Path Robotics closed a $300 million Series E on November 10 at a $2.45 billion post-money valuation, nearly quadrupling its previous price and bringing lifetime funding to roughly $610 million1,2. The round, led by Georgian with new investors D.E. Shaw Ventures and NAVER Ventures joining existing backers including Tiger Global and Yamaha Motor Ventures, is a major bet on industrial-robotics AI—though it is dwarfed by the capital flowing into adjacent AI infrastructure, where Crusoe alone raised $3.9 billion in September and Anthropic priced an IPO at $100 billion the following day6,9. The valuation rests on a specific thesis: that an AI model trained on one of manufacturing's hardest problems, welding, can generalize across the factory floor fast enough to justify a price that implies Path is already a platform, not a point solution.
The company's own numbers make the bull case. Path said annual recurring revenue grew 17-fold over the past 18 months11. Its customer roster includes HII, Lockheed Martin, and autonomous-vessel maker Saronic10,13. A $600 million performance-based production agreement with HII, announced earlier this year, ties Path's technology to U.S. Navy shipbuilding programs. The company reported 97 percent year-over-year workforce growth and said it plans to reach 400 employees by year-end. CEO Andy Lonsberry said Path plans to begin expanding its technology beyond welding into other production-line tasks in 2027.
The core technology is Obsidian, a model that uses data from production welds to help robots identify joints, plan and execute welds, and adjust as conditions change. Welding is a useful proving ground precisely because it resists traditional automation: heat-induced metal distortion warps parts unpredictably during and after the process, and mass-produced components vary enough from blueprint to blueprint that static programming often fails. Path has also introduced Rove, a quadruped mobile welding system that runs the same Obsidian software, with 50 preorders reported. "We started with welding because it is one of the hardest jobs on the factory floor to automate, and we've proven that our technology can succeed in real production environments," said Andy Lonsberry, CEO and co-founder. "Our goal isn't to automate one step, but to automate the entire factory," Lonsberry said separately.
ANALYSIS That ambition is where the valuation's logic gets tested. A 17-fold ARR increase sounds extraordinary, but the base was small and the named customer roster skews heavily toward defense: HII, Lockheed Martin, and Saronic account for the only publicly identified production deployments. Path has not disclosed what share of ARR these contracts represent, so the degree of concentration is an open question rather than a settled fact. The $600 million HII agreement is performance-based, meaning revenue recognition depends on deployment milestones rather than upfront commitment. Lonsberry's stated plan to move beyond welding in 2027 means the diversification that would justify a platform multiple has not yet begun.
The broader context, however, explains why investors are paying up now. Google DeepMind introduced Gemini Robotics 2.5 on November 10, a vision-language-action model that achieved a 96 percent success rate when Apptronik's Apollo 2 humanoid robots sorted automotive parts on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany4,5[1]. DeepMind said it tested the model across more than 200 tasks, with success rates 30 to 100 percent higher than previous baselines depending on task type, and task completion 20 to 30 percent faster3,14. The system also posted a 33 percent improvement in safe task completion around obstacles and 43 percent near people. DeepMind used Gemini 4 Argon as a simulation engine to generate thousands of training runs from real-world camera footage. On safety, Google reported that Gemini 4 Argon achieved 99.5 percent resistance to adversarial prompts under its new open-source ASIMOV-Jailbreak benchmark, compared with 82 percent for the previous Gemini Robotics ER 2 baseline.
Separately, ABB Robotics expanded its OmniCore EyeMotion platform with AI Robot Trainer, a tool designed to let system integrators train, deploy, and improve AI-powered robot vision applications without building models from scratch7,8. "Physical AI fundamentally changes what robots can do, where they can operate and the value they can create," said Marc Segura, president of ABB Robotics.
Google is demonstrating that a foundation model can drive a humanoid through real production-line kitting. ABB is lowering the expertise barrier for deploying AI vision in existing robot fleets. Path is raising capital to extend a domain-specific model across new tasks and form factors. The common bet across all three is that AI models trained on physical-world data can now operate reliably enough to earn production-floor trust, not just pilot-program tolerance. But the tasks are not equivalent, and the competitive overlap is narrower than the shared narrative implies.
The Mercedes-Benz trial was a proof-of-concept, not a production deployment. A 96 percent success rate in parts kitting means roughly one failure in 25 attempts. On an automotive line running thousands of operations per shift, that error rate would require human oversight at a level that limits the labor savings the technology promises. Gemini Robotics ER 2's vendor-claimed progress-classification accuracy is 57.40 percent, and its vendor-claimed moment-finding accuracy is 91.30 percent, both as of November 10. Those figures are strong for a research model but leave considerable room before reaching the reliability levels automotive OEMs would need for unattended operation. DeepMind's own framing, describing the work as a way to let "technicians focus on complex build tasks" rather than replacing them, acknowledges this constraint.
Path faces a related but technically distinct challenge. Its welding AI works in production environments today, but the jump from welding to "the entire factory" requires Obsidian's perception-and-adaptation architecture to transfer across tasks with fundamentally different physics: cutting, grinding, assembly, inspection. Welding involves real-time weld-pool dynamics, heat-affected-zone prediction, and torch-path correction under thermal distortion. That domain specificity is Path's moat, but it is also its constraint: the skills that make Obsidian valuable for welding do not automatically transfer to pick-and-place kitting or visual inspection, just as Google's kitting model does not transfer to welding. The 2027 timeline Lonsberry cited for non-welding expansion is aggressive, and the company has not disclosed results from non-welding applications.
The investment logic is a bet on domain depth versus breadth. Path's $2.45 billion valuation prices in the platform it intends to become, not the welding specialist it is today. Without disclosed absolute revenue figures, a precise multiple is impossible to calculate. Georgian and its co-investors are betting that the head start in production data, the defense customer base, and the 17-fold ARR trajectory create a moat wide enough to hold while the technology generalizes. The risk is not that Google's foundation models will solve welding, but that platform-level AI lowers the barrier for new entrants in the adjacent tasks Path needs to capture to justify its valuation.
What to watch: whether Path discloses non-welding pilot results before mid-2027, which would validate the platform thesis on the timeline Lonsberry set. Whether Mercedes-Benz or another OEM converts the Gemini Robotics 2.5 proof-of-concept into a contracted production deployment. Whether Gemini Robotics ER 2's vendor-claimed progress-classification accuracy of 57.40 percent, as of November 10, improves enough to support that conversion. Whether Path's ARR growth rate holds through 2027 as it moves past the initial HII ramp. And whether Path's customer base diversifies beyond defense: the named roster (HII, Lockheed Martin, Saronic) is narrow, and the 17-fold ARR claim will mean more once the company discloses how many customers contribute to it. Path's round and Google's demo both confirm that physical AI is reaching production environments, but they also confirm that welding and kitting remain separate technical problems, not interchangeable proofs of a single capability.