Bottom Line Up Front
- Applying 25% Section 232 semiconductor tariffs to data centers would represent a 15.6% tax on data center construction.
- When you tax something, you get less of it.
- Data center buildout represented 92% of U.S. real GDP growth in the first half of 2025.
- 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.
- This would cost the U.S. about $450 billion in capex and GDP between 2026-2030, or about $90 billion per year.
- That amounts to a roughly 0.29% GDP loss annually, and threatens 243,000 jobs.
Background
The U.S. artificial intelligence buildout of data centers and associated infrastructure is the single largest channel of productive investment in the American economy today. Information processing equipment and software accounted for roughly 92 percent of U.S. real GDP growth in the first half of 2025, per analysis by Harvard economist Jason Furman. McKinsey projects U.S. data center capital expenditures will reach $2.7 trillion between 2025 and 2030. The trajectory of the U.S. economy through the end of the decade depends, to a degree without modern precedent, on whether this buildout proceeds at its current pace.
Phase 2 of the Section 232 semiconductor tariff regime, if applied to data centers, would place that buildout at material risk. Under the January 14, 2026 Presidential Proclamation, the Administration imposed a 25 percent duty on advanced computing chips and certain derivative products but explicitly exempted “Covered Products for use in United States data centers.” If Phase 2 removes that exemption and extends duties to chips and all chip-containing derivative products, including servers, it will harm U.S. investment and cause ripple effects throughout the economy. There is a strong economic case for ensuring U.S. data centers and companies manufacturing in the U.S. avoid premature semiconductor tariffs while adequate domestic semiconductor capacity is further developed and scaled up.
Tariffs Cause a Cost Shock
Three primary parameters determine the magnitude of the shock the tariff would deliver to the U.S. data center buildout.
The first is the share of data center spending that goes to computing equipment. IoT Analytics, in its November 2025 Data Center Equipment & Infrastructure Market Report, finds that 78 percent of data center spending in 2024 went to IT infrastructure: servers (61 percent), networking (10 percent), and storage (6.5 percent). The transition to AI-optimized facilities has, if anything, increased this share, as Dell’Oro Group reported in March 2026.
The second is the share of that compute stack the United States imports. The World Trade Organization March 2026 Global Trade Outlook and Statistics estimated an import intensity of 70-90%. Trade data confirms this: U.S. Census Bureau figures show that the United States imports more than 95 percent of its servers, with Taiwan alone supplying $86 billion in 2025, and Taiwan’s Ministry of Economic Affairs reports that Taiwan produces 100 percent of U.S.-brand AI servers. The capacity to substitute domestically does not exist at scale in the relevant time horizon; Taiwan also accounts for 92 percent of global advanced logic capacity at the leading-edge nodes that AI training requires. The midpoint of the WTO range, 80 percent, is used as our central parameter for the data center channel.
The third is the tariff rate itself: assumed to be 25 percent under the current Phase 2 design, and assumed to cover both semiconductors and semiconductor derivative products such as servers.
Multiplying these together produces the effective tax on every dollar of U.S. data center investment: 78 percent * 80 percent * 25 percent = 15.6 percent. Given the highly consolidated supply chain and lack of domestic alternatives, we assume the pass-through is expected to approach 100%. This estimate also assumes no double-taxation in cases where a semiconductor is imported, then shipped out of country to be packaged into a derivative product like a server, and then re-imported.
Applied to McKinsey’s $2.7 trillion buildout total for 2025 to 2030, multiplied by ⅚ to account for 2025 being out of tariff coverage, the gross five-year tariff cost is $351 billion, or roughly $70 billion per year, before accounting for project cancellations in response to tariffs. That figure is on par with the entire annual capital expenditure budget of any single U.S. hyperscaler. It is a cost shock without modern parallel directed at the single most economically significant investment channel in the country.
What the Cost Shock Would Do
The 15.6 percent capex tax does not, however, get evenly distributed across the buildout and absorbed by everyone. It pushes a fraction of currently planned projects from workable to unworkable. Cancellation, delay beyond the 2030 window, or relocation to non-U.S. jurisdictions are all possibilities explored in our impact scenario.
For self-funded hyperscaler projects, which represent roughly 60 percent of the buildout and are paid from operating cash flow and board-approved capex envelopes (A single hyperscaler’s FY25 envelope alone is $80 billion), a 15.6 percent shock slows the pace of later-phase additions and may shift the locations of planned facilities. It will rarely push a specific facility already underway from go to no-go. Affected rate: roughly 10 percent of self-funded hyperscaler planned 2026-2030 data center capex gets delayed outside the window or relocated outside the United States.
For merchant developers and AI-native operators like Stargate, CoreWeave, Crusoe, Lambda, Vantage, and similarly situated entities, which represent roughly 30 percent of the buildout, the math is very different. These projects are debt-financed against equity IRR projections commonly in a 11 to 19 percent range. A 15.6 percent capex shock consumes roughly the entire equity IRR cushion on a typical merchant data center deal. For example, the Stargate Abilene 600 MW cancellation in March 2026 occurred under financing strain before any Phase 2 shock, showing that shocks can and do lead to cancellation, relocation, or scope reduction for large scale data center projects. Affected rate: roughly 40 percent of merchant developer and AI-native operator planned 2026-2030 data center capex gets delayed outside the window, relocated outside the United States, or canceled outright.
For the enterprise, colocation, and offshoring tail representing the final 10 percent of the buildout, a 15.6 percent U.S.-specific cost premium makes Canadian, Irish, Nordic, and Gulf alternatives more attractive, since these locations house the same hyperscalers’ existing campuses. Affected rate: roughly 20 percent of enterprise, colocation, and offshoring planned 2026-2030 data center capex gets delayed outside the window or relocated outside the United States.
Weighting the segments produces a central estimate that approximately 20 percent of planned U.S. AI data center capacity between 2026 and 2030 (about $450 billion in capex) would be cancelled, delayed beyond the 2030 window, or relocated abroad under a 15.6 percent effective tax. Bloomberg has already 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.
The translation of foregone investment into GDP is not a one-for-one identity but the product of two well-anchored adjustments that roughly offset. Of every dollar of foregone U.S. data center capex, approximately 65 cents is direct U.S. value-added after netting out imported equipment, but adding back the design margins U.S. firms capture on chips fabricated abroad (NVIDIA’s FY2025 gross margin was 75 percent on GPUs physically built in Taiwan; SIA, the SIA 2025 Factbook, and BCG find U.S. firms capture roughly half of global semiconductor value despite producing only 10–12 percent of physical wafers). Applying a conservative 1.7 multiplier for indirect and induced effects, which is near the low end of BEA RIMS II nonresidential construction multipliers and below the 2.0 to 2.5 range data-center-specific studies report, yields a total GDP impact of approximately $1.10 per dollar of foregone capex. The 1:1 capex to GDP ratio used below is therefore slightly conservative.
Spreading $450 billion over five years produces an annual GDP loss of approximately $90 billion, representing about 0.29 percent of U.S. nominal GDP, which the Bureau of Economic Analysis reports at $31.4 trillion as of Q4 2025. Translated through Okun’s Law (the empirical relationship between output gaps and unemployment, “a one percent deviation of output from potential causes an opposite change in unemployment of half a percentage point,” per Ball, Leigh & Loungani 2017) and applied to the U.S. labor force of 170 million per the BLS, this implies an unemployment increase of 0.14 percentage points and approximately 243,000 jobs at risk.
How Policy Design Sensitivities Reshape the Harms
Three policy variations would meaningfully reduce these effects, although none of these policy modifications would be as beneficial as simply continuing the exemption for U.S. data centers.
Lower the rate from 25 percent to 10 percent: This cuts the effective capex tax to 6.2 percent. Roughly half as many merchant projects cross the IRR cliff; the offshoring tail thins; hyperscalers adjust pacing. The affected fraction falls to approximately 10 percent of the buildout, GDP loss to $45 billion annually, and jobs at risk to roughly 122,000.
Carve servers out at the 25 percent rate: Because servers carry the majority of imported compute value, exempting them reduces the tariff base far faster than reducing the rate. CSIS has documented that exempting the inverse (semiconductors but not derivatives) reduces tariff burdens by only 20 percent, implying that exempting servers (the most common derivative) reduces it by roughly 70 percent. The effective tax falls to about 4.7 percent. GDP loss falls to roughly $36 billion annually, with about 98,000 jobs at risk.
Both: 10 percent rate AND servers carved out: Effective tax of approximately 1.9 percent. Only the most fragile merchant projects fall below viability. GDP loss falls to roughly $16 billion annually, jobs at risk to approximately 43,000.
Implications
Each of these alternative policy design scenarios reduces the economic damage. None eliminate it. Only retaining the existing data-center-end-use carveout that already exists under the Phase 1 proclamation eliminates the data center channel cost entirely.
That carveout exists in current law because the Administration recognized, in January 2026, that tariffing the inputs to the most important productive investment channel in the U.S. economy was not in the national interest. Nothing about the underlying facts has changed since then. Domestic chip manufacturing cannot grow quickly enough to meet U.S. AI buildout needs through 2030. TSMC’s Arizona facility has reached only a fraction of the volume needed to substitute for imports, and at production costs 50 percent higher than Taiwan. The dependence on Taiwan for advanced logic, on Korea for high-bandwidth memory, and on Taiwan-and-Mexico assembly for AI servers remains unavoidable through 2030 even if allied country firms build entire supply chains including CoWoS packaging at scale in the U.S. A tariff cannot conjure the missing domestic supply chain into existence at scale by 2030; it can only impose a cost on the buildout and slow U.S. leadership in AI while we wait for that supply chain to emerge.
Taxing the AI infrastructure buildout while domestic capacity catches up is a recipe for an economic slowdown, a decline in U.S. AI leadership, and compute shortages that slow AI deployment, adoption, and productivity gains. It is better to let the AI buildout proceed and use the productive capacity it creates to anchor that domestic capacity over time; these data centers will need to replace chips when they become depreciated or obsolete, and that provides a guaranteed demand base for U.S. semiconductor production in the future. Retaining the data center carveout in Phase 2 is the analytically and economically optimal path to achieve the strategic aims of maximizing long-term U.S. leadership in the AI space and value chain while preserving current U.S. economic growth and jobs.
