Nuclear Technician Postings Rose 68%. BLS Said the Job Was Shrinking.

Job postings for nuclear technicians rose about 68% in 2025 over 2024. The same year, the Bureau of Labor Statistics projected that the occupation would shrink 8% over the coming decade. One of those two signals was wrong, and it was not the postings: in its next annual update, the BLS flipped the forecast to +1%.

That reversal is the sharpest example in a new Brookings analysis by Rachel Lipson and Allison Schwartz, published September 10, 2026, arguing that the labor market data most policymakers rely on (NAICS industry codes, SOC occupation codes, O*NET task lists, the monthly jobs report) was built for consistency, not for spotting the jobs that new technologies are creating right now. If you are wondering whether a job that did not exist five years ago is real, this post walks through why the official numbers may not be able to tell you, and what they miss along the way.

One caveat before anything else. This site maps AI exposure onto 1,016 ONET occupations. Every blind spot the authors describe in SOC and ONET is a blind spot in our data too. We come back to that at the end.

The AI debate is about job losses because nobody can count the new jobs

[Claim] The authors open with a diagnosis rather than a statistic: public debate about AI and work is dominated by displacement, even though the historical pattern is that new technologies create more jobs than they destroy on net. Their explanation is that we have far less clarity about what the new jobs are, partly because the instruments used to count jobs were not designed to see them in real time.

[Fact] The piece draws on Lipson's research for a forthcoming book on job training in what she calls "frontier regions," places that were early to deploy critical and emerging technologies, and on Revelio Labs data covering more than 1 billion worker profiles plus job postings. Brookings is a think tank, not a statistical agency, and this is an analytical article rather than a peer-reviewed paper. Read the numbers below as the authors' own tabulations of private data, not as official statistics.

Blind spot 1: A data center company can be a landlord, a software publisher, or a retailer

[Fact] Under NAICS, firms self-classify by the activity that generates the most revenue. The authors show what that does to data centers. CyrusOne, which leases facilities to hyperscalers, could plausibly sit under 531120 (Lessors of Nonresidential Buildings). CoreWeave, which sells AI compute through a software platform, could sit under 513210 (Software Publishers). Meta, Google, and Amazon, which build and operate some of the largest data centers in the United States, are filed under their parent codes: social networking, software publishing, retail.

[Fact] The code most often used for data centers, 518210, captures computer occupations and office support, but not the construction and maintenance work needed to build and run the facilities. The authors report the same problem for quantum technology, where ten quantum companies could reasonably be assigned to roughly ten different NAICS codes.

[Claim] The practical consequence, in the authors' framing: pick a different industry code and you get a completely different list of jobs to train for. For a community college designing a data center program, the official data cannot say whether the demand is for property managers, software engineers, or electricians.

Blind spot 2: Supply chain jobs triple the semiconductor picture

[Fact] Most CHIPS Act workforce analyses use NAICS 3344, semiconductor and other electronic component manufacturing. The authors expanded their count of semiconductor job postings from 2023 onward to include thirteen supplier codes: industrial gases, basic inorganic chemicals, semiconductor machinery, printed circuit boards, electronic connectors, machinery repair, and others. Including the supply chain nearly tripled the number of postings compared with the core manufacturing code alone.

[Fact] They note that this lines up with a Brookings Paper on Economic Activity by Erten, Stiglitz, and Verhoogen, which found the CHIPS Act's estimated job impact roughly doubles once upstream suppliers and construction are counted.

[Fact] The sequencing point matters for workers. Factories have to be built and equipment installed before a fab produces a single chip, so supplier and construction jobs show up first. The authors heard the same from the nuclear industry: microreactors and small modular reactors are designed to be assembled in factories, so the earliest hiring is for welders, machinists, and operators in fabrication shops that serve multiple industries and will never appear under the nuclear power generation code, 221113.

Blind spot 3: The job with no SOC code doubled anyway

[Fact] The authors borrow the term "frontier jobs" from David Autor and colleagues for roles that did not exist before a technology did. Their example is the "biomechatronics technician," a worker who understands biomanufacturing processes and can also troubleshoot robotic liquid handlers and automated bioreactors. There is no SOC code for it. In government data the person is filed as an industrial machinery mechanic, a mechanical engineer, or a bioengineer.

[Fact] Using Revelio's resume-based clustering rather than SOC codes, the authors count workers focused on machine maintenance inside pharmaceutical and biopharmaceutical manufacturing at roughly 26,000 in 2008 and 57,000 in 2025, more than double.

[Estimate] Spread across those 17 years, that is a compound growth rate of about 4.7% a year, sustained through a recession and a pandemic. That is our arithmetic, not the authors'. It also is not a headline that any SOC-coded series could have produced, because the series does not exist.

[Fact] In North Carolina, employers including Eli Lilly, bioMérieux, and Fujifilm are already posting for this combination of skills, most listing a high school diploma as the minimum and many preferring an associate degree. Wake Technical Community College is launching a dedicated training program. The demand exists on the ground; the category does not exist in the data.

Blind spot 4: Same title, different job

[Fact] The authors' fourth category is "retooled jobs," roles that keep an existing title but require new skills tied to emerging technology. Their case is a journeyman electrician posting at Base Power, a Texas home battery company that raised $2 billion in venture funding over the past year. The electrician still needs a traditional license, but the job now involves systems where power flows both ways between grid and battery. O*NET's task list for electricians does not fully capture that.

[Claim] Here the article turns explicitly to AI. It cites OpenAI's finding, from work-related ChatGPT conversations mapped to O*NET, of significant "task crossover": workers using AI to do tasks outside the traditional boundaries of their stated occupation. We covered that report when it came out. The authors treat AI as an accelerant of the retooling problem, not its only source; their electrician example is about batteries, and they also mention automotive programs still teaching carburetors.

Blind spot 5: The forecast said minus 8. The postings said plus 68.

[Fact] Official occupation projections are 10-year forecasts, and the monthly jobs report tracks payroll employment by industry, not occupation. Neither is a real-time hiring signal. JOLTS, the openings series, is also by industry.

[Fact] The nuclear example makes the gap concrete. In August 2025 the BLS projected nuclear technician employment would decline 8% from 2024 to 2034. In the same period, private investment in advanced reactors hit record levels, Google and Amazon signed deals to power AI data centers with small modular reactors, and the federal government made nuclear a priority. Revelio postings for nuclear technicians were up about 68% in 2025 versus 2024. In its next annual update, the BLS reversed course to +1% for 2025 to 2035.

[Estimate] That is a swing of nine percentage points in the ten-year outlook within a single revision cycle. For a worker deciding in late 2025 whether to enter a two-year nuclear technology program, the official number pointed the wrong way for a full year.

[Claim] The authors are careful not to overclaim in the other direction. They cite the 2017 OECD projection of more than 4 million truck driver jobs lost across the United States and Europe by 2030 on the assumption of rapid driverless-truck deployment, which has not materialized. Postings can be wrong too. The argument is for using private signals as a complement to official data, not a replacement.

What this means if you are the worker in the job center

The article is written for education and workforce planners, but three of its risks land directly on individuals.

If funding for training is tied to lists of "in-demand" occupations built from historical data, the emerging fields are the ones most likely to be locked out. The authors give the example of advanced photonics manufacturing in western Massachusetts, where training dollars are tied by statute to placement rates that a new program cannot yet show. If you are looking for a subsidized path into a frontier field, the subsidy may not exist yet precisely because the field is new.

The opposite failure is real too. Programs for electric vehicle fleet technicians were stood up in anticipation of jobs that did not appear locally and had to close or pivot. A university leader described the same cycle during the Obama-era wind and solar push. A program existing is not proof the demand is real.

And the quietest risk is the curriculum lagging the job. The authors' point about electrical training still centered on lighting and outlets while solar and battery work goes untaught applies to any field where the title is stable and the tasks are not.

Their five practical steps for planners translate into questions a worker can ask a training provider: Which announced investments in this region are you tracking? Are you using job postings or only BLS projections? Have you mapped the supply chain, not just the headline industry? Which employers gave input on this curriculum, and when? Is this occupation on the state's in-demand list, and if not, is there a petition process like the one Texas recently added?

Where our own data sits in this critique

This is the part we owe you. Our occupation pages are built on ONET, and ONET is built on SOC. The authors' argument is that this structure cannot see frontier jobs, undercounts supply chain hiring, and lags retooling within existing titles. All three apply here.

Concretely: there is no page on this site for nuclear technicians, machinists, or semiconductor processing technicians, three of the occupations in this article, because our 1,016-occupation coverage does not include them. There is no page for biomechatronics technicians because no such SOC code exists anywhere. The pages we do have for electricians, industrial machinery mechanics, mechanical engineers, biomedical engineers, welders, and nuclear engineers describe the "typical" version of each job, which, as the authors put it, is not necessarily what the job looks like at the technological frontier.

[Claim] Our reading is that this cuts both ways for AI exposure scores specifically. A score attached to "electrician" measures the O*NET task bundle for electricians. If the retooled version of the job has shifted toward bidirectional power systems and diagnostics, the exposure of the actual work may differ from the exposure of the coded work, in either direction. We do not know which. Neither does anyone else using the same classification, which is nearly everyone.

[Fact] The authors close on the institutional side: the Department of Labor is standing up an AI Workforce Research Hub under the 2025 America's AI Action Plan, has issued a request for information on modernizing O*NET including faster AI-enabled task updates, and the Census Bureau is adding AI adoption questions to its business surveys. Some states are adding occupational codes to unemployment insurance wage records. If those changes land, the data underneath sites like this one gets better. Until then, treat any occupation-level number, including ours, as a description of the job as it was coded, and check it against what employers near you are actually posting.

Sources

• Lipson, R., & Schwartz, A. (2026). Traditional labor market data isn't keeping up with jobs in the 'frontier economy'. Brookings Metro, September 10, 2026. https://www.brookings.edu/articles/traditional-labor-market-data-frontier-economy/
• Related coverage on this site: BLS 2025-35 projections and the new AI exposure categories and OpenAI's 2026 workplace ChatGPT report, which is the task crossover source the authors cite.

The Brookings article is copyright The Brookings Institution; quotations here are short and attributed, and none of its tables or figures are reproduced. Job posting and worker profile figures are the authors' tabulations of Revelio Labs data and are not official statistics.

This article was produced with AI-assisted analysis and reviewed for factual accuracy against the source article. Statements tagged [Fact] are taken from the article, [Estimate] are our own calculations from figures it reports, and [Claim] are interpretations, ours or the authors', that the data do not establish on their own.