Someone has to own the hall
ResearchA German contract manufacturer that automated its own machining hall with AI entered insolvency in August. In the same weeks, a British network that mostly signs up machine shops other people own doubled its site count and opened its largest building yet. A distributed network does not make the fixed cost of a factory go away. It decides whose balance sheet carries it.
2026-10-12 · Field notes · 5 min read · By Sondre Hegerland KristiansenThis site has written about both companies before. In July it named Daedalus as a working example of AI-run production, and corrected that on 2 October once the insolvency was found. In August it described Isembard as a network that grows by converting existing shops. This note puts the two next to each other, with a third company that answers the same question differently, because together they say more than any one of them does alone.
An automated hall, owned outright, in insolvency
Daedalus was founded by Jonas Schneider, who worked on robotics at OpenAI. It sold precision metal parts as a contract manufacturer: customers uploaded drawings, and software derived the process plan, the tools, the machining order and the setup, while robot arms loaded the machines. In February 2024 it raised a USD 21 million Series A led by NGP Capital, with Addition and Khosla Ventures, to develop that software and scale its production facilities in Germany.
A local profile from 29 July 2026 described a 4,600 m² hall supplying customers across Europe and, on the company's own figure, more than 150 employees. Three days later, on 1 August, the Karlsruhe district court opened insolvency proceedings for inability to pay, in self-administration. On 1 September the court-appointed supervisor gave notice that the estate is expected to be insufficient to meet the obligations of the proceedings themselves, under § 208 of the German Insolvency Code. Sifted reported the insolvency on 9 October from the court filings.
No public statement of the cause has been found, and this note does not supply one. An insolvency can come from a lost customer, a failed round or a lease as easily as from the model. What can be said is structural. Daedalus owned the whole stack: the software, the machines, the hall and the payroll. When revenue fell short of that fixed base, all of it sat on one balance sheet.
A franchise puts the hall on someone else's books
Isembard also machines precision parts, for aerospace and defense, and also runs on its own software, MasonOS. The difference is who owns the floor. On the company's own description, franchisees own and run their factories, finance them through Isembard's banking partners, and can either start a new site or convert an existing workshop. Isembard supplies the order pipeline, the operating system, the standards and the route to accreditation.
That division moves most of the fixed cost off the network's balance sheet. The lease, the machines and the operators belong to the franchisee. Isembard carries demand, software and standards, which scale far more cheaply than halls. On 14 September the company said its network had grown from 4 factories at the start of 2026 to 16, that it had added 10 franchise operators, and that revenue had risen tenfold in a year. Those are the company's figures.
The model has a cost of its own that is easy to miss. When a franchised site fails, the debt is the franchisee's, and no court notice carries the network's name. An owner's failure is public. A franchise network's failures are quiet by construction, so its survival rate is harder to see from outside than Daedalus's was.
The network's flagship is a building
The headline of Isembard's September was not a conversion. It was a 160,000 sq ft headquarters and factory in Southwark, with 35 staff and a stated plan for 200 by the end of 2027. The target also changed: the 25 factories by the end of 2026 announced in March has been replaced by 100 owned and franchise factories across the UK by the end of 2028.
The count moves too. When this site read Isembard's locations listing on 2 October it showed 16 sites, 11 marked Live and 5 Onboarding. On 12 October the homepage counter read 15 and the London site at Park Royal was no longer listed. Why is not known, and a listing change is not a closure. It is a reminder that a network's size is a number the network publishes.
The pattern that fits both companies is not owned against franchised. It is that concentrated defense demand still wants a large owned node at the center, and conversion fills in the edges. A network can push the edge halls onto other balance sheets. It has not, so far, avoided building one of its own.
Or make the hall small and fast
Reframe Systems builds homes in its own microfactories and raised USD 40 million in a round led by Energy Impact Partners, announced on 31 August. Its new site, FAB1, was set to begin operations on 5 October with less than USD 5 million of equipment, reaching full readiness in under 70 days from receiving the keys, and is designed for up to 500 multifamily or 250 single-family units a year. The company has completed 10 homes. All of these are the company's figures, the capacity is designed rather than demonstrated, and no report confirming that FAB1 started on schedule had been found by 12 October.
Reframe owns its halls, like Daedalus. Its answer to the fixed cost is to make each one small and quick enough that a bad site is a bounded loss. That is construction, not conversion, but construction priced so that it behaves more like a lease than a plant.
What this bears on, and what would change it
P-24 in the register predicts that certified networks scale by converting existing factories, not by building new ones. The evidence this month points both ways. The best-funded conversion network built its largest node from scratch. The owner-operator that built its own automated hall is in insolvency. The fastest new capacity in the layer is being built, not converted, but built cheaply. One insolvency with no public cause is an anecdote, not a trend, and this note scores nothing.
P-15 predicts that manufacturing capacity comes to work like cloud capacity. Cloud computing worked because buying hardware was separated from selling capacity, and someone with a very large balance sheet agreed to own the servers. In manufacturing the hall is the server. Daedalus tried to be both the owner and the seller on a venture balance sheet. Isembard is trying to sell capacity while other people own most of it. Neither has yet shown which arrangement survives a bad year.
What would change this reading: a published cause for the Daedalus insolvency, or a buyer that keeps the hall running · Isembard publishing how many franchised sites have closed, or what a franchisee earns · FAB1 confirmed producing at anything near its designed rate.
What this is built on
Amtsgericht Karlsruhe, 60 IN 541/26: Daedalus GmbH insolvency notices ·Sifted: Ex-OpenAI robotics lead's factory startup in insolvency proceedings, 2026-10-09 ·karlsruhe.digital: profile of Daedalus, 2026-07-29, with insolvency update ·Tech.eu: Former OpenAI lead's startup Daedalus secures $21M in NGP Capital-led round ·Startbase: Daedalus Series A, 2024-02-08 ·Isembard: Franchise launch, 2026-03-06 ·The Manufacturer: Isembard launches manufacturing franchise model ·Automation News: Isembard opens largest new central London factory since Second World War, 2026-09-14 ·Isembard: homepage and locations listing, read 2026-10-12 ·Robotics & Automation News: Reframe Systems raises $40 million, 2026-09-01 ·HousingWire: Reframe Systems raises $40M to scale modular microfactories, 2026-08-31 ·Moduloa: The bottleneck was never the machines ·Moduloa: The box beats the body
The insolvency dates, case number and the § 208 notice come from the court's published notices, and Sifted independently reports the proceedings from court filings. The hall size and headcount are from the July profile and are the company's own figures. The franchise terms are Isembard's own description, also reported by The Manufacturer. Network size, revenue growth, staff and targets are Isembard's claims as reported by Automation News. The 2 October site count is this site's own reading of Isembard's listing, recorded in its monthly company radar, and the 12 October count was read the day this note was published. Reframe's figures are the company's, reported independently by Robotics & Automation News and HousingWire.
Add to this
Corrections, evidence, and disagreement are welcome. This is knowledge in the open. Anyone can read; sign in with GitHub only to post or react, and it appears here instantly.