The ASML Killer?

Using Holograms to Make Chips w/ Omar Durrani from Cnuic

Hello friends, colleagues and enemies.

The next generation of computing will look tres different from the last. CMOS silicon is basically one material, patterned in two dimensions, optimised over half a century around a single device: the transistor. What’s coming is 3D, multi-material, heterogeneous, and photonic. And the fabs that could build it at scale don’t really exist yet.

To see why, start with the one machine every fab is built around: the lithography scanner. Every chip in every device you own was patterned by one.

Here’s how it works:

1. Take a silicon wafer.

2. Coat it with a light-sensitive chemical called photoresist.

3. Project UV light through a stencil — a mask — onto the resist. (Remember this step.)

4. Where the light hits, the resist either hardens or dissolves.

5. That pattern decides which bits of silicon get etched away.

And this is your Dutch monopoly, Europe’s ace in the hole. Our dinner ticket. ASML’s EUV scanners cost north of $300 million each, and TSMC, Samsung and Intel buy almost all of them. The entire global compute supply chain runs through one car park in Veldhoven. And every one of those machines is optimised, end to end, for a single job: shrinking transistors.

But.

Photonics needs something different. Larger structures, sized to the wavelength of light, hundreds of nanometres rather than tens. + materials beyond silicon. 3D geometries. None of which the world’s most expensive machines were built to make. Interesting…

Which brings me to Cnuic, a Scottish startup betting on a different lithography stack entirely. Co-founded by Omar Durrani and Ben Szutor, they’re building a multi-beam laser interference lithography system. No step 3.

Instead, multiple laser beams interfere inside the photoresist itself, writing a three-dimensional pattern directly — what they call a virtual mask. Or, if you prefer, a hologram. Think Tupac at Coachella, or ABBA Voyage.

Disclosure: I am not an investor. So I am not talking my book. So when I do talk my book, maybe you will give me some more god damn respect when I do.

What did I learn?

• Photolithography is a three-way optimisation between throughput, feature size, and feature complexity, and every existing tool wins on two and loses on the third.

  • Immersion has throughput and feature size and cannot do truly 3D structures.

  • E-beam has feature size and complexity and is glacially slow.

  • Nano-imprint has throughput and a kind of complexity but is contact-based, struggles with slanted or pyramidal geometries.

  • The open question is whether you can win on all three at once.

• The unlock isn’t smaller features, it’s geometries existing tools cannot make. Classic “enabling tech”. Photonic devices don’t need 3nm pitch. They need 3D control, multi-material patterning, and structures sized to the wavelength of light, which is 100s of nanometres rather than 10s. Maskless, non-contact, three-dimensional patterning is the lever. Multi-material is a capability the tool can support, but bringing dissimilar materials together under one process roof is a separate problem. The unresolved question is whether that scales to production volume rather than getting stuck at prototyping.

• The bottleneck on a category-defining UK photonics company isn’t physics, it’s UV laser supply and talent geography. Cnuic buys components built for other purposes, optical benches from Germany, lasers from the US, Japan, Korea. Talent works, partly because the global talent visa is genuinely effective at importing engineers from Korea, Ireland, and elsewhere in the UK. Go UK!

---

The State of the Future Show

What up Omar. I have a very technical question to start. Why photolithography?

Honestly, we fell into it. Ben tried to get a job at Rolls-Royce three times. They rejected him three times. So he did a PhD in laser physics with a quantum focus, ended up at a laser company, became its CEO. While he was selling lasers to customers he noticed they were using them for interference lithography setups. Benchtop, fiddly, poor yield. His angle was that people don’t actually want benchtop setups, they want a turnkey system, and they’ll pay for it. So we’d package one. That was the initial idea. It evolved.

Ben and I had a habit of going to evening lectures on time machines, antimatter reactors, and space elevators while we were at university, trying to figure out how you might design any of them. So when we started looking at lithography, it didn’t seem as bad as those.

ASML is one of the most extraordinary companies ever built. What’s the actual gap, what does ASML not do well that you do?

We’re not playing the Moore’s law game. Some lithography startups are still trying to drive transistor density and shrink to ever more compact structures. We aren’t.

ASML are genuinely optimised for two things we are deliberately not. The first is barrier to entry. EUV machines are extraordinarily expensive, even DUV machines are expensive, and big companies sometimes end up with bench-top setups internally just to prototype, because they can’t get the scanner time they need. The second is structures. Semiconductors as an industry is super-optimised for 30-centimetre silicon wafers, planar geometries, and tens-of-nanometres features. What we’re doing is a different wedge. Different materials beyond silicon. Larger structures, hundreds of nanometres rather than tens. And, increasingly, three-dimensional geometries.

If you think about photonic crystals, lattices, multi-material modulator stacks, three-dimensional control is inherent. Projection lithography is by definition two-dimensional. To get 3D you have to do all sorts of post-processing, etch trickery, layer stacking. We don’t.

Talk us through the approach. Multi-beam laser interference lithography. What is it, what’s a virtual mask?

The standard projection system has a light source, extremely complex for EUV where you fire lasers at molten tin droplets, bounce the resulting EUV through ultra-polished mirrors, project a beam down through a photomask onto photoresist on the wafer. The mask is a stencil. The pattern on the wafer is the shadow of the stencil.

What we do is a bit different. The easiest way to visualise it is to think about a hologram. The way I tried to explain it to my family was Tupac Shakur appearing as a hologram at Coachella, a three-dimensional image. Instead of projecting a two-dimensional picture, we generate a hologram inside the photoresist using multiple coherent laser beams. The virtual mask engine is what generates that hologram. It’s a complex piece of work but it does two useful things. One, it removes the mask step entirely. Two, because the beams overlap over a 3D volume, we create true depth of focus, well over a micron in size, which is much deeper than projection systems. This 3D volume is where we can add our secret sauce to not just create deep features, but also add variation to the vertical profile in a single exposure, within less than 10 seconds.

Sidebar: What is a photomask?

In conventional lithography, a photomask is a quartz plate patterned with chrome that acts as a stencil. UV light projects through the mask (DUV) or reflects off it (EUV) onto photoresist on the wafer, and wherever light hits the resist, the resist either dissolves or hardens. Modern EUV masks cost around $300,000 each, take 15-16 weeks to fabricate using e-beam writers, and represent one of the largest single fixed costs in lithography. A maskless system removes that fixed cost entirely.

What does that actually unlock at the device level that you can’t do today?

A few things matter. Maskless means you can iterate without waiting 15-16 weeks for a mask set. Non-contact means you don’t get the contamination or geometry-limit issues nano-imprint has. Three-dimensional control means you can pattern slanted or pyramidal structures, photonic crystals, layered metamaterials in a single exposure rather than dozens of stacked passes.

The clearest example is flat optics, sometimes called metasurfaces and metalenses. These are miniature versions of existing optical components made of glass or plastics. The advantage is that they can achieve the same function over much less volume, and can theoretically enable much quicker and cheaper production. The problem is that it doesn’t currently work well at production scale, especially for visible wavelengths, like RGB.

Sidebar: What is a meta-lens?

A meta-lens is a flat optical element that focuses light using sub-wavelength structures patterned across its surface, instead of curved glass. Apple’s Face ID uses a related kind of optic. Meta-surfaces extend the same idea, controlling phase, polarisation, or amplitude of light by structure rather than material thickness. The promise is replacing bulk optics in everything from phone cameras to AR glasses with much thinner, much lighter components. The bottleneck has always been fabrication. At the scales meta-lenses need (tens to hundreds of nanometres, multi-material, sometimes three-dimensional), nothing existing patterns them well at volume.

The known alternatives, direct laser write, e-beam, nano-imprint, each hit a wall. Which walls do you actually break, and where do the alternatives still win?

Each one found its niche for a reason. E-beam is great for resolution and precision, also extremely slow. You cannot mass-produce with it. Its niche is mask-making and prototyping. Direct laser write has similar throughput limitations that scale with complexity. Nano-imprint is more flexible on materials than immersion, it’s the technique behind roll-to-roll polymer production, but it’s a contact method. It struggles with slanted sidewalls.

Immersion dominates because of the reliability, defect rates, and throughput it’s achieved, around 80% of the patterning market. People who prototype on e-beam end up putting the device through immersion for production and find it doesn’t quite work the way they designed it. Light is more fickle than electrons. Line-edge roughness, sidewall roughness, wavelength-scale loss all matter, and immersion was optimised for transistor performance rather than photonic performance. To optimise all of this, companies have to spend millions on design iterations, where the design and procurement of photomasks, setup costs and so on pile up.

Our bet is to complement, not replace. Immersion still does the high-density CMOS work. A tool like ours does the multi-material, 3D, subwavelength-but-not-10-nm work, meta-lenses, photonic packaging, advanced interconnect, the geometries the existing tools can’t really do.

Multi-material patterning is the bit that doesn’t get discussed enough. Why does the next wave of optical devices need it?

Different photonic components want different materials. Modulators perform best in one material. Lasers need a different material, indium phosphide for active gain, for example. Waveguides want yet another. To build a truly performant photonic chip, you really want to start from materials specialised to each component.

What’s happening today is foundries take silicon-photonics platforms and bolt on post-processing trickery, hybrid integration, transfer printing, all sorts of clever workarounds, to get different materials onto a fundamentally silicon-optimised wafer. It works. It’s also slow, expensive, and yields trail what they could be. The vision is a foundry where multi-material is the default rather than a retrofit.

You list AR/VR, displays, meta-lenses, ultrafast lasers, roll-to-roll polymer. Five segments is a lot for a pre-seed. Which one is real today, and which are roadmap?

Fair question. We’re a pre-seed and we need flexibility, so I’ll be honest about it. The real pull is in consumer electronics, focusing on flat optics. A lot of large companies are exploring these and struggling with the fabrication. 3D fabrication of optical components in the subwavelength regime unlocks new consumer applications, including better filters for screens to achieve higher brightness, or less glare, enhanced camera functionalities and so on. This is our entry. The technology then scales to more complexity, for example optical interconnect technologies, higher speeds in telecoms, and quantum and optical computing.

What changed in the last twelve months that wasn’t true before?

Three things that enable us to move forward with this idea.

The first is laser power. The lasers we use had an output of 100x less a few years ago. That’s what makes the technique viable for production-scale exposure rather than benchtop.

The second is compute. Generating the hologram pattern is computationally expensive. AI has made GPU-accelerated compute available at a scale where the calculation is feasible.

The third is supply chain. The consolidation of the global photonics supply chain over the last few years has made building a system from off-the-shelf parts much easier. It doesn’t sound exciting but it’s a real unlock.

Sidebar: Why is laser power the constraint?

In photolithography, the dose, total energy delivered per unit area to the photoresist, sets how quickly you can expose a wafer. A 1 mW source will expose a wafer hundreds of times slower than a 1 W source. Throughput is dose-limited, and pushing UV laser sources from milliwatts into the watt regime is what moves a technique from “useful in academic settings” to “plausibly competitive on wafers per hour.”

Laser miniaturisation is a horizontal trend that benefits a much wider set of applications than just lithography. Helion is shrinking the laser cavity for fusion. Photonic chips are getting on-chip laser sources. AR/VR sensing. Where does that go?

Yes. Ben spent a lot of his career working on more compact lasers. Even short of fully on-chip cavities, just shrinking them means you can fit more in a system. The footprint of semiconductor equipment in general should improve as the sub-systems shrink. We’re not personally bottlenecked on laser footprint, but the broader industry is going to benefit from miniaturised laser sources for photonic ICs, metamaterials, optical interconnects. The miniaturisation of UV sources is one of those upstream trends nobody talks about because it’s an input rather than a product, but the downstream impact is wide.

Building photolithography hardware in the UK in 2026. Where are you actually bottlenecked on supply chain?

We’re buying primarily off-the-shelf components, with all the limitations of off-the-shelf for an application it wasn’t built for. Optical benches and most of our optics come from Germany. Lasers come from the US, Japan, Korea. The UK has UK-based suppliers in the sense that you order through them, but most of what you actually receive is shipped from elsewhere.

There are also contract issues with US suppliers, some semiconductor-related applications you’re not allowed to use their components for, depending on export licensing. The UK is more of a services layer. There are companies here providing services around photonics, integration, testing. There aren’t many actually building components for the photonics or lithography industry. That’s the gap.

How is talent in Scotland?

Talent has been surprisingly easy. The global talent visa works very well for attracting high-quality international engineers. We’ve hired from Korea, Ireland, other UK cities. People are open to relocating. Scotland has a slightly different lifestyle, some people describe it as Scandi-adjacent. I don’t fully agree but you get the idea, and people have been quite happy to move. We have a core team of 12. Once you’re trying to hire 50, 100, or a thousand it might prove much harder, which we expect.

Debrief

What does the next photonic foundry actually look like? Phanofi gave one answer, work inside existing foundries, build optical components compatible with what’s there. James Lee at Wave Photonics gave another, build the design tools that let smaller fabless companies use the existing foundries (e.g. especially GlobalFoundries Fotonix) more effectively. Mintneuro pointed at heterogeneous integration on mature CMOS nodes for a market that doesn’t need leading-edge.

Cnuic is asking the question one layer up. If you actually wanted to make 3D photonic structures at subwavelength scales — meta-surfaces, photonic crystals, layered metamaterials in a single exposure — what would the lithography tool look like?

The answer they’re working toward is “not what ASML built.” Different physics, different optimisation surface, different machine. And probably a different upstream supply chain, because UV laser sources, multi-beam optics, and SLMs (spatial light modulators) for hologram generation matter more than EUV light sources and reflective masks.

I don’t know if the bet works. The biggest failure mode is throughput at production volume, almost everyone who tried interference-lithography techniques in the past stalled at the prototyping stage. The second is whether multi-material patterning, which in principle drops out of the technique, is constrained by photoresist chemistry, by adhesion between dissimilar layers, by post-exposure processing in ways that erode the theoretical advantage. Meta-lenses are a real market and not yet a TSMC-scale market. The wedge has to start narrow.

What strikes me, though, is the bet only became coherent in roughly the last 24 months. Watt-class UV lasers, AI-grade GPU compute for hologram calculation, supply-chain consolidation. As part of my work with Lunar Ventures, from which this project was birthed, initially tried to explore technologies where a supply-side change had recently occurred. And here is one: the miniaturisation of photonic components.

My pet thesis on UK photonics, which I’ve written about before, is that the credible path to semiconductor sovereignty is to specialise in places where the leading-edge logic race is already lost. It’s lost. Photonics is structurally different, earlier, more multi-material, less optimised, more open to new tooling. The UK has world-class research and a small set of commercial entities. What it doesn’t have is a critical-mass cluster.

A photonic city is half-joke, half-real proposal. The barriers aren’t financial. They’re industrial-policy, geographic, the way UKRI distributes funding regionally for political reasons that produce a sub-optimal cluster. Solving those is harder than the technical problem Cnuic is solving, which is saying something.

The conversation I want to have in five years is about what the foundry looks like once it exists. For now, the bet is that the lithography tool comes first. Cnuic is working on the tool.