Mackenzie Bowes
Menu

Vibe manufacturing: the evidence behind each answer

By Mackenzie Bowes · mackenzie@mackenziebowes.com · @_mackenziebowes on X

The tables behind the vibe manufacturing answer: design benchmarks, small-order prices, past cost collapses, small-shop output and labour data.

Contents
Drafts are cheap; most of them stop at the checks. On HWE-Bench, a set of 300 board projects, the best model passed every check 8.15% of the time, drawn here as 8 in 100.

Design: drafting got cheap, checking didn't

The checks that stop most drafts are deterministic software, and a person can put them in the loop so what they catch goes back to the agent. The via is from one documented attempt; the heap and the chute are illustrative. The price is one vendor's advertised starting price.
BenchmarkWhat it measuresResult
MUSE[5]Text to editable CAD assemblies: code runs, geometry valid, design intent metClosed models averaged 60.4% / 44.5% / 30.6%; the best reached 77.4% / 68.9% / 52.4%. The best closed model met about 33% of tolerance items
Text2CAD-Bench[6]600 models in four levelsGPT-5.2 gave 11.1% invalid output at level 1 and 68% at level 3
HWE-Bench[7]300 board projects, static rules plus simulationBest full pass 8.15% (Claude Sonnet 4.5); others 4.4-6.6%
SchGen[8]LLM schematics82% valid; 60.5% expert-verified correct
IoT-SkillsBench[9]Firmware on real microcontrollers, 42 tasks41 or 42 of 42 with expert-written skill notes; model-written notes were inconsistent

Benchmarks that score past "does it run". MUSE's last stage uses a model judge.

In one documented agent-designed badge, the model nearly used a 555 timer on a 3 V coin cell; a person caught it.[11]

Another author cut a redesign from about 10 days to about 2, and says the model made no architectural decisions and didn't check electrical correctness.[12] A text-to-CAD vendor says its own tools can't "guarantee manufacturability or structural safety".[10]

Making: cheap to start, priced all over the map

The same aluminium bracket was quoted from $65.55 to $166.81.
RouteAdvertisedSource
JLCPCB boards and assemblyBoards from $2.00 for 5; assembly $8 setup plus $0.0016 per solder joint; about a week typicalVendor[13]
SendCutSend sheet metalAs little as 24 hours; one laser-cut aluminium piece $9.41 single, $2.24 each at 100Vendor[14]
One bracket, nine platforms$65.55 (RapidDirect, before an estimated landed cost of about $112) to $166.81 (eMachineShop)One test, 1 Mar 2026[17]

For a US buyer, JLCPCB's own tariff FAQ gives about 35% to 92.5% on Chinese parts as of March 2026.[18] Demand at the quoting platforms is growing: Xometry has 89,557 active buyers,[16] and Proto Labs grew 10.6% on larger customers.[15]

The last time a hardware cost collapsed

Of 381 successful Kickstarter projects, 24.9% delivered on time.
EpisodeWhat happenedSource
Kickstarter design and technology, 201224.9% of 381 successful projects delivered on time; 3.6% failed outrightMollick 2014[25]
144 funded consumer-hardware campaignsNo more than 32% delivered on timeDesign Science 2018[27]
MakerBotStratasys paid $403M in stock; MakerBot printers sold fell from 79,906 (2014) to 18,673 (2015)Hackaday, Stratasys 20-F[28][29]
QuirkyRaised $169.5M over four rounds; Chapter 11 in September 2015Wikipedia[30]
ShapewaysBankrupt on 2 July 2024; all staff laid offCG Channel[32]
3D RoboticsAbove 350 staff at peak, below 80 after pricing against DJIRobot Report[31]

The one practice tested against delay: crowdfunders with a full project schedule before launch were late less often.[26]

Small shops and the 3-5x claim

Receipts per employee rise with firm size; the smallest shops sit near $177K.
EmployeesFirmsReceipts per employee
10-141,690$176,620
15-19966$183,581
50-99565$205-225K
200-29970$230,446
5,000+32$311,985
All sizes17,130$204,602

US machine shops (NAICS 332710), 2022.[19]

Our arithmetic on those figures: three times $176-184K is about $530-550K per employee, and five times is about $880-920K. The largest size class averages $312K. Receipts include pass-through material, so it's a rough comparison, and it's still nearly three times the top class.

The best documented case is Pro Products: 39 staff and nine robots, sales doubled over about nine years, and setup programming cut from 3-4 hours to 15-30 minutes.[22] Alexandria Industries' first robot cell failed; a simpler one raised output by more than 30%.[23]

Small manufacturers name robot cost (71%) and lack of automation experience (61%) as the barriers.[24]

Jobs: the labour premise

Data-centre construction rose 73.2% in a year while general office fell 9.7% and manufacturing fell 19.8%.

Capital is moving to data centres, not factories. Private data-centre construction ran at $84,950M a year in August 2026, up 73.2%; manufacturing construction fell 19.8%.[1]

Series (thousands employed)EarlierSep 2026
Information sector3,061 (Jan 2023)2,739
Computer systems design2,482.9 (Nov 2022)2,355.4
Manufacturing12,903 (Jan 2023)12,652
Construction7,932 (Jan 2023)8,364

BLS all-employee series.[4]

The Yale Budget Lab finds no clear AI disruption in the labour market as of 15 September 2026.[2]

The signal is in the young: 22-25 year olds in the most exposed jobs fell 16% relative to others, and young software developers nearly 20% from peak.[3] Manufacturing has 522,000 openings,[20] and 97,000 of 309,000 machinists are 55 or older.[21]

Sources

  1. US Census Bureau (2026). Monthly Construction Spending, August 2026 (release CB26-158), private construction put in place, seasonally adjusted annual rate. www.census.gov/construction/c30/xlsx/privsa.xlsx
  2. Yale 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
  3. Brynjolfsson, Chandar & Chen (Stanford Digital Economy Lab, 2025). Canaries in the Coal Mine. ADP payroll data. digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
  4. BLS Current Employment Statistics, all-employee series (manufacturing, construction, information, computer systems design). api.bls.gov/publicAPI/v2/timeseries/data/
  5. MUSE: a Text-to-CAD benchmark for complex editable assemblies (arXiv 2605.28579, 2026). arxiv.org/html/2605.28579
  6. Text2CAD-Bench (arXiv 2605.18430, 2026). arxiv.org/html/2605.18430v1
  7. HWE-Bench (arXiv 2603.18102, 2026). 300 board-level projects checked with static rules and LTspice simulation. arxiv.org/abs/2603.18102
  8. SchGen (arXiv 2605.30345, 2026). arxiv.org/html/2605.30345v1
  9. IoT-SkillsBench (arXiv 2603.19583, 2026). arxiv.org/html/2603.19583
  10. Zoo (KittyCAD), frequently asked questions. A vendor describing its own limits. zoo.dev/docs/faq
  11. StationX (Oct 2026). AI PCB design: an agent designs a coin-cell badge through KiCad scripting. app.stationx.net/articles/ai-pcb-design
  12. tinycomputers.io. Redesigning a PCB with Claude Code and open-source EDA, part 1. tinycomputers.io/posts/redesigning-a-pcb-with-claude-code-and-open-source-eda-part-1.html
  13. JLCPCB, PCB assembly service page. Vendor prices. jlcpcb.com/pcb-assembly
  14. SendCutSend, home page. Vendor prices. sendcutsend.com/
  15. Proto Labs, second-quarter 2026 results (SEC exhibit 99.1). www.sec.gov/Archives/edgar/data/0001443669/000144366926000034/prlb-2026731xexx991.htm
  16. Xometry, second-quarter 2026 results (4 Aug 2026). www.globenewswire.com/news-release/2026/08/04/3338170/0/en/xometry-reports-record-second-quarter-2026-results.html
  17. Otron Atlas (1 Mar 2026). Is Xometry expensive? One STEP file quoted on nine platforms. www.otronatlas.com/price-comparison/is-xometry-expensive/
  18. JLCPCB, US tariff policy FAQ (effective 17 Mar 2026). jlcpcb.com/help/article/us-tariff-policy-faq
  19. US Census Bureau, Statistics of US Businesses 2022, detailed firm-size tables (NAICS 332710). Receipts are gross shipments. www2.census.gov/programs-surveys/susb/tables/2022/us_state_naics_detailedsizes_2022.txt
  20. BLS Job Openings and Labor Turnover Survey, table 1 (August 2026, preliminary). www.bls.gov/news.release/jolts.t01.htm
  21. BLS Current Population Survey, table 11b (2025 annual averages). www.bls.gov/cps/cpsaat11b.htm
  22. Modern Machine Shop. Same headcount, double the sales: successful job shop automation (2023). www.mmsonline.com/articles/same-headcount-double-the-sales-successful-job-shop-automation
  23. Modern 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
  24. Messina (2024). NIST GCR 24-054, quoting a McKinsey 2022 survey. tsapps.nist.gov/publication/get_pdf.cfm?pub_id=958147
  25. Mollick (2014). The dynamics of crowdfunding: an exploratory study. Journal of Business Venturing. leeds-faculty.colorado.edu/bhagat/Dynamics-of-Crowdfunding.pdf
  26. Mollick & Kuppuswamy (2014). After the campaign: outcomes of crowdfunding. leeds-faculty.colorado.edu/bhagat/AfterCampaignCrowdfunding.pdf
  27. Identifying challenges in crowdfunded product development. Design Science (2018). www.cambridge.org/core/journals/design-science/article/identifying-challenges-in-crowdfunded-product-development-a-review-of-kickstarter-projects/F2659234936B8349EAF5AFCFBA2199E8
  28. Hackaday (2016). The MakerBot obituary. hackaday.com/2016/04/28/the-makerbot-obituary/
  29. Stratasys Ltd., Form 20-F for fiscal 2015. www.sec.gov/Archives/edgar/data/0001517396/000120677416005045/stratasys_20f.htm
  30. Quirky (company). Wikipedia. en.wikipedia.org/wiki/Quirky_(company)
  31. The Robot Report. 3D Robotics undergoing layoffs and other challenges. www.therobotreport.com/3d-robotics-undergoing-layoffs-and-other-challenges/
  32. CG Channel (2024). Shapeways files for bankruptcy. www.cgchannel.com/2024/07/3d-printing-firm-shapeways-files-for-bankruptcy/