Short answer: South Bay startup Neuron Industries raised a $10M seed on Oct 29, 2026 for the Cortex AIC (AI Industrial Controller), which replaces the programmable logic controller (PLC) and generates "deterministic real-time Python" from plain English, verified by unit tests in a simulator (Neuron release). Our thesis is that automating software across the stack moves factories toward closed-loop production, or software-defined factories, and eventually dark (lights-out) factories. The direction looks right. The speed is overstated: code is only part of commissioning, safety needs certified hardware, and physical changeover and inspection still bind. The credible wedge today looks like retrofitting obsolete controllers on legacy lines.
What Neuron announced
- Round: $10M seed co-led by FCVC and Obvious Ventures, with Y Combinator (S26) and others (Obvious). Founded in 2026 by Dennis Ren (co-founder and CEO) and Kenneth Rhee (co-founder and CTO) (Neuron).
- Product and pricing: Cortex Mini at $2,950 (Q1 2027) and Cortex at $6,450 (Q2 2027), per the company.
- Location: the release is datelined Gardena; Neuron's site says it builds in El Segundo (Neuron). Either way, it's South Bay.
The proof case
Neuron's one detailed case is CureWood, in British Columbia: a 1991 Siemens S5 line, decoded, ported and commissioned in about three engineering hours and eight hours of downtime (Neuron case study). The running program differed from the documentation in 15 places. These are company-reported figures from one retrofit. "Months to days" is also the company's framing.
The 15 discrepancies may be the most telling detail. On old lines, the real logic often lives in the machine, not the documents.
How close are dark factories?
Further off than the headline suggests:
- Code is a bounded slice. A review citing a 2005 estimate puts commissioning at 15% to 20% of project time, with up to 70% of that lost to software errors (InTech). By our arithmetic, that's roughly 10% to 15% of a project. Big, but not the whole job.
- Safety needs certified hardware. Neuron's site lists "UL + CE (planned)." We found no functional-safety rating (SIL or PL) or plan for one. Whether Cortex leaves safety functions on separate certified hardware is unknown.
- Physical steps still bind. Changeover, inspection and maintenance aren't solved by better code.
A sharper version of the thesis: automated software could make reconfiguring a factory nearly as fast as its hardware allows.
The missing layer
Between a build plan and a running controller sits a handoff: sequence, tags, interlocks, test limits. Whoever defines that handoff might own the factory software stack (our view). Arda Industries, which we covered on Oct 24, speeds up planning upstream. There's no public Arda-to-Neuron handoff.
What it means for you
- Early-stage industrial automation founders: retrofits of obsolete PLCs look like a fundable wedge, with a clear map of the hard parts: safety, physical commissioning and plant-floor adoption.
- Controls engineers: Python in place of IEC 61131-3 languages could widen who can build automation. It may clash with maintenance staff trained on ladder logic (our view).
- Investors: a fresh South Bay seed deal. Ask about the safety architecture first. Safety certification may stay with incumbents such as Siemens, Rockwell and Pilz, even if newcomers win standard control.
- COOs: a retrofit pilot on a legacy line is the low-risk way to test the claims.
Open questions
- Will Cortex keep safety on separate certified hardware, and is a SIL/PL rating planned?
- Is there a public, machine-readable spec for handing a build plan to a controller?
- Has the bounded-tool safety pattern from lab agents been applied to a factory cell?
The question we're asking: what machine-readable spec should carry a build plan to a controller, and who certifies the safety layer?
Your turn: if you run legacy PLC lines or certify machine safety, reply and tell us how you'd answer it. Follow LA Robotics Network for more on software-defined factories.

