Atomic Machines Matter Compiler: AI-Native Fab for Micro-Machines and the PrimeSwitch Relay

On October 7, 2026, a company called Atomic Machines left stealth mode with a bold claim: its Matter Compiler produces "functional micro machines using only computer code," with no per-product tooling. The New York Times ran a profile by Cade Metz the next day, and the news spread through Techmeme and social feeds. A skeptical reaction followed, because the company has shown no public video of the machine at work.

This post reads the company's own pages, which are the primary sources, and separates what Atomic Machines says from what anyone has verified. It also explains why a relay is the first product, and why the idea of "prompt-to-product" manufacturing matters to people who work with AI.

TL;DR: Atomic Machines at a glance

What is the Matter Compiler?

The company's Matter Compiler page calls it "an AI-native fab that turns digital designs into working micro-machines. And learns with every build." The page says it "designs, builds, measures. And learns."

Three properties define it in the founder's essays.

• All-digital with zero hard tooling. No molds, no masks, no fixtures. The machine runs on symbolic instructions, so the marginal cost of a new device design is "the cost of writing it down."
• Composable primitives. Its operations take design intent as input, instead of recipes rediscovered by trial and error for each part.
• Measured operations. Every operation is measured, so errors are caught and corrected before they compound. The founder says this is what lets construction programs "run deep."

The company's pitch is a workflow with three steps: imagine, design, create. A prompt starts it. The system "designs, simulates, and writes its own process," and produces a build-ready device with validated CAD, a validated process graph, optimized time and cost, and met performance targets. A human approves and releases it to production. The page says this collapses cycles that "took years" into "weeks, or days." The stated aim is "one-shot Prompt-to-Product."

How is this different from a chip fab or MEMS?

Founder Jeff Holden's essay The Micro-machine Universe makes the argument. Semiconductor fabrication, he writes, is "hyper-optimized for one cell" of a grid of machine types: the electrical domain with an information purpose, meaning transistors. The grid has seven physical domains: mechanical, fluidic, thermal, electromagnetic and optical, charged-particle, chemical and electrical. It has two purposes: delivering energy and matter, or encoding information.

MEMS, the micro-electro-mechanical systems field, has spent forty years coaxing chip fabs into building small machines. The results include phone accelerometers, micro-mirrors, inkjet nozzles, microphones and thermal cameras. Holden calls these "genuine marvels" but shallow. Each is "a quasi-planar structure in the fab's own materials," with its own custom process. The field's maxim, he writes, is "one product, one process."

The Matter Compiler is meant to break that rule. The company says it is multi-process and multi-material, and it builds true 3D geometries from materials that fabs do not handle, such as magnetic cores and metals and alloys used in full-size machines. The company's grid marks most of the interior of the map as "open," meaning nothing exists there at microscale today.

It is a large claim. The honest comparison is with 3D printing, where one process makes one kind of part. The company says its system is the multi-material version, and that "bits and raw materials go in and complete, functional micromachines come out." Whether the process can hold yield at volume is not shown.

Why is the first product a relay?

PrimeSwitch is a power relay, and the company explains the choice with an old trade-off. Electromechanical relays conduct metal to metal, the most efficient path, but open in milliseconds. Solid-state switches open in microseconds but conduct through semiconductors, with losses and heat. The page says: "Pick your poison: Fast or Efficient."

The PS-150 spec sheet claims both:

The speed, the company says, comes from the scale: "A contact with almost no mass, traveling microns, can open in microseconds." In the founder's words, PrimeSwitch "isn't a relay made small. It's a relay made fast, because it's small." He compares the opening time to a conventional contactor and says it is about a thousand times faster.

Credit: Atomic Machines product imagery.

The timing ties to AI infrastructure. Holden writes that "the move to 800-volt DC in AI data centers has made it acute." A fast, efficient switch protects a high-power bus without a constant heat penalty. That is a real pain point, and explainx.ai has tracked its power side in AI energy math for model selection and data center power pricing. The claim here is the company's own. No customer has been named in the sources we could read.

What does "prompt-to-product" mean?

Holden argues that the Matter Compiler's language is symbolic, which makes it "AI's native medium: the same data type generative models read and write." The destination, he writes, is "an experience like working with a modern AI coding agent, except that some of the output tokens are physical."

This is the part that matters for readers who build with AI. Coding agents work because code is text, the build step is fast, and tests give feedback. If a fab can accept a symbolic design and return measured results quickly, an agent loop becomes possible for hardware. That loop would be design, simulate, build, measure, revise, much like the loops described in explainx.ai's loop engineering guide.

The same pattern appears in science. The Genesis Mission pledges and self-driving labs pair models with physical systems that give "granular, closed-loop control" and feedback. The company's own company page says exactly this: generative design "needs more than a model: it needs physical systems that give the AI granular, closed-loop control over the manufacturing process and feedback from it."

What is the origin story?

In the origin story post, Holden writes that civilizational progress rests on three axes: the command of energy, of information and of matter. He argues matter is the one the others run on. His benchmark is biology's ribosome, which reads a digital string and builds a working molecular machine. In human manufacturing, he says, "the design lives in the hardware": a mold, a mask, a die, a retooled line. In the ribosome's system, "the design lives in the data."

He says Atomic Machines spent six years in stealth building "the first manufacturing system in history with the same architecture as biology's: a combinatorial constructor whose unit of output is the working machine." The name nods to Neal Stephenson's The Diamond Age. The company's stated mission is "on-demand universal command of matter."

The company page lists possible micro-scale uses: medical implants that shrink from the size of a hockey puck to a grain of rice, surgical micro-robots, diagnostic chips that leave the lab bench, and processors cooled by pumps inside the chip package. These are aspirations, not products.

What are people saying?

The NYT profile, as summarized by Techmeme, describes Emeryville-based Atomic Machines as "training AI on materials and designs to revamp building microelectromechanical systems (MEMS)." The Times article itself is paywalled, so this post relies on Techmeme's one-line summary for that framing.

What developers are saying

A Hacker News thread on the company's site had 6 points at the time of research and a short discussion. One commenter wrote: "Well, that completes the grey goo bingo card." A user who says they work at the company replied: "No grey goo, but lots of robots. My mental model for the long run goal is the Feed from Diamond Age." Those are unverified comments from individual users.

HuggingNews reported skepticism on social media about the missing demo. One commentator joked that a "demo making piles of paper clips" would have been an effective marketing move. That reaction reflects a real gap: the company gives specs but no footage of the fab running.

What is not shown?

The limits are significant.

• No independent data. PrimeSwitch specs are company claims. Contact resistance, cycle life, contact welding and failure modes are not published in the pages we could read.
• No fab demo. The sources include product renders and diagrams. They do not include video of the Matter Compiler producing a device.
• No yield or cost numbers. The economic claim that marginal cost falls toward materials is a thesis. No unit price, volume or yield is given.
• No funding or customer disclosure. Investors, headcount and named customers are absent from the pages we read.
• "Sampling now" is not shipping. It means early units may be available to some buyers.
• The AI role is described, not shown. The company says the system designs, simulates and writes its own process. It does not say which models it uses or how much a human engineer does.

A relay is also a modest first product. It is a good test, because specs like on-resistance and opening time are measurable by any lab. If outside engineers test PS-150 samples and confirm 200 microohm and 50 microseconds, the claim about the process becomes much more credible.

What should builders and learners watch?

1. Third-party measurements of PrimeSwitch samples.
2. A fab demo or a customer reference.
3. The second device. One product shows a capability. A second in a different physical domain would show the "language" the founder describes.
4. Design tooling. If the symbolic language is exposed to AI agents, it could become a new target for code-generation models.

For a wider view of how AI is reaching into physical design and making, read Figure F.02's trip through an arc furnace and explainx.ai's coverage of NVIDIA's N1X chip, where chip supply constraints meet AI demand. Learners who want to build agent loops can look at explainx.ai workshops.

What people are asking

Is this nanotechnology? No. The company says it starts at the micro scale, features of microns, and plans later editions "as we descend the feature length scale." It talks about nano and atomic scale as future goals.

Can I buy PrimeSwitch? The product page says "Sampling Now." Contact the company for terms.

Is it a 3D printer? The company says no: a 3D printer typically makes one material with one process. It says the Matter Compiler is a multi-process, multi-material system.

Related reading

• Model selection and AI energy math
• Virginia data center power tax and AI pricing
• Data center water lawsuits
• Genesis Mission: who pledged what
• Loop engineering for coding agents
• Figure F.02 and the arc furnace
• NVIDIA N1X Arm chip

Primary sources: Atomic Machines home; Matter Compiler; PrimeSwitch; The Micro-machine Universe; Origin story; Company; NYT profile via Techmeme.

Specifications and claims reflect Atomic Machines' public pages as of October 9, 2026. They are company statements and may change.