Atomic Machines wants to build tiny machines straight from code, and its first switch opens 1000x faster

For two years the big promise of AI has been that you describe what you want and software appears. Write a sentence, get an app. People call it vibe coding. What nobody has managed is the same trick for physical things. You still cannot describe a tiny motor in plain words and get a working one out of a machine a few days later.

A startup called Atomic Machines says it is now building exactly that, at least for very small devices. This week the company came out of six years of stealth with 250 million dollars raised to date, a manufacturing system it calls the Matter Compiler, and a first product with a name that sounds like a streaming plan: PrimeSwitch. The New York Times profiled the company under the headline that it wants to build "microrobots" with AI. The real story is a little less sci fi and a lot more interesting, because the first thing it builds is aimed straight at one of the hardest engineering problems inside today's AI data centers.

Who is behind it

The founder is Jeff Holden, a name that will ring a bell for people who follow tech history. He joined Amazon as its tenth engineer, built Amazon Prime and rose to senior vice president. Later he became Uber's first chief product officer and started the company's self driving, AI labs and air taxi efforts. In other words, he has built very large systems before, both in software and in hardware heavy moonshots.

The investor list is long and fairly serious: Gigafund, OneIM, XTX Ventures, KAS Venture Partners, Valor Equity Partners, Tru Arrow Partners, Construct Capital, Sozo Ventures and even the Regents of the University of California. The company is based in the San Francisco Bay Area with offices in Emeryville and Santa Clara. It did not say whether the 250 million dollars came in one round or several, and it did not share a valuation.

The problem: the world cannot build small machines

To see why this matters, it helps to know where today's factories stop. Classic manufacturing such as casting, molding and machining works well down to a few tens of microns. A human hair is roughly 70 microns thick, so that is already small, but not small enough for a whole class of devices people dream about: motors and gears the size of a grain of sand, robots that could work inside the body, or cooling channels built directly into the chips that run AI.

Below that size, the world's main tool is the semiconductor fab. Chip fabs are astonishing, but they were built to do one thing: pattern flat electronic circuits in silicon. Engineers have squeezed some moving parts out of that toolset. The field is called MEMS, and it gave us the accelerometers and gyroscopes in every phone, plus the tiny mirrors in some projectors. But MEMS has a famous rule of thumb: one product, one process. Every new kind of device needs years of custom process development before it can be made reliably. That is why, after decades, there are only about a dozen big commercial MEMS device types.

Atomic Machines argues that tiny machines have been blocked not by physics but by manufacturing. If you could make a new device simply by changing the instructions, the bottleneck would disappear.

What the Matter Compiler does

The company describes two breakthroughs. The first is that the Matter Compiler is a general purpose constructor. One system builds complete three dimensional machines with moving parts, several materials and assembled subsystems, with features down to single digit microns. There is no hard tooling: no masks, no molds, no product specific production line. Code alone decides what comes out.

The second is that the whole factory sits inside an AI loop. Every operation is measured with high precision metrology, errors are corrected in real time, and the AI runs its own experiments on its own hardware to check what simulations propose. Each build teaches the system something it carries into the next one.

Holden compares it to 3D printing. 3D printing, he says, made shapes programmable: any geometry straight from a file. The Matter Compiler is meant to make machines programmable: many materials, moving parts, sealed and wired, straight from a file. "When the marginal cost of a new design is the cost of the design file, you get an innovation explosion," he said in the announcement. Luke Nosek of Gigafund went further and called it the start of "vibe manufacturing".

The name is a nod to Neal Stephenson's novel The Diamond Age, in which matter compilers assemble objects atom by atom from digital plans. Atomic Machines is not working at the atomic scale, of course. But the ambition is clearly similar: the company says its long term aim is one shot "prompt to product".

There is also an economic argument. Normally flexibility and low cost fight each other. Cheap parts come from committing to a fixed process and expensive tooling, and every design change reopens that bill. If a factory needs neither, then making one unit carries no extra engineering charge, and a fab that builds one design can build any other. High mix and high volume could, in theory, happen on the same machines.

The first product: a switch for AI power

All of that would be just a nice slide deck without a real device. That device is PrimeSwitch PS-150, and the choice is revealing. It is not a medical microrobot or a toy gear. It is a power relay, a switch that cuts electricity, designed for the new 800 volt DC power systems that AI data centers are moving to.

Why does that matter? Modern AI racks draw enormous amounts of power, and the industry is shifting to higher voltage direct current to feed them efficiently. The trouble starts when something goes wrong. At 800 volts, a fault can release destructive energy within microseconds, and direct current creates an arc that does not put itself out the way alternating current often does.

Today engineers have two imperfect tools. Solid state switches built from semiconductors are fast and do not arc, but they waste energy as heat while current flows through them, and they cannot create a true physical gap in the circuit. Mechanical contactors conduct almost without loss and open a real gap, but they need milliseconds to move, which is far too slow at these voltages, and at high power they open into an arc of their own. So many designs end up using both, with the cost, cooling and space that implies.

PrimeSwitch tries to close that gap. Because its moving contact has almost no mass and only travels a few microns, it can open in 50 microseconds. Atomic Machines says that is about a thousand times faster than a conventional contactor. The company's own phrase is that PrimeSwitch "isn't a relay made small, it's a relay made fast, because it's small".

The listed specifications are impressive on paper. The PS-150 carries 150 amps continuously, has an on resistance of just 200 micro ohms, provides 1,500 volts of galvanic isolation and holds its state with zero power, because it is bistable. All of that fits in a sealed surface mount package 9.5 millimeters across and 3 millimeters thick, about the size of a shirt button.

In practice the switch would be used in a hybrid breaker next to a small solid state bypass. For the few microseconds while the metal contact is opening, the silicon carries the current. Then the contact is open and the circuit is truly isolated. In Atomic Machines' words: the silicon interrupts, the metal carries and isolates, and the breaker never arcs.

The company also claims that this device can only be made in its own fab. Semiconductor fabs cannot process PrimeSwitch's materials or assemble its moving parts, and conventional tools cannot reach its feature sizes. That is the core of the pitch: the first product exists because the Matter Compiler exists.

Where things stand

PrimeSwitch is shipping to unnamed early access customers for evaluation. It will be shown in public for the first time at the Open Compute Project Global Summit in San Jose from October 12 to 15, which is exactly where data center hardware buyers gather. Atomic Machines has not published a price or a production volume.

What to watch, and what to doubt

It is worth being clear about what we know. Every number in this story comes from Atomic Machines itself. Nobody outside the company has published independent test results for PrimeSwitch yet, and early access evaluation is not the same as mass production. A single relay, however clever, also does not prove that the Matter Compiler can really build "any" micro machine from code. MEMS history is full of brilliant first devices that took years to scale.

There is also a quiet tension in the pitch. The grand vision is microrobots, tiny motors and chip cooling. The first product is a sensible, sellable component for a market with huge demand right now. That is probably smart business. It also means the more exciting promises are still promises.

Still, the timing is striking. AI is eating electricity at a pace that forces data centers to rethink their entire power design, and better protection for 800 volt racks is on the industry's own roadmaps. If a startup can deliver a switch that combines the speed of silicon with the efficiency and isolation of metal, that alone would be a meaningful business.

And if the bigger idea works, it could matter far beyond data centers. Software got its explosion when the cost of trying a new idea fell to almost zero. Atomic Machines is betting that the same can happen for physical devices, as long as they are small enough. The next few months, starting with the first public demo in San Jose and the first customer feedback, will show whether vibe manufacturing is a real shift or just a very good slogan.

Sources

• Atomic Machines announcement via ManufacturingTomorrow: Atomic Machines emerges from stealth with 250 million dollars to launch the Matter Compiler
• Atomic Machines: The Matter Compiler
• Atomic Machines: PrimeSwitch
• The New York Times: The start up that wants to build microrobots with AI
• Mercury News: The startup that wants to build microrobots with AI