I was traveling to Japan to visit Toshiba's Oita fab and Shinagawa office. I had just landed (Tue evening there), got to the hotel, turned on the TV and there it was, couldn't believe what I was watching. All flights back to US were canceled. I had an extra bag, delivery for a friend of friend at US naval base near Tokyo. The based also locked down. So I dragged the bag with me the rest of the week, while we tried to work out how to get back to US. Finally things opened up towards the end of the week. I handed off the bag, and got on a flight home. It's the emptiest flight I've ever been on, everyone had a row to themselves. Not many people wanted to fly I guess. When I got home my cellphone was full of messages from my parents. I went to college in NYC. Since my phone didn't work in Japan, they couldn't get in touch, and thought I might have gone to NYC to see friends and something happened, and freaked out. Thankfully no one I know directly from college in NY were hurt.

🧵1/ An interesting EUV lithography paper just dropped from the Department of Radiophysics at the University of Nizhny Novgorod. It is not a tweak to ASML’s architecture. It is a radically different theory of the lens: stop counting mirrors, and start counting how many times each photon is allowed to bounce.

This is a big deal in the lithography world. Photoresist platforms, and the developer materials that go with them, almost never change. In the entire history of semiconductor lithography, it has happened once: the shift from Novolac/DNQ resists (broadband and i-line) to chemically amplified resists (CAR). CAR then carried every subsequent scanner generation: i-line, KrF, ArF, and EUV. These are essentially acrylic polymers plus a photoacid generator and a basic quencher. The prevailing assumption was that high-NA EUV would break that streak. CAR would not work; a new platform, metal oxide resist (MOR), would be required. High-NA scanner data comparing the two appeared to support that view. In my opinion, that comparison was never apples-to-apples. CAR is built on a long polymer chain with a defined molecular weight. MOR uses a much smaller metal-oxide molecule, a fraction of CAR’s MW. Across the generations CAR has spanned, which is all of them, that molecular weight was never static. As printed geometries shrank, polymer chain length was reduced accordingly. Most CAR-vs-MOR comparisons did not use a resist optimized for the new high-NA pattern sizes. imec, working with materials suppliers, did. With a newly optimized CAR, the data shows that chemically amplified resists can work. Why this matters: changing photoresist platforms is a major undertaking. MOR in particular brings serious high-volume challenges, metal contamination of the tool set, and the fact that it only works in negative tone. CAR is 100% organic, already established in production, and most importantly operates in positive tone, with a negative-tone option available as well. If an optimized CAR can perform anywhere near MOR, it will be the preferred material of choice. Bad news for JSR and LAM who have made big investments in the metal oxide materials and dep/etch tooling.

Qnity Electronics was liable here for $55 million, which is a huge hit for a materials supplier. That's about 20000 gallons of their immersion photoresist, enough to coat 64 million wafers.

Logic Foundry view of the EUV photomask supply chain: Some important downstream companies are left out bc they are second order off the foundry, such as Veeco who makes the deposition tools for the blanks. Source: Samsung

TSMC may appear to get the credit for forcing ASML and the mask suppliers to enable the 6×12 HNA reticles, but it was Intel (and others) that had been working with the suppliers for the last 3-4 years.

TSMC is the biggest buyer (by far) of advanced photomasks, so it has the leverage to push the whole supply chain toward larger masks. That is what makes High-NA cost-effective, and why TSMC had not committed a node to it until now. This is standing up an entire ecosystem and involves hundreds of suppliers.

Frank Rohmund, CEO of Zeiss, on China replicating their DUVi/EUV litho optics: "decades-old [trade secrets], it's extremely hard to copy...the protection of this know-how is very important to us."

Here's a fresh paper on a new litho source using gadolinium. Same basic idea as mainstream EUV, but it drives a plasma with Gd instead of tin, and the multilayer mirrors use lanthanum/boron instead of molybdenum/silicon. That drops the source wavelength from 13.5 nm to 6.76 nm. Compare that resolution against today's EUV lens sizes and an introductory Gd-based system lands roughly where ASML's hyper-NA EUV is aiming. If a larger lens is even possible on top of that, you're talking about pushing into the 3 nm realm. The hard part now with introducing a new lithography tool isn't just the physics. To improve resolution, you either shorten the wavelength or increase the lens NA. Both are monumental changes, and both require standing up an entire ecosystem. That wasn't always true for NA, but in the EUV world we live today, it is. There are already process techniques that split pitch and print much smaller features without changing the scanner: self-aligned double patterning (SADP) and self-aligned quadruple patterning (SAQP). Look at where even a large-lens Gd source sits on the scaling chart, and the raw resolution looks impressive. But a fab could still take a low-NA scanner available today and reach the same place by putting SAQP on top of it. That's also the challenge for high-NA EUV and why we see a split in philosophy between Intel and TSMC/Samsung. There are multiple ways to skin this cat. The low-NA tools already in every EUV fab can apply the same multipatterning tricks and beat high-NA resolution hands-down without taking on the ecosystem complexity of the bigger-lens tools.

🚨 TOK raises guidance by double digits on AI-driven demand for high-purity chemicals (adv. photoresist). Watch the materials vendors. They see the fabs’ wafer-start forecasts first—it’s a long supply chain: Wafer starts forecast. → Materials vendors. → Batch planning → Raw material sub-suppliers

28nm is not an easy node to jump into even if a fab buys the IP. It was the first true immersion lithography made device. It was the node that busted TI out of the leading edge club 20 years ago, and they are only starting to rekindle that now with the acquisition of Micron's Utah fabs (2021) and then Silicon Labs' design IP (2026). 28 nm has huge portfolio value in the foreseeable future. It will continue to make money for decades. When I see startups put out roadmaps saying they'll ramp in a year or two on 28nm I take it with a grain of salt. Easier said than done! This is a long game.

I like how imec reframed their logic device scaling roadmap around the track layout for front/backside power and now includes the embedded memory interposer. Its complicated to visualize it all.