Foundry Robotics' $100M Bet: The Defence Bottleneck Is the Assembly Line

Adarsh Kulkarni spent his early career teaching machines to walk. He was employee number four at Ghost Robotics, which built legged robots for military buyers and was later acquired for about $400 million. He went on to run robotics at Scale AI.

The problem he talks about now is smaller and harder. A screw.

At a 130,000-square-foot plant in Emeryville, California, Kulkarni's company, Foundry Robotics, runs a grid of robotic cells that fasten, place, cable and inspect complex hardware. On 28 October it announced an oversubscribed $100 million Series A led by Marlinspike Partners, with Khosla Ventures, Red Glass Ventures, SE Ventures, Wing, Axiom Ventures, Hanabi and Zero Shot Fund participating.

The pitch is not another drone. It is the factory that builds one, then builds something else on Wednesday.

The constraint moved to the last hundred metres

Foundry's argument starts with a supply-chain observation. Designing defence hardware has become fast. Assembling it has not.

“Production capacity is a national constraint on getting unmanned systems to the fleet,” said Duane Fotheringham, president of HII Unmanned Systems, in the funding announcement. HII is Foundry's first publicly named customer.

Traditional contract manufacturers build a line around a single product. Retooling that line for a new design takes months and millions of dollars, according to Foundry and its early investor Garuda Ventures. For hardware that iterates quarterly, that cadence is the bottleneck.

Kulkarni puts the shift plainly. Assembly, not the forming of parts, has become “more and more of the bottleneck,” he told Tectonic, and it is also where robotics foundation models earn their keep.

The market has noticed the same seam. Anduril has said it will build an automated shipbuilding facility, Arsenal-2, aimed at submarine production constraints. In Europe, Argotec raised €50 million this month to lift satellite output from 52 a year to 80 — a capacity argument one layer up the supply chain. As we wrote in November about Argotec's satellite factory, the factory, not the design, is where the schedule now lives.

Foundry's answer is to own the factory and sell time inside it. Customers bring a design; Foundry handles assembly, testing and traceability. The company says it can move a product from design to production in weeks without the customer funding a new line or a retooling cycle.

One cell, three products, one operating system

Foundry does not sell robots. It owns the building, the cells and the software, and charges customers for assembly. The company calls the model manufacturing-as-a-service.

The coordination layer is FoundryOS, which links robotic cells, production schedules and cloud systems. Foundry trains vision-language-action models on assembly primitives — fastening, placement, cabling — and runs them across cells that can be retasked in software rather than rebuilt in steel.

A vision-language-action model fuses three inputs: camera images, a natural-language instruction and the robot's own action history. Training happens end to end on demonstration trajectories, so the model learns to recover from small errors instead of failing at each handoff in a classical perception-planning-control stack.

“We can build drones on Monday, batteries on Tuesday, satellites on Wednesday, and that's the whole point,” Kulkarni said.

That breadth is the bet. A fixed line optimises for one product; a general line trades peak efficiency for utilisation across many. Foundry says its Emeryville site, twenty times larger than its previous footprint, is already shipping to paying customers in robotics, energy, telecom, autonomous vehicles and defence.

The company is ITAR registered and says it meets CMMC 2 requirements, the certification baseline for handling controlled defence data. It also promotes a method it calls Design for Robotic Assembly: simplify parts and steps so robots can build the product repeatably, cutting assembly time and bill-of-materials cost.

The demand signal behind the round is real, if uneven. North American companies ordered 10,989 robots worth $723.5 million in the third quarter of 2026, up 24.8% in units and 26.1% in value year on year, according to the Association for Advancing Automation. Electronics orders rose 97.6%. Automotive components rose 90.4%.

The humans the pitch leaves out

The word “autonomous” carries a lot of weight in this story. The paperwork is more careful.

Foundry has also acquired Avea Robotics, a software company whose Sentinel platform lets operators control robots remotely in real time. Foundry announced the deal in early October. The law firm Goodwin, which advised on both transactions, disclosed it in a 29 October notice. Terms were not disclosed.

That acquisition changes what kind of factory this is. Vision-language-action models generalise better than classical controllers, but a well-tuned classical controller still wins on precision in a fixed, controlled cell. Zero-shot adaptation to a genuinely new part geometry is not solved at production reliability. Research through 2026 has argued over discrete-token versus diffusion-based action decoding without settling which works better for high-precision industrial work.

Sentinel is the hedge. The model handles the repeatable majority. A human takes the novel minority.

Every supervised run produces another labelled demonstration, which is exactly the data that trains the next model. The factory keeps shipping while the autonomous envelope widens. That is a clever loop, and it is also an admission: the fully autonomous factory is a direction, not a current product.

The architecture echoes a pattern we covered this week. As we wrote in our Shield AI and Leonardo piece, buyers are separating the autonomy layer from the airframe. Foundry is doing the same on the factory floor — autonomy as a layer, with a human override where the layer thins out.

Five customers, no marquee

For a company roughly ten months old, the customer list is the strongest part of the story. Foundry says it has five paying customers. HII is the only one it has named publicly.

“We don't really have, like, one big marquee customer who uses 80 percent of the facility,” Kulkarni told Tectonic. “It's actually pretty distributed.”

The product mix spans quadcopters, unmanned underwater vehicles, quadruped robots, autonomous mobile robots, telecom and compute hardware, and wire harnesses. That range is unusual for a contract manufacturer, and it is central to the model: utilisation comes from diversity, not volume on one line.

The team is built for both halves of the problem. Co-founder Emmanuel Lamacchia ran Tesla's Model Y programme, a line that produced roughly 1.5 million vehicles a year. The broader leadership includes former Tesla software staff and military veterans who worked at Raytheon and Northrop Grumman. Marlinspike, the lead investor, is a defence- and national-security-focused fund.

“The next era of industrial strength will belong to nations that can turn new designs into finished hardware the fastest,” said Mislav Tolusic, managing partner at Marlinspike. “Defence and industrial equipment manufacturers urgently need more domestic assembly capacity.”

What the company has not disclosed matters too. There is no valuation, no revenue figure and no full customer list. Its own site lists ITAR registration and CMMC 2 compliance, not throughput, defect rates or how many products the Emeryville floor runs at once.

The incumbents are not idle. ABB, FANUC, KUKA and Rockwell already sell automation into factories, and a wave of physical-AI startups is chasing the same capital. Foundry's differentiator is vertical integration: it owns the building, the cells, the operating system and now the remote-control layer.

That stack is expensive to build and hard to copy. It is also the reason the company cannot hide behind a software multiple. A factory that does not ship is a factory that does not earn.

What would make this real

The round is a bet on a specific claim: that flexible assembly capacity, not another robot design, is the scarce input in defence hardware. The evidence so far is a large facility, five customers and a name-brand shipbuilder. The evidence still missing is operational.

A $100 million cheque buys time, not proof. The test is whether a general assembly cell beats a dedicated line on cost per unit once changeovers are counted. If it does, the bottleneck Foundry targets — the last hundred metres between a finished design and a delivered system — is a durable business. If it does not, the round funds a very expensive demonstration.

The factory is the product. Everything else is instrumentation.