[Veliladon]

You would be correct if investment were the bottleneck. It's not.

The bottleneck on advanced processes is how fast ASML can build its EUV fabrication machines and they're extremely complicated pieces of kit with lots of specialized parts that are also extremely complicated pieces of kit. Even with all our modern production ability we're limited in how quickly we can assemble these machines that can actually do the lithography and everybody wants them.

ASML currently has a backlog of 100 machines. Intel, Samsung, TSMC, they all want them. They sell for $200m+ a piece. Everyone has the cash for them. ASML just can't produce them fast enough.

[bhouston]

Why can't ASML scale? They are siting on $20B of orders. That is insane. What is ASML's delivery throughput of these machines?

[hoorible]

My understanding is that it could, but these machines take a really long time to build. Probably a simplification, but it’s likely not unfair to say that by the time they scaled 3nm production, we’d already have moved on to 2nm, etc

Asianometry YouTube channel did a few really good videos on chip fabbing and the unique challenges. Scale is addressed IIRC. Highly recommend.

[Veliladon]

They make about 50 a year. One of the parts in these machines is a 30cm mirror that is the flattest that humans can make with our current level of technology.

If you have a way to churn them out please let Zeiss SMT know.

[cortesoft]

Do whatever they are doing now, but twice?

What in the process wouldn’t work with this method?

[jsnell]

They can't simply choose to do the same thing twice. They might need to double the machines, facilities, and workers trained for the production of those components. Oh, wait. Not just that component, but every other specialized component of the lithography machine, and every specialized input for one of the specialized components. If there's a single one that can't ramp up production and they can't find a substitute for, there's no point in any of them doing that investment.

Presumably a lot of the machinery needed to build these components is incredibly specialized too. That machinery won't be purchasable off the shelf. It will be ordered custom with long lead times, and (again!) could have bottlenecks in production that make it simply impossible to deliver in large quantities.

I'd bet (with no supporting evidence, just a gut feeling) that this industry needs to know five years in advance what the demand for chips in a given process node is going to be just to set up the supply chains to manufacture the right number of lithography machines. If that estimate is wrong, it'll take so long to build up that infrastructure that you might as well not bother, and just try to get things right for the next node.

[sdwr]

I got first-hand (virtual) experience with this playing Dyson Sphere. Late-game tech requires a whole web of dependencies. Doubling from 1x <shiny item> per minute to 2x <shiny item> means doubling the entire supply chain all the way down the line. It's not just doubling the factory making the thing, it's doubling the inputs, and the inputs to the inputs.

For example, the mirror might require extremely high purity silica, made in a top-notch cleanroom environment. To produce 2x, you don't just need 2x the raw material, you need 2x the air filtration systems, purifying machines, trained staff.

Which is fine, that large footprint can be doubled! But maybe this whole process is only going to be state-of-the-art for 5 years, and after that won't command top dollar.

[jb1991]

Peak HN. Surely a casual brainstorm by a casual reader of the thread somehow unlocks all of the challenges that ASML is facing.

[cortesoft]

I figured there was something I was missing, I was asking to try to get an answer. Why won’t the naive scaling method work?

[tekno45]

3nm.Do()

[tcmart14]

I feel like we are missing a GOTO in this one

[lr1970]

> What in the process wouldn’t work with this method?

Testing of EUV lithography machines is very complicated and slow process. You cannot scale it without magically cloning all the test engineers and technicians.

[gamblor956]

Doing twice as much of it wouldn't scale.

Making a super-flat mirror is extremely difficult and requires state-of-the-art facilities, and state-of-the-art equipment. And highly trained, experienced employees. And there aren't enough of any of that to scale rapidly.

It's not just about scaling 1 input. You have to scale the entire supply chain.

[cortesoft]

Right, when I said “do everything twice” I was including the entire supply chain.

If there is enough money sitting on the table, it seems like it would be worth it to do that. But it sounds like it would take too long with uncertain returns?

[manuelmoreale]

Maybe there isn’t enough money on the table? They might be sitting on 20B of orders but maybe doubling the entire supply chain would cost a lot more? I’m sure there must be a compelling reason.

[MagicMoonlight]

But then they’d get less money

[notaustinpowers]

I'm not familiar on the throughput, but their EUV lithography machines are some of the most complex machinery humanity has developed to date. We're lucky this even works, let alone dream of scaling it.

[renewiltord]

The answer is quite simple, actually. It's a hardware business so they have to right-size their equipment to projected demand and the tool-up time in specialty hardware is long because of the need for specialty components.

i.e. they need to project out demand x years in advance because their things are expensive and slow to build, and their components vendors have the same problem with the additional problem that their components vendors only have one customer for some of their products: them.

There is no other equivalent use for much of what they need so if they fuck it up they've got too much equipment too fast.

This is a classic hardware business problem.

[michelb]

Here's a good docu about ASML that shows some of the challenges in making these machines: https://www.vpro.nl/programmas/tegenlicht/kijk/backlight/asm...

[Timshel]

A bit late to the party but the 100 backorders is in fact quite small if you consider they make 50 a year (as mentionned by another comment).

Even if it was possible to double the production for free in a year then without new orders when this come online you only have 50 orders left to process ...

If you look some plane production backorders are in the decade I believe atm. Looking at the A320 orders you can see how they ramp up production (and this ramp up might production line switching from aircraft type not new lines) : https://en.m.wikipedia.org/wiki/List_of_Airbus_A320neo_famil...