"Imagine a car factory where after a couple of cars roll off the line, a laptop emerges, then a bicycle, then a refrigerator." That is how Jeff Holden answered people who asked whether his new manufacturing system is just a 3D printer, according to Peter Diamandis, founder of XPRIZE and host of the Moonshots podcast. Holden's company, Atomic Machines, had just come out of stealth, meaning it had worked without public announcements, after six years.
On the episode, recorded October 13, 2026, Diamandis and his co-hosts used the launch to argue about a much bigger idea: machines that can build copies of themselves. They disagreed about how that would happen, not about whether it would.
What Atomic Machines says it has built
Diamandis, who called Holden a friend and said he is or was on his XPRIZE board, described Holden's career: Amazon's 10th engineer, the builder of Amazon Prime, later a senior vice president there, head of product at Groupon and chief product officer at Uber. Holden had always been fascinated by nanotechnology, Diamandis said, and had put his wealth and his team behind it. Diamandis said the company is backed by a quarter of a billion dollars from investors including Valor Equity.
Holden's stated mission, as Diamandis read it, is "on-demand universal command of matter." The first step is what the company calls the Matter Compiler. Diamandis described it as an AI-native manufacturing system that builds working micromachines from code alone, with no new tooling or process development for each product: "Different code builds different machines." He said it has made 3D machines with moving parts and several materials, with features down to single-digit microns. A micron is a thousandth of a millimeter. Diamandis noted that the name comes from Neal Stephenson's science-fiction novel The Diamond Age. Holden's goal "at maturity," Diamandis said, is to go "from prompt to product."
The company's own pages describe a narrower present. Atomic Machines describes itself as starting with microscale devices, with possible uses including tiny pumps, surgical tools, satellite thrusters and grippers that handle cells. It presents later systems working at smaller scales, and prompt-to-product manufacturing, as where it is heading. The Matter Compiler page describes a system that builds from digital designs, measures each finished device and feeds the results into the next attempt. A new product means changing software rather than making new masks (the patterned templates used in chipmaking) or retooling a line. The page illustrates its prompt-to-product aim as a three-step flow of imagining, designing and creating: the system designs, simulates and writes its own process, and the build-ready device is then approved and released to production. That is the company's illustration of the goal; the page does not say who performs the approval or how much of this runs today. One-shot operation from a single prompt is the company's stated aim, not a description of how it works today.
The first product: a very fast switch
The first product is PrimeSwitch, a relay, which is an electrically operated switch for cutting or connecting power. Diamandis described it as a power relay for 800-volt AI data centers that opens in 50 microseconds, "a thousand times faster than conventional approaches." The company's specification page confirms the 50-microsecond opening time, to 1,500 volts of isolation. It lists a continuous current of 150 amps and a package 9.5 millimeters across and 3 millimeters thick, and says samples are available. The relay is bistable, meaning it holds its position without drawing power. The company says the speed comes from a very light contact that travels only microns, and pitches the device as combining the low losses of metal-to-metal contact with switching speeds usually associated with semiconductors. The 800-volt data-center use and the thousand-fold comparison came from the discussion; the page does not state them in those terms.
Dave Blundin, founder and general partner of Link Ventures, reacted immediately: "Oh, my God. What an unlock for robotics and all physical business plans that would be." He said Elon Musk is "on the same path," with laying down individual atoms becoming a big part of the future. That would start with compute and data centers, Blundin said, and then spread to all physical devices.
"One big hairball"
Alexander Wissner-Gross, a computer scientist and founder of Reified, said he spent "a good chunk of my PhD trying to solve nanotech," without the benefit of today's AI. He thinks most of the world is "sleeping on" what he considers highly likely: that the technologies of science fiction are "basically one big hairball" and will arrive more or less together. He listed nanotech, LEV (longevity escape velocity, the hoped-for point where medicine adds life expectancy faster than people age), superintelligence and the AI-driven solving of math. "Because these are all so interrelated, they're all just going to get pulled out at the same time," he said.
What he most wants from nanotech is self-replicating machines. He nodded to Eric Drexler's Engines of Creation and its "gray goo" scenario, often painted as dystopian, and suggested "friendly smart dust" instead. "We haven't seen that yet," he said. Wissner-Gross said Neil Gershenfeld, who pioneered the fab-lab movement and was one of his research advisors at the MIT Media Lab, has been urging people to solve self-replicating machines and factories. He predicted that this would become "the equivalent of recursive self-improvement for the physical domain," comparing it to the idea of AI systems improving the research that builds better AI.
Gershenfeld's group has built real hardware toward this. A 2022 MIT account describes robots made from modular parts that assemble structures from identical building blocks called voxels, which can carry power and data as well as weight. The research also studies when it pays to build more robots, or bigger ones. The same account treats a fully autonomous system that builds structures and also reproduces its own robot workforce as a goal still ahead.
Two machines that make each other
Blundin offered a different picture of how replication might arrive, drawn from a tour of a Gigafactory that he called "a big eye opener." He had always imagined a self-replicating machine as one machine making a copy of itself. The alternative, which he said Musk is pursuing, is for one machine to make a manufacturing device that then makes a copy of the first machine. "It goes back and forth between the two, which is actually a lot easier than the exact self-replicating machine," he said, with "the same exact outcome."
His reasoning is that specialist machines are easier to build than generalist ones, and "a network of 10 or 20 specialist machines can make the next generalist machine." He pointed to the talk of humanoid robots building other humanoid robots. Touring humanoid makers such as 1X or Figure, he said, makes clear that "there's no way that that big robot hand is going to make these tiny little like surface mount devices," the small electronic components soldered onto circuit boards. The answer is that the factory already has a surface-mount machine. The humanoid only has to operate it, and that machine makes parts for the next humanoid. "That mode of operation is just totally imminent in my mind," Blundin said.
A factory the size of a cell
Wissner-Gross said both views could be true at once, and used biology to show how. A single cell, he said, can be viewed as "a bag of dedicated machines called proteins." Most proteins are specialists, yet together they let the cell copy itself. A factory, or a manufacturing city like Shenzhen, can be seen the same way. He said people he knows have tried to turn this into a quantitative scaling law: how big an industrial ecology, with all its people and machines, must be before it can copy itself. Today, he suggested, the answer might look like a Shenzhen.
He expects that size to shrink "relatively quickly." His forecast is that within five to ten years "at the outer bound," humanity will be able to produce a self-contained industrial ecology the size of a eukaryotic cell, the kind of cell that makes up plants and animals, that can copy itself.
The assembler dream
Diamandis traced the field to Richard Feynman's 1959 Caltech lecture "There's Plenty of Room at the Bottom" and to Drexler's 1986 book. Feynman, he said, asked why we couldn't arrange atoms precisely the way we want. Diamandis explained Drexler's assembler as a device that picks and places individual atoms. In his illustration, an assembler builds copies of itself from atoms in your skin cells. Then you drop one into the soil, ask for an electric Lamborghini, and it pulls atoms from the ground and builds the car atom by atom: "basically an anything factory." He compared it to an oak seed, which pulls atoms from soil and air using the sun's energy, only slowly. Nanotechnology's vision, he said, is to speed that up, and it "is going to fall on the back of superintelligence."
Wissner-Gross imagined how that might happen: someone, somewhere, will "one prompt" a request like "build me a Drexlerian nanorobot molecular assembler," presumably without running afoul of guardrails, "and it will just get designed."
Salim Ismail, founder of Open ExO, offered a low-tech version. About 15 years ago, he recalled, someone built an art project: a solar-powered cart with a magnifying glass that rolled through the desert. It scooped sand into a mold and focused sunlight to melt it into glass, leaving glass bowls behind as it went. Ismail called it "a self assembly machine that just kept going," and a visual of what is coming as technology reaches the molecular level. Wissner-Gross had one suggestion: "Let's do that to the moon."