Charts and data via the Financial Times (US Census Bureau and UN Comtrade).
US imports of data-center hardware quadrupled in two years — and the customs data, charted by the Financial Times from US Census Bureau and UN Comtrade figures, shows exactly where all that hyperscaler capex physically lands.
What Happened
The clearest measure of the AI buildout may not be a capex line or a financing deal — it may be a customs form. According to Financial Times analysis of US Census Bureau and UN Comtrade data, US imports of data-center hardware filed under tariff code HTS 847150 — the enterprise and AI server category — more than quadrupled from $37 billion in 2023 to $165 billion in 2025. Taiwan remains the single largest source at roughly $86 billion; Mexico has climbed to a close second at roughly $80 billion, or about 40% of total US imports. Both have risen several-fold since 2021. The CSIS analysis underlying the data makes clear that this is a structural shift, not a rounding error.
The more arresting data point sits at the country level: in the first five months of 2026, Mexico’s exports of AI and computing hardware reached a record $105.8 billion — up 84.5% year on year — surpassing its car exports of about $60.5 billion for the first time in the country’s modern export history. HTS heading 8471 shipments were up roughly 145% in 2025 alone, knocking automobiles off the top spot. The assembly is carried out by contract manufacturers — Foxconn, Flex, Jabil, Pegatron, and Sanmina — operating in Mexican border-state industrial parks, with finished systems trucked north under USMCA. Manufacturing FDI into Mexico reached $40.9 billion in Q3 2025, a figure that reflects the factory-floor consequence of decisions made in hyperscaler boardrooms.
Hold the headline carefully, though. Around 95% of the value of Mexico’s server exports consists of imported components — predominantly Taiwanese — so Mexico captures the labor of final assembly, not the margin of silicon. Taiwan is not being displaced; it remains the largest US supplier and owns the guts of every machine trucked north. HTS 847150 also lumps AI accelerator servers together with standard enterprise hardware, and a meaningful share of the recent surge reflects importers front-loading shipments ahead of possible tariffs on exactly this category. Customs data double-counts components that cross borders twice — first as chips, then again inside finished systems. The signal is real; the vertical line on the chart still borrows partly from the future.
The key insight: Mexico overtaking its own car exports with server shipments is revealing not because Mexico out-innovated Taiwan — Taiwan still supplies the chips and remains the largest US source — but because the AI infrastructure boom has grown large enough to reorganize an entire nation’s export base around assembling someone else’s silicon. That is what a trillion-dollar capex cycle looks like when it hits the physical economy.
The Structural Read
Think of this as the AI capex boom made physical. The AI Capex Map frames the trillion-dollar compute spending cycle in financial terms — hyperscaler balance sheets, structured-finance SPVs, the $35 billion compute securitization vehicles assembled by Apollo, Blackstone, and Broadcom around Anthropic and others. The customs data is what that abstraction resolves to at the border: $165 billion worth of pallets of servers, in a single year, flowing into US data centers. The financing vehicle is the paper; the HTS 847150 import line is the iron.
The supply topology of the compute floor runs in three stages. Stage one: chips designed and increasingly fabricated in Taiwan — TSMC’s 3nm ramp remains the chokepoint of the entire AI stack, and Taiwan’s direct US server imports have risen several-fold since 2021. Stage two: those components shipped to Mexican border-state facilities where Foxconn, Flex, Jabil, Pegatron, and Sanmina perform final assembly — bolting Taiwanese silicon into finished rack-ready systems, capturing the labor value but not the semiconductor margin. Stage three: finished servers trucked north under USMCA, arriving inside the tariff wall, close to the hyperscale data centers that ordered them. The geography is a nearshoring topology, not a technology transfer.
This is also where the limits of NVIDIA’s moat analysis become visible in physical form. The chip design and advanced fabrication layer — where the real margin lives — stays in Taiwan. The assembly layer — commodity labor, thin margins, policy-exposed — moves to Mexico. Mexico’s export miracle is real in volume terms and fragile in value terms: 95% imported content means the country is one tariff decision or one supply-chain reroute away from seeing those numbers reverse. The CXMT DRAM build-out in Beijing is a reminder that the memory layer of this same stack is being contested separately — the compute floor has multiple pressure points, and Mexico sits at the most exposed one.
Business Engineer — The Capex Boom Made Physical
“The abstraction of a $35 billion compute SPV resolves, at the border, into pallets of servers. The AI buildout is now large enough to reorganize a nation’s entire export base around assembling someone else’s chips — and the customs data is the most honest account of where all that capital actually lands.”
Three Implications
IMPLICATION 1 — NEARSHORING IS STRUCTURAL, NOT OPTIONAL
The USMCA assembly topology — Taiwanese silicon, Mexican labor, US data centers — is not an accident of comparative advantage. It is the deliberate output of hyperscaler procurement strategies optimizing simultaneously for tariff exposure, delivery speed, and geopolitical hedging. Foxconn, Flex, Jabil, and Sanmina expanding Mexican capacity is the contract-manufacturing layer of the AI capex cycle institutionalizing itself. Even if tariffs land on this exact hardware category, the assembly infrastructure will not simply relocate overnight — the $40.9 billion Q3 2025 manufacturing FDI figure represents sunk costs that anchor the topology.
IMPLICATION 2 — THE TARIFF FAULT LINE IS THE REAL RISK VARIABLE
HTS 847150 — AI and enterprise servers — is precisely the category US trade policy has signaled it may target. Part of the vertical line on the import chart is front-loaded demand: importers pulling shipments forward before duties land. That means some portion of the 2025–2026 surge borrows from future quarters rather than reflecting pure demand growth. If tariffs land as signaled, the flow will not stop — Taiwan and Mexico remain structurally necessary — but the timing and cost structure of hyperscaler procurement shifts materially. The companies most exposed are the contract manufacturers with thin margins and fixed Mexican footprints; the companies least exposed are the chip designers in Taiwan, whose product has no substitute.
IMPLICATION 3 — VALUE CAPTURE STAYS IN THE SILICON, NOT THE ASSEMBLY
Mexico’s servers-overtaking-cars story is volume-first and margin-thin. With ~95% of export value as imported content, the country’s GDP gain is a fraction of the headline trade figure — it is capturing assembly wages, not semiconductor economics. The structural lesson for any nation watching this play out is the same one the smartphone supply chain taught a decade ago: the entity that designs and fabricates the leading-edge chip captures the durable margin; the entity that bolts the chip into a chassis captures a tariff-exposed slice of labor value. Taiwan’s position at the top of the US import table, even as Mexico assembles more finished units, is the clearest expression of that asymmetry.
