The Carveout Is the Story: Data Center Tariff Policy Reaches Its Most Dangerous Inflection Point as Anthropic Ships Sonnet 5

The Real Deadline Is Today

Anthropic's release of Claude Sonnet 5 this week is the product story. The infrastructure story is a Commerce Department report due to the White House today, July 1, 2026, that will determine whether imported AI chips flowing into U.S. data centers stay exempt from the 25% Section 232 semiconductor tariff.

By July 1, 2026, the Secretary of Commerce must provide the President with an update on the market for semiconductors used in U.S. data centers, so that the President may determine whether it is appropriate to modify the tariff imposed in Proclamation 11002. That proclamation, signed January 14, imposed a 25% ad valorem duty on advanced AI accelerator chips including the NVIDIA H200 and AMD MI325X, but broad exemptions exist for U.S. data centers, startups, R&D operations, and public sector applications, meaning the practical impact of the Phase 1 tariff falls on a narrow slice of commercial imports that do not qualify for these carve-outs.

The carveout is not guaranteed to survive Phase 2. In the second phase, after trade negotiations concluded, the Commerce Secretary recommended broader tariffs on semiconductors at a rate of duty that is significant, accompanied by a tariff offset program to enable companies investing in U.S. semiconductor production to obtain preferential tariff treatment. What that means for the data center exemption remains unresolved.

Why the Carveout Matters: The $90 Billion Arithmetic

The Computer and Communications Industry Association ran the numbers on what happens if the exemption disappears. Applying 25% Section 232 semiconductor tariffs to data centers would represent a 15.6% effective tax on data center construction. The mechanism is straightforward: servers, storage, and networking equipment represent approximately 52%, 12%, and 1.5% of total data center capital expenditures respectively, and within the server category, semiconductors account for roughly 81% of value in traditional data centers and up to 87% in AI-optimized facilities.

A 15.6% tax on data center construction would lead to the relocation, cancellation, or delay of about 20% of planned 2026-2030 data center buildouts in the U.S., costing about $450 billion in capex and GDP between 2026 and 2030, or about $90 billion per year, and threatening 243,000 jobs.

Take this as a signal of magnitude, not gospel. The CCIA has a direct stake in the outcome. CSIS, with no such interest, reached similar conclusions independently: a 100% tariff on all semiconductors and products containing them would likely impose an additional $1.4 trillion burden on the buildout, and even more moderate tariff scenarios would function as a tax on U.S. AI ambitions.

The Supply Chain Cannot Be Substituted

The policy tension is structural. U.S. Census Bureau figures show that the United States imports more than 95% of its servers, with Taiwan alone supplying $86 billion in 2025, and Taiwan's Ministry of Economic Affairs reports that Taiwan produces 100% of U.S.-brand AI servers.

Domestic chip manufacturing cannot grow quickly enough to meet U.S. AI buildout needs through 2030. TSMCs Arizona facility has reached only a fraction of the volume needed to substitute for imports, and at production costs 50% higher than Taiwan. A tariff designed to force faster substitution doesn't create domestic supply. It creates a cost shock and deferred investment.

One worrisome policy concern is that while President Trump's AI Action Plan prioritizes data center construction, his approach to tariffs makes this effort far more expensive. Alternative U.S. suppliers that could reduce costs are also limited, highlighting the inconsistencies between AI national goals and policy implementation.

Who Gets Hurt First

The tariff risk cuts differently across operators. For self-funded hyperscaler projects, which represent roughly 60% of the buildout and are paid from operating cash flow and board-approved capex envelopes, a 15.6% shock slows the pace of later-phase additions and may shift the locations of planned facilities, but will rarely push a specific facility already underway from go to no-go. Roughly 10% of self-funded hyperscaler planned 2026-2030 data center capex gets delayed outside the window or relocated outside the United States.

The pain concentrates downstream. For merchant developers and AI-native operators like Stargate, CoreWeave, Crusoe, Lambda, Vantage, and similarly situated entities, which represent roughly 30% of the buildout, the math is very different. These projects are debt-financed against equity IRR projections commonly in the 11 to 19% range. A 15.6% effective capex tax collapses those projections. Marginally viable projects stop.

Equipment vendors show strain. Lam Research and KLA each face an estimated $350 million annual impact from the broader tariff environment. ASML withdrew its 2026 revenue growth guidance mid-year, citing tariff uncertainty.

Supply chain friction predates any tariff decision. Medium-voltage switchgear now carries a 22-month lead time, while high-capacity transformers take 18 months, and globally, the average lead time for critical components is 33 months. Bloomberg has reported that almost half of U.S. data centers planned for 2026 are facing delays due to infrastructure supply chain strain, and that is before adding a Phase 2 semiconductor tariff shock to the existing strain.

Where Claude Sonnet 5 Sits in This

Claude Sonnet 5 launches with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, and is built to be the most agentic Sonnet model yet, capable of making plans, using tools like browsers and terminals, and running autonomously. Its performance is close to that of Opus 4.8, but at lower prices.

The model is technically sound and competitively priced. What Anthropic cannot control is whether the inference infrastructure needed to serve it at scale remains economically viable to build inside the United States. Anthropic, like xAI and every other AI-native company without a hyperscale balance sheet, depends on merchant and co-location operators building that capacity. Those operators are precisely the ones most exposed to a tariff-driven IRR collapse.

For compliance teams, the practical implication is clear: the current narrow tariff may expand significantly in the second half of 2026. Today's Commerce Department report is the pivot point.

What to Watch

Step 1 (today): Monitor whether the Commerce Department's July 1 data center semiconductor market update is published or leaked. The content will either confirm the carveout is safe, signal Phase 2 expansion, or hand the decision to the President with ambiguous language. Ambiguity itself carries risk for anyone with an open capex decision.

Step 2 (this quarter): Watch Phase 2 tariff scope. The relevant modeling questions are whether Phase 2 raises the rate above 25%, whether it extends scope to networking ASICs or a broader definition of derivative products, and whether it imposes country-specific differentials such as a higher rate on South Korean HBM than on Taiwanese logic chips.

Step 3 (ongoing): Track data center partnership announcements from Anthropic and competitors. Offshore or non-U.S. co-location deals signal hedging. Domestic partnerships signal confidence in the carveout surviving.

Step 4 (Series B/C founders): If your infrastructure cost model assumes current carveout terms hold through 2027, remodel with carveout removal and present both scenarios to your board before your next raise. Investors will ask.