Applying Semiconductor and Server Tariffs to Data Centers, Under Section 232 or Section 301, Would Now Cost the U.S. About $180 Billion a Year

Ashburn, Virginia, United States: An aerial view of a data center being constructed inside "Data Center Alley" in Ashburn. Northern Virginia is the largest data center market in the world.

An update to CCIA’s June 2026 analysis: the Section 232 data center exemption remains under review, a Section 301 excess capacity determination could be weeks away, and the buildout at stake has grown significantly in size

Bottom Line Up Front

• Refreshing the June model with a larger capex base, a 2027-2031 window, and current GDP and labor force figures: a 25% tariff on semiconductors and servers amounts to a 15.6 percent effective tax on data center construction and would put about $900 billion of planned 2027 to 2031 U.S. data center investment at risk, costing the U.S. roughly $180 billion per year in capex and output. That’s about 0.56 percent of U.S. GDP at risk, and about 474,000 American jobs jeopardized.
• The widening U.S. goods deficit with Mexico is sometimes attributed to a surge in server imports. The surge is in value per unit (average price), not an increase in the quantity of units being imported from Mexico. Census data show the number of digital processing units imported from Mexico fell 9 percent between 2019 and 2025 while the average value per unit rose three and a half times, and through August 2026 units were down a further 24 percent year over year while value more than doubled.
• In short, the rising goods deficit from Mexico associated with the data center buildout is driven by servers becoming more expensive, not by an increase in the number of machines imported from Mexico. Tariffs, which would make servers even more expensive, are an exceptionally counterproductive tool to use. If the policy response to rising price per unit of an import is to impose a tax (tariff) that makes the price to the user rise even further, the expected outcome over multiple iterations of the policy response process is a price spiral. Using tariffs to respond to the rising price per unit of an import from a country from which the quantity imported is actually declining fails to achieve any objective other than raising prices further and putting U.S. frontier technology businesses at a competitive disadvantage relative to China.
• In addition, while America’s compute buildout is powering a strong economy with new investments, manufacturing, and good-paying jobs, the supply of critical inputs is unable to keep up with growing demand. Thus, if Section 301 excess-capacity tariffs are meant to address a surplus, compute inputs face a shortage, so tariffs would only raise costs and slow the buildout in an already supply constrained environment. 
• We can eliminate this significant threat to the compute buildout  by protecting and expanding the data center end-use exemption. Specifically, we must: (1) Preserve the exemption established in the Phase 1 proclamation of the Section 232 semiconductor investigation for all future actions. (2) Expand it to cover all U.S. data center inputs, removing the restriction that limits it only to facilities above 100 megawatts. (3) Apply this same exemption to any upcoming Section 301 excess capacity trade actions.

The Context for These Revised Findings

• In June, CCIA estimated that extending 25 percent Section 232 semiconductor tariffs to U.S. data centers would cost the U.S. about $90 billion a year in lost American investment and output and put about 243,000 American jobs at risk.
• Since then, the administration has reportedly moved from treating a Phase 2 tariff as a possibility to actively developing one that would hit data center servers, with the January data center exemption possibly not carrying over. Commerce Secretary Lutnick confirmed on September 2 that a new round of semiconductor tariffs is coming.
• A second tariff authority now potentially points at the same goods. USTR’s Section 301 investigation into structural excess capacity covers sixteen economies, including Taiwan, Mexico, Korea, Vietnam, and Japan, lists semiconductors and electronics among its sectors of concern, and is nearing a determination: Ambassador Greer said on October 1 that countermeasures would be announced in the coming weeks. A Section 301 tariff on semiconductors, servers and related data center inputs would impose the same cost as a Section 232 tariff if applied at the same rate on the same goods.
• The buildout the exemption protects has grown. Dell’Oro Group now projects worldwide data center capital spending above $3 trillion by 2030, nearly double its January forecast, and U.S. imports of computers and servers reached $363.5 billion over the twelve months to July 2026.
• Because a Phase 2 proclamation has not yet been issued, a phase-in is reportedly under discussion, and a Section 301 remedy ordinarily follows a proposed action and comment period, neither tariff is likely to take effect before 2027; this update therefore models the impacts over 2027 through 2031. 

What Has Changed Since June

The policy is no longer hypothetical

When CCIA published its analysis on June 11, Phase 2 was a possibility written into the January 14, 2026 Presidential Proclamation, which imposed a 25 percent duty on certain advanced computing chips and products containing them, exempted “Covered Products for use in United States data centers,” and directed the Commerce Department to report by July 1 on the market for data center semiconductors to inform whether the tariff should be modified. That report has reportedly been completed but not published.

On August 27, Politico reported, citing eight people familiar with the discussions, that the administration is weighing a second round of semiconductor tariffs that would broaden duties from a narrow set of advanced chips and the products containing them to a much wider range of products built with chips, including data center servers, laptops and gaming consoles, with relief tied to companies’ commitments to manufacture chips in the United States and a phase-in period under discussion (CNBC). Commerce officials have reportedly indicated in private talks, according to Politico’s sources, that the exemptions attached to the January action, including the data center exemption, may not carry over.

On September 2, Secretary Lutnick confirmed on CNBC’s Squawk Box that the administration is developing “targeted, thoughtful” semiconductor tariffs on a single principle: “if you build here, you don’t pay, but if you don’t build here, expect to pay” (see also Law360). Rates, product lists and timelines have not been finalized. Data center operators are primarily buyers of chips and servers rather than chip manufacturers. How relief keyed to domestic fabrication capacity would flow through to them remains uncertain, and it would not automatically replicate a broad end-use exemption.

A second tariff authority now points at the same goods. On March 11, USTR initiated Section 301 investigations into structural excess capacity and production in sixteen economies, among them Taiwan, Mexico, Korea, Vietnam, Japan and Malaysia, the economies that supply most of the chips, memory and assembled servers entering U.S. data centers. The Federal Register notice lists semiconductors and electronics among the sectors it considers affected, and it cites Taiwan’s semiconductor and information technology exports and Singapore’s semiconductor exports as evidence. Public hearings ran from May 5 to May 8. On October 1, Ambassador Greer told Nikkei Asia that the United States would announce countermeasures against overproduction in the “coming weeks” and was “not going to sit there and take” inflows of excess production. No determination or proposed remedy had been published as of mid-October, and Greer has said that USTR will take the tariff caps negotiated in trade deals into account when it considers outcomes, and that he did not want to “get ahead of the report and proposed action.”

For a U.S. data center buyer, the authority under which a tariff is issued does not change what the tariff costs. A Section 301 duty of 25 percent on semiconductors, servers and related inputs would produce the same tariff bill at the border, the same effective tax on each dollar of data center investment, and the same response from the projects that bear it as a Section 232 duty of 25 percent on the same goods, for as long as the duty remained in force. The two authorities differ in who issues the tariff, how it can be challenged in court, and how long it lasts; the economic impacts are essentially identical. The estimates in this update therefore apply to either.

The data center question could be decided in either proceeding. 

The buildout is larger than the June baseline assumed

The June analysis used McKinsey’s projection of roughly $2.7 trillion in U.S. data center capital expenditure between 2025 and 2030, more than 40 percent of a $6.7 trillion global figure published in April 2025. The market has moved well past that. On August 18, Dell’Oro Group reported that worldwide data center capital spending is now forecast to surpass $3 trillion by 2030, with its 2030 outlook having “nearly doubled” since its January 2026 forecast of $1.7 trillion. Dell’Oro attributes the revision to higher hyperscale capex guidance, higher projections for global data center power capacity, and higher commodity costs, and it says the top four U.S. hyperscalers alone could represent about half of global data center capex. Worldwide data center capex grew 92 percent year over year in the second quarter of 2026 and is expected to exceed $1 trillion for the full year.

Company guidance tells the same story. After second-quarter earnings, Alphabet raised its 2026 capital expenditure guidance to $195 billion to $205 billion, and the four largest U.S. hyperscalers together now guide to roughly $735 billion of 2026 capital spending, with Amazon at about $220 billion, Microsoft at about $175 billion, and Meta at $135 billion to $145 billion. 

Trade data confirm both the scale and the import dependence of the compute stack. U.S. imports of computers and units (HS 8471) totaled $363.5 billion in the twelve months to July 2026, 91 percent more than a year earlier and more than three times the 2023 total, and July’s $43.8 billion was the largest month on record. Digital processing units alone (HTS 8471.50), the line that carries servers, reached $83.9 billion in May through July 2026, nearly double the year-earlier figure. In May through July 2026, Mexico supplied a record 43.7 percent of the value of U.S. computer imports under HS 8471, with Taiwan at 28.1 percent; Taiwanese firms control roughly 90 percent of the global AI server market and are shifting final assembly to Mexico and the United States. Taiwan’s exports to the United States more than doubled to $27.5 billion in September on AI server and chip demand. The World Trade Organization’s October update found that AI-enabling goods such as semiconductors and servers accounted for 47 percent of growth in the value of global merchandise trade in the first half of 2026, with trade in those goods rising 67 percent year over year.

The buildout is now visible in the national accounts. In its September 30 release, the Bureau of Economic Analysis revised second-quarter GDP growth up to 2.2 percent, with nonresidential structures among the leading contributors to the upward revision in private fixed investment, led by commercial and health care structures, “mainly data centers.”

Server imports from Mexico have risen in value per unit while quantity of units imported from Mexico have fallen

Some proponents of extending tariffs to data center inputs point to the widening bilateral goods deficit with Mexico and attribute it to surging imports of AI servers assembled there. The deficit did widen, from $99.4 billion in 2019 to $197.0 billion in 2025 according to the Census Bureau, and imports of digital processing units (HTS 8471.50, the tariff line that carries servers) from Mexico rose from $25.2 billion to $80.5 billion over the same period, 57 percent of the increase in the deficit.

However, the number of machines imported did not rise. Census Bureau import records for this line show 15.2 million units imported from Mexico in 2019 and 13.9 million in 2025, a decline of 9 percent, while the average value per unit rose from about $1,660 to about $5,800. The trend has steepened in 2026: through August, units were 24 percent below the same months of 2025 while value had more than doubled to $99.4 billion, already more than the whole of 2025, and July and August together brought $39.0 billion of imports on 1.96 million units, about $19,900 each. A standard decomposition of the change in value into a quantity effect and a unit-value effect attributes more than the entire increase since 2019 to value per unit; the change in unit counts subtracts from it. Had unit values stayed at 2019 levels, with other trade flows unchanged, the rise in the bilateral deficit would have been roughly $40 billion rather than $98 billion.

The pattern is national rather than Mexican. For all suppliers combined, the trailing twelve-month average value per unit of HTS 8471.50 imports has risen more than sevenfold since the end of 2019, reaching about $11,300 in August 2026, while cumulative unit counts for January through August 2026 are about 16 percent below the same period of 2025. The deficit reflects what is inside each machine, above all the accelerators designed by U.S. firms, which CSIS notes enter the country already embedded in servers, and the memory and networking around them, rather than more machines crossing the border from Mexico. A tariff aimed at the server would tax the U.S. design margin embedded in it along with the Mexican assembly it is nominally directed at.

The chart below puts the four series on one scale, each indexed to its 2019 level. The goods deficit with Mexico has roughly doubled and the value of server imports has risen more than fivefold on an annualized basis, while the number of machines crossing the border under this line remains below its 2019 level. The line that tracks the import value is value per unit. The server-related import bill is growing because each machine is worth more.

Figure 1: Imports from Mexico on the Server Tariff Line Rose in Value per Unit, Not in Unit Count

Notes and Sources: U.S. goods trade deficit with Mexico, HTS 8471.50 imports from Mexico, units imported and value per unit, indexed to 2019 = 100. Underlying data sourced from U.S. Census Bureau.

The Cost Shock: Same Rate, Larger Base

The three parameters that determine the effective tax on data center buildout that tariffs would impose are unchanged from June. IoT Analytics finds that 78 percent of data center equipment and infrastructure spending goes to IT infrastructure, led by servers. The surge in memory and storage prices that Dell’Oro reports lifted server average selling prices in 2026 pushes that share higher, so 78 percent is a conservative figure. The 70 to 90 percent import intensity for computer equipment and recent AI investment reported in the World Trade Organization’s March 2026 Global Trade Outlook, with the 80 percent midpoint used here, is consistent with the import data above. The tariff rate is assumed to remain 25 percent, the rate enacted in Phase 1, applied to both semiconductors and derivative products such as servers. The same rate applied to the same goods under Section 301 would have the same effect. Multiplying the three yields the same 15.6 percent effective tax on every dollar of U.S. data center investment, and near-complete pass-through is assumed given the concentration of supply. NVIDIA’s gross margin was 75.0 percent in the quarter ended July 26, 2026, consistent with the design-margin add-back used in June.

What has changed is the base and the window. The June analysis covered 2026 through 2030. With no Phase 2 proclamation issued as of mid-October and a phase-in reportedly under discussion, and with any Section 301 excess capacity remedy ordinarily preceded by a proposed action and public comment period, a tariff on data center imports is unlikely to bind before 2027, so this update covers 2027 through 2031. Taking Dell’Oro’s anchors of roughly $1 trillion in 2026 and more than $3 trillion in 2030 implies compound growth of about 32 percent a year through 2030. Dell’Oro’s published forecast ends in 2030; this analysis holds 2031 spending flat at the 2030 level rather than assuming further growth, a conservative choice. On that path, worldwide data center capex over 2027 to 2031 is roughly $11.3 trillion. Applying McKinsey’s estimate that more than 40 percent of global data center investment will occur in the United States gives a U.S. base of about $4.5 trillion for 2027 to 2031, roughly double the $2.25 trillion used in June. Dell’Oro’s published scope centers on IT equipment, networking and physical infrastructure, the spending a tariff on chips and servers would hit; applying the same 78 percent IT share to it is a modeling assumption carried over from June.

On that base, the gross five-year tariff cost is about $707 billion, or roughly $141 billion per year, before any project is canceled in response. That annual figure is about two-thirds of the entire 2026 capital budget of the largest single hyperscaler. Few, if any, prior tariff actions have placed a cost of this size on a single investment channel, let alone one that the Bureau of Economic Analysis now cites by name in its GDP release.

What the Cost Shock Would Do: June Versus October

The segment structure of the June analysis is unchanged. Self-funded hyperscaler projects, roughly 60 percent of the buildout, would see about 10 percent of planned 2027 to 2031 capex delayed beyond the window or relocated abroad. Merchant developers and AI-native operators, roughly 30 percent of the buildout and debt-financed against equity returns in a range that a 15.6 percent capex shock largely consumes, would see about 40 percent of planned capex delayed, relocated or canceled. The enterprise, colocation and offshoring tail, the remaining 10 percent, would see about 20 percent move abroad. Dell’Oro’s finding that the AI-specialized cloud segment of model builders and neoclouds is the fastest-growing customer segment, at nearly a 60 percent compound annual rate, means the most tariff-sensitive part of the buildout is gaining weight over the window.

Weighting the segments again produces a central estimate that about 20 percent of planned U.S. data center investment between 2027 and 2031 would be canceled, delayed beyond 2031 or relocated abroad. On the updated base that is about $906 billion of capex, or roughly $180 billion per year. Using the same conservative one-for-one conversion of foregone capex into GDP (the fuller estimate of $1.10 per dollar, combining 65 cents of direct U.S. value added with an assumed 1.7 multiplier for indirect and induced effects under the BEA RIMS II framework, would be higher), the annual GDP loss is about $180 billion, or 0.56 percent of the $32.6 trillion U.S. economy reported by BEA for the second quarter of 2026. Translated through Okun’s Law, under which a one percent output gap moves unemployment by 0.4 to 0.5 percentage point per Ball, Leigh and Loungani (the 0.5 figure used in June is retained here), and applied to a civilian labor force of 170.3 million, this implies an unemployment increase of about 0.28 percentage points and roughly 474,000 jobs at risk.

The table below sets the June and October figures side by side. Every change in the bottom rows flows from the larger capex base, the later five-year window, and the newer GDP and labor force denominators; the method and judgment parameters are the same.

The fragility that the affected-rate assumptions capture is now documented in the pipeline data. Bernstein reported in October a record 51.7 gigawatts of deferred U.S. data center capacity, 11 percent of a 456 gigawatt project pipeline, with 27 states restricting or reviewing new development. PJM’s independent market monitor told the Federal Energy Regulatory Commission in September that 4,017 megawatts of data center load had been canceled or delayed, 42 percent of the data center demand growth PJM had forecast between the 2026/27 and 2028/29 delivery years (see also Forbes). Dell’Oro’s October reading of the physical infrastructure market describes permitting as a major near-term risk to the compute buildout. Each of these constraints is already generating project cancellation risk before any new semiconductor tariff shock is added to it, making the estimated project cancellation risks from tariffs both highly plausible and fairly conservative.

How Policy Design Sensitivities Reshape the Harms

The three design variations examined in June reduce the damage on the updated base as well, and none of them matches simply continuing the exemption. Each variation applies in the same way to a Section 301 action: a 10 percent Section 301 rate on the same goods produces the same figures as a 10 percent Section 232 rate, and a server carve-out removes the same share of the burden under either authority.

Lower the rate from 25 percent to 10 percent: The effective capex tax falls to 6.2 percent and the affected share to about 10 percent of the buildout. Annual GDP loss falls to about $91 billion and jobs at risk to roughly 237,000.

Carve servers out at the 25 percent rate: Because servers carry most of the imported compute value, exempting them shrinks the tariff base faster than cutting the rate. CSIS has documented that exempting semiconductors but not derivatives removes only about 20 percent of the burden, and that roughly 72 percent of GPU import value arrives inside servers. The server-exemption sensitivity here assumes that exempting servers removes roughly 70 percent of the modeled burden. The effective tax falls to about 4.7 percent, annual GDP loss to about $72 billion, and jobs at risk to roughly 190,000.

Both a 10 percent rate and a server carve-out: An effective tax of about 1.9 percent, with annual GDP loss of about $32 billion and roughly 84,000 jobs at risk.

The table below summarizes the scenarios. The three parameters that set the effective tax, and the share of the buildout each scenario is judged to affect, are the assumptions that vary; the annual GDP loss and jobs at risk are the outcomes. All scenarios run on the same $4.5 trillion 2027 to 2031 base.

Implications

Each alternative reduces the economic damage without eliminating it. Only a data-center-end-use exemption that retains the carve-out in the Phase 1 proclamation and extends it to all U.S. data center inputs, and that is carried into any Section 301 excess capacity action, eliminates the data center channel cost entirely. The existing exemption is written for facilities requiring more than 100 megawatts of new load dedicated to AI inference, training, simulation or synthetic data generation; a Phase 2 that broadened server coverage would need to cover the enterprise and colocation facilities below that line as well.

USTR has already excluded key data center inputs from another Section 301 action. The forced labor action that took effect on July 24 exempts digital processing units under HTS 8471.50 and the integrated circuit lines under HTS 8542 from its 10 and 12.5 percent duties, alongside goods already subject to Section 232. The same exclusion belongs in any excess capacity remedy.

The case for the exemption is stronger in October than it was in June, for two reasons. First, the investment it protects has grown: the forecasters, the hyperscalers and the trade statistics all describe a buildout whose tariff-relevant spending over a 2027 to 2031 window is roughly double what the June baseline assumed for 2026 to 2030, so the same tariff design now puts about $180 billion a year of U.S. output at risk rather than $90 billion. Second, domestic capacity remains far from sufficient to replace imports at the scale required. TSMC announced on July 16 an additional $100 billion for Arizona, bringing its planned investment there to $265 billion across ten fabs, which is welcome. But only the first fab is in volume production, on the 4-nanometer node, and TSMC said on October 15 that the second fab will begin volume production in the third quarter of 2027. Taiwan still produces more than 90 percent of the world’s most advanced chips, Korea remains the principal source of high-bandwidth memory, and Taiwanese firms assemble roughly 90 percent of the world’s AI servers in Taiwan, Mexico and the United States. A tariff cannot bring that capacity onshore at the scale the buildout requires within this window; it can only raise the cost of the buildout and slow U.S. leadership in frontier technology while the domestic supply chain is built.

The administration recognized in January that taxing the inputs to the most important productive investment channel in the U.S. economy was not in the national interest. The facts that supported that judgment have only strengthened. Letting the compute buildout proceed, and using the demand base it creates to anchor domestic semiconductor production over time, remains the optimal path, and carrying the data center carve-out into Phase 2 and into any Section 301 excess capacity action is the way to stay on it.