Back to Research
ResearchReference document

Vibe manufacturing: given the premises, what can one do?

Take as given that AI makes physical design cheaper and that making becomes the constraint. What does the evidence say about the routes open to a person with software skills, little capital and no factory? Every figure attributed.

Summary

What the evidence supports

This document starts from four premises taken from public posts on one day and does not try to prove them. It asks what a person with software and agent skills, little capital and no factory could do if they hold. It collects what was found about each route and says where the sources are thin.

  • Making simple parts is already cheap and fast. Vendors advertise boards from $2.00 for five and assembly from $8 setup, and the one independent quote test priced the same part from $65.55 to $166.81, the low figure being before an estimated landed cost of about $112.[24][25][28] Difficulty concentrates in complexity, tooling, certification and the move from prototype to production.
  • AI design output runs more often than it is right. On the benchmarks that score past "does the code run", results drop steeply, and none of the end-to-end electronics write-ups read reported a fabricated and tested result.[16][18][22] The evidence that design has become the cheap step is thin; the scarce step looks like verification.
  • Earlier cost collapses did not reward the people who entered the cheap stage. In crowdfunded hardware, consumer 3D printing and open drones, failures clustered at making and delivery.[39][41][43][46]
  • The 3-5x shop claim is untested. No source shows a 10-19 person shop reaching it at constant headcount. The best documented case doubled sales over about nine years at 39 people.[35]
  • The labour premise has weak support so far. The one clear signal is for workers aged 22-25 in exposed jobs.[7][8] No measured flow of displaced office workers into factories was found.

Section 6 draws a conditional reading from this. It is inference, not a finding, and is labelled as such.

Section 1

The premises

These come from four posts and one reply on 4 October 2026. Each author has a stake in the claim, which is noted so the reader can weigh it.

Premises, who said them, and what they sell
PremiseSaid byContext
Design of physical products becomes available to many more people, and "the bottleneck moves from designing things to actually making them"@MatthiasWagner[3]Works at Flux, an AI hardware design company
"We don't need more office workers. we need more data centres and factories to build robots"@Karmedge[2]Posts about physical-world operations; quotes a construction-spending post
Shops with 10-19 employees could make 3-5x their output without matching headcount@ThaaatColin[5]No basis for the multiple is given in the post
Doubt: owners would already have done it, every product vertical differs, some businesses cannot scale@OSHBuilt[4]Runs a job shop
"The winners in physical AI won't just have better models. they will have the infrastructure to actually feed them"@_joe_harris_[1]Names its own product in the post
Section 2

What the numbers say about them

2.1 Capital: data centres and offices

The Census construction-spending release for August 2026 reports private construction put in place at a seasonally adjusted annual rate.[6]

Private construction spending, August 2026 (preliminary), $ million annual rate
LineAug 2026Change vs Aug 2025
Data center84,950+73.2%[6]
General office45,809-9.7%[6]
Office, all (includes data centers)134,119+29.8%[6]
Manufacturing168,209-19.8%[6]

The post's figures match the release only for the right lines. "Office down 10%" is true of general office buildings. The Census headline office line contains data centres and rose. Manufacturing construction, which is building shells and not equipment, fell by about a fifth, which does not look like a factory-building boom.

2.2 Labour

  • The Yale Budget Lab update of 15 September 2026 "does not provide clear evidence of labor market disruption associated with AI". Occupational churn, exposure among the unemployed and usage data "remain flat, lie within historical ranges, or continue along pre-AI trends".[7]
  • The clearest signal is for the young: ages 22-25 in the most AI-exposed occupations had a 16% relative employment decline, and software developers aged 22-25 were down nearly 20% from peak. Experienced workers were stable.[8]
  • BLS still projects software developers to grow 10% over 2025-35.[14]
  • Retrainees from high-exposure jobs moved into less exposed occupations, "particularly in healthcare and transportation". The sample is low-income workers in a federal training programme, not laid-off software workers.[15]
US employment, thousands (BLS all-employee series)
SeriesEarlierSep 2026
Information sector3,061 (Jan 2023)2,739[9]
Computer systems design2,482.9 (Nov 2022)2,355.4[9]
Manufacturing12,903 (Jan 2023)12,652[9]
Construction7,932 (Jan 2023)8,364[9]

Manufacturing employment fell from early 2023 and has been roughly flat since late 2025, so it is not absorbing workers at scale in this series. These are totals and do not say why any series moved. No source reached measured a flow of displaced office or software workers into manufacturing, data centres or robotics.

2.3 Design: how far has the cost fallen?

Benchmarks that score past "does it run"
BenchmarkWhat it measuresResult
MUSE[16]Text to editable CAD assemblies: code runs, geometry valid, design intent metClosed models averaged 60.4% / 44.5% / 30.6%. The best model reached 77.4% / 68.9% / 52.4%. Fine-grained criteria were lower: the best closed model met about 33% of tolerance items and about 41% of manufacturability items
Text2CAD-Bench[17]600 models in four levelsGPT-5.2 produced 11.1% invalid output at level 1 and 68% at level 3. A domain model had the lowest invalidity but the worst geometry error: runs, wrong shape
HWE-Bench[18]300 board projects, static rules plus simulationBest full pass rate 8.15% (Claude Sonnet 4.5); the other models listed 4.4-6.6%
SchGen[19]LLM schematics82% valid circuits; 60.5% expert-verified functionally correct
IoT-SkillsBench[20]Firmware on real microcontrollers, 42 tasks41 or 42 of 42 passed with human-written expert skill documents; LLM-written skills were inconsistent

MUSE scoring on the last stage uses a model judge. Static-rule scores in HWE-Bench average much higher than the full pass rate, so the headline is the lower figure.

Zoo, a text-to-CAD vendor, says its tools cannot "guarantee manufacturability or structural safety".[21] Of the end-to-end electronics write-ups read, none reported a fabricated and tested result at publication. One agent-designed badge nearly used a 555 timer on a 3 V cell; its author logged other errors and the board had not arrived.[22] Another author reports about 2 days against about 10 for a redesign, and says the model made no architectural decisions and did not check electrical correctness.[23] The stronger results pair a model with a deterministic checker or expert guidance (SchGen, IoT-SkillsBench).

2.4 Making: access, price and friction

Prices below are vendors' own advertised starting figures, not typical ones.

What making advertises for small orders (vendor claims unless noted)
RouteAdvertisedSource
JLCPCB boards and assemblyBoards from $2.00 for 5; assembly $8 setup plus $0.0016 per solder joint, from 2 boards, about a week typical; sourcing outside its stock adds 9-15 daysVendor[24]
SendCutSend sheet metalAs little as 24 hours; one example laser-cut aluminium piece $9.41 single, $2.24 each at 100Vendor[25]
Same aluminium bracket, nine platformsQuotes from $65.55 (RapidDirect, before an estimated landed cost of about $112) to $166.81 (eMachineShop); Xometry $118.10, Hubs $151.58, SendCutSend $150.54One test, 1 Mar 2026[28]
  • Demand at the quoting platforms is growing. Xometry's second-quarter 2026 revenue was $229.3M, up 41%, with 89,557 active buyers.[27] Proto Labs' was $149.3M, up 10.6%, with growth attributed to larger customers.[26]
  • For a US buyer of Chinese parts, JLCPCB's own tariff FAQ gives an overall range of about 35% to 92.5% as of March 2026.[29]
  • Why first-time customers fail is described only as practitioner opinion in the sources found (incomplete specifications, over-tight tolerances, minimum-order mismatch). No measured failure rate was found.
Section 3

Small shops and the 3-5x claim

Premise 3 is a claim about shops. The Census holds the structure of the industry.

US machine shops (NAICS 332710), 2022: firms and gross receipts per employee
EmployeesFirmsReceipts per employee
10-141,690$176,620[30]
15-19966$183,581[30]
50-99565$205-225K[30]
200-29970$230,446[30]
5,000+32$311,985[30]
All sizes17,130$204,602[30]
  • US manufacturing has 239,265 employer firms. 36,132 have 10-19 employees and hold about 491,000 jobs.[30]
  • Receipts per employee barely rise with size in machine shops. A 10-19 person shop is not structurally less productive per head than one ten times its size.[30]
  • Arithmetic on those figures, not a finding of any source: three times $176-184K is about $530-550K per employee and five times is about $880-920K. The largest machine-shop size class averages $312K.[30] Receipts include pass-through material, so this is a rough comparison. It does say 5x would put a small shop above every size class in its own industry.
  • The claim in the post that about 80% of shops with under 10 employees disappear in a decade has no source. Firm counts for under 10 employees fell 3-5% between 2017 and 2022 (all manufacturing, under 5 employees: 102,242 to 97,530).[30][31] Counts are by size class, not cohorts, and the window includes 2020-22.

Labour and capacity

  • Manufacturing job openings were 522,000 in August 2026 against 416,000 a year earlier.[32]
  • BLS projects machinist employment roughly flat over 2025-35 (-1% for machinists and tool and die makers together), with about 30,400 openings a year, all from replacement.[10] Median machinist age is 45.7, and 97,000 of 309,000 employed machinists are 55 or older (the 97,000 is the sum of the 77,000 aged 55-64 and 20,000 aged 65 and over).[33]
  • Deloitte and The Manufacturing Institute projected 3.8 million manufacturing jobs needed over 2024-2033, with as many as 1.9 million possibly unfilled.[57]
  • Manufacturing capacity utilization was 75.7% in August 2026, against a 1972-2025 average of 78.2%.[34]

Documented cases

  • Pro Products (39 employees) says its sales doubled over about nine years with nine robots, and that setup programming fell from 3-4 hours to 15-30 minutes. A trade magazine chose it as a success story.[35]
  • Alexandria Industries' first robot cell failed. A simplified cell with one dedicated machine raised spindle uptime and output by more than 30%.[36]
  • The Census Annual Business Survey found no net employment effect from robotics: 9.5% of adopters reported more workers, 8.1% fewer.[37]
  • NIST cites a McKinsey 2022 survey for adoption barriers in small manufacturers: robot capital cost 71% and lack of automation experience 61%.[38]
Section 4

When a cost collapsed before

Hardware design and prototyping got cheaper before. The sources below say what happened after.

Earlier episodes and what is documented
EpisodeDocumented outcomeSource
Kickstarter Design and Technology, 2012 sampleOf 381 successful projects with clear outcomes, 24.9% delivered on time, 3.6% failed outright. Founders named manufacturing problems, shipping complexity, changes in scale and certificationMollick 2014[39]
Funded consumer-hardware campaigns, 144 reviewedNo more than 32% delivered on timeDesign Science 2018, abstract[41]
Same crowdfunding survey, successful projectsOver 90% still ongoing ventures; the survey response rate was about 39%, so this is probably overstatedMollick and Kuppuswamy[40]
Consumer 3D printing, 2013-2016Stratasys paid $403M in stock for MakerBot, then booked goodwill impairments of $102.5M (2014) and $942.4M (2015) across all its reporting units; MakerBot printers sold went 79,906 (2014) to 18,673 (2015)Hackaday, Stratasys 20-F[42][43]
QuirkyRaised $6.5M, $16M, $68M and $79M over four rounds; Chapter 11 in September 2015. A student case study blames lack of focus and uneven qualityWikipedia, Harvard case[44][45]
ShapewaysFiled for bankruptcy on 2 July 2024 and laid off all staff; the marketplace was not revived in its original formCG Channel[58]
3D Robotics (open drones)Peaked above 350 employees, cut to below 80 after 2015, priced against DJI; its co-founder says the organisation changed at each customer shiftRobot Report[46]

Software after LLM coding

  • DORA 2025 reports that AI adoption now improves delivery throughput but still increases delivery instability, and that without strong foundations gains are "often lost to downstream chaos".[47]
  • METR's 2025 randomized trial of 16 experienced developers found tasks took 19% longer with AI, against an expected 20-24% faster; the authors say it does not generalise beyond that group.[48] Nothing found measured distribution or demand as the new software bottleneck.

Where the gains went

Bessen estimated that employment grew for a century in cotton, steel and autos because demand was elastic. Estimated price elasticity of demand in autos fell from 6.77 in 1910 to 0.15 in 2007.[49] Whether a cost collapse enlarges a market depends on that elasticity, and the paper is about employment in an industry, not about who captures the profit.

Section 5

Routes in use

Concrete moves open to someone with software and agent skills, with what the sources say about how each is going.

Routes, and the evidence found for each
RouteWhat the evidence saysLimits
Small-batch hardware through on-demand makingAccess is cheap for simple parts (section 2.4). Failures in earlier cases cluster at production, delivery and certification (section 4)No measured failure rate for first-time customers; no data on how many cheap prototypes reach volume
Software for manufacturersCapital-heavy builders are funded: Hadrian raised $1.37B at a $7.87B valuation.[50] In a vendor survey, about half of small manufacturers rate their digital adoption as manual or basic and 8% of firms under 50 staff are beyond basic; budget is the barrier for every firm with 10 or fewer staff.[53] Census puts AI use at 19.8% of all firms in May 2026[59]No sales-cycle figure and no data on small bootstrapped teams in this niche; the survey is a vendor's
Physical AI and robotics data toolingFoxglove raised a $40M Series B and has 50 staff.[51] XDOF raised $70M and its CEO says collecting data takes "a warehouse of hundreds of thousands of square feet with hundreds of robots".[52] Open building blocks (ROS 2, MCAP, MuJoCo) are named in the source post[1]No revenue evidence for individuals or small teams; entry costs are capital and hardware
Brokering between designers and shopsXometry and Proto Labs are growing (section 2.4). Shapeways failed (section 4)No data on small independent brokers or integrators
Trades and technician workMedian pay and 2025-35 projections: electricians $63,190, +9%;[11] industrial machinery mechanics $64,520, +18%;[12] CNC programmers $68,120, +6% of only 28,600 jobs; CNC operators -9%.[13] Software developers' median is $135,980[14]Growth is in people who maintain, wire and program automated equipment; pay is about half the software median
Open-source and agent-run workflowsNothing found runs an agent from design through quote to a real shop orderThe code-CAD libraries are mature; agent layers are single-author experiments
Section 6

What follows

  • Where it is cheap to be wrong, try it. Making is cheap and fast for small electronics and flat sheet-metal parts (one agent-designed board was about $2 for five[22]), so a first order is small and teaches the failure points that no source measured.
  • The gap named most often is verification. AI drafting runs ahead of checking, the stronger results pair a model with a deterministic check, and no pass-rate data was found for AI designs run through instant-quote manufacturability checks. A tool that checks is a smaller bet than one that designs.
  • Be cautious about stages that need capital. Quirky and Shapeways, which carried making or a marketplace at scale, both went bankrupt.[44][58] The funded players in software for factories and in robot data are well capitalised.
  • The labour premise is a reason to watch, not to move. Nothing found shows a flow of office workers into factories, and manufacturing employment is flat.
  • Treat the 3-5x shop claim as a question the data cannot yet answer. It matters because small-shop capacity decides how much making is available.
Section 7

The term

None of the posts above uses "vibe manufacturing". Where it appears:

  • An Automation Alley executive's op-ed (18 July 2025) uses it for manufacturing done with AI assistance, with no figures.[54]
  • The association's playbook (5 March 2026) uses it for AI plus additive manufacturing inside existing companies.[55]
  • "Vibe coding for hardware", a related phrase, is used by a Danish start-up for AI-written microcontroller code.[56]

No source found uses it for an individual with agents getting physical goods made, which is the sense the posts above imply. A Medium piece reported by search summary uses it for AI tech packs; it could not be read.

Section 8

Not found

  • Any study that took LLM-generated CAD or boards through fabrication and counted how many worked.
  • A measured first-time-customer failure rate, and local job-shop prices, minimums and lead times.
  • Any 10-19 person shop reaching 3x or more at constant headcount, and any independent measure of AI quoting or CAM gains.
  • A measured flow of displaced office or software workers into manufacturing, data centres or robotics.
  • Sales-cycle length for software sold to machine shops, and survival data for small brokers or integrators.
  • Failure rates for Amazon FBA, print-on-demand or Shopify private-label brands that trace to a dataset.
References

Sources

Every figure above is attributed. Numbered citations in the text resolve to these.

  1. 01@_joe_harris_, post on the 2026 physical AI and robotics data stack (4 Oct 2026). Names its own product, Alloy, in the post. x.com/_joe_harris_/status/2106734413941792835
  2. 02@Karmedge (Robert Lukoszko), post on offices, data centres and factories (4 Oct 2026). Quotes a @KobeissiLetter post on construction spending. x.com/Karmedge/status/2106631643037577685
  3. 03@MatthiasWagner, post on AI and the number of people who can design physical products (4 Oct 2026). His profile says he works at Flux, a company that sells AI hardware design. x.com/MatthiasWagner/status/2106594428479742165
  4. 04@OSHBuilt (Caleb, CEO of OSH Cut), reply doubting a 3-5x capacity gain at the same headcount (4 Oct 2026). x.com/OSHBuilt/status/2106741418354250114
  5. 05@ThaaatColin, post claiming shops with 10-19 employees could make 3-5x their output without matching headcount (4 Oct 2026). The post gives no basis for the multiple. x.com/ThaaatColin/status/2106580602359103960
  6. 06US Census Bureau (2026). Monthly Construction Spending, August 2026 (release CB26-158), table of private construction put in place, seasonally adjusted annual rate. Release text at census.gov/construction/c30/pdf/release.pdf. www.census.gov/construction/c30/xlsx/privsa.xlsx
  7. 07Yale Budget Lab (15 Sep 2026 update). Evaluating the Impact of AI on the Labor Market: Current State of Affairs. budgetlab.yale.edu/sites/default/files/page_to_pdf/1419/publication_1419.pdf
  8. 08Brynjolfsson, Chandar & Chen (Stanford Digital Economy Lab, 2025). Canaries in the Coal Mine. Payroll data from ADP. digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
  9. 09US Bureau of Labor Statistics, Current Employment Statistics, all-employee series CES3000000001 (manufacturing), CES2000000001 (construction), CES5000000001 (information), CES6054150001 (computer systems design), read through the public API. api.bls.gov/publicAPI/v2/timeseries/data/
  10. 10BLS Occupational Outlook Handbook, machinists and tool and die makers (2025 data, projections 2025-35). www.bls.gov/ooh/production/machinists-and-tool-and-die-makers.htm
  11. 11BLS Occupational Outlook Handbook, electricians. www.bls.gov/ooh/construction-and-extraction/electricians.htm
  12. 12BLS Occupational Outlook Handbook, industrial machinery mechanics, maintenance workers and millwrights. www.bls.gov/ooh/installation-maintenance-and-repair/industrial-machinery-mechanics-and-maintenance-workers-and-millwrights.htm
  13. 13BLS Occupational Outlook Handbook, metal and plastic machine workers (CNC tool programmers and operators). www.bls.gov/ooh/production/metal-and-plastic-machine-workers.htm
  14. 14BLS Occupational Outlook Handbook, software developers, quality assurance analysts and testers. www.bls.gov/ooh/computer-and-information-technology/software-developers.htm
  15. 15Hyman, Lahey, Ni & Pilossoph. NBER Working Paper 34174 (revised Aug 2026). Returns to workforce training for workers from AI-exposed jobs (WIOA training spells, 2012-2024). www.nber.org/system/files/working_papers/w34174/w34174.pdf
  16. 16MUSE: a Text-to-CAD benchmark for complex editable assemblies (arXiv 2605.28579, 2026). Code check, geometric check and design-intent alignment, scored by a model judge. arxiv.org/html/2605.28579
  17. 17Text2CAD-Bench (arXiv 2605.18430, 2026). 600 human-curated models in four difficulty levels. arxiv.org/html/2605.18430v1
  18. 18HWE-Bench (arXiv 2603.18102, 2026). 300 board-level hardware projects checked with static rules and LTspice simulation. arxiv.org/abs/2603.18102
  19. 19SchGen (arXiv 2605.30345, 2026). LLM schematic generation with expert verification. arxiv.org/html/2605.30345v1
  20. 20IoT-SkillsBench (arXiv 2603.19583, 2026). Firmware tasks on real microcontrollers, with and without expert-written skill documents. arxiv.org/html/2603.19583
  21. 21Zoo (KittyCAD), frequently asked questions. A vendor of text-to-CAD tools describing its own limits. zoo.dev/docs/faq
  22. 22StationX (Oct 2026). AI PCB design: an AI agent designs a coin-cell badge through KiCad scripting. The board had not arrived at publication. app.stationx.net/articles/ai-pcb-design
  23. 23tinycomputers.io. Redesigning a PCB with Claude Code and open-source EDA, part 1. One author's account; fabrication and test are in an unpublished second part. tinycomputers.io/posts/redesigning-a-pcb-with-claude-code-and-open-source-eda-part-1.html
  24. 24JLCPCB, PCB assembly service page. A vendor's own prices and lead times. jlcpcb.com/pcb-assembly
  25. 25SendCutSend, home page. A vendor's own prices and lead times. sendcutsend.com/
  26. 26Proto Labs, Inc. second-quarter 2026 results (exhibit 99.1, SEC filing). www.sec.gov/Archives/edgar/data/0001443669/000144366926000034/prlb-2026731xexx991.htm
  27. 27Xometry, second-quarter 2026 results (4 Aug 2026 press release). www.globenewswire.com/news-release/2026/08/04/3338170/0/en/xometry-reports-record-second-quarter-2026-results.html
  28. 28Otron Atlas (1 Mar 2026). Is Xometry expensive? One STEP file quoted on nine platforms. One part, one date, publisher not otherwise vetted. www.otronatlas.com/price-comparison/is-xometry-expensive/
  29. 29JLCPCB, US tariff policy FAQ (effective 17 Mar 2026). A vendor's own summary. jlcpcb.com/help/article/us-tariff-policy-faq
  30. 30US Census Bureau, Statistics of US Businesses 2022, detailed firm-size tables (NAICS 31-33 and 332710). Employer firms; size is enterprise-wide employment; receipts are gross shipments, not value added. www2.census.gov/programs-surveys/susb/tables/2022/us_state_naics_detailedsizes_2022.txt
  31. 31US Census Bureau, Statistics of US Businesses 2017, detailed firm-size tables. www2.census.gov/programs-surveys/susb/tables/2017/us_state_naics_detailedsizes_2017.xlsx
  32. 32BLS Job Openings and Labor Turnover Survey, table 1 (August 2026, preliminary). www.bls.gov/news.release/jolts.t01.htm
  33. 33BLS Current Population Survey, table 11b (2025 annual averages): employed persons by detailed occupation and age. www.bls.gov/cps/cpsaat11b.htm
  34. 34Federal Reserve, G.17 Industrial Production and Capacity Utilization (September 2026 release). www.federalreserve.gov/releases/g17/current/default.htm
  35. 35Modern Machine Shop. Same headcount, double the sales: successful job shop automation (2023). A trade magazine's chosen success story. www.mmsonline.com/articles/same-headcount-double-the-sales-successful-job-shop-automation
  36. 36Modern Machine Shop. If at first your CNC machine shop doesn't succeed with robots, automate again. www.mmsonline.com/articles/if-at-first-your-cnc-machine-shop-doesnt-succeed-with-robots-automate-again
  37. 37US Census Bureau (2025). Annual Business Survey, technology module: how firms say new technology changed employment. www.census.gov/library/stories/2025/09/technology-impact.html
  38. 38Messina (2024). NIST GCR 24-054, on robotics in small manufacturers. Its barrier percentages are quoted from a McKinsey 2022 survey. tsapps.nist.gov/publication/get_pdf.cfm?pub_id=958147
  39. 39Mollick (2014). The dynamics of crowdfunding: an exploratory study. Journal of Business Venturing. leeds-faculty.colorado.edu/bhagat/Dynamics-of-Crowdfunding.pdf
  40. 40Mollick & Kuppuswamy (2014). After the campaign: outcomes of crowdfunding. leeds-faculty.colorado.edu/bhagat/AfterCampaignCrowdfunding.pdf
  41. 41Identifying challenges in crowdfunded product development: a review of Kickstarter projects. Design Science (2018). Abstract read. www.cambridge.org/core/journals/design-science/article/identifying-challenges-in-crowdfunded-product-development-a-review-of-kickstarter-projects/F2659234936B8349EAF5AFCFBA2199E8
  42. 42Hackaday (2016). The MakerBot obituary. A secondary account that cites company figures; its causal claims are opinion. hackaday.com/2016/04/28/the-makerbot-obituary/
  43. 43Stratasys Ltd., Form 20-F for fiscal 2015 (SEC filing). Goodwill impairments for all reporting units, not MakerBot alone. www.sec.gov/Archives/edgar/data/0001517396/000120677416005045/stratasys_20f.htm
  44. 44Quirky (company). Funding rounds, layoffs and Chapter 11 filing. Wikipedia. en.wikipedia.org/wiki/Quirky_(company)
  45. 45Harvard Business School digital initiative, student case study. Quirky: the failure of invention crowdsourcing. A student analysis, not a controlled finding. d3.harvard.edu/platform-digit/submission/quirky-the-failure-of-invention-crowdsourcing/
  46. 46The Robot Report. 3D Robotics undergoing layoffs and other challenges. Includes a quote from its co-founder. www.therobotreport.com/3d-robotics-undergoing-layoffs-and-other-challenges/
  47. 47Google Cloud DORA (2025). State of AI-assisted software development. Self-reported survey associations, not an experiment. services.google.com/fh/files/misc/2025_state_of_ai_assisted_software_development.pdf
  48. 48METR (10 Jul 2025). Randomized trial of AI tools with 16 experienced open-source developers. metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
  49. 49Bessen (2018). Automation and jobs: when technology boosts employment. NBER Working Paper 24235. www.nber.org/system/files/working_papers/w24235/w24235.pdf
  50. 50Tech Startups (6 Aug 2026). Hadrian raises $1.37B at a $7.87B valuation. A news article. techstartups.com/2026/08/06/defense-tech-startup-hadrian-raises-1-37b-at-nearly-8b-valuation-to-scale-ai-powered-factories/
  51. 51The Robot Report. Foxglove raises a $40M Series B. www.therobotreport.com/foxglove-raises-40m-scale-data-platform-roboticists/
  52. 52TechCrunch (17 Jun 2026). Robot training-data collection at XDOF. techcrunch.com/2026/06/17/collecting-robot-training-data-is-dirty-unglamorous-work-some-ai-labs-are-already-paying-xdof-to-do-it/
  53. 53Lasso Supply Chain Software (Q1 2026). State of digital manufacturing: small and mid-sized companies in the United States. A vendor survey. lassosupplychain.com/research/q1-2026-state-of-digital-manufacturing-report-small-midsized-companies-in-the-united-states/
  54. 54Muzumdar, Automation Alley (18 Jul 2025). Vibe manufacturing is coming and Michigan companies must seize the moment. Op-ed. automationalley.com/2025/07/18/opinion-vibe-manufacturing-is-coming-and-michigan-companies-must-seize-the-moment/
  55. 55Integr8 playbook, Automation Alley (5 Mar 2026). Vibe manufacturing: where AI meets additive. integr8series.com/vibe-manufacturing-where-ai-meets-additive/
  56. 56TechCrunch (14 May 2026). Atech, backed by Lovable, brings vibe coding to hardware. techcrunch.com/2026/05/14/lovable-just-backed-a-company-thats-looking-to-bring-vibe-coding-to-hardware/
  57. 57Deloitte and The Manufacturing Institute (Apr 2024), press release. Projected manufacturing labour need to 2033. www.prnewswire.com/news-releases/us-manufacturing-could-need-as-many-as-3-8-million-new-employees-by-2033--according-to-deloitte-and-the-manufacturing-institute-302105892.html
  58. 58CG Channel (2024). Shapeways files for bankruptcy. www.cgchannel.com/2024/07/3d-printing-firm-shapeways-files-for-bankruptcy/
  59. 59US Census Bureau (May 2026). Business Trends and Outlook Survey: AI use by businesses. www.census.gov/library/stories/2026/05/ai-use-businesses.html
Colophon

Revision notes

  • Created 2026-10-04.
  • How it was gathered. Five research passes ran in parallel on one open question and each was told to report evidence against the premises. The premises themselves were taken as given and not tested.
  • Rechecked against the source: the Census August 2026 levels (data center 84,950; office 134,119; manufacturing 168,209) and the HWE-Bench 8.15% pass rate in its abstract. The MUSE abstract confirms the three-stage drop but not its percentages.
  • Not rechecked: every other figure was read by a research pass through a summarising fetch tool, not as raw page text. The Census year-on-year changes, the SUSB tables, BLS series and the crowdfunding numbers are the ones most worth rechecking before reuse. BLS and SSRN pages blocked a direct recheck.
  • Vendor figures (JLCPCB, SendCutSend, Lasso) are the vendor's own and are labelled where they appear.
  • Left out because they could not be traced: failure-rate claims for Amazon FBA and print-on-demand, a "97% of hardware startups fail" figure, Quirky's product count and cash burn, a MakerBot-only impairment, and any figure seen only in a search snippet.
  • Known tension. The CNC operator median pay was $47,700 in one read of the BLS page and $50,690 in another; the page does not quote it.
  • Post time. The posts in section 1 were read through a mirror on the day they were published and their reply threads were not retrieved.