# Data-center manufacturing jobs in the monthly payroll data: method and replication This package rebuilds a two-panel chart from the BLS monthly payroll survey (Current Employment Statistics, CES). - **Left panel:** the "Data Centers Index" adapted to payrolls data from Zoffer & Sampson, *AI data centers: A hidden driver of U.S. jobs* (SIEPR Policy Brief, October 2026). It is plotted with all manufacturing, durable goods, and both of those excluding the data-center industries. - **Right panel:** the change in payroll employment in the same industries since January 2025. Run `python3 replicate_side_by_side.py`. It needs Python 3.9+ and pandas, and no API key. It downloads the data from the BLS API and writes the CSV files to `output/`. The script is commented step by step; this note explains the two choices behind it: **how the weights work** and **how the brief's industries were mapped to payroll industries**. --- ## 1. How the weights work **Where the weights come from.** - SIEPR's model estimates the U.S. manufacturing jobs needed to supply this year's data-center buildout, about 304,000. The estimate starts from projected 2026 U.S. data-center capital spending (Goldman Sachs). It splits that spending into components using component cost shares from CSIS, keeps the share of each component made in the U.S., and converts dollars of output into jobs using Census data. - Each component's share of the 304,000 jobs is its weight. The weights sum to 1. - The weights are hard-coded in the authors' notebook. The spreadsheet model that produces them is not published, and the brief does not give the data years for the domestic-share and jobs-per-dollar inputs. So the weights can be used, but not rebuilt from scratch. **What the weights are not.** - They are not each industry's size. - They are not the share of each industry's jobs that serve data centers. A component gets a large weight if a lot of data-center job creation runs through it. For example, structural steel for building shells gets 17% and chillers 14%. GPUs get 4%, because most are made abroad. **How the index uses them.** This follows footnote 28 of the brief. 1. Each industry's employment is expressed relative to its own January 2019 level, so January 2019 = 100. 2. The index is the weighted average of those industry indices. Index(t) = Σ weightₖ × 100 × employmentₖ(t) / employmentₖ(Jan 2019) **Worked example.** Fabricated structural metal (weight 0.168) went from 91,400 jobs in January 2019 to 108,100 in September 2026. Its industry index is therefore 118.3, and it contributes 0.168 × 118.3 = 19.9 points to the index. Adding up the 15 components gives 111.3. **What follows from this.** - **The weights are fixed.** The 2026 weights apply to every month back to 2015. The index asks how the industries that matter most to today's data-center buildout have grown since 2019. It does not track changes in the mix over time. - **Industry size does not matter.** A 50,000-job industry and a 400,000-job industry with the same weight move the index equally for the same percent change. The index is a weighted growth rate, not a job count. - **An equivalent way to see it.** Scale each industry so that its share of a January 2019 total equals its weight, add the scaled series, then index the total. This gives exactly the same numbers; the script computes both ways and checks that they match. So in January 2019 each component's share of the index equals its weight, and after that the shares drift with growth. - **Each industry counts in full.** The index tracks *all* employment in these industries, including work unrelated to data centers. It measures data-center job growth only if the data-center jobs in each industry grow at the industry's overall rate. - **The brief's job count rests on extra assumptions.** Its figure (about 158,000 through December 2025) multiplies the index's lead over manufacturing by the industries' total January 2019 employment. In effect it treats every job in them as a data-center job. That total also counts three nested industries twice, which the brief does not mention: 332312 sits inside 3323, 333415 inside 33341, and 334111 inside 33411. Counting each once gives a January 2019 base of about 970,000, not 1.26 million. **Two components share one industry.** Switchgear and power distribution units both map to NAICS 335313, so that industry carries two weights (0.069 and 0.054). That is how the authors' notebook enters it. **Check against the brief.** Run on QCEW data with the brief's industries, this method matches the notebook's printed results exactly: all manufacturing −1.44%, an index lead of 12.50 points, and about 157,256 jobs through December 2025. --- ## 2. Crosswalk: from the brief's QCEW industries to payroll (CES) industries The brief uses the Quarterly Census of Employment and Wages (QCEW). QCEW covers nearly every employer and is published about five months after each quarter ends. The monthly payroll survey (CES) is timelier, but it publishes less industry detail. Each of the brief's 14 industries is therefore mapped to the closest industry that CES publishes: - **8 exact matches:** CES publishes the same industry. For 33341 and 33411, the CES 4-digit industry (3334, 3341) contains only that one 5-digit industry, so it is identical. - **6 parent industries:** CES does not publish the brief's 6-digit industry, so the smallest published industry that contains it is used. These cover 7 of the 15 components and 37% of the weight. | Component | Weight | Brief's industry (QCEW NAICS) | Payroll industry (CES series) | Match | What else the payroll industry includes | |---|---|---|---|---|---| | Building shell and containment | 0.168 | 332312 Fabricated structural metal | CES3133231201 Fabricated structural metal | Exact | — | | Chillers and cooling towers | 0.143 | 333415 AC, refrigeration and forced-air heating | CES3133341501 same | Exact | — | | Computer room air conditioners and air handlers | 0.137 | 33341 Ventilation, heating, AC and commercial refrigeration | CES3133340001 NAICS 3334 (same industry) | Exact | — | | Diesel backup generators | 0.072 | 335312 Motors and generators | CES3133530001 Electrical equipment (3353) | Parent | Transformers; relays and industrial controls; switchgear | | Switchgear | 0.069 | 335313 Switchgear and switchboard apparatus | CES3133530001 Electrical equipment (3353) | Parent | Transformers; motors and generators; relays and controls | | Server racks | 0.065 | 332999 All other misc. fabricated metal products | CES3133299901 same | Exact | — | | Power distribution units | 0.054 | 335313 Switchgear and switchboard apparatus | CES3133530001 Electrical equipment (3353) | Parent | as for switchgear | | Security and monitoring systems | 0.053 | 334290 Other communications equipment | CES3133420001 Communications equipment (3342) | Parent | Telephone apparatus; broadcast and wireless equipment | | Fiber optic cable | 0.051 | 335921 Fiber optic cable | CES3133590001 Other electrical equipment and components (3359) | Parent | Batteries; other wire and cable; wiring devices; misc. electrical equipment | | UPS systems | 0.050 | 335999 All other misc. electrical equipment | CES3133599001 All other electrical equipment (33599) | Parent | Carbon and graphite products | | Pumps and valves | 0.045 | 33391 Pumps and compressors | CES3133391001 same | Exact | — | | GPUs and accelerator boards | 0.044 | 334111 Electronic computers | CES3133411101 same | Exact | — | | Raised floors, ceilings, other metal building parts | 0.025 | 3323 Architectural and structural metals | CES3133230001 same | Exact | — | | Fire protection systems | 0.023 | 332919 Other metal valves and pipe fittings | CES3133291001 Metal valves (33291) | Parent | Industrial valves; fluid power valves and hose fittings; plumbing fixture fittings | | Networking (interconnects and switches) | 0.001 | 33411 Computer and peripheral equipment | CES3133410001 NAICS 3341 (same industry) | Exact | — | The weights and NAICS codes come from the authors' notebook, which lists only codes. The component names are inferred from the brief's text and the NAICS titles, and the two cooling components could be the other way round. `output/components.csv` has the same table, with payroll levels for January 2019, January 2025 and the latest month. **Each component keeps its weight.** When a component is mapped to a parent industry, its SIEPR weight is applied unchanged to the parent's index. The parent is larger and includes unrelated products, so the weight then applies to a broader, likely slower-moving industry. This **dilutes** the signal. Making the same swap in QCEW, where both the brief's industries and the parents are available, shows how much: | QCEW, March 2026 | Index lead over all manufacturing | |---|---| | Brief's industries | 15.0 points | | Parent industries swapped in for the 6 unpublished ones | 11.9 points | The payroll version therefore probably **understates** the index's lead, by about a fifth on this test. Alternative fixes were tested on QCEW: subtracting the unrelated sibling industries, or dropping the parents and re-weighting. Both did worse, because some of the excluded industries, such as transformers and batteries, are themselves growing fast. Using the parents as they are came closest. Other differences between the two sources: | | QCEW | CES payrolls | |---|---|---| | Seasonal adjustment | Not seasonally adjusted | Seasonally adjusted | | Coverage | A count of nearly all employers, from unemployment-insurance records (private sector used here) | Estimates from a monthly sample survey of businesses (private sector) | | Benchmarking | Itself the benchmark | Re-benchmarked to QCEW each March; months after March 2025 are not yet benchmarked | | Timing | About five months after each quarter | A month after the headline jobs report for industry detail; recent months are preliminary | --- ## 3. The right panel: summed payrolls, no weights The right panel answers a simpler question: how many payroll jobs did these industries add? **What it adds up.** Payroll employment in 9 CES industries, about 1.44 million jobs in September 2026. These are the 13 payroll industries above, minus the 4 that sit inside another one on the list, so no job is counted twice: - fabricated structural metal sits inside architectural and structural metals; - AC and refrigeration sits inside HVAC; - electronic computers sits inside computer and peripheral equipment; - all other electrical equipment sits inside other electrical equipment. **What it shows.** The change in that total from January 2025 to the latest month. There are no weights. **The "ex data-center industries" bars.** These are the published totals for manufacturing, or for durable goods, minus the same sum. Durable goods contains every one of these industries. **How to read it.** - The bar equals the change in data-center jobs only if **all** of the change in these industries came from data-center demand. - Other demand, and the parent industries' unrelated products, push it the other way in either direction. The bar can overstate or understate the data-center effect. - Large industries dominate the sum. Architectural and structural metals alone is 429,000 jobs, so the summed growth (+5.7% since 2019) is much smaller than the weighted index (+11.3%). --- ## Sources - Zoffer, J. and Sampson, A., "AI data centers: A hidden driver of U.S. jobs," SIEPR Policy Brief, October 2026, and the accompanying notebooks. - BLS Current Employment Statistics, via the BLS public data API. - BLS Quarterly Census of Employment and Wages. It is used in the QCEW comparison above, not in this script.