Atomic Machines Turns Code Into Working Machines. A New Era, or a Very Fast Switch?

01/ The short answer

The short answer: not an ASML killer, and not yet an era

Three questions follow this launch around, and they deserve straight answers before the story begins. Does it threaten ASML? Is it a new era of manufacturing? And is this where the robot age really starts?

What Atomic Machines is really going after is older, and in its way more interesting, than a lithography monopoly. It is the rule that has made new kinds of tiny machines slow and expensive to bring to market for forty years: each one needs its own manufacturing process. If that rule breaks, the people who should worry make sensors, relays and circuit breakers, not chip scanners. The rest of this piece is about whether it has broken, and what the very first product tells us about the bet.

02/ The founder

The man who led Amazon Prime went quiet for six years

Jeff Holden has a habit of building things that people stop noticing because they simply work. He wrote trading software for Jeff Bezos at D. E. Shaw in the early 1990s, joined Bezos at Amazon in 1997, and rose to senior vice president in charge of the consumer website. His alma mater, the University of Illinois, credits him with leading the development of Amazon Prime[9], the subscription that taught a generation to treat free two-day delivery as a birthright.

Then came Pelago, a city-discovery start-up, a spell as Groupon's senior vice president of product, and in 2014 the job that made his name in Silicon Valley: chief product officer at Uber. There he founded the group that built Uber's self-driving technology, created its AI Labs, and spearheaded Uber Elevate, the company's plan for urban air taxis[9]. It is a career spent on the border where software meets the physical world, nearly always approached from the software side.

In 2019, according to Crunchbase[8], he started Atomic Machines, and then said almost nothing in public for six years. Recruiting pages sketched the outline for anyone paying attention: engineering offices in the East Bay and in Santa Clara, a long-term goal of atomically precise manufacturing, and a first step at the scale of single-digit microns[28]. On 7 October 2026 the company finally spoke, with about $250 million raised in total from backers including OneIM, Gigafund, XTX Ventures, Valor Equity Partners and the Regents of the University of California[7][8]. Gigafund's Luke Nosek, a PayPal co-founder and early SpaceX investor, called it “the start of vibe manufacturing”[6]. Holden's own pitch was shorter still: “We're making machines programmable.”[7]

The name explains the ambition. In The Diamond Age, Neal Stephenson's 1995 novel, matter compilers are household appliances that assemble objects from a feed of raw molecules. Holden borrowed the name with some care, describing his system as a real but bounded first step toward that vision rather than the vision itself[3], and he traces the worldview behind it to watching Star Trek at the age of five. On X he squeezed the entire idea into four words: different code, different machine.

03/ The problem

Why tiny machines never got their Moore's law

In the last days of 1959, Richard Feynman stood up after dinner at an American Physical Society meeting in Pasadena and told a room of physicists that there was plenty of room at the bottom. He offered $1,000 to anyone who could build a working electric motor that fit inside a cube one sixty-fourth of an inch on a side[19]. Within a year, and to Feynman's surprise, a Caltech graduate named William McLellan collected the prize with a motor made through meticulous craftsmanship and conventional tools[20]. It met the rules and advanced nothing, because nobody could make a million of them.

That gap between making one and making many is the whole history of micro-machines. In 1967 Harvey Nathanson's team at Westinghouse published the resonant gate transistor, one of the very first devices to put a moving part on a chip[21]. In May 1982 Kurt Petersen of IBM Research published “Silicon as a Mechanical Material”, the review that pointed chip fabs at building machines and is still treated as required reading[22]. The field took a name, MEMS, for microelectromechanical systems, and it earned real triumphs. Larry Hornbeck's digital micromirror device, invented at Texas Instruments in 1987, flips microscopic mirrors up to 10,000 times a second, lit more than eight in ten digital cinema screens by 2015, and won him an Oscar statuette that year[23]. The sensor that rotates your phone screen is a MEMS device. So is the microphone in your earbuds.

And yet the industry never compounded the way chips did. Yole Group counted 31 billion MEMS units shipped in 2024, worth $15.4 billion in total, and forecasts growth of just 3.7% a year, to $19.2 billion by 2030[18]. ASML alone reported €32.7 billion of sales in 2025[26]. The reason has a name inside the trade, sometimes called the MEMS law: one product, one process[15]. Every new kind of device tends to need its own sequence of steps, its own tuning and its own years of development, so the learning from one product rarely carries cleanly into the next.

The engineers who live with this describe the cost plainly. A 2018 paper by Alissa Fitzgerald and colleagues, whose consultancy has completed more than 400 MEMS projects[17], warns that turning a working prototype into a manufacturable product can take years and millions of dollars before commercial production begins, with the transfer into a production foundry started at least a year ahead and budgeted at no less than $500,000 a year for a 150-millimetre fab and $1 million a year for a 200-millimetre one[16]. Packaging, assembly and test can make up as much as 75% of a finished unit's cost[17]. Even Bosch, the largest MEMS maker, which says its standard sensors now run on largely standardised processes, concedes that demanding new designs still need substantial technology development[15].

Holden's diagnosis goes a step further. MEMS, he argues, only ever covered the devices that fit tools designed for transistors: mostly flat structures, in a narrow set of materials. Magnetics, real force, sustained pressure and true three-dimensional assembly stayed out of reach[4]. Below roughly the low tens of microns, he writes, conventional manufacturing falls apart, and the one process that survives down there, the chip fab, was built for machines that process information[3]. In between sits a whole universe of devices that physics allows and factories cannot make.

04/ The machine

What a Matter Compiler actually is, and what it is not yet

Strip away the science fiction and the Matter Compiler, as the company and The New York Times describe it, is a chain of 13 automated machines that build and test the components of a device[5]. Atomic Machines calls it a self-driving fab: it designs, builds, measures and learns, then carries what it measured into the next design[2]. The features reach single-digit microns, the parts can move, and one device can mix several materials, including the metals and alloys that full-size machines rely on[4][7].

Two claims do the heavy lifting. The first is that nothing in the line is specific to one product. “No masks to make. No line to retool,” as the company's own page puts it, so every iteration becomes a change in software rather than a change in the factory[2]. The second is that the system measures each step with high-precision metrology and corrects in real time, while an AI layer mines thousands of experiments, combining materials, heating them across temperature ranges, assembling them in different ways and testing the results, to work out how to build each design[5][7].

Holden's favourite analogy is biological. A ribosome reads a string of genetic letters and assembles a working protein, and across four billion years of evolution the cost of trying a new design has never gone up. Human factories, by contrast, keep their designs in hardware: in moulds, masks, dies and fixtures. His essay compresses the difference into one line worth stealing: “An additive repertoire is a list; a combinatorial repertoire is a language.”[3] A 3D printer can print any shape, but only a shape. The bet is that many processes and many materials, driven by instructions, add up to a vocabulary for machines rather than a catalogue of parts.

Outsiders are encouraged, if careful. Eugene Chow, an electrical engineer at Lawrence Livermore National Laboratory, told the Times that attempts to remake manufacturing this way are rare and very difficult, and that the United States “desperately needs more of this”[5]. Nobody outside the company has yet published measurements of the process, its yields or its throughput, and Atomic Machines has not said how many devices the line can make in a year[8].

05/ The switch

Why a universal factory's first product is a switch

Thomas Edison lost the war of the currents in the 1890s, largely because alternating current was far easier to step up and down for long-distance transmission. Direct current also carried a quieter handicap that never went away: it is hard to switch off. Alternating current passes through zero a hundred or more times a second, which gives an arc between opening contacts a natural moment to die. Direct current never passes through zero, so if you interrupt it under load the arc can simply keep burning.

More than a century later, AI is handing DC its revenge, and the old handicap comes with it. In May 2025 NVIDIA laid out a plan to move its AI factories to 800-volt DC distribution from 2027, because the 54-volt power systems inside today's racks start hitting physical limits beyond about 200 kilowatts, and a one-megawatt rack fed at 54 volts would need up to 200 kilograms of copper busbar[10]. The same document lists the protection this new world needs, from fuses and safety breakers to emerging solid-state devices, and concedes that “innovations will still be required”[10]. If you want to know why racks are heading toward a megawatt in the first place, our field guide to NVIDIA's Vera Rubin NVL72, whose GPUs draw 1,800 to 2,300 watts each, shows where the watts go.

That is the gap Holden says he chose six years ago, reasoning from first principles[3]. Today's options both hurt. An electromechanical contactor carries current through metal contacts with very little loss, but it takes milliseconds to open. A solid-state switch opens in microseconds, but it conducts through a semiconductor channel, so it turns a slice of every amp into heat. The New York Times opens its story with an engineer who builds power equipment for data centres describing exactly that dilemma: a mechanical switch the size of a baseball that is very slow, or a silicon one that wastes power[5].

PrimeSwitch claims the empty corner. Its product page reads like a datasheet: a bistable relay that holds its state with zero power, rated for 150 amps continuous, AC or DC, with 200 micro-ohms of on-resistance, 1,500 volts of galvanic isolation and a 50-microsecond opening time, in a hermetically sealed surface-mount package 9.5 millimetres across and 3 millimetres high[1][7]. It needs no active cooling, the company says, and it is sampling now[1].

Two pieces of arithmetic the launch did not do

First, the resistance. Two hundred micro-ohms is contactor territory rather than a breakthrough in itself: TE Connectivity's KILOVAC EV200, a workhorse 500-amp DC contactor, lists a typical contact resistance of 0.2 milliohms, the very same number, in a package that weighs 430 grams[11]. The achievement is getting that figure out of something the size of a shirt button. At its full 150 amps, PrimeSwitch would turn about 4.5 watts into heat (current squared, times resistance), and all of it has to leave through the circuit board underneath. No active cooling is plausible, but the board is doing real work.

Second, the speed. Atomic Machines says 50 microseconds is about 1,000 times faster than a conventional contactor[4]. The EV200 datasheet gives a maximum release time of 12 milliseconds with arcing included, which makes PrimeSwitch roughly 240 times faster than that particular part[11]. To reach 1,000, the comparison contactor has to take around 50 milliseconds. Two hundred and forty times is still a startling gap. It is also a reminder to read every multiplier with its baseline attached.

Speed alone does not stop an 800-volt arc, and Atomic Machines does not pretend it does. It pitches PrimeSwitch as half of a hybrid breaker: a small solid-state bypass carries the current during the few microseconds the relay needs to open, so the contacts part without arcing, and the open relay then provides true galvanic isolation[7][8]. The idea itself is old. In 2012 ABB unveiled a hybrid breaker for high-voltage DC grids that paired very fast mechanics with power electronics to interrupt a fault within five milliseconds, and presented it as clearing a barrier that had stood for a century[14]. What would be new here is the scale: a hybrid you solder onto a circuit board, with a mechanical half that opens in microseconds.

There is precedent on the MEMS side as well, and it is humbling. Menlo Microsystems spun out of GE in 2016, after twelve years of research there, to sell metal MEMS switches[13]. In 2018 it announced a 200-volt, 10-amp relay with less than 10 milliohms of on-resistance and, yes, a 1,000-fold improvement in switching speed over traditional relays[12]. PrimeSwitch's genuine claim to novelty is scale: fifteen times the current, at a small fraction of the on-resistance, in a single device. If that holds up under independent testing, it is a real step. If it does not, the 1,000× headline will have been heard before.

And the choice of first product is the sharpest decision in the whole launch. AI data centres are the one market where power density is climbing fast enough, and the cost of a fried rack is high enough, that a buyer will evaluate a strange new part from a start-up. PrimeSwitch makes its public debut at the Open Compute Project Global Summit in San Jose, 12 to 15 October, and early-access customers are already evaluating samples. Their names, the price and the production volume have not been disclosed[8].

06/ The ASML question

So, can it threaten ASML?

No, and the reason is worth understanding, because it says something about both companies. ASML's EUV scanners print the patterns of transistors with light of 13.5-nanometre wavelength, resolving features of 13 nanometres on its standard machines and 8 nanometres on the newest High-NA tools, at well over a hundred wafers an hour[24]. The Matter Compiler works at single-digit microns, roughly a hundred to a thousand times coarser, and it builds three-dimensional machines with moving parts rather than flat patterns.

The deeper difference is what each machine is built to do. ASML makes the most perfect printing press in history: one pattern, carried on a mask, repeated billions of times, as fast and as small as physics allows. The Matter Compiler is built for the opposite job, making many different things without a new mask or a new line, and Atomic Machines has described its own process in recruiting material as direct-write[28]. Chipmaking already knows what direct-write is good for, and that explains why the two worlds do not collide: electron-beam writers draw the masks themselves, precisely because they are far too slow to pattern production wafers that carry tens of billions of transistors.

ASML is the ultimate printing press. Atomic Machines wants to be a compiler. Presses win on copies; compilers win on variety.

The money tells the same story. Atomic Machines' entire $250 million would buy about two-thirds of a single High-NA EUV scanner, which costs around $380 million[25]. A genuine challenger to ASML is somewhere else: China's state-backed EUV programme, or Substrate, the American start-up that raised $100 million in 2025 to attempt X-ray lithography powered by a particle accelerator[27]. We weighed both in our deep dive into how one company came to control every advanced chip on Earth, and nothing in this launch changes that conclusion.

What the Matter Compiler does threaten, if it works, is a business model. The MEMS industry's moat is process know-how: years of tuning that a rival cannot copy quickly. A line that makes new devices without new processes turns that moat into a puddle. The companies with the most to think about are the large MEMS makers and the foundries that sell custom process development, followed by the makers of relays, contactors and DC breakers, whose data-centre business PrimeSwitch targets directly. Power semiconductor makers sit somewhere in between, because a hybrid breaker still needs a solid-state bypass, just a smaller one.

There is one thing ASML and Atomic Machines genuinely share, and it may be the most interesting line in the comparison. ASML's scanners are wrapped in metrology and computational software that measure errors and correct them continuously; precision at that scale is a feedback loop as much as an optical feat. Holden is trying to build an entire factory on that principle. If ASML is the proof that closed-loop control can push manufacturing to its physical limits, the Matter Compiler is a bet that the same idea can push it sideways, across many kinds of machine.

07/ The robot question

Is this the start of the robot age?

It is tempting to say yes, and the temptation is half right. Every robot is a stack of sensors, actuators and power electronics wrapped around software. Many of the sensors are already micro-machines: the accelerometers and gyroscopes that keep a drone level or a humanoid upright come out of MEMS fabs by the billion. What MEMS never delivered in quantity is the other half, small actuators with real force: micro-pumps, valves, magnetic micro-motors, the moving parts that Holden says transistor-era tools could not make[4].

Atomic Machines is explicit that this is where it wants to go. Its launch materials talk about gears the size of a grain of sand, robots small enough to work inside the human body, and cooling built straight into AI chips[6][7], and Holden told the Times he imagines microrobots for surgery and manufacturing[5]. A factory that could compile custom versions of those parts in days rather than years would widen the design space for robot hands, tactile skins and medical devices, much as the foundry model Morris Chang built at TSMC widened who could design a chip.

Now the cold water. Most of what limits a humanoid robot today is the size of a fist, not a grain of sand: motors, gearboxes, batteries, hands, and the AI that drives them. A micro-machine fab does not make a better knee. And Holden reached for the most cautionary example available when he described what a matter compiler could make: a device that runs medical tests on a few drops of blood, which is exactly what Theranos promised. His point, as the Times reported it, is that such a device needs a different kind of manufacturing technology[5]. He may well be right. It is also a reminder that the distance between a beautiful manufacturing story and a validated product is measured in years of data, not in launch videos.

So the honest answer is that a robot age running on compiled micro-machines is possible, and it is not what launched this week. What launched is a switch. The robotics story starts on the day a second and a third device, built on different physics, come off the same line without a new process. That is the experiment worth waiting for.

08/ Claims versus evidence

What is proven, what is claimed, and what to watch

Launch weeks are when claims are loudest and evidence is thinnest. Here is the ledger as it stands on 10 October 2026, with every line traceable to the sources at the end of this piece.

What would change our minds

09/ The verdict

The verdict: the first real chapter of an old dream

Atomic Machines has done something rare: it has taken a word from science fiction and attached it to a product with a part number. The Matter Compiler is real in the sense that matters at launch: a working line, a working device and a hungry market paying attention. It is not yet universal, it is not yet hands-free, and nobody outside the company has measured it.

It is also not an ASML story. Nobody will print a GPU on this line, and nobody is trying to. If Holden is right, the casualty will be a forty-year-old rule, one product, one process, and the companies whose moats were dug around it. That would be a bigger shift than a new lithography rival, because it would change who gets to build small machines at all.

Feynman's prize was won by a craftsman, and it changed nothing. The test for Atomic Machines is the mirror image: not one tiny machine, but a hundred different ones, with no craftsman in the loop.

Until then, the most honest thing about this launch is its first product. A universal factory that begins with one excellent, unglamorous switch for the hungriest market on Earth is not an anticlimax. It is a company that knows exactly where the money is while it tries to prove the dream. Different code, different machine: we would very much like to see the second one.

FAQ/ Questions people actually ask

Questions people actually ask about Atomic Machines

10/ Sources

Sources

Twenty-nine sources, with company material labelled as such. Where a figure comes only from Atomic Machines, the text says so.

1. PrimeSwitchAtomic Machines product page. 150 A continuous, 50 µs opening, 200 µΩ, 1,500 V isolation, 9.5 by 3 mm, bistable, sampling.
2. Matter CompilerAtomic Machines product page. The self-driving fab, no masks or retooling, and the one product, one process framing.
3. Atomic Machines origin storyJeff Holden, company essay, 7 October 2026. The Diamond Age naming, the ribosome analogy, the choice of first product.
4. The micro-machine universeJeff Holden, company essay, 7 October 2026. The limits of MEMS, seven physical domains, the 1,000× contactor comparison.
5. The startup that wants to build ‘microrobots’ with AIThe New York Times, 7 October 2026, as syndicated by The Star. The 13-machine chain, the manual steps, Eugene Chow, the Theranos remark.
6. Atomic Machines emerges with $250M to build micro-machines from codeThe Next Web, 7 October 2026. Funding, investors, Luke Nosek's comment, target applications.
7. Atomic Machines emerges with $250 million and AI-native manufacturing platformcitybiz, 7 October 2026. Launch details, Holden's quote, the hybrid breaker, the hermetic package, investors.
8. Atomic Machines ships its first Matter Compiler relay for evaluationRuntimeWire. Crunchbase founding date, cumulative funding, the hybrid configuration, what was not disclosed, the OCP debut.
9. Jeff Holden, alumni award profileSiebel School of Computing and Data Science, University of Illinois. D. E. Shaw, Amazon Prime, Groupon and Uber roles.
10. NVIDIA 800 V HVDC architecture will power the next generation of AI factoriesNVIDIA Technical Blog, 20 May 2025. The 2027 transition, the 200 kW limit at 54 V, 200 kg of copper, protection devices.
11. KILOVAC EV200 contactor datasheetTE Connectivity, via Farnell. 500 A continuous, 0.2 mΩ typical, 12 ms maximum release including arcing, 0.43 kg.
12. Revolutionary MEMS switch handles 200V, 10AeeNews Europe, 19 March 2018. Menlo Micro's relay, under 10 mΩ, and its own 1,000× switching-speed claim.
13. GE spins out MEMS startup, wants cheaper power switcheseeNews Europe, 12 December 2016. Twelve years of GE research behind Menlo Micro.
14. ABB develops hybrid circuit breakerRenewable Energy World, 7 November 2012. A hybrid HVDC breaker that interrupts within five milliseconds.
15. MEMS: one product, one process?SEMI, 15 October 2018. Interview with Udo Gómez of Bosch on the MEMS law and standardisation.
16. Translational engineering: best practices in developing MEMS for volume manufacturingFitzgerald, Jackson, Chung and White, Sensors and Materials 30(4), 2018. Time and cost of moving MEMS into production.
17. MEMS product development, webinar slidesAMFitzgerald for NCCAVS, 24 June 2021. Packaging, assembly and test up to 75% of unit cost; more than 400 projects.
18. MEMS market regains momentumYole Group press release, 2025. $15.4B and 31 billion units in 2024, $19.2B forecast for 2030, Bosch first.
19. Feynman's 1959 lecture and his two $1,000 prizesAPS News, This Month in Physics History, November 2016.
20. There's Plenty of Room at the BottomWikipedia. William McLellan's prize motor, November 1960.
21. Technical brief on the early history of MEMSIEEE Electron Devices Society Newsletter, January 2020. Nathanson's resonant gate transistor, 1967.
22. Kurt E. PetersenEngineering and Technology History Wiki. “Silicon as a Mechanical Material”, Proceedings of the IEEE, May 1982.
23. Texas Instruments Fellow Larry J. Hornbeck wins the OscarTexas Instruments, 9 February 2015. The DMD invented in 1987; DLP Cinema's share of digital screens.
24. ASML's roadmap for chipmaking lithography tools examinedTom's Hardware, 1 May 2026. 13 nm and 8 nm resolution, 13.5 nm light, scanner throughput.
25. ASML's High-NA chipmaking tool will cost $380 millionYole Group industry news, citing Reuters and the Taipei Times.
26. ASML reports €32.7 billion total net sales and €9.6 billion net income in 2025ASML results press release.
27. US startup Substrate raises $100M to take on ASML and TSMCBits&Chips, 29 October 2025.
28. Atomic Machines company profileWellfound recruiting page. The atomically precise manufacturing goal, the single-digit-micron first step, direct-write fabrication.
29. Menlo Micro's ‘Ideal Switch’ gains traction with $150 million fundingAll About Circuits, 14 March 2022. Reported switching speeds of 10 to 20 microseconds.