Twelve days ago the Commerce Department was due to hand the President a report that will decide whether the United States begins taxing the semiconductors its own bid for artificial-intelligence dominance is built on. The July 1 deadline set in Proclamation 11002 has passed with no public verdict, which means the decision is live, imminent, and undecided.¹ The administration now faces a contradiction of its own construction: it has declared advanced chips essential to national security, then built a tariff machine whose logical next move is to make those chips more expensive to import. The narrow 25% duty imposed in January was labeled Phase 1 for a reason. Phase 2 is the question, and the honest answer is that broad semiconductor tariffs would tax the one input the United States cannot yet make at scale and cannot afford to slow down.

Tariffs Are a Tax on the Past

Tariffs have almost always been a tax on the past: on the sunset industry a government is trying to shield from a rising competitor. The steel duties of the 1980s protected mills that were losing to more efficient foreign producers. The logic was always the same, and always the same trap. Protection buys time for the protected and raises costs for everyone downstream who buys what the protected industry makes. A steel tariff is a subsidy to steelmakers paid for by carmakers, appliance manufacturers, and the construction trades. The economics profession has spent a century documenting that the downstream bill usually exceeds the upstream benefit.

What makes 2026 different is that the government is proposing to run the same play on a sunrise industry. Semiconductors are the foundational input to the fastest-growing capital formation story in the modern economy: the hyperscale build-out of AI infrastructure by Google, Amazon, Meta, and Microsoft, which collectively planned to spend more than $350 billion on data centers in 2025 alone.² The tariff instrument suits industries with abundant domestic capacity and thin foreign competition.

Semiconductors have neither. The United States consumes roughly a quarter of the world's chips and fully manufactures only about 10% of what it needs.^[1:1] A domestic industry at the scale demanded does not yet exist, so protection has nothing to protect. The tariff can only fall on the imports standing in for it.

The Tariff Is Built to Miss Its Target

The mechanics of Proclamation 11002 reveal the contradiction with unusual clarity. On January 14, the President invoked Section 232 of the Trade Expansion Act, the national-security tariff authority previously used for steel, aluminum, and autos, and imposed a 25% ad valorem duty on a deliberately narrow category of advanced logic chips: the class exemplified by Nvidia's H200 and AMD's MI325X. The duty took effect the next morning.^[1:2] This is assessed with high confidence; it is a matter of published record in the Federal Register.³

The revealing part is what the proclamation exempted. The 25% duty explicitly does not apply to chips imported for use in United States data centers, for domestic research and development, for startups, for public-sector applications, or for consumer and civil-industrial uses outside data centers.^[1:3] Read that list again. It exempts nearly every economically significant use of an advanced AI chip on American soil. The tariff, as written, is engineered to miss its largest target. The drafters knew what they were doing. Taxing the compute going into American data centers would raise the cost of the very build-out the administration has made a centerpiece of its technology strategy, so they carved it out. A tariff that must exempt its own most important use case is a negotiating posture wearing the costume of a trade barrier.

The negotiating logic is confirmed by the Phase 2 architecture. Commerce recommended, and the proclamation preserves, the option of broader tariffs at a rate it described only as "significant," paired with a tariff-offset program rewarding companies that invest in domestic fabrication.^[1:4] Secretary Lutnick has publicly warned that South Korean and Taiwanese firms declining to build in the United States could face duties of up to 100%.⁴ The day after the proclamation, the United States and Taiwan signed a framework letting Taiwanese firms building domestic capacity import up to 2.5 times their planned output duty-free during construction.⁵ The tariff, in other words, is leverage to extract fab commitments, with TSMC's $165 billion Arizona expansion as the template.⁶ This is estimated with moderate-to-high confidence, drawn from the proclamation's own text and the Secretary's statements; no grand strategy has been stated outright.

The cost of getting Phase 2 wrong is quantifiable and large. CSIS analysis estimates that a 100% tariff on semiconductors could raise AI-server costs by as much as 75% and impose $75 to $100 billion in additional AI-infrastructure costs over five years, equivalent to fifteen to twenty fewer hyperscale facilities.^[2:1] Memory prices are already projected to climb 30 to 50% through mid-2026 on AI demand alone, and Nvidia has already raised high-end accelerator prices.^[2:2] Layering a broad tariff on an existing supply-driven price surge raises the cost of the compute the United States is racing to deploy. Fabs take three to five years to build, so reshoring arrives long after the decisive window has closed. Speed is the entire competitive thesis, and the tariff spends it.

A Strong Hand, Played Badly

The strategic risk is a self-inflicted wound disguised as toughness. The United States holds a genuine and durable advantage in the semiconductor value chain: design software, high-performance architectures, and the lithography chokepoints its allies control. That advantage compounds with each generation precisely because it sits upstream, where switching costs are highest. A broad Phase 2 tariff attacks the problem from the wrong end. It leaves the upstream chokehold on China untouched and raises the downstream cost of American compute. It is possible to hold the strongest hand in the game and still lose money playing it badly.

There is a second-order risk in the tariff-as-leverage strategy itself. Leverage works only if the threat is credible and the counterparty believes you will use it. A 100% duty on chips bound for American data centers is a threat the administration cannot execute without taxing its own AI ambitions, and sophisticated counterparties in Taipei and Seoul know it. The exemption list telegraphs the bluff. The leverage retains real value, since reshoring commitments matter, but it has a ceiling: the administration can press only so hard before the threat becomes self-harming, and that ceiling sits lower than the rhetoric suggests.

The divergent view deserves a fair hearing. A serious analyst can argue that dependence on foreign fabrication, concentrated overwhelmingly in Taiwan, is a national-security vulnerability so acute that near-term cost is worth paying to force reshoring, and that only a credible tariff threat moves capital at the required speed. There is merit here: the Taiwan concentration is real, and export controls alone have not built domestic capacity. Where this view strains is on timing. Tariffs raise costs immediately and build fabs slowly, and the AI race is being run on a two-to-three-year clock. The vulnerability is real, and the instrument is mismatched to it.

Watch Which Instrument Ships First

The recommendation is restraint disguised as patience: keep the narrow Phase 1 tariff and the data-center exemptions intact, publish the tariff-offset program to reward domestic investment with a carrot, and decline the broad Phase 2 duty that would tax American compute during the exact window when compute is the competition. The offset program does the strategic work, pulling fab investment onshore, without the self-inflicted cost of a general tariff that reshores nothing on a relevant timeline. If the President wants leverage over Taipei and Seoul, the framework deals already provide it; the threat of Phase 2 is more useful held in reserve than spent.

For readers positioned in this economy, the decision framework is simpler than the politics. Anyone building or financing AI infrastructure should treat a broad Phase 2 tariff as a low-probability, high-severity event and price the option accordingly: bulletproof end-use documentation to secure the data-center exemption, sourcing diversified across allied fabs, and contracts that allocate tariff risk explicitly rather than by default. The tell to watch is the offset program. If it ships before a broad tariff does, the administration has chosen the carrot and the exemptions will hold. If a "significant" duty lands first, the bluff was real, and the cost of American compute is about to rise. Watch which instrument the administration reaches for first. That choice is the policy.