A Startup Leaves Stealth With $250 Million to Print Tiny Machines

Atomic Machines spent six years saying nothing. This week it came out of stealth with $250 million raised and an unusual claim: that it can turn a software description straight into a working machine barely larger than a grain of sand.

The Berkeley company, founded in 2019, is run by Jeff Holden, who was Amazon's tenth engineer and built Amazon Prime before serving as Uber's first chief product officer. Its system, reported by The Next Web, is called the Matter Compiler, after a device in Neal Stephenson's novel The Diamond Age. The idea, in the company's words, is that bits and raw materials go in and complete, functional micro-machines come out, with moving parts, several different materials, and features measured in single-digit microns.

A switch for the data center

The first thing it has chosen to build is mundane by comparison, and that is the point. PrimeSwitch is a relay, an electrical switch, that the company says opens in 50 microseconds. Its target market is the AI data center, where power is now the binding constraint. Faster, more precise switching helps route and protect the enormous electrical loads these buildings draw.

That framing is shrewd. The grid problem around AI is no longer theoretical. The memory crunch now runs to 2028, operators are walking away from gigawatt power deals they can no longer justify, and readers keep asking whether the boom will crash the power grid. A company that sells a better switch into that panic has a clear buyer, which is a more sober place to start than selling grain-sized gears.

The longer bet

The gears come later. Holden's stated ambitions run to cooling built directly into AI chips and to robots small enough to work inside the human body. Those are the kind of promises that six years of stealth and $250 million are supposed to buy time for, and the kind that are easy to make and hard to ship. Listed backers include Valor Equity Partners and NJP Ventures, according to a CB Insights profile. For now the test is narrower and more telling: whether a machine that compiles other machines can produce one humble relay at a price and volume data centers will actually pay for.