Version 4

AI compute demand has driven hedonically-adjusted semiconductor import prices to historically anomalous levels — up 3.6% in May 2026 alone and 14.4% year-over-year — reflecting genuine chip scarcity rather than a shift to higher-performance products.[3][4] High-bandwidth memory now consumes roughly 23% of global DRAM wafer capacity,[5] and the shortage has spread from AI data centers to consumer electronics, with Apple's Tim Cook stating memory-driven price hikes are unavoidable.[6][7][8] A 25% Section 232 tariff on advanced semiconductor imports adds a policy-driven cost layer for U.S. buyers,[12][13] while U.S. export controls on AI GPUs have created a mirror effect in China: the Nvidia DGX B300 Blackwell server now sells there for over $1.1M, roughly 2.75 times its $400K U.S. retail price.[16] New items this pass hint at additional export control structure — an H200 sales cap and case-by-case BIS review policy for China — but lack extracted claims sufficient to update the narrative.[17][18]

Two decades of steadily falling compute costs have reversed, and the pressure runs in both directions geographically: U.S. buyers face tariff-inflated import costs on restricted chips, while Chinese buyers face export-control-driven premiums on the same finished hardware. The HBM shortage constraining AI infrastructure is the same shortage now repricing consumer electronics, and new fab capacity cannot relieve it quickly.

For roughly two decades ending around 2020, hedonically-adjusted import prices for computers and semiconductors fell steadily — by about 52% in total — as Moore's Law reliably delivered more compute per dollar with each chip generation.[1] Hedonic adjustment matters here: the price index accounts for improvements in chip speed and capacity, so a rising index cannot be explained by buyers simply upgrading to faster hardware.[2] When the index rises, buyers are paying more per unit of compute, independent of product generation. That index is now rising sharply: in May 2026, semiconductor and computer import prices increased 3.6% for the month and 14.4% year-over-year.[3] SemiAnalysis attributes this to genuine supply shortages driven by AI demand outpacing Moore's Law cost reductions — not a statistical artifact of product-mix shifts toward higher-end chips — and notes the data is calculated pre-tariff, meaning actual buyer costs are higher still.[4]

A specific bottleneck within the broader shortage is high-bandwidth memory. HBM is the memory architecture used in AI accelerators, and it now consumes roughly 23% of global DRAM wafer capacity,[5] crowding out DRAM production for other applications. The shortage has spread visibly to consumer electronics: Apple's Tim Cook has confirmed that memory-driven price increases on consumer products are unavoidable.[6][7][8] Korea's June memory semiconductor exports were simultaneously on track to set an all-time record, reflecting sustained demand even as supply strains persist,[9] and Micron's stock crossed $1,200 as investors identified DRAM and HBM suppliers as direct beneficiaries of the crunch.[10][11]

A compounding policy layer runs in parallel. On January 20, 2026, the White House imposed a 25% Section 232 tariff on certain advanced semiconductor articles, justified on national supply-security grounds.[12][13] Because the import price data SemiAnalysis cites is calculated before tariffs apply, the 14.4% year-over-year increase reflects market prices only; total landed costs for U.S. buyers exceed that figure by the tariff margin.[4] Industry opposition has organized: a CTA-led coalition wrote to Commerce Secretary Howard Lutnick in May 2026 arguing against the tariffs,[14] and the Information Technology and Innovation Foundation published analysis in June 2026 arguing the tariffs risk harming U.S. economic growth.[15]

U.S. export controls on AI hardware are producing a parallel and inverse effect in China. Financial Times reporting shows that Nvidia AI GPUs restricted from export to China have approximately doubled in price in the Chinese market: the DGX B300 Blackwell server rose from roughly $550K to over $1.1M, even though its U.S. retail price is approximately $400K — meaning Chinese buyers now pay nearly three times the U.S. price.[16] Additional reporting suggests the export control regime may have a more specific structure, including a reported cap on H200 unit sales and a BIS shift to case-by-case review for certain chips, though extracted claim detail on these items is limited.[17][18] The two policy levers — Section 232 tariffs and export controls — therefore operate in opposite geographic directions: one raises input costs for U.S. importers of semiconductor components, the other restricts outbound access for Chinese buyers of finished AI compute hardware.