The robot was never the expensive part

When MicroFactory raised its pre-seed in September 2025, it was covered as a dog-crate-sized box with two robot arms that learns tasks by watching people. Coverage of the round put it at around USD 5,000, with more than 100 paid reservations and first commercial units expected in early 2026. TechCrunch reported shipping about two months out.

The company's own page, read on 26 September 2026, now offers a one-arm cell starting from USD 11,950 and a two-arm cell starting from USD 19,900, each listed as installed and configured. The headline no longer sells a cheap robot. It sells "robotic cells for high-mix, low-volume tasks with a built-in AI integrator", and the benefit line is to replace weeks of integration work and the integrator bill. A 30-day money-back guarantee starts after the robot is fully delivered. This note found no customer anywhere who says they have received one.

The two prices are not like for like. A reservation price for a kit is not a starting price for an installed cell, and "starting from" leaves the top open. That is the point, not a caveat to it. The product moved from a thing you buy to a job someone does for you, and the price moved with it, before the first public delivery.

ProvenMetal assembles and routes circuit boards for US customers. Its founders' launch post on Hacker News, dated 6 August 2026, is the clearest first-person account this beat has found of trying to own small capacity. They started in a garage with prosumer pick-and-place, X-ray and rework equipment, on the theory that owning the process would let them automate it. Instead they were "spending 90% of our time assembling circuit boards instead of growing the business", and software was "not the binding constraint at low volumes".

So they stopped assembling as the main business and moved to coordination: sourcing the bill of materials, checking designs against each manufacturer's capability sheet, and routing orders through partner shops. Their site now opens with the claim that more than half of lead time is not manufacturing but quoting, design review and sourcing, and describes the factory as a network. The traction they report, about USD 70,000 across 11 orders in six weeks, is their own figure and unverified.

The same post says where they expect this to end: "at some volume the only remaining move is adding physical capacity." The retreat was a stage, not a verdict on owning machines. What it was a verdict on is owning machines before you can afford the people and hours that make them produce.

Neither company discovered anything the literature did not already say. Boston Consulting Group's April 2026 analysis of physical AI states that in traditional robotics deployments, roughly 75% of total cost of ownership is tied to initial setup and reengineering: configuring workflows, adapting systems to new products, integrating them into existing operations.

A NIST contractor report on manufacturing robotics, published in June 2024, makes the same point from the other side. Robots have become inexpensive and integration has not, and it cites MIT research finding that integration can cost up to five times the robot itself. The same report notes that about 93% of the roughly 240,000 US factories had fewer than 100 employees in 2020, and that small firms cannot afford dedicated robotics engineers. A cheap arm does not change that sentence. [Correction, 2026-10-03: the up-to-five-times figure is not an MIT research finding. It is one robot maker's estimate of four to five times, given in interviews MIT published in 2020 (Sanneman, Fourie and Shah, pp. 35-36), which the NIST report cites on p. 4 and restates as a general finding. The 93% counts firms, not factories. NIST takes it from the National Association of Manufacturers, whose own figure is of firms, and Census Bureau business statistics for 2020 put 93.13% of 239,607 US manufacturing firms under 100 employees. Counted by establishment, meaning a single physical location, Census County Business Patterns for 2022 puts 90.96% under 100.]

MicroFactory's repositioning is a bet that frontier models can do the integrator's job: design the gripper, write the vision code, collect training data. That is the right target, and BCG's own estimate says software-defined approaches can cut setup and reengineering costs by up to half.

Run that number to the end. If setup is 75% of lifetime cost and it halves, setup is still 37.5 points against 25 for everything else, the hardware included. Even the optimistic case leaves integration as the largest line. And MicroFactory's own page is candid that some tasks need human input or iteration, and some are too complex for it yet.

For the thesis, this cuts both ways. It supports the claim that the durable value sits in the layer around the robot: qualification, training, the record of what a cell has been proven to do. It also means that layer is paid for in hours, and hours are exactly what the small shops this model depends on do not have.

P-03 in the register predicts humanoid unit cost falling well below roughly USD 20,000 by 2031. This note does not bear on whether that happens. It bears on what it would be worth: if hardware is at most a quarter of a cell's lifetime cost, a cheaper body moves the smaller share. P-14 predicts a new profession combining machines, robots, fixtures, software, quality and data. Every case here is a shortage of exactly that person.

P-24 predicts a network that scales by converting existing factories. Conversion inherits an existing shop's people, which is the scarce input, and that is quiet support for the route over building new. P-28 asks whether Karkhana.io takes a further owned-capacity step. ProvenMetal is not Karkhana, and an intention is not a step, so it scores nothing there. It is a second router saying capacity is where routing ends up. Nothing here is scored and the register is untouched.

What would change the reading: a customer publicly confirming a MicroFactory cell running a task it was not sold with, deployed without a human integrator. Or independent figures showing setup falling below half of lifetime cost for small deployments. Either would mean the hardware price is back to mattering.

MicroFactory: product page, read 2026-09-26 ·TechCrunch: This $30M startup built a dog crate-sized robot factory that learns by watching humans ·Robotics and Automation News: MicroFactory raises USD 1.5 million to scale affordable general-purpose manufacturing robot ·Hacker News: Launch HN, ProvenMetal (YC S26) ·ProvenMetal: company site, read 2026-09-26 ·BCG: How Physical AI Is Reshaping Robotics Today ·NIST: Research Opportunities for Advancing Measurement Science for Manufacturing Robotics, GCR 24-054 ·Sanneman, Fourie and Shah (MIT): The State of Industrial Robotics, interview RM1 on pp. 35-36 (added 2026-10-03) ·US Census Bureau: Statistics of US Businesses, 2020, manufacturing firms by size (added 2026-10-03) ·US Census Bureau: County Business Patterns, 2022, manufacturing establishments by size (added 2026-10-03) ·Moduloa: The bottleneck was never the machines ·Moduloa: Speed came back, price did not

MicroFactory's current prices, positioning, guarantee and stated limitations are taken from its own page on 26 September 2026 and are vendor statements. The 2025 price of around USD 5,000 and the reservation count come from coverage of the funding round, which appears to rest on the company's announcement. Two outlets reported shipping expected within months of the round. The absence of a public delivery is a negative finding from searches on the date of writing, not proof that none has occurred. ProvenMetal's history and figures are the founders' own words, with date, score and text taken from the Hacker News API. The 75% figure is BCG's, and the up-to-five-times figure is MIT research as cited by NIST, not an independent measurement by this site. Correction, 2026-10-03: that figure is one interviewee's estimate in the MIT study, not a research finding, and the 93% counts firms, not factories. Both are corrected where they appear above, and the original sentences are left as written.