Every manufactured object is the result of materials, machinery, people, and places working together in systems that have been refined over generations. These systems determine what we can make, how quickly we can make it, and what it costs.
AI gives us new ways to improve these systems. Realizing its full potential means building on generations of manufacturing knowledge and experience, both to improve existing factories and to create new technologies, companies, and factories that show what’s truly possible.
I sat down with John Hart the other week to talk about this. John is a professor at MIT and head of MIT’s Department of Mechanical Engineering. He also co-leads MIT’s Initiative for New Manufacturing. John has also helped found and advise several manufacturing startups, working on everything from technical strategy to team building. He describes his work simply as “trying to make things better at scale.”
In our conversation, we discussed what it means as AI gives manufacturers new capabilities: how to connect those tools to production, learn from what gets made, and build on the factories, suppliers, and expertise already around us.
As John put it: “Manufacturing is about technology, people and organizations.”
Absorbing AI’s full potential
“We can build faster than we were able to build before,” John told me early in our conversation.
Almost every step in the manufacturing process has been, or can be, impacted by AI. This includes everything from planning the process and understanding the costs, to programming the equipment and integrating it with software. And as these individual tasks improve, there’s a lot more room to rethink how a factory operates.
John’s view on what that should look like has shifted. He used to think that automated design and manufacturing might mean one giant model where you could ask it to machine a part and it would handle everything. “Now I think more about how the workflow can be decomposed into the different operations represented by tailored models and agents, and orchestrated at scale,” he said.
For example, if you’re making an injection mold for a plastic part, someone has to design the part, design the mold, simulate the process, and generate the toolpaths that tell a CNC machine how to cut it. Each of these steps requires different tools and different expertise. Breaking up the workflow gives AI more clearly defined tasks and makes its performance easier to evaluate.
“I think good-enough tools along with enough guardrails and prompting for each of those steps already exist,” John said. “But I don’t know if we’re fully absorbing that potential yet.” Connecting capabilities to processes requires a lot of engineering work. “One challenge,” John said, “is understanding how these new tools can fit into existing manufacturing workflows, to drive adoption that allows the truly breakthrough opportunities to unlock.”
And the measure of success is whether improvements in individual tasks translate into better parts, lower costs, or faster production at scale.
Capturing value in manufacturing
Being able to make something faster or cheaper creates an opportunity. But how a company then captures the value of that opportunity depends on what it sells.
I asked John whether selling the output of a manufacturing process is generally a more attractive business than selling the machines themselves. “Maybe. It depends on the industry and the use case,” he said.
John pointed to producers of high-quality forged materials for aerospace and medical applications as an example of how specialized expertise can support attractive margins. “That’s the narrowest point in the supply chain,” he said. “That’s where the specialized knowledge is.”
A company can capture value through both specialized expertise in individual processes and the ability to combine them to deliver a finished component.
VulcanForms is one example of a company that brought multiple processes in-house to deliver finished components to its customers. When John and Martin Feldman started the company, they developed an industrial metal printing technology with the goal that “our business model would be to sell the parts, not to sell the machines.” They brought CNC machining and heat treatment in-house because printing alone didn’t deliver the finished component their customers needed.
“For a lot of customers, we need to own even more of the value chain to be able to deliver them a product that then they can integrate and sell,” John said.
Learning from what actually gets made
This is one reason I find vertical integration in manufacturing so exciting. Owning more of the process can make it faster and easier to establish a feedback loop between design and production. After all, a design expresses what should happen and production shows what actually happened. This means you can identify more precisely what needs to change and then more easily feed those findings into the next design or production run.
John is excited about how collecting data will let us “learn so much more about the relations between design and manufacturing.” He sees particular opportunity in capturing the knowledge generated during the difficult transition from a successful prototype to reliable production.
“How do we accelerate the process from prototype to scale through getting data from the systems and people and processes that have to go through that extreme pain to bring something to full-scale production?” he asked.
But owning the process doesn’t automatically create that learning loop. It’s really about how you get the data, and this needs to be a focus of building the operation from the start.
“This is what enables you to be able to differentiate on product performance while understanding manufacturing so deeply – you can always win on cost and quality,” John said.
Rebuilding the base
A company can’t reach this goal alone. It’s also dependent on the suppliers around it. Even the most highly integrated manufacturers rely on other factories. In fact, it’s the network of these factories that determines what a vertically integrated manufacturer can make and how reliably it can deliver.
John visited factories around Shenzhen in January and was struck by the connectivity of the ecosystem and how extensively automation was being deployed. “In all the conversations I had, there was never the comment, ‘Oh, we’re not automating that yet because it’s too expensive,’” he told me. “Of course, that happens, but I didn’t hear it.” Instead, he saw that if companies weren’t automating something yet, it was because the capability wasn’t there yet.
“The density of what I would say is true innovation was really, really high… and the experience was both daunting and inspiring,” John said. He saw ingredients of a more automated, flexible manufacturing economy that he believes can succeed in the US. Developing those capabilities here will take investment in the entire ecosystem.
As I wrote about here, manufacturing has a tip and a base. The tip includes factories in aerospace, defense, and automotive industries. The base consists of the smaller suppliers, fabricators, and distributors.
Much of where I see the excitement and venture dollars going these days is concentrated at the tip, but there is a lot of untapped opportunity to improve the base, the suppliers that enable the tip.
“Those companies are also sources of innovation and opportunity,” John said. Supporting them requires several things to come together, including talent, which is in part why the Initiative for New Manufacturing at MIT has a technologist training program (TechAMP). It’s designed to help level up experienced technicians.
“This is a necessary but not sufficient ingredient,” John said. “The factory also needs to be able to go get the business to be able to justify the investment.” Training, financing, and customer demand have to reinforce one another.
When I asked whether all manufacturing capability should be brought back to the US, he immediately said, “Oh, no. Definitely not. And it never will be or could be economically.” Instead, it’s about building companies and supply chains that balance cost, resilience, automation, and human work.
That will take time. “It’s a long game,” John told me. “We’re talking about a multidecade rebirth.”
What “new manufacturing” means
That broader ambition helps explain MIT’s Initiative for New Manufacturing. Their definition of “new” includes improving the factories and processes already operating around us. “So much manufacturing is established, and the beauty is figuring out how to upgrade and refine existing manufacturing practice,” John said.
AI gives us new ways to build on that existing knowledge. We can learn more from production and use those learnings to inform designs. We can make better use of the expertise and talent pools currently available. We can build denser supplier networks to improve price competitiveness and lead times in the US.
But doing all of this will take time, investment, and changes to how factories operate.
Toward the end of our conversation, John told me that he tells all of his students to “look around and imagine all the things that have been in factories.” Every object reflects a chain of decisions around materials and processes that were made by a group of experienced people over years or decades.
The things we make next will draw on generations of knowledge about materials, machinery, and production. The promise of “new” manufacturing is to build on that experience, giving existing factories new ways to improve what they do, and to build the new factories and companies that define the future.
What we invest in now will determine how much new capability we can create from what we already have.
Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.
Part design, mold design, simulation and toolpath generation are different skills, and treating injection molding as one task hides where the engineering work sits: in connecting the steps, which is the part John says the field isn't yet fully absorbing.
That connecting work is where the evidence runs thin in quality inspection right now. Siemens and NVIDIA's claim of up to 85 percent less engineering effort at the Siemens Erlangen factory turns out to be a project target, and an analyst cautions it has not been proven at mixed-vendor sites. John's point that good-enough tools already exist for each decomposed step is probably right; what's missing is evidence that orchestrating them across a real, messy factory floor delivers what the vendor slide promises.