Atomic Machines exits stealth with $250M and an AI fab that builds micro-machines from code

What did Atomic Machines announce on 7 October?

Atomic Machines announced itself on 7 October 2026 through a Business Wire press release and two essays by founder and chief executive Jeff Holden on the company blog. The release says the company spent six years in stealth building the Matter Compiler, which it describes as the first in a new class of AI-native, all-digital manufacturing systems that build working machines directly from code, and that it has raised $250 million to date.

Holden is a known name in technology. The release says he joined Amazon as its tenth engineer, built Amazon Prime and rose to senior vice president, then became Uber's first chief product officer and founded its self-driving, AI labs and air-taxi groups. Named investors are OneIM, Gigafund, XTX Ventures, KAS Venture Partners, Valor Equity Partners, Tru Arrow Partners, Construct Capital, Sozo Ventures and the Regents of the University of California. The company did not break the total into rounds or disclose a valuation.

Alongside the fab, Atomic Machines announced its first product, a power relay called PrimeSwitch. The release says it will be shown publicly for the first time at the Open Compute Project Global Summit in San Jose from 12 to 15 October, the main annual gathering for data-centre hardware designers.

The announcement lands in a busy week for AI infrastructure; for context on how chip and system limits shape AI, see our explainer on what actually matters in AI chips, and for another large early-stage bet this month, our report on TypeSafe's $870M Series A.

How is the Matter Compiler supposed to work?

The company's pitch starts with a gap in manufacturing. Conventional processes such as casting, molding and machining run out of precision in the low tens of microns. Below that, the dominant industrial capability is semiconductor fabrication, which is superb at patterning flat electronic circuits in silicon but poor at building three-dimensional devices from many materials with moving parts. Holden argues that the MEMS industry, which makes accelerometers, gyroscopes, micro-mirrors and inkjet nozzles, has managed only about a dozen commercial device categories because each one needs years of custom process development, captured in the field's maxim 'one product, one process'.

The Matter Compiler is meant to break that rule. Atomic Machines describes it as a general-purpose constructor: a single system in which code alone determines which device gets built, producing complete, functional 3D machines with moving parts, multiple materials and assembled subsystems, with features down to single-digit microns and no hard tooling such as masks, molds or fixtures. A new device is supposed to be a change in software, not a new production line.

The second claim is that the factory sits inside an AI loop. According to the release, every operation is measured with high-precision metrology and corrected in real time, and the AI runs its own experiments on the hardware to test what simulation proposes. The company's product page describes the workflow as a prompt describing what a device should do, a design and simulation stage in which the system writes its own process, and then a physical build, with the stated aim of 'one-shot prompt-to-product'. It says design cycles that took years can shrink to weeks or days.

Holden frames this with an analogy to biology. In his origin-story essay he contrasts human manufacturing, where each new artifact needs its own mold, mask or retooled line, with the ribosome, which builds endless proteins from one fixed machine and a stream of instructions. He says the Matter Compiler has three properties that make that possible: it is all-digital with zero hard tooling, its operations are composable primitives that take design intent as input, and every operation is measured so that errors are corrected rather than compounding. Investor Luke Nosek of Gigafund called the idea 'the start of vibe manufacturing'.

What is PrimeSwitch and why start with a relay?

PrimeSwitch is an electromechanical relay, a switch in which a physical metal contact opens and closes. Atomic Machines says it picked this device six years ago because power systems face an awkward trade. Mechanical relays and contactors conduct almost without loss through metal-to-metal contact, but they take milliseconds to open and, at high power, open into an arc. Solid-state switches open in microseconds but conduct through semiconductors that waste energy as heat for as long as current flows, and they cannot provide galvanic isolation, a true physical gap in the circuit.

The company says AI data centres have made that gap acute as they move to 800-volt DC power distribution to feed denser racks. At those voltages, the release says, a fault can deliver destructive energy within microseconds and form an arc that direct current will not put out by itself. Its proposed answer is a hybrid breaker in which a small solid-state bypass carries current for the few microseconds while the PrimeSwitch contact opens, so that silicon interrupts while metal carries and isolates the current.

The trick, according to Atomic Machines, is scale: a contact with almost no mass that moves only microns can open far faster than a hand-sized mechanism. The company says semiconductor fabs cannot process PrimeSwitch's materials or assemble its moving parts, and that it exists only because of the Matter Compiler. That claim is central to the business case, and outsiders cannot yet check it.

Who can get it, and when?

Not many people, yet. The release says PrimeSwitch PS-150 is shipping to early-access customers for evaluation, and the product page lists availability as 'sampling now'. There is no published price, datasheet download, volume production date or named customer. The Matter Compiler itself is not offered as a service; the company's site invites people to get in touch about micro-machine ideas, partnerships and press, and is recruiting.

For data-centre engineers, the practical next step is the OCP Global Summit, where the relay will be shown in public for the first time. That will be the first chance for people outside the company and its customers to see the device and ask about test conditions, lifetime, reliability under repeated faults and certification, none of which the launch materials cover in detail.

Why does this matter for AI?

There are two links to AI. The first is the customer. AI training and inference clusters are pushing rack power so high that operators are redesigning power distribution, and protection hardware is one of the parts that has to keep up. If PrimeSwitch performs as claimed, it could shrink or remove some of the silicon, cooling and separate contactors that hybrid protection designs need today. That is a niche, but a lucrative and fast-growing one.

The second link is more ambitious. Atomic Machines describes the Matter Compiler's instruction set as symbolic, which it says is 'AI's native medium', the same kind of data generative models read and write. Holden's stated destination is manufacturing that feels like working with an AI coding agent, 'except that some of the output tokens are physical'. That fits a wider push to connect AI models to the physical world, from robotics to lab automation, which we have tracked in stories such as AWS's open-source Physical AI Toolchain. The company lists future targets including motors and gears the size of a grain of sand, robots small enough to work inside the body and cooling built directly into AI chips.

What is still unproven?

Almost everything rests on company statements. The PrimeSwitch numbers come from Atomic Machines' own product page and press release; we found no independent test data, peer-reviewed paper or third-party teardown. The 'about a thousand times faster than a conventional contactor' comparison depends on which contactor is used as the baseline, which the release does not specify.

The bigger claim, that one tool-free system can build arbitrary micro-machines across mechanical, fluidic, thermal, optical and chemical domains, is so far demonstrated by a single product. The company has not published yields, throughput, unit costs or how many distinct device types the Matter Compiler has actually built. Its marketing describes 'prompt-to-product' as an aim, not a current capability, and readers should keep that distinction. Six years and $250 million buy a lot of engineering, but scaling a new manufacturing process is where many hardware start-ups stall, and the OCP showing and early customer evaluations will be the first real evidence either way.