Bloomberg’s customs data shows AI-related products reached an estimated 23% of all US imports in 2025 — and the bottleneck has moved from chips to the electrical hardware America cannot manufacture fast enough.
AI IMPORT BOOM — KEY FIGURES (2025–2026, REPORTED ESTIMATES)
According to Bloomberg’s New Economy newsletter, customs data shows AI-related products accounted for roughly 23% of all US imports in 2025, up from an estimated 15% in 2023, with computer hardware making up approximately half that total. These are reported estimates built on a broad classification — reasonable analysts would draw the line differently — but the direction and scale are not in dispute. The AI buildout is one of the largest single forces now moving through the US trade account.
The scale of the underlying investment makes the import figure coherent. Meta, Alphabet, Microsoft, and Amazon collectively spent roughly $130 billion on capital expenditure in Q1 2026 alone. Total AI investment could reach around $725 billion in 2026, though that is a projection, not a settled number. Almost none of the physical hardware behind that spending is manufactured in the United States at the required pace or volume.
What is newer than the chip dependency — and more operationally consequential — is where the bottleneck has moved. The shortage is now in transformers, switchgear, and grid-scale batteries: the unglamorous electrical infrastructure that actually powers a data center. The US cannot produce this equipment domestically at the pace the buildout demands, and an increasing share is sourced from China. The result is a construction delay with a measurable footprint: by one estimate, only about one-third of the 12 gigawatts of data-center capacity planned for 2026 is actually under construction.
HOW THE DEPENDENCY SHIFTED
2023
AI-related imports ~15% of US total (est.). Bottleneck: advanced chips, concentrated in Taiwan (TSMC) and Korea (memory).
2024–2025
Assembly nearshored to Mexico; server rack production scales via Foxconn and Flex. Chip dependency persists. Electrical-gear shortages begin to surface globally.
2025
AI-related imports reach ~23% of US total (est.). Transformers, switchgear, and batteries — sourced increasingly from China — become the new hard constraint.
2026 (projected)
~$725B AI investment projected. Only ~1/3 of 12GW planned data-center capacity under construction. The physical layer is the binding constraint on the buildout’s pace.
The key insight: The AI boom has moved from a chip-import story to a whole-supply-chain-import story. Chips were always foreign-made. Now the transformers, switchgear, and batteries that keep data centers alive are too — and unlike chips, where domestic alternatives are at least plausible at multi-year timescales, US electrical-equipment manufacturing capacity has no near-term path to the required scale. That is a harder constraint, not just a different one.
AI-related products rose to roughly 23% of all US imports in 2025 from about 15% in 2023, with computer hardware accounting for roughly half of that, as the data-center buildout pulled in servers, chips and increasingly the electrical equipment – transformers, switchgear, batteries – the US struggles to make at home. These are reported trade-data estimates and ‘AI-related’ is a broad classification. Sources: Bloomberg New Economy; trade-data analysis.
The Structural Read
Front-load the hedges, because precision matters here. The 23%/15% figures are reported estimates from customs data and the “AI-related import” classification is deliberately broad — the Bloomberg analysis is honest about this. The claim that imports mute AI’s GDP contribution is one framing of ordinary macroeconomics, not a debunking: AI spending is genuinely additive to US growth, and the Congressional Budget Office and private forecasters credit it with adding roughly 0.73 percentage points to GDP. The import share offsets part of that — money that shows up as investment also shows up as a subtraction when the hardware crosses the border — but the net is still positive. The transformer and switchgear shortage is a global capacity problem driven by simultaneous demand from data centers, grid modernization, and the energy transition everywhere, not solely a China story or solely an AI one. And import dependence is not automatically fragility; the US has run large tech trade deficits for decades while capturing the dominant share of value in software, services, and platform economics on top.
What survives all those caveats is a structural fact worth internalizing: the AI buildout’s demand has become one of the largest single forces in US trade, and it is now pulling in not just chips but the unglamorous electrical hardware that actually powers a data center. Both the economic gain and the strategic control of that hardware sit disproportionately offshore.
Map the geography of leakage and you get a clear picture. Assembly has been nearshored to Mexico — Foxconn and Flex are building the compute floor there, which is a different and separable story. Advanced silicon is still made in Taiwan. Memory is concentrated in Korea — SK Hynix’s Chongqing decisions trace exactly this tension. And now the electrical equipment — transformers, switchgear, grid batteries — is increasingly sourced from China. Each layer of the physical stack is import-dependent, and each is controlled by a different geography.
Map of AI — Physical Layer
The Import-Dependent Physical Layer Is a Hard Constraint, Not Just an Accounting Entry
When the bottleneck moves from chips to transformers and grid gear the US does not make at scale, the buildout cannot simply be willed faster by increasing capex. The ~1/3 of planned 2026 data-center capacity sitting unbuilt is what an import-dependent physical layer looks like when demand outruns the supply chain. This is not a financing problem or a permitting problem alone — it is a manufacturing-geography problem. The reindustrialization narrative runs headlong into this: for now, America is importing its AI industrial base, and the value embedded in making that hardware sits elsewhere.
The Nvidia-Lancium powered-land bottleneck is the same structural problem from a different angle: the constraint on the buildout’s pace keeps moving up the stack, from silicon to power delivery to the electrical gear that connects the two. Each time the bottleneck shifts, it lands on something the US makes less of than it needs.
The deeper point — developed in Beyond NVIDIA’s Moat — is about where durable competitive position actually lives. Whoever ends up making the transformers, the grid gear, and the chips — not just assembling or financing them — captures the lasting margin. Today, that is mostly not the United States. The customs data is the least glamorous and most honest read on the buildout, and it confirms the industrial base is still somewhere else.
Three Implications
IMPLICATION 1 — GDP ACCOUNTING NEEDS A SUPPLY-CHAIN LENS
The ~0.73 percentage-point AI contribution to GDP is a gross figure. The import share of that investment — hardware crossing the border — registers simultaneously as a subtraction in the national accounts. The net domestic contribution is real but smaller than the headline capex numbers imply. This is standard macroeconomics, not a debunking, but it matters for anyone modeling AI’s macro impact: the leakage is structural, not cyclical, and it grows as the buildout scales.
IMPLICATION 2 — THE ELECTRICAL-GEAR SHORTAGE IS A DIFFERENT KIND OF RISK
Chip dependency was always a known variable; the industry has been building policy and investment responses to it for years. The transformer and switchgear shortage is newer, less visible, and harder to fix quickly — US manufacturing capacity here is thin and lead times are long. With only ~1/3 of planned 2026 data-center capacity under construction, the electrical layer is now actively rate-limiting the buildout. This is a global capacity problem, but the US’s domestic position within it is weak.
IMPLICATION 3 — MAKING BEATS ASSEMBLING FOR DURABLE MARGIN
The Map of AI’s physical layer has a consistent rule: the entity that manufactures the component — not the one that assembles, finances, or deploys it — captures the durable margin. Assembly has moved to Mexico; silicon stays in Taiwan; memory stays in Korea; electrical gear is increasingly in China. The US captures enormous value in the software, services, and platform layers on top — that is the historical pattern with tech trade deficits — but the physical-layer geography sets the terms of supply-chain leverage and strategic risk for everyone building on top of it.
Business Engineer Framework
The Map of AI — Physical Layer & Where Value Is Captured
The Map of AI traces 9 layers across 200+ companies and identifies where durable competitive position actually lives in the stack. The import boom is a physical-layer story: chips, electrical gear, and assembly are the substrate everything else runs on — and today, most of that substrate is made somewhere other than the United States. Understanding which layer you occupy, and who controls the layers below you, is how you read the buildout’s real risk map.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
