Macro Check: The Mark AI Leaves in Official Statistics

Orange is 334, computer and electronic products, where AI-related orders land; blue is all durable goods excluding 334. New orders, year on year, three-month smoothed. In June 2026 the raw monthly figures were about 17% against 7%. In the past the two lines pulled apart mostly because the blue line collapsed in a recession (2009, 2020); this time the orange line is surging on its own: the AI pull is concentrated and the traditional economy is mild.

Left: nominal 334 new orders at $31.1B a month, a new cycle high but still below the $41.2B dot-com peak of June 2000 (dashed line), the measure of 26 years of offshoring plus hardware deflation; this is the US-factory view, not the whole of AI demand. Right: unfilled orders at a record $157.4B, nearly 30% above the 2019 average, orders received faster than they can be worked off; the semiconductor and electronic components production index is nearly 70% above 2019, so domestic capacity is running flat out too.

Left: BEA information-processing equipment investment ran at an annualized $798B in the second quarter of 2026, up about 31% year on year; it includes imported equipment, so it is more sensitive than factory orders, and it is the direct footprint of AI capex in the GDP accounts. Software investment was $815B, up about 9%. Right: the four hyperscalers’ all-in investment (CapEx + new leases), annualized, placed on the same chart: $726B against the national $798B, almost the same size. The two are not the same scope (the four include overseas spending and leases, BEA counts only US equipment), so this is not a share but a cross-check: the steep line in the macro accounts is mainly these four companies.

Left: private data-center construction spending, latest Aug 2026 at an annualized $85B, up 73% year on year, and above traditional offices (grey) since late 2025. It is the earliest official read on physical data-center investment, ahead of power draw and equipment delivery. Right: unfilled orders for electrical equipment (NAICS 335, transformers, switchgear and the like) at $56B, far above the 2019 average; orders that cannot be worked off are the grid-equipment bottleneck. The blue line is monthly new orders.

Most servers are assembled overseas, so the part US factory orders miss shows up at customs. HS 8471 imports reached $41B in the latest month, up 75% year on year; Mexico assembles the GB racks and Taiwan ships boards and complete systems, and the two together are most of the total. The series moves with BEA equipment investment and is the direct evidence for the supply chain sitting abroad.

The 50 states plus DC split into two groups: states where Epoch-tracked large AI sites reach at least 500MW of peak IT power, online plus under construction, count as data-center states (Texas, Wisconsin, Indiana, Ohio, Tennessee, Louisiana, New Mexico and others), the rest are other states. The construction shows: construction employment is up 2.4% in data-center states against 1.5% elsewhere (Sep 2026). The state economy does not: real GDP 1.6% against 2.5% (2Q26) and nonfarm payrolls 0.5% against 0.6% (Sep 2026) are no better. The reasons: data centers employ few people, construction is only a few percent of state GDP, and the output once running is booked where the cloud provider is headquartered; Texas and Ohio in this group also carry oil, gas and manufacturing drags, while the other group holds New York, California and Florida. So the claim that data-center states have unusually strong economies holds, in official data, only for the construction slice.

Orange marks states with AI data centers built or under construction. The orange bars sit in the middle; the leaders are there for other reasons, New York on finance, Idaho on in-migration, California on the output of tech headquarters. Single quarters carry oil, gas and farm noise, so read the distribution rather than any one state.

The orange line is the year-on-year change in semiconductor exports from the Korea Customs Service monthly report, latest 260.3% (Sep 2026), positive for well over a year with a record single month, carrying both the price rises and the volume of HBM and server memory. The blue line extends the series back to 2014 with UN Comtrade exports under HS 8541 plus 8542, a narrower definition about six months behind, latest 42.5% (Dec 2025); only on that length does it show this wave running at more than three times the 2017 to 2018 memory supercycle. It is the cleanest official number at the other end of the supply chain, available without waiting for company earnings.

Orange is manufacturing PMI new orders (SA) minus the makers’ own inventories, latest 2.2 (Sep 2026); blue is new orders minus customer inventories, latest 12.9, where the customers’ stock is the real downstream level, and blue turning negative first usually means customers have started destocking. A widening positive gap means demand runs ahead of stocking; a turn negative means inventories are catching up. Vertical lines are the past buy and sell-1 dates under this rule: buys in 2013, 2016, 2019 and 2023, sell-1 in 2015, 2018, 2021 and 2024; the April 2026 dip below zero has since reversed and is listed as unconfirmed. This chart tracks the cycle position, not the AI trend.