There is a boring, expensive fact at the center of how physical things get made, and Foundry Robotics has decided to build a company around it. The fact is retooling. When a factory that makes one thing wants to make a slightly different thing, it stops. Engineers swap fixtures, reprogram machines, and rebuild the line, and for a while nobody makes anything. This is fine if you are stamping out ten million identical widgets. It is agony if you want to make a few thousand drones this quarter and a few thousand battery packs the next.
Foundry's pitch - and it is genuinely a pitch, the company is young and much of this is still ambition - is that retooling does not have to be a physical event. If a robot is general enough, and the software driving it is smart enough, then switching products becomes closer to a software update than a construction project. The company calls the result "The Everything Factory," which is the kind of name that is either going to look prophetic or adorable in five years. Possibly both.
The mechanism they are betting on is a class of AI models with an unwieldy name: Vision-Language-Action models, or VLA. The short version is that these models take in what a robot sees and a description of what it should do, and output the motions to do it. It is the same broad idea that has made large language models useful, pointed at the physical world instead of a text box. Foundry's wager is that VLA models, plus modern robotics hardware, plus a lot of unglamorous manufacturing software, add up to robots that can assemble products they were not specifically programmed to assemble.
"An AI+robotics-first, assembly-focused, dual-use contract manufacturer."
— How Foundry describes itself
The un-obvious part of the plan
Here is the detail that makes Foundry more interesting than a robotics-arms vendor: the company does not primarily want to sell you robots. It wants to be the factory. Foundry is set up as a contract manufacturer - you bring the product, they assemble it - which means the robots and the VLA models are the means, not the merchandise. This is a slower, more capital-intensive path than shipping robot arms to other people's plants. It is also a more defensible one, because the hard-won knowledge of how to actually assemble complicated things at low-to-mid volume stays inside the company and compounds.
The products they are aiming at are telling: drones, robot dogs, battery packs, satellite components. These are complex, relatively low-volume, high-mix items - exactly the category where traditional automation is too rigid and human labor is expensive or scarce. They are also, not coincidentally, things that are frequently made abroad and that both companies and the U.S. government would prefer to make closer to home. Foundry describes its initial targets as products in high demand from both commercial customers and the Department of Defense, which brings us to the word doing a lot of work in the company's self-description: dual-use.
Dual-use, meant literally
"Dual-use" in defense circles usually means a technology that could serve civilian or military ends. Foundry means it more concretely: the same robotic assembly cells that build a commercial product can build a defense one. One factory, two demand curves, and a supply chain that does not run through an ocean. In June 2026 the company registered a federal lobbying representative, which is the sort of paperwork that tells you a startup is taking the government-customer half of its thesis seriously rather than as a footnote.
Whether all of this works is genuinely unknown, and worth saying plainly. Generalized robotic assembly is hard in the way that hardware is always hard - reality does not grade on a curve, and a robot that assembles something 95% correctly has assembled nothing. Foundry is a small team taking on a problem that much larger and better-funded companies have circled for years. But the shape of the bet is clear, and the people making it have done adjacent versions of it before.
