The numbers tell a story that most manufacturing executives would rather not hear. The American Welding Society projects a need for approximately 80,000 welding professionals per year from 2025 to 2029, totaling 320,500 by the end of that period. According to BLS, in 2025, the workforce was 573,000, down from 592,000 in 2024—and more than 157,000 of those are nearing retirement. Do the math, and the problem becomes uncomfortably clear.
Which is why, on shop floors from California to Virginia, a new generation of machines is learning to weld. Not following pre-programmed paths the way industrial robots have for decades, but adapting in real time to the messy realities of fabrication work: inconsistent fit-ups, varying materials, the high-mix chaos that defines most American manufacturing. The question has never been whether robots can weld. The question is whether they can finally do it without requiring a programming PhD to set them up.
Enter physical AI—a term that's rapidly moving from Silicon Valley pitch decks to the language of defense contractors and job shop owners. These aren't your grandfather's robotic welding cells. They're machine learning systems that see, adjust, and improve with every arc struck. And they're arriving at precisely the moment when the alternative—finding enough human welders—has become something close to impossible.
A Structural Problem, Not a Cyclical One
The welder shortage is considered by many experts to be a structural problem rather than a cyclical one—it's not going away when the economy shifts or when the next recession hits. Bureau of Labor Statistics data shows employment levels dropping year-over-year even as demand surges across infrastructure projects, semiconductor fabrication plants, shipyards, and the relentless expansion of data centers. The median welder salary is modeled at approximately $55,895 in 2026, and shops still can't fill positions.
They're competing for talent with CHIPS Act-funded semiconductor facilities, the sprawling Infrastructure Investment and Jobs Act buildout, and an $85 billion data center construction market that continues expanding despite mounting concerns about power constraints. It's a fight most fabricators are losing.
The welding equipment market itself is enormous: $21.66 billion globally in 2025, projected to reach $32.53 billion by 2033. Yet traditional robotic welding has always suffered from a crippling bottleneck. Programming a conventional welding cell for high-mix, low-volume work—the daily reality for most job shops—meant weeks of painstaking teach-pendant programming for each new part. Fine for automotive production lines running millions of identical welds. Impossible for a shipyard building one-off assemblies.
Something has shifted. The robotic welding segment is forecast to grow from $9.7 billion in 2025 to $29.9 billion by 2035, an 11.9% compound annual growth rate. That's not incremental improvement. That suggests a fundamental change in what these systems can do—and who can use them.
Three Technologies Converge
The transformation rests on the convergence of three distinct technological advances, each of which has matured just as the labor crisis reached critical mass.
First: computer vision and real-time seam tracking. Modern systems can identify weld joints, track them dynamically, and adjust torch position mid-weld when parts don't fit perfectly. Which, in the real world, they rarely do.
Second: machine learning models trained on thousands of welds can predict optimal parameters, detect defects during the arc, and close the loop with post-weld inspection. The robots aren't just executing commands; they're learning what works.
Third: the cost of computing has fallen far enough, and cloud infrastructure has become accessible enough—Y Combinator now offers $2 million in OpenAI credits to every startup in its batches—that embedding AI directly into welding cells has become economically viable.
The labor crisis provides the immediate catalyst. But policy tailwinds matter too. The Build America, Buy America Act and IRA domestic content bonus credits, which offer graduated thresholds from 40% to 55% manufactured product content by 2027, are pushing fabrication back onshore. More welding, in other words, just as the welder pool shrinks.
U.S. industrial robot installations fell 9% year-over-year in 2024 to roughly 34,200 units, according to the International Federation of Robotics. But metal industry installations grew 16% in the same period. Welding, it turns out, is where automation is actually happening—driven by necessity rather than aspiration.
From Hawthorne to Hampton Roads
Advanced Metal Research exemplifies the new breed of physical AI startup. A three-person team in Y Combinator's Spring 2026 batch, the company is building what it describes as "machine intelligence for American Welding"—robotic cells with real-time seam tracking and post-weld inspection in a closed-loop system that learns from every arc.
The founding team is telling: Angus Muffatti, the CEO, previously built a startup to $1 million in revenue and has published machine learning papers alongside liquid bi-propellant rocket research. Stephen Lin came from NASA, where he built systems used by mission control for 19 spacewalks. Julian Fried was a welding foreman in eastern Pennsylvania. It's the kind of unusual combination—domain expertise married to AI research—that distinguishes physical AI ventures from pure software plays.
At a different scale entirely, Path Robotics launched Rove in April 2026: a mobile welding robot on a legged platform that brings its Obsidian physical AI model directly to large assemblies and field installations. Think less factory floor, more shipyard or construction site.
Huntington Ingalls Industries, the largest military shipbuilder in the U.S., signed a memorandum of understanding with Path in February 2026 to integrate physical AI welding into both manned and unmanned shipbuilding. Eric Chewning, HII's executive vice president, positioned it as a throughput multiplier, targeting a 15% increase through workforce augmentation. Saronic, a Louisiana-based shipbuilder, became an early Rove customer. Path's CEO believes that America faces a shortage of 600,000 welders—a figure higher than AWS projections, though directionally aligned—and positions autonomous welding as the only scalable solution.
Miller Electric and Novarc Technologies formed a strategic partnership in early 2025 focused on AI-enhanced collaborative robots and pipe welding. By September, Novarc had unveiled its SWR+TIPTIG Autonomy system, claiming fully autonomous gas tungsten arc welding for pipe spools. Case studies from deployments like W.W. Gay Mechanical report shop productivity doubling and weld productivity increasing twelvefold, though the specifics on part mix and comparison baselines aren't always disclosed. (Welcome to the world of vendor-supplied case studies.)
The established equipment manufacturers are moving too. FANUC demonstrated AI-enabled welding with its CRX-3iA collaborative robot at Automate 2026 in May, showcasing vertical-up weld profiles designed to replicate skilled welder techniques. Universal Robots partnered with Scale AI to launch the "UR AI Trainer" at GTC 2026 in March, an imitation learning system meant to accelerate AI model training from laboratory to factory floor. Lincoln Electric discussed autonomous welding solutions using vision and AI during its February 2026 earnings call, alongside positive commentary on automation backlog.
Oak Ridge National Laboratory and Lincoln Electric are collaborating on large-scale wire arc additive manufacturing datasets and process control research. ORNL released co-registered in-situ and ex-situ datasets in early 2025, the kind of foundational work that often precedes commercial breakthroughs by a year or two.
The Framing Problem
How you describe these systems matters. The narrative of "physical AI as apprentice" circulated through industry publications like the American Welding Society's Welding Digest, positioning the technology as a tool to eliminate programming barriers rather than replace human welders outright.
It's a framing that plays well with an industry anxious about labor shortages but wary of full automation. Whether it's entirely accurate is another question.
The economics are increasingly compelling, at least on paper. Collaborative robots now represent roughly 18% of North American robot units, and welding ranks among the fastest-growing applications. Reported payback periods range from 12 to 36 months depending on labor rates and shift utilization. Robotics-as-a-Service models—hourly or monthly rentals rather than six-figure capital purchases—are gaining traction as well, lowering the barrier for skeptical shop owners.
But deployment reality tempers the hype. Integration challenges, connectivity issues, and operator training often determine ROI more than raw algorithmic capability. Path's own product page notes that only 50 Rove units are shipping in 2027, suggesting demand exceeds near-term supply in these early market cycles. Mobile autonomous welding for shipyards and large field assemblies remains, to put it mildly, unproven at scale.
The regulatory landscape is evolving in parallel, though few outside the industry pay attention. The ANSI/A3 R15.06-2025 standard—the U.S. adoption of ISO 10218-1/2:2025—was published in September 2025, updating industrial robot safety standards and integrating collaborative robot provisions that were previously separate guidance. Mobile robot systems fall under ANSI/A3 R15.08-2:2023. AWS D1.1:2025 represents the latest structural welding code for steel.
These aren't trivial details. They define liability, insurance requirements, and what's permissible on job sites that include both human welders and autonomous systems working side by side.
The Path Forward—Assuming It Exists
The near-term outlook hinges less on technology than execution. Management consultancies like Deloitte and BCG have positioned 2026 as an inflection year for embodied industrial AI, with manufacturers moving from pilots to scaled deployments. Foundation models and simulation-to-real transfer are reportedly cutting training times significantly and expanding the range of automatable tasks, though such claims deserve healthy skepticism given how context-dependent they tend to be.
Several trends seem likely to accelerate regardless. First, the shift from "programming" to "teaching" robots through demonstration and imitation learning should make automation accessible to smaller fabricators without robotics engineers on staff. Second, real-time quality assurance—acoustic sensors that detect defects during the weld, vision systems that catch problems immediately—will likely become baseline features rather than premium options. Third, data platforms from Lincoln CheckPoint, ESAB WeldCloud, Fronius WeldCube, and others will enable fleet-level analytics, turning every weld into training data for the next one.
For founders, the opportunity is clear but intensely competitive. Physical AI for welding sits at the intersection of robotics, machine learning, and deep manufacturing domain knowledge. A rare combination. The companies gaining traction aren't just building better algorithms; they're solving integration, safety, and usability problems that keep these systems from reaching production environments in the first place.
For manufacturers, the calculus has shifted. With 80,000 welder jobs opening annually through 2029 and a shrinking labor pool, the traditional model of hiring and training simply doesn't scale. The question has become less "can we afford automation?" and more "can we afford not to?" Whether to buy off-the-shelf cells from established suppliers, lease through RaaS models, or integrate custom systems from newer entrants represents the next decision point.
For policymakers, the challenge is ensuring that domestic content requirements and manufacturing reshoring initiatives don't outpace the labor and automation capacity to support them. Data center construction, semiconductor fabs, shipbuilding for the naval industrial base—all of these programs assume welders exist. When the Secretary of the Navy says shipbuilders need to hire 250,000 workers over the next decade, it's unclear where those workers are supposed to come from. Physical AI welding offers a potential path forward, but only if deployment keeps pace with demand.
The Best Welders Stay
The welder shortage won't disappear. Not completely, anyway. The robots aren't replacing welders so much as multiplying their output and taking on the repetitive, high-volume work that's hardest to staff. The best welders—the ones who can tackle critical joins, complex repairs, and the kind of adaptive problem-solving that no current AI system can match—will remain essential.
But the floor is being raised. The shops that figure out how to integrate human expertise with autonomous systems will have a decisive advantage in a market where labor, not capital or technology, has become the binding constraint.
Perhaps that's the real story here. Not that robots are learning to weld, but that they've finally become good enough to matter—just as America discovers it doesn't have enough humans left to do the job.
