.png)
Look at almost any object around you and there are a dozen ways it could have been made. It could be milled, molded, cast, stamped, printed or welded, out of metal, plastic, glass or ceramic. Above roughly 50 microns, a bit thinner than a human hair, manufacturing is a deep toolbox.
At the scale of a few microns, the options narrow considerably. The dominant method is the semiconductor process. The sequence is still built on the same four steps it started with — deposit a layer, pattern it, etch it, flatten it and repeat.
That process made modern computing possible. But it was built to make transistors on silicon wafers. Almost everything else that needs to be very small has had to squeeze itself into that mold, or it simply doesn't get made.
We started Construct in 2020 to back founders transforming the foundational industries of the economy: manufacturing, logistics, energy and critical infrastructure. Manufacturing and production have always been central to that thesis. We look for companies that change how the work gets done.
Today, after six years in stealth, Atomic Machines emerged with $250 million in funding to launch the Matter Compiler™. We're proud to be among the first investors backing the company because Atomic Machines changes not just how things are made, but what can be made in the first place.
Think about devices that depend on parts at the scale of a few microns: micromotors and micropumps, microfluidic chips that run a lab test on a drop of blood, millimeter-sized robots, implants that combine electronics, drug reservoirs and sensors, and optics with real curved surfaces positioned at precise angles.
Many of these have been shown in research labs for decades. Few have reached the market, because there is no good way to build them. They need materials that standard silicon processes cannot accommodate, true three-dimensional shapes and parts made separately and then assembled. A wafer process offers none of that. So each product needs its own bespoke process development, and many never make it out of the lab.
Atomic Machines built the Matter Compiler, an AI-native, all-digital manufacturing system that builds working micromachines directly from code. The system is made of Nodes: automated digital machines, each running one class of manufacturing process, all working together. Lasers machine parts. Fluids get deposited in picoliter drops. Electrochemical steps plate and polish. Robots move work between Nodes and assemble components. The materials library spans metals, polymers, glasses and ceramics, none of which depend on a silicon wafer.
The Matter Compiler goes beyond making a shape from a file and instead combines processes and materials to build three-dimensional machines with moving parts. In the company's words, bits and raw materials go in, and complete, functional micromachines come out.
The name Matter Compiler is deliberate. A software compiler turns code into instructions a chip carries out the same way every time. A Matter Compiler has a much harder job. At the micron scale, every physical step is noisy, and its result depends on everything that came before. Writing rules for every possible variation is difficult. Atomic Machines uses a learned world model of how each manufacturing step changes the material to turn a design into a working part.
Atomic Machines built its fab around that world model. Every operation is measured before and after it runs. A part that drifts out of spec is caught at that step and corrected, instead of moving on to ruin the steps after it. And every device the fab makes becomes labeled training data for reinforcement learning. The model learns from the hardware's results and uses that feedback to guide manufacturing.
This is what makes Atomic Machines a physical AI company in a way few companies are. Designing and building a new micro-device today often takes years of design, test runs and revisions. When AI can design a part, build it and measure the result without a human in the loop, that cycle shrinks to days. Over time we believe that could make it easier to turn a plain description of what you want into a physical thing you can hold.
A platform company needs a product that proves the platform. Atomic Machines picked one aimed at the biggest constraint in AI right now, which is power.
AI data centers are moving to 800-volt direct current, because conventional rack power can't keep up with megawatt-scale racks. Higher voltage at higher current raises a hard safety question: how do you cut off a fault before it does damage? Conventional electromechanical relays carry current with very little loss, but they open far too slowly. Solid-state switches are fast, but they waste energy as heat. Engineers have long known that a hybrid breaker, combining the two, would get the best of both. What was missing was a mechanical relay fast enough to make it work.
PrimeSwitch is that relay. It is an ultra-fast, ultra-low-resistance power relay that makes the hybrid breaker practical. The PrimeSwitch PS-150 opens in 50 microseconds with 200 micro-ohms of on-resistance, carries 150 amps continuously, and fits in a 9.5 mm × 3 mm hermetically sealed package. It is shipping to early-access customers today.
PrimeSwitch could only have been made on the Matter Compiler. Building it also gave the platform its first body of training data, so every device that comes next starts from what PrimeSwitch taught it.
A company like this needs a founder who has built systems at scale and who isn't intimidated by a problem with no playbook. Jeff Holden, Atomic Machines' founder and CEO, has done that three times.
After computer science degrees at the University of Illinois, Jeff joined D. E. Shaw in New York, where his first manager was Jeff Bezos. Five years later he followed Bezos to Amazon as engineer number ten. He inherited a modem and a shell script and built them into Amazon's global supply chain and fulfillment technology. He went on to join Amazon's senior leadership team, the S-Team, and led the development of Amazon Prime.
He then founded Pelago, which Groupon acquired, and became Groupon's SVP of Product through its IPO. At Uber, where he was Chief Product Officer, his first project was letting riders enter a destination in the app. He went on to start Uber's self-driving program, Uber Elevate, AI Labs and UberPool. Our co-founder Rachel Holt worked alongside Jeff at Uber. She saw firsthand how he takes an idea most people consider impossible and turns it into something that runs every day, at enormous scale.
Jeff has assembled a team to match the ambition. Atomic Machines brings together AI researchers, physicists, MEMS and process engineers, roboticists and manufacturing leaders under one roof. Building a learning factory takes all of those disciplines working as one.
The semiconductor industry offers a useful precedent for what could come next. In early 1988, about a year after TSMC opened its doors as the world's first dedicated chip foundry, its CEO pitched the idea to investors. Their big question was whether anyone needed all that capacity. Today, chip designers that own no fabs at all sit among the largest companies in the world, and the semiconductor market reached $796 billion last year.
We think micromachines are standing where chips stood in 1987. The MEMS industry, which uses many of the same tools as chipmaking, was $15.4 billion in 2024 and is forecast to grow less than 4% a year through 2030. Microfluidics adds roughly $25 billion more. We believe those numbers reflect, in part, how difficult these devices have been to manufacture.
A general-purpose way to manufacture at the micron scale could make these categories much larger. Over the long run, we believe the ability to command matter at the microscale will be worth as much to the world as the chip itself. The goal is to make producing a new micromachine as simple as specifying a design and letting the manufacturing system work out how to build it.
People will look back at 2026 the way the chip industry looks back at 1987. Atomic Machines is a turning point for micromachines, as much as the foundry model was for chips. Atomic Machines spent six years in stealth building the factory. Now the world gets to see what comes out of it. We are proud to back Jeff and the Atomic Machines team alongside a remarkable group of co-investors.