RESEARCH · AI INFRASTRUCTURE
Data centers just
overtook chip fabs
For the first time in nearly five years, America is putting more construction money into data centers than into semiconductor plants. The fab-construction boom has been unwinding for two years; the data-center line hasn't blinked once. Every number here resolves to a SHA-256-verified Census cell.
Two construction lines carry the American AI buildout. One is the fab line — the CHIPS-era factories meant to make the chips. The other is the data-center line — the buildings the chips compute in. In the Census Bureau's own monthly ledger of construction put in place, the fab line just surrendered a lead it had held since June 2021. As of April 2026, America spends more building data centers than chip fabs — and May widened the gap.
Two lines, five years, one handoff.
In April 2026 private data-center construction ran at a $58,977M seasonally-adjusted annual rate against $58,954M for computer/electronic/electrical manufacturing plants — the Census line that carries the semiconductor fabs. A $23M edge: a whisker, well inside revision range. Then May reported: data centers $59.3B, fabs $57.0B — a $2.3B gap, and the whisker became a lead. The last time the data-center line sat on top, it was May 2021 and both lines were under $10B.
source Census C30 · Value of Construction Put in Place, private · series Data center vs. Computer/electronic/electrical · basis seasonally-adjusted annual rate · window Jan 2020 → May 2026 · sha256-verified raw · census.gov ↗
The fab-construction boom is two years into reverse.
The fab line's arc is violent in both directions. It ran near $8B/yr at the end of 2020. The CHIPS Act was signed in August 2022 with the line already sprinting — $45B that month, $70B by December, $121B by November 2023. It peaked at $126.4B in June 2024 — and then fell in 21 of the next 23 months, down to $57.0B in May 2026. That is 55% off the peak, and the lowest reading since September 2022 — one month after the CHIPS Act was signed.
Read the measure carefully: construction put in place counts the shell work happening now — concrete, steel, cranes. A falling line doesn't mean fabs are being abandoned; it means the giant pours of 2023–24 are finishing, and new starts aren't replacing them at anything like the same pace. The buildout's shell phase is ending. What it doesn't say alone is whether the projects moved to the next phase — that shows up in a different dataset, below.
The data-center line has never had a bad year.
Against that round trip, the data-center series is almost monotonous. It held near $9–11B/yr through 2020–21, then compounded: $26.4B in January 2024, $45.3B a year later, $59.3B in May 2026 — up 125% in 29 months, roughly 6x its 2021 pace, and an all-time record in each of the last six months. No boom-bust arc, no CHIPS-style inflection — just a line that has climbed through every narrative cycle of the AI trade.
One honest caveat: the rotation is not additive. Combined, the two lines peaked in June 2024 at $161.7B and now run $116.3B — down 28%. Total AI-shell construction is shrinking; what changed is who's doing the building.
The tools kept coming after the concrete stopped.
If fab construction were collapsing outright, you'd expect the machines to stop arriving too. They haven't. US imports of semiconductor-manufacturing equipment (HS-8486) — the lithography, deposition, and etch tools that fill a finished shell — have run near $1B a month for three years, through the entire construction unwind. They sagged to a trough of $692M in November 2025, then jumped to $1.53B in April 2026 — the strongest month in a year and a half, up 122% off the trough.
source Census International Trade · HS-8486, all countries total · measure general imports, customs value · window Jan 2024 → Apr 2026 · sha256-verified raw · census.gov ↗
That is the reading that ties this piece to the last one. The payroll census showed the fab buildout producing buildings and paychecks but not yet production jobs. The construction ledger now shows the shells finishing. The trade data shows the tools still flowing in — into buildings that are done being built. Three federal datasets, one buildout moving through its phases: concrete, then tools, then — still pending — the jobs. Meanwhile the data-center side skipped the debate entirely and just kept pouring.
The crossover is a $23M whisker. Verify it yourself.
A claim this fine lives or dies on the exact cells. Both April 2026 figures resolve through get_source_evidence_v1, which re-opens the raw Census workbook server-side, re-checks its SHA-256, and hands back the cell.
Built with the live exascale.build AI-infrastructure API — the agent-ready OSINT layer for the US machine-economy buildout. Figures are US Census Bureau C30 "Value of Construction Put in Place," private construction, seasonally-adjusted annual rate — what the current monthly pace of shell work annualizes to, not spending that month. "Data center" is Census's named subcategory under Office. "Computer/electronic/electrical" is the manufacturing line that carries the semiconductor-fab buildout; Census publishes no finer split, so it also includes non-semiconductor electronics plants. Construction put in place counts buildings, not the equipment later installed in them. May 2026 values are Census-preliminary (May-26p) and April 2026 revised (Apr-26r) — the $23M April edge is within revision range; May's $2.3B gap is the firmer reading, and future revisions could move the crossover month. Construction data as_of 2026-05. Equipment imports are Census International Trade, HS-8486 general imports (customs value, all-countries total), which also includes flat-panel-display machinery — Census does not split it out; as_of 2026-04. Interpretations of cause are the author's reading; the figures are the source's.