procurement Teams Should Be Doing Right Now

Regardless of which scenario materialises, several actions are warranted.

Map your tariff exposure precisely. The technical parameter band is specific. Work with your chip vendors to confirm which SKUs you are purchasing fall inside or outside the 14,000–17,500 processing performance range and the 4,500–5,000 GB/s HBM bandwidth range. Products outside the band are not tariffed under the January proclamation. If next-generation chip generations from your vendor roadmap fall outside the band, that changes the total cost of ownership (TCO) calculus for a hardware refresh cycle.

Understand how your contracts handle tariff pass-through. Enterprise hardware contracts negotiated before January 2026 may or may not contain tariff pass-through clauses. If yours do not, you may have a window to renegotiate or lock current pricing before Phase 2 takes effect. If they do contain pass-through language, ensure your cost models are using the correct base price — the tariff applies to import value, not list price.

Assess your exclusion petition eligibility. Torres Trade Law and Cassidy Levy Kent have both published frameworks for evaluating exclusion eligibility. Large importers — including hyperscalers and system integrators importing directly — are the most likely candidates for exclusion petitions. If you are buying through an OEM, ask whether your vendor has filed petitions on the products you purchase.

Flag the 1 July review in your planning calendar. The second Commerce Department review is not a distant event. At 78 days from the April 14 deadline, it represents the next hard policy trigger. Teams building hardware procurement plans on 12- to 24-month horizons should treat 1 July as a second scenario branch point.

The July 1 Data Centre Review: Second Wave of Potential Action

The January 2026 proclamation initiated a separate, parallel Commerce review specifically examining the data centre semiconductor market. That review’s report is due 1 July 2026. Its scope has not been publicly defined, but the White House fact sheet accompanying the proclamation described the data centre review as examining “the national security implications of AI compute infrastructure” — language that suggests scope beyond individual chip products to the broader infrastructure build-out. This is an inference from public framing; the review’s formal scope will be defined by the Commerce Department’s terms of reference when published.

The practical risk is scope expansion. Networking silicon — the switching and routing ASICs that underpin the fabric of large AI clusters — currently sits outside the proclamation’s performance parameters. Memory controllers, high-speed SerDes components, and advanced packaging materials are similarly not currently targeted. If the 1 July review recommends tariff action on any of these categories, the cost impact on a full AI cluster build could be materially larger than the GPU and HBM tariff alone. A 10,000-GPU cluster includes thousands of network switches, cables, and supporting silicon at significant cost — components that have been priced without tariff exposure until now.

What to Watch

Report publication timing. The statutory deadline was 14 April, but the report may not be published simultaneously or in full. Watch for an executive summary or fact sheet from the White House, which has been the administration’s preferred initial communication format for this proclamation series.

Country-specific differentiation. Any signal that Phase 2 rates will differ between Taiwan, South Korea, and Japan — rather than applying uniformly — will have immediate implications for HBM versus logic chip procurement priorities.

Exclusion approval rates. The first wave of exclusion petition decisions is expected in Q2 2026. Approval rates will indicate how narrowly or broadly Commerce is applying the tariff band.

Chip vendor response. Neither Nvidia nor AMD has issued public guidance on how the January proclamation’s parameter thresholds may influence product specification decisions — as of the April 14 report deadline — though this is a question investors and enterprise customers are likely to raise on upcoming earnings calls. If next-generation products are spec’d to fall outside the band, expect product roadmap communications to reference tariff classifications — a new variable in what has historically been a pure performance conversation.

1 July scoping document. Any Commerce publication that defines the scope of the 1 July review before the report itself will be the earliest signal of where the second wave is aimed.

This article was produced with AI assistance and reviewed by the editorial team.

Further reading: AI chip supply chain vulnerabilities | RISC-V as a tariff hedge for AI accelerators