The rise of physical AI: can robots save US manufacturing?

On the assembly lines at GE Appliances’ 1.1mn sq ft cooker plant in the city of LaFayette, Georgia, humans and robots work side by side. Large automated arms flip stoves upside down so workers can wire their bases, racing together to produce one unit every 15 seconds.

The entire operation is monitored by sensors and cameras gathering data displayed on big screens overhead and fed to managers eking out every last drop of productivity in the notoriously low-margin consumer goods business.

“A human sitting there inspecting thousands of units each day, their mind tends to drift and it’s challenging, it’s a very numbing job, and quite frankly, they’re not that great at it,” says Bill Good, head of manufacturing at the Chinese-owned company. “But an AI vision system doesn’t ever get tired, and it never misses.”

What really excites Good, however, is what he’s working on next: a live digital simulation of every process at every one of the company’s plants that will allow him to diagnose and address problems as they arise. “[It can’t] turn a wrench, but what it can do is tell me specifically what’s wrong with it so that I can go fix it.”

Good is not alone in his enthusiasm about so-called “physical AI”, a concept that encompasses everything from humanoid robot workers to automated “smart factories” that can adapt their own operations in real time.

Nvidia founder Jensen Huang has predicted that “every industrial company will become a robotics company” as AI expands from the digital to the physical realm, while Elon Musk is pumping tens of billions of dollars into Tesla’s pivot from electric vehicles to robotaxis and humanoids.

The prospect is exciting Wall Street and Silicon Valley, with manufacturing start-ups that offer “robotics-as-a-service” attracting valuations redolent of software companies.

Deepu Talla, Nvidia’s vice-president of robotics and edge AI, says the company expects its physical AI-related revenues to rise from $10bn to $100bn over the next decade. “Ninety per cent of the world’s actions need to happen in the physical world, not just in the digital world — that’s why we believe the opportunity for physical AI is arguably an order of magnitude larger.”

The promise of physical AI is also feeding hopes in Washington that America’s supremacy in tech will ride to the rescue of a manufacturing sector long in decline, spurring a revival in the country’s competitiveness.

“The only way the US is going to regain its role in manufacturing is by applying AI more effectively than China,” says Chris Miller, an economic historian at Tufts University.

A particular focus for the Pentagon is its potential to revitalise the US defence industrial base, which is struggling to expand production amid a severe shortage of munitions and intensifying competition from China.

Leading defence contractors under pressure from AI-savvy rivals like Anduril and Palantir are rushing to embed physical AI in their factories, while defence-focused start-ups offer everything from 3D printing of composite materials to gecko-like robots used to inspect the hulls of submarines.

But labour unions worry the trend could have a devastating impact on blue-collar workers, as the most valuable tasks are automated and the fruits of productivity gains are absorbed by shareholders.

“Dangerous, de-skilling and harmful uses of AI are occurring with too much frequency,” says Ed Wytkind, senior adviser to the American Federation of Labor and Congress of Industrial Organizations’ Tech Institute.

“We must ensure that workers, not billionaire tech companies or investors, determine the future of jobs in the US and around the world.”

In Cherokee, Alabama, deep in the country’s south-eastern manufacturing region, a physical AI start-up called Hadrian is building a new facility on the site of an old railcar factory.

Founded by Chris Power, a 35-year-old Australian, Hadrian uses AI software to co-ordinate automated machines performing tasks ranging from welding and fabrication to machine tooling, helping manufacturers bypass their traditional reliance on highly skilled technicians. It also has units embedded in factories operated by Lockheed Martin and the US Army.

Power notes that a chronic shortage of machinists and other technicians is constraining production in the aerospace and defence industries as they struggle to address ballooning backlogs.

“The vast majority of manufacturing work of any defence programme is actually done by a hundred thousand small businesses that on average are run by 65-year-olds with less than $10mn in revenue,” he says. “Those owners and workers are retiring and ageing out and no one is coming behind them to replace them.”

Because the key technical knowledge is contained within Hadrian’s automated systems, most of his workforce has no prior manufacturing experience, and can be put to work after just 30 days’ training, he says.

“The backgrounds we’re seeing [are] you’re driving a forklift at Home Depot, you were a diving instructor, you’re a marine,” says Power. “You have a lot of 40-year-olds that are coming out of bus-driving jobs.”

Hadrian, which was founded in 2021 and is based in Southern California, this month closed a $1.37bn funding round at a valuation of almost $8bn — more than four times its valuation following its previous funding round in January.

Its investors include JPMorgan Chase’s Strategic Investment Group, software pioneers Andreessen Horowitz and Founders Fund, as well as Donald Trump Jr’s investment firm 1789 Capital.

“I think if you’d said to someone on Wall Street even three years ago that the largest investment banks are seriously leaning into manufacturing, they would have laughed you out of the room,” says Power.

Talla of Nvidia says excitement is building because of dramatic progress in AI-powered general-purpose reasoning, which allows robots and automated systems to figure out solutions even to problems they have not been trained to deal with.

An automated system trained to perform a specific task is “brittle, because the moment it sees things that it was not trained on in the past, it doesn’t know what to do,” Talla says. But now, software advances are allowing systems to be trained to perform multiple tasks in increasingly sophisticated virtual environments.

The closer to the real world these virtual environments get, the cheaper it is to train a robot and the more problems it can solve for itself, lowering costs and allowing them to be adopted by more and more small and medium-sized businesses, which account for the majority of US manufacturers.

This will lead to the proliferation not only of “billions of robots” globally, but also of “outside-in” systems that use cameras and sensors to monitor and co-ordinate all the different automated activities into a coherent whole, says Talla.

America’s tech strengths, including its lead over China in AI and chip design, give it an advantage in this area, says Miller of Tufts, who singles out western dominance of “electronic design automation” software, traditionally used to design semiconductors but which is now increasingly being deployed in a wide range of other manufacturing sectors.

The US ability to finance innovation is also peerless, he adds. “The speed at which the US has deployed capital to build data centres is astounding. It illustrates that if indeed we do see a lift-off in physical AI over the coming years, capital may indeed be mobilised more quickly than in China.”

For Power, radical automation is the only chance the US will ever have of competing with China’s manufacturing heft. He argues the ultimate goal is to preserve the pillars of American financial and military power, and through it the west’s global pre-eminence.

“The next transition is very simple — either America reindustrialises itself, or the world will be led by China.”

Some experts warn that the physical AI evangelists are getting ahead of themselves, noting the first wave of industrial robotics in the 1980s also led to predictions of a golden age in American manufacturing that never materialised.

Ken Goldberg, a professor and roboticist at UC Berkeley, cautions that the technological obstacles to mass adoption remain formidable because manufacturing requires such a detailed understanding of the physical world.

Whereas large language models are trained on information already accumulated on the internet, equivalent data sets for the physical world do not exist and will take years to gather.

Goldberg adds that people impressed by displays of humanoids performing backflips and scripted martial arts routines — such as those shown off at the World Humanoid Games in China this month — underestimate humans’ distinct physical advantages.

“We can predict the motion of an asteroid a million miles away, but we still don’t fully understand the physics of a fingertip interacting with another deformable surface like a shoelace,” says Goldberg, co-founder of Ambi Robotics, which is developing AI-enabled robotic software for logistics and ecommerce companies. “That gap in understanding is why we still don’t have a robot that can tie up a pair of shoes, or to pile up wooden blocks as skilfully as a toddler.”

He notes that mistakes that might be accepted as an easy fix in an LLM would not be acceptable in a manufacturing context. “When something is screwed up on the assembly line, that can cost you $100,000 a minute.”

Analysts see signs that Wall Street is also recalibrating its expectations. Shares in Nasdaq-listed Symbotic, which deploys AI-enabled robots to operate highly automated warehouses on behalf of large customers like Walmart, are down more than 30 per cent this year after its operating margins undershot investor expectations as it struggles to convert its order backlog into revenue.

Goldberg adds that venture capital firms that have ploughed money into humanoid robotics start-ups are “getting impatient — they want to see something work, and that’s been a challenge because there’s no one really producing any revenue yet.”

Denise Hall, founder of workforce development consultancy Peak Performance in Cleveland, Tennessee, says that persuading small-scale manufacturers to deploy their limited capital in untested technologies will also be difficult.

“The majority of US manufacturing does not look like an Elon Musk factory,” says Hall, whose Smart Factory Institute helps prepare manufacturing companies and their employees for the coming wave of automation.

“They’re still doing lots of paperwork, they’re still doing a lot of manual handling,” she adds. “You might have a plant manager who’s driving the forklift because Joe called out sick today — so how can they find time to be strategic?”

The degraded and highly fragmented nature of America’s industrial base also casts doubt on its ability to compete with China’s scale and level of integration.

“Because so many spheres of manufacturing have migrated to other regions, there’s no American ecosystem at all in certain sectors,” says Miller of Tufts. “When it comes to the challenge of actually building the physical robots, there’s nowhere better in the world than Shenzhen, just because of the depths of the consumer electronics ecosystem there.”

There is also the question of what this process of radical automation might mean for jobs and wages.

GE Appliance’s head of manufacturing Good says that automation liberates workers from “dull and difficult” tasks, allowing them to concentrate on areas where they can be more productive and allowing management to keep more production and jobs in the US. He notes the company added 4,000 jobs between 2016 and 2025 “even though we have put in a ton of fixed robots and mobile robots”.

Companies also point out that human workers with the necessary skills are very often simply not available across the US. Rather than humans being displaced, they say, more opportunities will be created as a result of bottlenecks constraining activity being unblocked.

When Scott Samples joined the GE Appliances cooker plant in LaFayette as a high school graduate, his job was to plug in wires on the assembly line. Sixteen years on, he oversees a team of technicians operating robots performing logistical tasks ranging from stocktaking to moving parts around the factory floor.

“The coolest thing about the job now is that my kids dig it,” he says. “They think that Dad gets to play with robots all day, so it kind of makes me feel alive.”

Power of Hadrian argues that his company’s ability to put inexperienced employees to work after just 30 days’ training lowers the barriers to entry into the manufacturing workforce.

But David Autor, a labour economist at MIT, says that while individual companies might increase headcount as they grow more productive, it is a fallacy to assume this will be replicated across the manufacturing sector as a whole, noting that previous waves of automation in the US have all driven large-scale declines in blue-collar employment.

He questions how many opportunities there would be in practice for workers to replicate Samples’ career path as a “robot wrangler” at GE Appliances. “There can’t be that many such jobs — you can’t have everyone being a manager of everyone and everything else.”

There may be opportunities elsewhere, however. Elizabeth Anderson, executive vice-president and chief academic officer at Georgia Northwestern Technical College’s campus down the road from GE Appliances in LaFayette, notes that “even though AI is being incorporated to increase productivity, we still have to train the technical skills of knowing how to make AI do what it needs to do”.

That could enable human workers to retain some degree of leverage in wage negotiations with their employers. But if the skills required to supervise and maintain automated systems are even more difficult to acquire than those needed at present, physical AI may struggle to deliver on its promise of circumventing America’s shortage in technical skills, as the bottleneck shifts from a shortage of one set of skills to another.

Hall of the Smart Factory Institute in Tennessee cautions that as long as the necessary skills are in short supply, workers will continue to enjoy bargaining power in the job market. Over time, however, “as the machines become smarter and more self-correcting and more self-aware, yes, skills will be devalued and leverage will be lost”.

Autor of MIT notes that whatever the consequences for employment, the US still has a strategic interest in reviving — or at least protecting — its manufacturing sector for reasons of economic and national security. “I do not think of manufacturing as an employment programme.”

Nevertheless, some are eyeing up opportunities in the emerging world of physical AI. Braylon Camp, a 17-year-old student at Georgia Northwestern Technical College, simultaneously works as a mechatronics intern at a plant in nearby Rock Spring run by Japanese auto component maker Astemo.

“I originally got put into welding but I said no, I want to do the hardest, most complicated thing I can, so other people can’t easily compete with me for a job,” says Camp. “That’s really what I strive to do, to be the best mesh of skills I can be — I’m just like, I’m not gonna fail, that’s all it is.”