Your Data Center's Tariff Bill Is $6.2M a Megawatt • The Weighted Average

The most consequential number in AI infrastructure this week is not a model score or a funding round. It is 15.6 percent — the effective tax that a 25 percent semiconductor tariff would impose on every dollar of American data center construction if Washington lets its current carve-out lapse, according to the Computer and Communications Industry Association’s June analysis. Politico reported this week that a sweeping Phase 2 of the Section 232 regime could land in weeks or months, and the industry’s reaction, relayed by Ars Technica, was blunt: one trade official called it “the single dumbest way imaginable to pursue American dominance in AI.”

Translate the percentage into the unit operators actually buy. Capacity, not chips.

The exemption everyone budgeted around

The tariff already exists. The January 14, 2026 presidential proclamation imposed a 25 percent duty on advanced computing chips and certain derivative products, and then carved out “Covered Products for use in United States data centers.” That document also set out the two-phase design explicitly: a narrow first tranche while trade negotiations run, then “broader tariffs on semiconductors, at a rate of duty that is significant,” paired with an offset program for companies investing in US production. The proclamation records the Commerce Secretary’s finding that the United States “currently fully manufactures only approximately 10 percent of the chips it requires.”

The carve-out was always provisional. The Next Web’s reading of the mechanics notes that the Commerce Department report determining whether the exemption survives was due July 1 and has not been published, so “every operator planning American capacity is working from an assumption rather than a settled rule.” That is the actual exposure: not a rate, but an unpriced binary sitting inside every 2027 capacity plan.

The arithmetic behind the 15.6 percent is deliberately simple, which is both its strength and its weakness. The CCIA multiplies the share of data center spending that goes to computing equipment — 78 percent, from IoT Analytics’ data center infrastructure market work, which splits that spend into servers at 61 percent, networking at 10, and storage at 6.5 — by the share of that compute the United States imports, 80 percent, the midpoint of a World Trade Organization range, by the 25 percent rate. Seventy-eight times eighty times twenty-five gives 15.6. The import-intensity assumption does the most work, and it is an estimate rather than a customs return.

That server share is the load-bearing detail for anyone modelling exposure, because it determines whether a duty on “derivative products” is an inconvenience or a repricing. Chips alone are the smaller half of the bill; assembled systems are the larger one.

Two structural facts make it hard to argue down. US Census figures cited in the Coalition for a Prosperous America’s trade analysis put server imports above 95 percent of domestic consumption, with Taiwan supplying $86 billion in 2025. And CSIS’s assessment of tariffs on the AI buildout finds TSMC’s Arizona output at a fraction of substitution volume, at production costs roughly 50 percent above Taiwan. You cannot tariff your way to a supply chain that takes five years to pour concrete for.

Run the number in megawatts, not percentages

Here is the figure nobody has published, and it is the one to carry into a capacity meeting.

The CCIA annualizes its cost at roughly $70 billion by applying 15.6 percent to McKinsey’s $2.7 trillion 2025-2030 buildout, scaled by five-sixths for the untariffed 2025. But that denominator is a five-year average, and the buildout is not flat. Dell’Oro Group’s March 2026 report found worldwide data center capex rose 57 percent in 2025 — 76 percent among the top four US clouds — and forecast full-year 2026 capex to surpass $1 trillion, driven by well over 10 million high-end accelerators. Apply the same 15.6 percent effective rate to a trillion-dollar year and the gross tariff cost is about $156 billion in 2026 alone, roughly 2.2 times the CCIA’s own annualized figure. The five-year average understates the near-term shock precisely because the near term is where the spending is concentrated.

Now convert to the operator’s unit. This paper derived a build cost of about $40 million per megawatt of AI capacity from AM Intelligence’s 200-megawatt, $8 billion Hyderabad program — silicon, racks, cooling, interconnect, shell. A 15.6 percent effective tax on that all-in figure is $6.24 million per megawatt. A 100-megawatt campus carries a $624 million tariff line that did not exist in the January model. A gigawatt carries $6.2 billion.

The scaling exercise deserves its caveat. Dell’Oro’s trillion-dollar figure is global while the tariff is American, so the $156 billion is an upper bound on what a worldwide buildout would owe if all of it were built in the United States under Phase 2. The point is the ratio, not the total: the CCIA’s own $70 billion annual average is drawn against years in which spending is a fraction of what 2026 alone will absorb, so the near-term shock is materially larger than the headline suggests.

That is the number that decides siting. It is also why the CCIA’s headline consequences read the way they do: about 20 percent of planned 2026-2030 US AI data center capacity, roughly $450 billion of capex, cancelled, pushed beyond 2030, or relocated abroad; $90 billion in annual GDP loss, about 0.29 percent of the $31.4 trillion economy the Bureau of Economic Analysis reports; and 243,000 jobs at risk.

The distribution matters more than the total. Hyperscalers absorbing a 15.6 percent shock inside an $80 billion capex envelope slow their pacing; the CCIA models roughly 10 percent of their planned spend slipping or relocating. Merchant developers and AI-native operators — the CoreWeave, Crusoe, Lambda, Vantage cohort financing against equity IRRs in an 11 to 19 percent band — see roughly 40 percent affected, because a 15.6 percent capex shock consumes essentially the entire equity cushion. The tariff does not tax everyone evenly. It taxes the leveraged.

Softening the chip tariff shrinks the damage but never removes it

Projected annual US GDP loss by tariff design, $ billions

CCIA analysis of Section 232 Phase 2 · Jun 2026

The ways this thesis breaks

Start with the source. The CCIA is a trade association whose members would pay the duty, so its analysis is an advocacy document with a methodology attached. Every parameter tilts one way: the 80 percent import intensity is the WTO midpoint rather than a measured figure, and the 1.7 GDP multiplier, while conservative against data-center-specific studies, is still a multiplier. Halve the affected-capex fraction and the annual loss lands near $45 billion — painful, not existential.

The trade groups have already told Treasury what they want instead. A May letter cosigned by about 20 associations asks for exclusions for products with incidental semiconductor content, a de minimis threshold, paperwork limited to documents firms already hold, and — the biggest ask — no stacking of duties on both a chip and the product containing it. Read as a negotiating position, it concedes that some tariff is coming and fights over the base.

Second, the policy may never arrive in its maximal form. Politico’s sources describe a phased rollout and a possible relief channel tied to foreign firms investing in US fabs, an approach Commerce Secretary Howard Lutnick reportedly favors. Under the duty-free-allowance design attributed to him, volume entitlements scale with domestic production pledges — which is a tariff for the unpledged and a subsidy for the pledged. One industry representative’s objection quoted by Ars is the operative critique: “the volume they’re talking about granting duty-free wouldn’t cover the hyperscalers alone.”

Third, the design levers genuinely matter, and the CCIA quantifies them. Dropping the rate from 25 to 10 percent cuts the effective tax to 6.2 percent and the annual loss to about $45 billion. Carving servers out at the full 25 percent rate is more effective still, because servers carry the majority of imported compute value: effective tax about 4.7 percent, loss about $36 billion. Do both and you land near 1.9 percent and $16 billion. None of these reach zero, which is the analysis’s real argument — only retaining the existing carve-out eliminates the channel.

Fourth, and least discussed: the shock is not arriving into a healthy market. Memory and storage inflation is already in the price of every rack. Server prices were already climbing before any duty, as Nvidia’s guidance of a 15-percent-plus AI server price increase against a 3.6x rise in DRAM contract prices showed. Dell’Oro separately projects general-purpose server ASPs rising by high double digits in 2026 on memory and storage costs alone. A tariff stacks on top of an inflation the buildout is already absorbing.

What would change the verdict? Publication of the July 1 Commerce report with the data center exception intact; a Federal Register notice defining “derivative products” narrowly enough to exclude assembled servers; or a tariff-offset rule whose duty-free volumes actually cover hyperscale demand. Any of those turns a $6.2 million per megawatt line item back into a rounding error.

What to do before the rule lands

Treat the exemption as a live risk with a price, not a policy story. Three moves are available now.

Price the binary into every 2027 contract. A 15.6 percent effective capex tax is $6.2 million per megawatt at a $40 million all-in build. Ask whether your colocation or turnkey provider eats duty changes or passes them through, and get the answer in writing. If the contract is silent, you own it.
Know your derivative exposure, not just your chip exposure. The gap between a tariff on accelerators and one that reaches assembled servers is the difference between painful and site-changing, because servers are 61 percent of data center spending per the IoT Analytics split the CCIA uses. Inventory what crosses the border as a finished system versus a component.
Stress-test the leveraged parts of your supply. If your capacity comes from merchant developers rather than hyperscaler-owned campuses, you are exposed to the cohort the CCIA models at a 40 percent hit rate. Hedging that means a second source, an earlier commitment, or a non-US region — and the last of those is exactly the outcome the tariff is meant to prevent.

The strategic irony is durable regardless of how Phase 2 lands. Domestic fabs are the goal; imported chips are the only way to build the demand base that justifies them. Every megawatt built now is a customer for a future US fab, because these machines depreciate and get replaced. Taxing the buildout to accelerate the buildout is a policy that argues with itself, and the buildout is already the load-bearing wall of the economy — information processing equipment and software accounted for roughly 92 percent of US real GDP growth in the first half of 2025 on the Furman analysis the CCIA cites.

Meanwhile the demand does not wait. Global capex is compounding toward the 32 percent annual growth required to reach $3 trillion by 2030, memory capacity keeps getting committed years ahead, as SK Hynix’s Indiana packaging plant with H2 2029 production illustrates, and the software layer keeps compressing its own costs — Cursor’s router now ships a commit for $4.63 against $12.69 on a frontier model, while $26.5 billion of open-weight infrastructure deals chase a 6.1 percent adoption base. Efficiency at the top of the stack cannot offset a 15.6 percent tax at the bottom of it. Concrete and customs do not respond to better routing.

The exemption everyone budgeted around

Run the number in megawatts, not percentages

The ways this thesis breaks

What to do before the rule lands

Sources