Move Fast and Make Things

Revitalizing American manufacturing has become at once a bipartisan priority and a source of contentious debate in Washington. Some believe that tariffs, direct subsidies to industry, and similar industrial policies are the solution. Others argue these policies will cause more problems than they will solve. Some debate whether American manufacturing has declined as much as is commonly believed. Others point to shortages of key goods—semiconductors during the pandemic, electrical transformers for the grid, and munitions for the war in Ukraine—as evidence of a persistent and serious crisis in American manufacturing.

One empirical fact, though, cannot be disputed: The labor productivity of American manufacturing has declined in the past decade. American workers needed 106 hours of work in 2023 to produce what needed only 100 hours in 2013. And while policies to encourage investment in domestic industrial capacity or subsidize the purchase of American manufactured goods can help grow output, their ability to boost productivity is less obvious. Productivity enhancements, as a general rule, tend to come from increased and more sophisticated use of technology. Only with higher productivity can a sector compete globally while offering its workers better jobs and rising real wages.

Imbuing America’s sprawling and kaleidoscopically diverse manufacturing industry with contemporary and emerging technology is a tall order. It will require capital investment, regulatory reform, and the training of a new generation of advanced manufacturing workers—all subjects of intense focus for policymakers. Technology itself, however, will be the main driver of any American industrial resurgence. And the frontier technologies of today are more promising than at any time in recent memory. New companies are being founded by the month to develop and harness them.

What Happened to Productivity?

Why has manufacturing productivity stagnated? One problem is that American manufacturers use technology in less sophisticated ways than our competitors. In the 1970s and 1980s, American factories rushed to adopt industrial robotics, computers, and other then-emerging technologies. The maturation of these technologies drove productivity increases in the 1990s especially, but also led firms to often stand pat as the next wave of innovation moved forward.

Technological choices tend to be sticky. For example, the U.S. was very early to adopt credit cards when they were an emerging technology, building a nationwide payment infrastructure predicated upon the use of physical cards in every store. But with that infrastructure in place, the incentive to develop digital payment systems was weaker. Developing economies, ready to build out their own networks at the same time mobile devices were becoming ubiquitous, jumped straight to digital means of payment years before the U.S. did. A similar story played out in manufacturing, where American factories built to what was once cutting-edge standards now look woefully out of date. As of 2021, the U.S. ranked ninth globally in the density of its industrial robotics, closer to Slovenia than to China, and barely one-quarter the level of Korea.

Compounding this “leapfrog” effect, U.S. manufacturers began rapidly offshoring their capacity in the early 2000s, just as a new wave of investment would have been necessary. Instead, investment slowed significantly. Sector-wide productivity appeared briefly to spike as a result of composition effects—those segments with lowest productivity were likeliest to offshore first, making what remained domestically appear more productive by comparison. But once that dynamic played out, stagnation set in. A vicious cycle emerged: Low productivity in American manufacturing made it a low-prestige, relatively low-pay occupation, which thus struggled to attract talent, making productivity gains yet harder to achieve.

Obscuring the broader trend, and confusing many observers, there are also high-profile bright spots. SpaceX designs and builds the world’s most reliable rockets in the United States. These rockets collectively carry the vast majority of the equipment humanity puts into space every year. Key semiconductor manufacturing equipment is made in the United States. American firms lead the world in drilling for oil and gas, which they have transformed into a high-tech field, in turn transforming the United States into the world’s leading hydrocarbon producer.

But the productivity statistics do not lie. The gap between the most and least productive American manufacturers has widened during the same period that the aggregate productivity figure has stagnated. The best manufacturers do not drive the overall trend. If anything, these examples underscore the fundamental problem: U.S. policymakers and technologists alike often fixate on the frontiers of emerging technology, forsaking the basics. They obsess over flagship manufacturing facilities without building the infrastructure and nurturing the industrial commons necessary for those facilities to create the most value themselves and throughout a broader ecosystem.

When asked why the iPhone was not made in the United States, Apple CEO Tim Cook did not bemoan the lack of a high-dollar, publicly subsidized manufacturing plant. Instead, he talked about a lack of fundamental infrastructure and capabilities. “You could take every tool and die maker in the United States and probably put them in the room that we’re currently sitting in,” he noted. “In China you would have to have multiple football fields.” When Apple attempted to manufacture its top-of-the-line Mac Pro in Texas, it was thwarted not by a lack of STEM PhDs but by a lack of properly sized screws.

The bad news for policymakers is that rebuilding an atrophied, capital-intensive ecosystem takes much more than luring flagship investments from famous multinational corporations. The good news, though, is that the game of leapfrog calls for the jumped-over player to then do the jumping. And the oncoming wave of AI-enabled manufacturing technologies are well suited to U.S. advantages.

The Future of Manufacturing

What constitutes the basic industrial infrastructure of manufacturing? Generically, it is the ability to move raw materials through processes such as casting, machining, welding, and finishing, and then assemble those processed materials reliably into useful things. Basic as these steps may sound—in fact, partially because they sound basic—the U.S. lags behind in virtually all of them.

A new generation of firms and entrepreneurs is rebuilding our industrial base with fresh thinking and new technologies applicable at each step—first and foremost, artificial intelligence. In common usage, “AI” has quickly come to mean generative AI applications such as ChatGPT and Midjourney that produce text and images in response to prompts. Sure enough, those applications are already exhibiting a kaleidoscopic range of industrial uses. But AI is a much broader technology.

At its core, AI is software that can learn. When a child learns to throw and catch a ball, she doesn’t start by learning the equations and theorems of physics. Instead, she learns to model real-world physics from experience, so that when she sees a ball moving in some direction at some speed she can anticipate where it will land. Similarly, modern AI systems are not specifically programmed to do much of anything. Rather, programmers give them a structure to learn the patterns underlying a complex dataset, even complex aspects of reality that have evaded formal scientific approaches such as the structure of proteins or the state of hotter-than-the-sun plasma within a nuclear fusion reactor.

The potential uses for such a flexible tool are immense; most technologists believe we have just scratched the surface. In the context of manufacturing, AI has the potential to do everything from improving supply chain management to transforming radically the way that chemicals are made. It is far from the only technology with the potential to help define a new era of American manufacturing, but it is a kind of technological glue. There will be few foundational improvements to robotics without AI. The future of bioengineering industrial chemicals and pharmaceuticals is far dimmer without it as well.

Start with the raw materials. Both established firms and startups are using AI to do everything from improving equipment performance with predictive maintenance to discovering valuable mineral deposits more easily. A new company called Earth AI claims to have reduced the time to discover and verify a deposit by half.

More speculatively, in late 2023, Google’s DeepMind announced GNoME, an AI system that discovered 380,000 previously unknown materials. The researchers did so by representing the chemical structure of a material as a graph. They then generated millions of potential materials (some randomly, some based on modifications to existing materials) and, combining AI with a physics-based method called density functional theory, validated the stability of the resulting structures. Other scientists have disputed the methods of the paper, claiming that some of the discovered materials were already known. Still, if 90% of the 380,000 materials GNoME “discovered” are wrong, it would still mean the system had authentically found nearly 40,000 new materials—roughly equal to the total number of stable materials discovered in human history. Does the next Teflon, Nylon, or Velcro lurk somewhere in this discovery? Or something more exotic, like a new superconducting material, which could itself single-handedly transform the electronics industry? It will take years of scientific inquiry and industrial development to find out, but GNoME offers a promising start.

The potential for new materials becomes even more exciting when combining existing ones. Two metals mixed together are an alloy; when combined such that they do not dissolve into one another, the result is called a composite. Composites like carbon fiber (more accurately, carbon fiber reinforced polymers, or CFRPs) are important in many cutting-edge areas of manufacturing such as aerospace and automobiles because they give materials “best of both worlds” properties: strength, durability, and low weight, in the case of CFRPs. But manufacturing composite parts tends to be slow, ad hoc, and artisanal due to the industry’s combination of complex processes, stringent quality standards, and low volumes. This has also made it difficult for current composites firms to make the substantial fixed-cost investments in the software needed for automation. Companies such as Layup Parts, founded this year, are aiming to bring software-enabled automation to this burgeoning area of high-end manufacturing.

Once a material has been mined (or synthesized), a manufacturer must convert it into a useful form. Countless processes are involved, but one that stands out is machine tooling—what Brian Potter of the Institute for Progress describes as “the heart of industrial civilization.” Machine tools are the machines required to make most other machines. Almost every metal part in a commercial product—from an airplane to an iPhone—is made with machine tools.

Potter documents thoroughly how the American machine tool industry has become decrepit. As machine tools became digitized (with what is known as computer numerical control, or CNC machining), American firms, at the encouragement of the aerospace industry and the Air Force, concentrated on high-end, high-precision devices, leaving lower-end solutions to foreign competitors. Many of the firms that still exist in this field are mom-and-pop shops with aging workforces, long lead times, and an inability to operate at the scale achieved in other countries. Domestic knowledge of how to do cutting-edge machine tooling is gradually fading and the United States is at risk, as a civilization, of effectively forgetting how to do it.

Several companies have emerged to address this problem, including Fictiv, Xometry, Protolabs, and Hadrian. Their approach involves using software to automate large portions of the machining process, while relying on human workers to supervise and make key decisions. Hadrian’s CEO, Chris Power, has discussed using software to model the exact state of a tool (its precise dimensions, its level of wear and tear, and other factors) to allow for more efficient automation of the machining process.

Within the world of machine tooling, perhaps the most expensive bottleneck is tool and die making—the often-artisanal process of creating custom tools to mass-produce a specific good. This requires trial-and-error, deep understanding of the materials involved, iteration, and plenty of tacit knowledge. A Detroit-based company called Atomic Industries has developed a set of AI tools that allow substantial portions of this process to be automated, optimizing down to the level of the atom (as their name implies) for the complex combination of cost, efficiency, manufacturability, reliability, quality, and other factors that are essential for every manufactured product but vary widely across those different products. Ultimately, the objective is to make artisanal, labor-intensive industrial processes scalable.

Finally, there is assembly, where robotics plays the dominant role in improving productivity. While industrial robots have been around for a long time, engineers had to program those machines for each task they might handle—and, crucially, re-program them when a task changed even slightly. Such brittleness is exactly what contemporary approaches to AI stand to fix. In the last year, a flurry of robotics demonstrations have made headlines: Tesla’s Optimus and Figure’s 01 humanoid robots, as well as companies working on foundational AI models for robotics such as Google DeepMind and Physical Intelligence. While the demos tend to involve robots doing impressive gymnastics or basic household tasks, the initial uses for the technology are likely to be in the factory.

Traditionally, robotics training has been a manual process. Humans steer robots through the performance specific tasks. The task to be performed (“pick up the ball and move it to the right”), the state of the robot (its current position, its sensor data, and its camera data), and the exact trajectory it took to perform the task are all collected and become training data for an AI model. But at the end of the training, the robot has learned only how to move, as opposed to why it is doing the task. It has no information about what the object is or how its task relates to others. Not only do these constraints leave it incapable of adaptation, but they also preclude any interaction with humans in natural ways.

Language models like ChatGPT, by contrast, do have this more fluid understanding of objects in the world. They understand what an apple is, and that an apple is not a kind of meat. More importantly, they understand that an apple can be many things—a color, a flavor, an object, a company—and understand which kind of “apple” a user is referring to, based on context. It sounds trivial, but this kind of basic, generalized understanding of language was considered nearly impossible for AI until just a few years ago.

The constraint for language models, however, is that they have limited understanding of the physical world (though the nature and extent of their world models is an issue of intense debate). A robot with only a language model inside it could scarcely walk, much less play a role in advanced manufacturing, because it does not know how to move its legs.

What if an AI system could combine the robot action data described above with the internet-scale text, image, video, and audio data that is used to train chatbots? In a startling testament to the flexibility of deep learning, the approach to AI that has undergirded virtually all contemporary progress in the field, such combinations do, indeed, appear to work. The models seem to integrate knowledge of how to mechanically manipulate objects (learned from the robot action data) with the broader, ChatGPT-esque knowledge of the world.

In theory, then, general-purpose robots could potentially adapt to any environment without having to be explicitly programmed to do so, communicating about needed adaptations in human language. They could handle subtle changes in their environment (“the boxes are usually in one part of the factory, but today, they’re somewhere else”) in the same fluid way that a human worker does. Major AI players are taking note: After disbanding its robotics team in 2020, OpenAI has recently began pursuing this field again. Other companies, from Google DeepMind to startups such as Covariant and Physical Intelligence, are making serious investments as well.

Back to the Industrial Commons

These examples only scratch the surface of what AI can do for American manufacturing. What they have in common is the potential to deliver what the late scholar Clayton Christensen called “disruption.” Though the term is often used to mean “innovation,” it has a more precise meaning—new entrants to a market achieving lower costs through a fundamental technology or business process innovation. Chinese firms, built to employ large numbers of workers and optimized for the accompanying cost structure, may struggle to adapt. The United States, with its smaller manufacturing workforce, may be better positioned than China to deploy the next wave of automation.

Indeed, the American advantages might mirror those that have led to its leadership in the digital economy, including the development of AI. Imagine, for a moment, that the United States had leading AI development firms such as OpenAI, DeepMind, and Anthropic, but none of the data centers. All the brilliant machine learning innovations dreamed up by the scientists and engineers in the AI firms accomplish little without the so-called hyperscalers: Amazon Web Services, Microsoft Azure, Google Cloud, and similar operations. This computing infrastructure built over decades and with hundreds of billions of dollars enables startups and established companies alike to abstract away the enormous complexity of providing web services to millions of people. It is now as simple as clicking a button and paying a monthly bill.

Basic infrastructure of this kind is what American manufacturing needs most. A durable revitalization of manufacturing depends less on sexy semiconductor fabrication facilities or electric vehicle plants, and more on decidedly unsexy things like materials discovery and machine-tool manufacturing and maintenance. The successful startup stories of the past 15 years—Uber, Airbnb, and the like—were built upon the computing infrastructure of the hyperscalers. With the industrial groundwork we are laying today, a new generation of successes can be built—this time in the world of atoms rather than bits.

The author wishes to thank Austin Bishop, General Partner at Tamarack Global and Co-Founder of Atomic Industries, for his advice and feedback.