Foundational Industries raises $25M to build AI-run factories

The pitch from Foundational Industries, which Fortune reported exclusively has closed a $25 million seed round, is not about welding another robot arm onto a legacy assembly line. It is about writing the factory itself in software first and letting AI operate it.

The first physical product is data-center hardware, specifically custom rack enclosures aimed at data-center developers, neoclouds, and chipmakers whose new AI silicon runs at different voltages than the racks they inherit. CEO Jonathan Winer told Fortune the team has "built the entire factory in software already using software emulators," which is the core bet. If the factory exists as a working simulation before any steel gets cut, changing the product line means changing the code rather than retooling the shop floor.

Winer's broader argument, as reported, is that the U.S. cannot beat China on the terms China already dominates, meaning vast automated plants that need constant utilization to justify their subsidies. His counter is that America's real advantages are sophisticated AI models and abundant compute, and that AI-native factories can get cheaper and faster with each one built. The round was led by BoxGroup and Zigg Ventures, with Abstract Ventures, Adverb Ventures, Buckley Ventures, and Offline Ventures participating. Winer spent six years at Alphabet's Sidewalk Infrastructure Partners before founding the company, where per Fortune he deployed over $1 billion in capital.

The honest caveat is that "the entire factory in software" is the founder's characterization, and a seed announcement is not evidence of throughput, yield, or unit economics. What the reporting does not give you is a named first customer, a shipment timeline, a facility location, or how much of the $25 million is going into a physical line versus more simulation work. Take the specifics as reported, not settled.

What is genuinely interesting is the customer wedge. If the buyers really are chipmakers and neoclouds staring at rack designs that legacy contract manufacturers cannot iterate on fast enough for each new silicon generation, then software-defined enclosures are a plausible gap to fill, and the same approach could migrate outward to the rest of the data-center bill of materials as the AI buildout keeps accelerating.