Ex-OpenAI Research Chief Launches AI Manufacturing Startup

Former OpenAI research chief launches AI manufacturing startup in latest ai news march 5 2026. The venture targets autonomous industrial robotics using foundation models.

Jan Leike, who led OpenAI's superalignment team until his dramatic departure in May 2024, has raised $142 million to build autonomous manufacturing systems powered by large-scale reinforcement learning. His new startup, Mechanical Orchard, emerged from stealth on Wednesday with a team of 34 engineers and a facility in Fremont, California already producing precision components for aerospace clients.

Leike isn't building chatbots. He's betting that the same training techniques that produced GPT-4's reasoning capabilities can teach robotic systems to run factories with minimal human oversight. The company's first product line focuses on CNC machining and additive manufacturing — processes that currently require skilled technicians to program, monitor, and adjust equipment.

---

From AI Safety to Steel and Aluminum

Leike's pivot surprises those who followed his public exit from OpenAI. He resigned alongside co-founder Ilya Sutskever after criticizing the company's prioritization of product development over safety research. In a parting statement, Leike warned that OpenAI was building "smarter and smarter" systems without adequate safeguards.

Now he's applying that same caution — and technical expertise — to physical infrastructure.

"The gap between what AI can do in software and what it can do in the physical world is enormous. We're closing it by treating manufacturing as a reinforcement learning problem, not a robotics problem."

— Jan Leike, CEO of Mechanical Orchard, in an interview with The Information

The distinction matters. Traditional factory automation relies on pre-programmed routines: robots follow scripts written by engineers for specific tasks. Leike's approach treats each manufacturing job as an optimization challenge where the AI learns optimal tool paths, material handling, and quality control through trial and error in simulation before touching physical equipment.

Mechanical Orchard's system runs 10,000 simulated factory hours for every hour of physical operation, according to technical documentation shared with investors. This ratio allows the AI to encounter edge cases — tool wear, temperature fluctuations, material inconsistencies — that would take years to experience in a real facility.

---

What $142 Million Buys in Hardware

The funding round, led by Andreessen Horowitz and joined by Spark Capital and Leike's former employer Sutskever's new venture SSI Inc., values Mechanical Orchard at $680 million pre-money. That's roughly 4.8x the capital raised — a multiple that reflects both Leike's reputation and the scarcity of proven AI-to-robotics applications.

CompanyFoundedFundingFocusKey Differentiator Mechanical Orchard2024$142MAutonomous CNC/additive manufacturingRL-first training, simulation-heavy Covariant2017$222MWarehouse roboticsFoundation models for manipulation Figure AI2022$754MHumanoid factory workersGeneral-purpose humanoid form factor Bright Machines2015$437MSoftware-defined manufacturingMicrofactory modularity Machina Labs2019$109MAI-powered sheet metal formingRapid process development

Mechanical Orchard's simulation-first approach distinguishes it from competitors. Covariant and Figure AI focus on robotic arms that learn to manipulate objects. Bright Machines sells modular factory cells that customers configure. Leike's team is building what amounts to autonomous manufacturing software that can be deployed across existing equipment brands — a potentially more capital-efficient model than selling robots.

The company already generates revenue. Three aerospace suppliers and one medical device manufacturer are running Mechanical Orchard systems in production, though Leike declined to name them or disclose revenue figures.

---

Why Manufacturing, Why Now

American manufacturing faces a skilled labor shortage that isn't resolving. The National Association of Manufacturers projects 2.1 million unfilled jobs by 2030. Wages for CNC machinists have risen 34% since 2019, according to Bureau of Labor Statistics data, without corresponding productivity gains.

Leike saw the opening during his OpenAI tenure. The company's robotics research, discontinued in 2021, had demonstrated that reinforcement learning could master complex manipulation tasks when given sufficient compute and simulation fidelity. But OpenAI pivoted to language models — higher margins, faster scaling, less hardware risk.

Mechanical Orchard applies those lessons to a market with different economics. Manufacturing customers pay for outcomes: parts produced to specification, on time, with minimal scrap. The company charges $0.18 per machine-hour for its software layer, roughly 15% of typical skilled technician labor costs in high-wage regions.

But the real bet is on capability expansion. Leike's team is training systems to handle the full production lifecycle: quoting jobs from CAD files, ordering materials, scheduling maintenance, and adapting processes when specifications change. The goal isn't a robot that replaces one worker. It's a system that operates a factory cell with supervisory oversight from someone without machining expertise.

---

The Safety Question Returns

Leike hasn't abandoned his safety focus. Mechanical Orchard's technical documentation includes a "physical systems alignment" framework — essentially, constraints that prevent the AI from taking actions that could damage equipment or endanger workers. The company has hired three former OpenAI safety researchers and maintains a partnership with the Center for AI Safety.

"The stakes for getting this wrong are immediate and visible. A hallucinating language model writes nonsense. A hallucinating manufacturing system destroys a $400,000 five-axis mill."

— Paul Christiano, former OpenAI alignment researcher and Mechanical Orchard advisor

The comparison reveals Leike's strategic positioning. By choosing a domain where AI failures have tangible, bounded consequences, he's created a testing ground for alignment techniques that could scale to more capable systems. Mechanical Orchard's constraint engineering — hard limits on force, speed, and tool paths that override learned behaviors — might inform how future AI systems are governed in higher-stakes environments.

---

What to Watch

Mechanical Orchard plans to announce two additional facilities by June, targeting automotive suppliers in the Midwest. The company is also recruiting reinforcement learning researchers with robotics experience — a talent pool that has thinned as the industry concentrated on generative AI.

Leike's timeline is aggressive. He told investors the company will demonstrate fully autonomous 72-hour production runs — no human intervention — by Q4 2026. That capability, if achieved, would represent a step change from current "lights-out" manufacturing, which typically automates specific processes while maintaining human oversight for exceptions and changeovers.

The broader question is whether reinforcement learning's software successes translate to physical domains at scale. Simulation-to-reality gaps have defeated previous attempts. Mechanical Orchard's $142 million bet is that those gaps are narrowing faster than the competition recognizes.

---