Atum Works Is Building a 3D Printer for Semiconductors, and NVIDIA Already Wants In

The Macro: Semiconductor Manufacturing Hit a Wall, and Nobody Wants to Admit It

The semiconductor industry runs on lithography. ASML, the Dutch company that makes extreme ultraviolet (EUV) lithography machines, is one of the most valuable companies in Europe for a reason. Every advanced chip in the world passes through their equipment. Intel, TSMC, Samsung. They all depend on ASML’s machines, which cost upwards of $350 million each and take months to install.

Here is the thing nobody in the industry likes to talk about: lithography is fundamentally a 2D process. You deposit a layer, pattern it, etch it, repeat. Stack enough layers and you get a chip. But the structures themselves are flat. Every “3D” feature in a modern chip is really just a stack of 2D slices, painstakingly aligned.

This worked fine when the industry was shrinking transistors on a predictable schedule. But we are running out of room to shrink. The industry response has been to stack more layers (3D NAND, chiplets, advanced packaging), but the tools for doing this are slow, expensive, and constrained by the fundamental limitations of 2D patterning.

The opportunity for a true 3D fabrication technology is enormous. Not just for chips, but for photonics, MEMS, sensors, medical devices, and any application where you need complex metal structures at nanometer scale. The companies that currently serve this need are using focused ion beam (FIB) systems or two-photon polymerization, both of which are laboratory tools. They work beautifully for making one thing. They fall apart completely when you need to make a million things.

The gap between “we can make this in a lab” and “we can manufacture this at scale” is where fortunes get made.

The Micro: Caltech and NASA Engineers With a Printer and an NVIDIA LOI

Atum Works was founded by Lucas Pabarcius (CEO), Matteo Kimura, and Malcolm Tisdale. All three come from Caltech and NASA backgrounds, which is exactly the kind of pedigree you want when someone tells you they are going to reinvent how things get made at the nanometer scale. These are people who have worked on problems where failure means a satellite does not work, not where failure means the app crashes.

They came through Y Combinator’s Spring 2025 batch and set up their manufacturing facility and lab in Mountain View, California. The core technology prints arbitrary 3D metal structures with submicron resolution. To put that in context, submicron means features smaller than one millionth of a meter. Human hair is roughly 70 microns wide. They are working at a fraction of that.

The pitch is straightforward: they are building what they call “3D ASML.” Where ASML patterns chips layer by layer in 2D, Atum Works fabricates complete 3D structures directly. The cost advantage they claim is roughly one-tenth of current 2D lithography for comparable structures. If that holds at production volumes, the economics change the calculus for chip designers, photonics engineers, and anyone else who needs complex nano-scale geometries.

The early validation is strong. They have secured a letter of intent for co-development with NVIDIA. Their advisory board includes H.-S. Philip Wong, the former TSMC Chief Scientist, along with leaders from Intel and the silicon photonics space. When the former chief scientist of the world’s most important chip foundry agrees to advise your three-person startup, that tells you the technology is real enough to take seriously.

They are positioning as a foundry, meaning they manufacture parts for customers rather than selling printers. This is the right model. The machines are complex, the process expertise matters enormously, and customers in semiconductors want parts, not equipment headaches. TSMC built one of the most valuable companies in history on exactly this model.

The Verdict

I think Atum Works is one of the most ambitious companies I have covered. The technical risk is high. Going from a working prototype to a production foundry in nanomanufacturing is a multi-year, capital-intensive journey. The companies that have tried to disrupt semiconductor manufacturing from the outside can mostly be found in the graveyard.

But the founding team has the right background, the advisory board signals deep industry credibility, and the NVIDIA LOI suggests the technology performs well enough that a company with unlimited options is choosing to work with a three-person startup.

The competitive landscape is thin at the intersection of “3D” and “nanoscale” and “production volume.” Nanoscribe (now owned by BICO) does two-photon polymerization for research applications. Various FIB companies serve prototyping needs. Nobody is credibly offering what Atum Works is describing at foundry scale.

Thirty days, I want to see the NVIDIA co-development producing tangible results, not just a signed letter. Sixty days, I want to know if they can demonstrate repeatable quality at something approaching production throughput. Ninety days, the question becomes whether additional customers are lining up or whether NVIDIA is the only buyer. If the pipeline is filling, this could be generational. If it is one customer and a lot of promises, the timeline extends significantly.