Hamish Low

Research Associate @ IAPS -

[Romeo Dean @romeovdean]

Thank you to Hamish for this very well-written piece. For what it's worth, my overall view is that we should adjust the base rates downwards, because I see the reasons for things to go faster than ASML's history as stronger than the reasons for it to go slower. Especially the "larger economic stakes" point.

[Hamish Low @Hamish_Low5]

@jaygoldberg We did struggle to date the beginning of the Huawei-led effort well, and ultimately went for 2021 as the date we could best justify. Anything you'd point to for evidencing 2018 as the starting point? Would be quite helpful

[Hamish Low @Hamish_Low5]

When China can produce indigenous lithography machines is essential to the outlook for US-China competition on AI, yet despite its importance this question has attracted surprisingly few rigorous forecasts.

A new piece out today that I have coauthored with Joshua Turner and @romeovdean of @AI_Futures_ looks to take a first step to rectifying this. We use the experience of ASML developing lithography tools to develop base rates for when China will do the same. Doing so gives a baseline expectation of the late 2030s for extreme ultraviolet lithography, and the mid-2030s for deep ultraviolet immersion lithography.

In reality China’s indigenization experience will differ substantially from ASML’s development. Due, for instance, to substantial state support, poaching ASML employees or deploying increasingly advanced AI systems into R&D. The piece sets out these various arguments for why China’s efforts may move faster or slower than the baseline set by ASML, as well as placing where the current most advanced Chinese efforts at SMEE and Yuliangsheng sit.

It finishes with a set of research questions that would help us build better forecasts of China's indigenization process, as well as ones for how China's indigenization relates to larger strategic outlooks for the US-China compute gap.

[Hamish Low @Hamish_Low5]

@zephyr_z9 Ah okay gotcha, I wrongly had in mind a trade off between a 1980i and a Yuliangsheng DUVi, but see your logic there and agree

[Hamish Low @Hamish_Low5]

I know the export control context, you were the one that initially suggested they would continue to use a domestic/ASML mix which is what I took issue with. Obviously if you don't have any ASML tools available you would use domestic ones to plug gaps at every layer. But surely if you have a mix of ASML and domestic tools you use the ASML ones for the critical layer?

[Hamish Low @Hamish_Low5]

I would genuinely appreciate thoughts about why my analysis is wrong. Why do you think the gap from test machines into SMIC to volume commercial production would be so small? That has never been the case for previous litho platforms. Canon and Nikon struggle to get even their relatively mature litho platforms into fabs because of how tricky integration and yield learning is. What are you seeing that I’m not?

[Hamish Low @Hamish_Low5]

The closest analogue for these very small-scale, test volume machines is generously the NXT:1950i which ASML shipped in 2009, then the tool that China is currently using for 7nm is the NXT:1980Di which only shipped in 2015, and only began being used for 7nm at TSMC in 2018. Even if China progresses faster than ASML (despite working with novel platforms and weaker suppliers) that is still 4-5 years at best which takes you to 2030.

Lithography is really hard, even if you have some of the underlying tech figured out it takes years of iterative development from tools actually within fabs, churning out wafers, before you reach a tool capable of real high-volume, high-yield manufacturing.

Rubbish subsystems from inexperienced university spin-outs, poor reliability, or lukewarm adoption from fabs could easily push out timelines longer for actual 7nm capable DUV immersion tools producing commercial wafers at volume - which is the actual outcome we care about strategically not market share gains in advanced packaging or small-scale sale of test units into fabs.

[tphuang @tphuang]

We've known for a yr that YMTC was tuning up a fully Chinese production line for SOTA SSD this yr.

Multiple suppliers have reported getting revenue in 2025 from Arf lithography machines. Now, we have article confirming delivery to SMIC for 7nm production last yr.

Com'on, keep up with the times. This is just getting embarrassing if you are still one of those that think China is a decade away.

[Hamish Low @Hamish_Low5]

The record of state plans here is pretty terrible! Both the 02 special project and the Made in China 2025 plan fell way short of their goals in lithography, that is precisely why it has become such a critical bottleneck. This is also not an internal model, it is a pretty simple base rate calculation based on the closest reference class, which is ASML

[Hamish Low @Hamish_Low5]

@shashj We’ve just put out a new piece at IAPS on lithography indigenisation you’ll likely find interesting!

https://t.co/ef9Y8d4NJU

[Hamish Low @Hamish_Low5]

@s_pect_re That is why we don’t do this! We only use ASML’s timeline as a base rate to better calibrate a forecast, and then have a long section looking at all the differences between ASML’s experience and China’s that would cause you to move away from that ASML base rate

[Hamish Low @Hamish_Low5]

We do in fact have a section on where China’s efforts currently are that does flag up that ArFi tools are being tested by fabs. But it is a long way to go from 28nm level overlay tool being tested at a fab, to 7nm overlay capable tool being mass produced and deployed for HVM. That gap can easily take you into the 2030s

[Hamish Low @Hamish_Low5]

@zephyr_z9 Would be curious what your forecast for indigenous DUVi actually is, full line of SMIC N+2 running on domestic tools?

[Hamish Low @Hamish_Low5]

Logic folding is fundamentally an adaptation to a lack of EUV. I don't think it follows that the two would cannibalise on one another. Chinese firms can very much handling developing hybrid bonders + 3D EDA as well as litho in parallel. But Huawei will eventually need what better litho provides to raise density at the wafer level rather than just pursuing advanced packaging to compensate for weaker manufacturing inputs

[Hamish Low @Hamish_Low5]

We do discuss this in the piece in the section on alternative technological trajectories. It is true there are other paths that China can go, but these still require overcoming significant engineering hurdles and long R&D incubation periods to bring that tech to commercial maturity. There is also a reason that ASML + TSMC/Samsung/Intel landed on EUV as the key technological pathway, so it is the most likely one for Chinese firms to follow. It is very possible that they find other solution to the various technological challenges within EUV (power of the light source, overlay accuracy etc.) but I don't think the fact that they will do slightly different things changes the overall timeline trajectory in meaningful ways

[Hamish Low @Hamish_Low5]

Likely of interest: @Discoplomacy , @jordanschneider, @peterwildeford

[Hamish Low @Hamish_Low5]

Read the full piece at The Substrate: https://t.co/n7UJxtCMFx

[Hamish Low @Hamish_Low5]

While it’s true pre-training costs fall quickly due to algo progress I think this can easily be a bit of a red herring. Have been exploring this recently for an upcoming piece on when Chinese firms will train models as computationally intensive as Mythos, and the striking feature is that many Chinese firms, and indeed firms like Meta, Microsoft and Amazon easily have enough compute to train models with as many FLOP as Mythos likely used.

It is just that in practice these firms dedicate only some very small proportion of their overall compute to large pre-training runs. In some cases this is because this compute simply has commercial opportunity costs (you’re better off optimising reels/tiktok/ads) but for other cases it seems more like they don’t have sufficient human capital or the right organisational structures to take advantage of that compute in useful ways.

It doesn’t help that compute and algorithmic progress are highly intertwined, such that large amounts of compute are needed to conduct R&D at sufficient scale that you can find the right efficiencies and breakthroughs that make very large pre-training runs effective.

Europe’s issue would not just be building a large enough single cluster to train a ~2e26 model - that is fairly easy, just throw a couple hundred thousand blackwells together - it is building out a much larger than that compute portfolio and the right match of human capital and organisational drive (DeepMind has the former + lots of compute but perhaps not the latter?)

Does that mean Europe should just give up? Probably not. But it does mean that getting even to 6 months behind is really really hard and probably takes an extremely large quantity of compute.

One thing I’ve not seen people engage with much or good takes on is why xAI has seemingly rapidly fallen behind the frontier and what this means for other catch-up projects

[Rob Wiblin @robertwiblin]

To train a model 6 months behind the frontier only requires maybe a third as much compute.

The idea Europe as a whole could never muster a training run 1/3rd the size of 1 US AI company is crazy - it may be true, but it's a choice they'd be making, not a fact of nature.

In offence-dominant domains being 6 months behind still leaves you highly vulnerable.

But for defence-dominant ones, a model you trust that's 6 months behind could be very helpful.

(We're not sure which one cyber will be. It's probably offence dominant now, but eventually becomes defence dominant.)

[Hamish Low @Hamish_Low5]

I actually just do not understand the case against YouTube, it is clearly the most net positive social media platform and restricting it for under 16s seems just obviously net negative. Reels and TikTok sure, they are spiritual pollution, but come on, YouTube?

[Tom Harwood @tomhfh]

Including YouTube in the list of banned 'social media' websites is going to be the governments biggest single own goal.

More people watch YouTube than the BBC.

[Sam @Discoplomacy]

We have the most boring insular political commentariat in the G7