On 7 October 2026 the US start-up Atomic Machines presented the Matter Compiler, a system it says builds working micro-machines, with moving parts and multiple materials, from code alone and without per-product tooling. For now the details and figures are the company’s own, but the announcement brings back an idea with a long history, in science fiction and in engineering: describing a physical object in a language and letting a machine make it.

Stephenson’s matter compiler

The system’s name comes from The Diamond Age: Or, A Young Lady’s Illustrated Primer, Neal Stephenson’s novel published in 1995 by Bantam Spectra, winner of the Hugo and Locus awards in 1996. Stephenson was coming from Snow Crash, from 1992, one of the reference novels of cyberpunk, and with The Diamond Age he moves his attention from networks to matter.

In the novel, set in the twenty-second century, almost every object comes out of a matter compiler, a device that assembles products from basic molecules. Molecules and energy arrive through the Feed, a centralised network controlled by the neo-Victorians of New Atlantis. Against the Feed grows the Seed, an anarchic, decentralised technology: a seed that carries the instructions to grow whatever is needed by itself, without going through the network.

The nanotechnology of the novel closely follows what Eric Drexler imagined in Engines of Creation, from 1986. The tension between Feed and Seed also runs through maker culture and free software, and it concerns who controls the tools that turn an instruction into an object.

Stephenson was not the first. In the 1956 film Forbidden Planet the Krell machine materialises whatever its users think, including the monsters of the unconscious. In Damon Knight’s A for Anything, which came out in 1959 as The People Maker and was revised in 1961, the Gismo duplicates any object, even another Gismo, and the society that follows is a slave society, because the only thing that keeps its value is human labour. Since 1987 the replicators of Star Trek: The Next Generation have produced food, medicine and spare parts from a molecular structure on file. In Charles Stross’s Singularity Sky, from 2003, an unknown civilisation hands out cornucopia machines to a planet held at a nineteenth-century level of technology, and within days its economy and government fall apart.

These stories ask what remains scarce when making things becomes easy. For Knight it is labour, for Stross the power that holds the social order together, for Stephenson education and belonging to a community: the Primer of the subtitle is an interactive book that accompanies a girl as she grows up.

Hardware already written as code

The idea of describing hardware in a language and leaving the rest to a toolchain has a long history.

  • Chips. For forty years digital circuits have been written in hardware description languages such as Verilog and VHDL, and synthesis software turns them into the layout from which the foundry makes the masks. With Tiny Tapeout anyone can send a small design into production on processes with open development kits, SkyWater at 130 nanometres, IHP at 130 and GlobalFoundries at 180. The TTSKY25a run on SkyWater collected 237 designs from around the world.
  • Circuit boards. atopile describes a board in a declarative language, ato, made of modules, interfaces, units and tolerances. You write a requirement such as 10uF +/- 20% and the tool picks the part.
  • Sheet metal. Machina Labs forms sheet-metal parts with industrial robots that follow paths computed from CAD, without dies, measuring and correcting the part as it is formed.
  • 3D printing. In 2008 the RepRap Darwin printed more than half of its own plastic parts, with design and software under a free licence. A few years earlier MIT’s Center for Bits and Atoms had started replicating fab labs around the world, where two identical labs exchange a file and get the same object.

Tiny Tapeout, atopile, RepRap and the fab labs are on the side of the Seed, with open languages, documented processes and communities. Proprietary digital factories, for now, are on the side of the Feed.

Continuous integration for hardware too

If hardware is written as code, it can also be tested as code, with the same continuous integration and delivery processes (CI/CD). In May 2021 SpaceX’s software team described on Reddit how it does this. The pipelines start with fast, cheap tests and, if those pass, move on to long and complex ones. Alongside pure simulation there are hardware-in-the-loop testbeds, with copies of the flight computers and electronics running the real software, and developers can try a change on those testbeds before even merging it. For Starlink the testbeds are satellites taken off the production line and connected to the continuous integration system.

The latest digital factories bring the same loop into production: every operation is measured, and what is learnt goes into the next design. Testing sits inside the production process, at every step.

When the product is an autonomous machine, a robot or a micro-machine working inside the body, the pipeline continues after release. This is the subject of DebugABot, the research initiative I lead on debugging autonomous systems, both software and physical. It proposes nine primitives organised in three phases, identify, diagnose and intervene, and a kill switch implemented in hardware, on FPGA, below the layer where the model runs. It covers the part of the chain that follows release, while the machine is at work.

The cell compiles itself

The comparison with biology comes naturally. The ribosome is a fixed machine that reads an instruction, messenger RNA, and builds the corresponding protein by chaining twenty kinds of amino acids. A protein of 150 amino acids has 20¹⁵⁰ possible sequences, and according to the estimate that Jeff Holden, Atomic Machines’ founder, gives in the essay accompanying the announcement, in four billion years evolution has explored about 10¹⁶.

The figures for the ribosome are well known.

  • In Escherichia coli a ribosome adds about 20 amino acids per second.
  • The bacterial ribosome is made of three RNAs and about fifty proteins, and those proteins are built by other ribosomes, so the machine produces its own parts.
  • In favourable conditions E. coli doubles its mass in about 21 minutes, and according to the model by Rami Pugatch and Yinon Bar-On assembling a ribosome takes about 6 minutes.

The cell does in a single system what no factory does today: it reads the instructions, gathers the raw materials, builds its own machines, checks for errors and finally copies the instructions and itself. It is the machine John von Neumann described in the abstract between the late 1940s and the early 1950s, in work published posthumously in 1966: a universal constructor able to make any machine described on a tape, including a copy of itself with the tape.

Synthetic biology is retracing this path from the bottom up. In 2013 Michael Jewett’s group built working ribosomes in a test tube, transcribing ribosomal RNA and assembling it with the proteins in the same extract, one of the steps towards a synthetic cell that replicates. In March 2026 a University of Illinois group led by Zan Luthey-Schulten simulated in 4D the whole cycle of JCVI-syn3A, the minimal cell with fewer than 500 genes. The simulated cycle, of about 105 minutes, differed on average by two minutes from the one measured in the lab.

The differences from digital factories are large. The ribosome is about 20 nanometres across and works with atomic precision, while the finest industrial systems reach features of a few microns. The cell uses a small alphabet of standard components and leaves the shape to folding, while a factory works with metals and alloys. Finally, none of these factories builds itself, just as RepRap did not print its own motors.

What I think

I believe the future will be incredible. Three things that were separate until a few years ago are converging: models that design from a request in natural language, factories that reprogram themselves without changing tooling, and a synthetic biology that has started to build the parts of the cell from the bottom up. Going straight from request to product, prompt-to-product, follows the path that in software led from assembly to compilers and then to agents that write code.

When an object becomes code, the code brings its practices with it: versions, reviews, tests, signatures, supply chains to verify. A micro-machine meant for the human body, one of the declared targets in this field, falls under medical device regulation, and a design that compiles into a physical object raises security and dual-use problems different from those of software. It is governance work to be designed together with the technology.

Stephenson’s question remains: Feed or Seed. Whether the matter compiler becomes a centralised network or a set of open tools that anyone can use and study will be decided partly by the communities that built RepRap, the fab labs and the Tiny Tapeout chips.

We are working on it too.

Cover image: detail of the cover of the Bantam Spectra paperback edition of Neal Stephenson’s The Diamond Age — © the rights holders, all rights reserved. Reproduced as a quotation, for critical commentary.