Chapter 4 · Job accounting
How many jobs has AI created?
Not currently established by the evidence assembled. What exists are honest answers to narrower questions. They use different units, boundaries and counterfactuals, so they sit side by side here and are never added.
Employment in NAICS 518 (computing infrastructure, data processing, web hosting)
- Boundary
- NAICS 518
- Stock or change
- stock, n/a
- Channel
- direct; mixed
- AI-specific
- AI-adjacent industry
- Counterfactual
- None
- Overlap group
- operations
Permitted claim. Existing employment in the industry that houses data centers. Not jobs created by AI.
Census 'data center' employment 2016 to 2023
- Boundary
- Data-center industry (likely NAICS 518)
- Stock or change
- change, net
- Channel
- direct; mixed
- AI-specific
- AI-adjacent industry
- Counterfactual
- None
- Overlap group
- operations
Permitted claim. Net growth in a broad industry before and during the generative-AI period. Overlaps J1; never add.
Manufacturing jobs supported by data-center capex (SIEPR)
- Boundary
- Upstream manufacturing
- Stock or change
- stock, gross
- Channel
- indirect; unknown
- AI-specific
- AI-adjacent industry
- Counterfactual
- No substitution assumed
- Overlap group
- supply-chain
Permitted claim. Modeled employment supported at an investment level. Not new hires, not net.
SIEPR 'Data Centers Index' outperformance vs all manufacturing
- Boundary
- Manufacturing sectors supplying data centers (QCEW)
- Stock or change
- change, net
- Channel
- indirect; unknown
- AI-specific
- not AI-specific
- Counterfactual
- All manufacturing trend
- Overlap group
- supply-chain
Permitted claim. Supplier industries outgrew manufacturing overall; these industries also serve other buyers. Overlaps J3.
Headcount effect at high-intensity AI adopters (Ramp/Revelio)
- Boundary
- Mixed, mostly Information
- Stock or change
- change, net
- Channel
- direct; unknown
- AI-specific
- AI-specific
- Counterfactual
- Not-yet-treated firms in same intensity group
- Overlap group
- firm-level
Permitted claim. A percentage effect within a selected sample. Do not multiply by national employment.
Businesses reporting AI changed headcount (Census BTOS)
- Boundary
- All employer businesses
- Stock or change
- rate, n/a
- Channel
- direct; unknown
- AI-specific
- AI-specific
- Counterfactual
- Respondent judgment
- Overlap group
- firm-level
Permitted claim. Shares of businesses, not numbers of jobs.
Colossus 2 jobs (xAI statement)
- Boundary
- Data center operations
- Stock or change
- change, gross
- Channel
- direct; permanent
- AI-specific
- AI-specific
- Counterfactual
- None
- Overlap group
- local
Permitted claim. Company claim. Not realized employment.
Net national jobs created by AI
- Boundary
- All
- Stock or change
- change, net
- Channel
- mixed; unknown
- AI-specific
- AI-specific
- Counterfactual
- Would require economy-wide identification
- Overlap group
- none
Permitted claim. Not currently established by the evidence assembled.
Industry-boundary audit
The 'data center jobs' series is broader than data centers, and it has stopped growing.
Census reports data-center employment rising from 306,000 (2016) to 501,000 (2023), citing BLS. The article does not print its industry code; the scale matches NAICS 518 (computing infrastructure, data processing, web hosting), which also covers hosting and data-processing services with no physical data center. On BLS monthly data that industry peaked in early 2023 and was 454,800 in October 2026 (preliminary), while construction spending on data centers multiplied. Building data centers employs construction and manufacturing workers; operating them employs relatively few.
Employment in NAICS 518
Rose from about 331 thousand in Jan 2019 to about 485 thousand in early 2023, then edged down to 455 thousand.
Caveat. Includes web hosting and data processing, not only physical data-center operations. Not AI-specific.
Inspect the evidence
- BLS API, series CES5051800001
- No transformation
Text description. Rose from about 331 thousand in Jan 2019 to about 485 thousand in early 2023, then edged down to 455 thousand.
Private construction spending on data centers
Rose from about $1.6 billion a year in 2014 to $96.5 billion a year in September 2026.
Caveat. Measures construction put in place (buildings, power and cooling fit-out), not the chips inside. Latest months preliminary.
Inspect the evidence
- privsatime.xlsx, column 'Data center'
- Month labels parsed; p/r flags kept
Text description. Rose from about $1.6 billion a year in 2014 to $96.5 billion a year in September 2026.
Payroll jobs in AI-supply and everyday industries, January 2019 = 100
Specialty trade contractors and electrical equipment grew faster than manufacturing overall; semiconductors were roughly flat.
Caveat. These industries serve many customers besides data centers. Diverging lines are consistent with, but do not prove, an AI-infrastructure effect (SIEPR makes the same point about its own index).
Inspect the evidence
- Jobs in these industries attributable to AI demand.
- Construction job-years on specific projects.
- BLS CES series as labelled
- Each indexed to its own Jan 2019 level
Text description. Specialty trade contractors and electrical equipment grew faster than manufacturing overall; semiconductors were roughly flat.
Replication assessment
SIEPR: about 304,000 manufacturing jobs supported
What it is. A modeled level of manufacturing employment supported by one year of U.S. data-center capital spending at the 2026 run rate (Goldman Sachs forecast), split into components (CSIS shares), reduced for imports (Census trade and production, 2022), converted to jobs with Economic Census labor intensity.
Reproducible here? Partly. Every input is public in principle, but the brief does not publish its component-to-NAICS crosswalk or the capex dollar figure. Not yet reproduced; marked as an open task.
Audit. GPUs are 48% of capex but 4% of domestic jobs (import dependence). The estimate excludes substitution away from other buyers of the same inputs, which would lower it, and excludes deeper supply-chain and induced jobs, which would raise it. Sensitivity: about 150,000 at half the capex, 450,000+ at 1.5 times.
Permitted wording. Modeled jobs supported at an investment level. Not new hires, not net, not a count.
Study critique
Ramp and Revelio: heavy adopters hired more
Design. 21,559 linked firms, 2021 to Feb 2026. Callaway and Sant'Anna staggered difference-in-differences comparing adopters with later adopters in the same eventual spending-intensity group, with sector fixed effects.
Result. High-intensity adopters (top tercile of AI spend per employee) +10.2% headcount over 24 months (coefficient 0.097, SE 0.022); entry-level +11.3 log points. Low-intensity adopters: −0.6%, not significant.
What weakens it. Intensity is chosen by firms that may already expect to grow; 3 of 11 pre-periods are flagged for the headline result; the high group averages 26.7 employees and is 63.8% tech-adjacent; only Information is statistically significant (+13.4%). Industrial, logistics, health and consumer groups are near zero or imprecise. Workforce counts come from online-profile data (Revelio), which under-cover some occupations.
Permitted wording. An association in a selected sample of Ramp customers, concentrated in software and media firms. Do not scale to national employment.