US imports of AI-related hardware grown faster than domestic production in every major equipment category since 2019, according to data highlighted by venture capital firm a16z. Computer imports rose 366% over the period, while US production increased 46%, underscoring how heavily the country's AI infrastructure expansion relies on overseas supply.

The figures were published on October 25, 2026, as part of a16z's "Charts of the Week" series and span four categories central to the AI buildout: computers, communications equipment, semiconductors, and grid and electrical equipment.

Imports Outpace Production Across the Board

Communications equipment imports increased 175%, while domestic output grew 87%. Semiconductor imports climbed 131%, against production growth of 76%. Grid and electrical equipment showed the sharpest divergence: imports rose 69%, while US production fell 7%.

Those figures expose uneven domestic responses across the equipment categories supporting the AI expansion. The comparison measures changes in dollar values rather than providing a direct count of machines or components. Because the figures are denominated in dollars, they can move with prices as well as volumes, so the data shows how import and production values have diverged without establishing their respective shares of total supply.

Where Production Gains Came From

Manufacturers can raise output by adding capacity or by using existing facilities more intensively. According to a16z, those two responses have developed differently across the four categories. Separating the two helps show whether growth comes from newly built capacity or from existing plants running closer to their limits.

Computer manufacturers expanded capacity and increased utilization, meaning they used more of their available production capability. Together, those changes supported higher output.

Communications equipment makers added substantial capacity, but utilization declined. That combination suggests available manufacturing capability grew faster than its use.

Semiconductor manufacturers also expanded capacity, while utilization remained unchanged. Their production increase therefore came from additional capacity, rather than more intensive use of existing facilities.

The Electrical Equipment Gap

Electrical equipment followed a different path. Capacity declined despite higher utilization, leaving domestic production below its 2019 level — the only one of the four categories where output shrank.

That weakness matters because the AI buildout requires electrical infrastructure alongside computing hardware. Switchgear and related equipment support the grid expansion needed to power GPUs, and imports help meet demand when domestic output cannot keep pace. The divergence leaves the United States more reliant on foreign suppliers for the equipment that underpins that buildout.

The supplied figures do not establish why manufacturers avoided greater expansion. a16z suggests construction costs, investment risks, or competition from overseas suppliers could explain that hesitation. These remain possible explanations, rather than confirmed causes.

a16z summarized the trend in a post on X:

Since 2019, imports of AI hardware have grown faster than US production in every category

US production vs. imports: – Computers: +46% vs. +366% – Comms equipment: +87% vs. +175% – Semis: +76% vs. +131% – Grid & electrical equipment: -7% vs. +69%

Charts of the:… pic.twitter.com/wJN2meTEdu

— a16z (@a16z) October 25, 2026

(a16z on X)

GPU Rentals and Token Demand

Meanwhile, rental data points to continued demand for computing resources. According to Ornn Data, every tracked GPU model costs more to rent than at the beginning of the year, including older A100 chips. Rental prices stood between 9% and 143% above a benchmark assuming straight-line depreciation over six years. Straight-line depreciation spreads a machine's loss of value evenly across its assumed lifespan, so rates above that benchmark indicate rental pricing running ahead of what steady hardware depreciation alone would imply. a16z notes that older A100s and H100s — Nvidia data center chips still widely deployed for AI workloads — previously spent more time below that benchmark.

OpenRouter's figures also show rising token consumption across open and proprietary models. Tokens are the units of text that these AI systems process. a16z cautions that these datasets provide incomplete views of GPU pricing and token spending.

Labor Pressures Build

Labor adds another constraint. Citing payroll processor ADP, a16z reports that weekly hours per worker increased substantially across sectors supporting the AI buildout — a measure of how intensively existing employees are being used, separate from headcount growth. Job-switchers also commanded a five-percentage-point wage-growth premium over workers who stayed with their employers, meaning moving between employers currently comes with larger wage gains than staying put.

Taken together, the figures point to growing hardware demand, uneven factory expansion, and heavier workloads across the AI supply chain. Imports are helping bridge the gap, while electrical equipment remains the clearest domestic production weakness. Future installments of a16z's weekly chart series, along with updated readings from Ornn Data, OpenRouter, and ADP, offer a basis for tracking whether these patterns persist.