Hacker News new | past | comments | ask | show | jobs | submit login
Two startups tried to catch up to makers of advanced computer chips, and failed (wsj.com)
69 points by prostoalex 17 days ago | hide | past | favorite | 92 comments




Economical and performant chip and jet turbine production are two areas where China can't really take any shortcuts. They have to put in a lot of time and resources to rediscover closely guarded secrets that are almost impossible to reverse engineer. They are dumping in the resources, so I'm sure they will eventually get there, but it might be a few decades (in both fields, they can already do performant or economical, just not both at the same time).


NO manufacturing line is easily transplantable or built from scratch. This is not limited to semiconductors.

My favorite story about this is a paint manufacturer building a brand new, state-of-the-art manufacturing facility to increase their capacity. Everything was controlled, monitored, etc. to the nth degree.

Only one problem--paint from the new line no longer stuck to anything.

Cue a crash debugging with all hands on deck. The problem? The new line was too clean.

Paint relies on the correct ends attaching to the wall and then the other ends forming a kind of network to protect those. But there is a tradeoff--the outer paint surface (exposed to elements) vs the inner paint surface (attaches to the wall) have different needs. If you make the inner paint surface too dense (molecules all vertical) the paints sticks great, but then doesn't protect against the elements. If you make the outer paint surface dense (molecules all horizontal) then the paint protects well, but then won't stick to the wall.

The solution: add some "dirt". Adding some very fine, very small inert particles can change the angle that the paint molecules connect and fix the adhesion issues.

The lesson: nobody knows all the things necessary to run a manufacturing line. Ever.


Ah, here's the circuit design equivalent of that story: https://twitter.com/dave_universetf/status/14734753486542684...


I suspect some or most of this secret sauce is more like institutional knowledge and not one single person knows how the whole system works (to produce high quality chips at scale). Or maybe there IS one person, but that’s the CEO/founder, haha.

Regardless, it’s similar to software companies claiming to run agile, but always reverting to the “get it done, waterfall deadline” mode. It takes good effort and detailed knowledge of pitfalls to avoid to actually make a less intuitive, but more repeatable and productive process work. (Agile’s been around for over 10years and people still don’t do good versions well)


> Agile’s been around for over 10years and people still don’t do good versions well)

Agile software development has been around since the early 90’ies. Scrum is from something like 1995.

The reason people fail to implement it so often is that project management sort or doesn’t work in software development, and it doesn’t work because you can spend two hours looking for a spelling error or a missing semicolon, and you can create Godlike code in half an hour, and you never know which mode you’re going to be in when you get up in the morning.

Waterfall doesn’t work either, for the same reason, and for all the adaptive reasons that lead to agile.

The best way to deal with it is to take the methods that make sense for your current project and run with that. 99% or the time that’s a kanban board, and maybe some time registration software if you’re unfortunate enough to work in a place that bills by the hour.

The worst part about the project management methods in my personal opinion is that they all want to sell themselves as the universal solution, in a world where no two projects are alike but you also can’t be bothered to have 90 different project models in an organisation, because then what’s the point.


In my experience agile mostly fails because of unrealistic management expectations: Deliver a given set of features in time and quality. Especially C-Suite and one level below has this expectation with no idea what agile really means. From there, it trickles down...


> and it doesn’t work because you can spend two hours looking for a spelling error or a missing semicolon, and you can create Godlike code in half an hour, and you never know which mode you’re going to be in when you get up in the morning.

That's ADHD (maybe modulo some external factors like scattered meetings)


> That's ADHD (maybe modulo some external factors like scattered meetings)

That doesn't seem like ADHD to me. Getting to a super-productive flow state is notoriously difficult to do reliably for everyone, and getting stuck on a stupid typo you can't find even though you know it is somewhere in the 50 lines of code you're looking at is likewise a fairly common experience.

Folks with ADHD may actually be better at achieving a hyperfocused state of flow, but the flip side of that isn't hunting for a typo, it is falling down a rabbit hole of yak-shaving, procrastination (productive or otherwise), and other distractions.

Somewhere in between is the experience of trying to hold more and more context in your head at once, resulting in a feeling like your brain has been pummeled and leaking out your ears, staring at some code that has become entirely illegible. This can be the result of ADHD-driven yak-shaving, but more commonly is the result of needing to comprehend huge chunks of a badly architected system before you can make a change with any confidence.


Coming back to this comment after a few days, I find that what I was trying to say with the last bit is that yak-shaving can be either intrinsic (ADHD) or extrinsic (badly architected, deeply intertwingled code with leaky abstractions).

I can assure you the CEO would be the last person to know how the whole system works, in case of semiconductor chips, to the detail sufficient to reproduce it again from scratch should the entire company suddenly resign somehow.


My grandfather figured out how to grow large haloid crystals, which were used in camera polaroid Carl Zeiss "Bernotar" filters. He kept the process a secret, which died with him.

My family gets contacted now and then by researchers trying to figure out how to do it.


Wait, Bernauer is your grandfather? He was killed in the war..


Yes, on my mother's side. He was too old for the draft, but got pressed into the Volkssturm (old men and boys) to defend Berlin, and was mortally wounded in it.

Two pictures I have of him, one with a telescope the other with a microscope. Both sides of my family were nerds :-)

He also led a team to Iceland in the 1930s to set some surveying markers to determine if Iceland was expanding, which would prove the continental drift theory. After the war, other researchers went back and confirmed it. I'm kinda sad he died before completing this, and so others got the credit for continental drift.


So interesting! Any pictures?



Indeed, there is no free lunch. Just look at Intel's missteps on their 10nm process. There's a company with decades of experience building semiconductors and $77 billion in revenue last year that took close to 4 years to recover from initial mistakes with their transition from their 14nm process to a 10nm class process. If Intel has that much difficulty, just imagine how many orders of magnitude harder it is for a small upstart without all of that institutional experience and resources to recreate that process step from scratch.


Well, to be fair their 10nm is everyone else's 7nm and they were trying to do it without DUV litho, for reasons?

However, their style of process/chip design is very hierarchical with HUGE teams to get the implementation out in the 3-4year cycles (there are multiple overlapping teams). This was driven over the successes of decades of scaling. Then it failed. Now they are behind. If they had been able to pivot early they might be parallel with rather than behind TSMC, but they couldn't admit it wasn't working. The large group structure got in the way.

We'll see if they can fix it.


I do wonder at what point Intel will no longer be able to function being the sole source of products running through their own fabs. Companies like TSMC can amortize the cost of fab construction and process development over a much larger swath of the electronics industry. Combined with the fact that Intel no longer has Apple as a customer for new products, they could be facing enormous change by 2030. AMD couldn't support their own fab, and I think that's still a risk factor (albeit much smaller than it was for AMD) for Intel if the cost of new fabs keeps going up by factors of 10 or more every generation.


Interesting. What about hiring people from the leading companies? Isn't that how a lot of knowledge transfer happens in all industries? Are non-competes enforceable internationally?


Knowing that you fire microdroplets of tin with a specific shape into a chamber at 50khz and hit it with two laser pulses is one thing, doing it is an entirely different one.

How EUV works: https://youtu.be/5Ge2RcvDlgw


> Interesting. What about hiring people from the leading companies? Isn't that how a lot of knowledge transfer happens in all industries?

Maybe it's not that easy. I have no knowledge of semiconductor manufacturing (let alone advanced semiconductor manufacturing), but it strikes me as one of those areas that might have thousands of very specialized crazy hard problems that all need to be solved just right to get things working. If you hire away some guy from a leading company, at best he might have a thousandth of that company's solution (and maybe that thousandth of a solution is only valuable in a path-dependent context with all the other solutions that leading company followed).

> Are non-competes enforceable internationally?

Doubt it. Though I suppose in some cases disclosing trade secrets for advanced technology my violate other laws.


> (and maybe that thousandth of a solution is only valuable in a path-dependent context with all the other solutions that leading company followed).

Yeah, that's a real obstacle, not much of a "maybe" about it.

I have a feeling that this, more than any other factor, is what is driving large-scale international tech IP theft. You have to grab stuff wholesale in hopes of either (a) teasing out isolated nuggets that happen to be applicable to your own efforts, or (b) recreating the surrounding path-dependent context (which doesn't have a well-defined boundary) in order to have a working solution.


On some other forums/blogs that I follow, there are Turkish information operations guys who post about the triumphs of the Turkish defence industry. Their lead guy says that Turkish companies are throwing million-dollar salaries at experienced American engineers with things like aerospace experience. Take that statement with a pound of salt. But I did look at a few Turkish companies that seemed to be hiring foreign expertise, and their domestic arms industry has grown by leaps and bounds in the past decade, despite their broad economic problems.


They don’t pay high salaries to foreigners at all. In fact they lose local engineers to foreigners, especially the likes of Airbus, who poach them. Their only success stems from US and foreign unofficial embargoes and sanctions which has forced them to produce domestically. Turkey used to produce airplanes even in the 1940s which were then shut down by US allied governments.

https://istanbultarihi.ist/578-the-nuri-demirag-aircraft-fac...


Stuff like this isn't just trade secrets, in the US it's often ITAR-controlled and requires a security clearance. There are CFD algorithms that are classified. If you possess this information and plan to move to (or even visit) China, I imagine US immigration will look at you very closely and not simply allow you to leave to go work for Chengdu Aerospace Corporation.


> If you possess this information and plan to move to (or even visit) China, I imagine US immigration will look at you very closely and not simply allow you to leave to go work for Chengdu Aerospace Corporation.

I don't think the US border control system works like that. If you're a citizen, you never need an exit visa. IIRC, as long as you have a valid passport, you can exit and enter as you please, but if there's probable cause to suspect you of a crime they can arrest you.

China's system, on the other hand, does work like that. IIRC, anyone with any authority or access to secrets isn't allowed to have free control over their passport, and they have systems for denying people exit permission.


Do you actually understand how this works or are you theorizing? The no-fly-list exists. Security clearances are tracked.


Anyone who has been through civics 101 in the US knows this. Unless there is a warrant out for your arrest, the US government cannot stop a US citizen from leaving the country.

Even if somehow you got on the no-fly list, you can take a boat. Or drive/walk across one of the land borders. Perfectly legally.

Security clearances are a red herring. The folks working in Intel/TSMC/TI/etc fabs don't have or need security clearances. This is private technology.


> Anyone who has been through civics 101 in the US knows this

Anyone who has been through civics 102 in the US knows that there's lots of ways to get your passport taken away, and without one you're not getting far. It looks like Canada's requiring either passports or fancy drivers licenses (only available from some states), and Mexico's at least as strict.


> Anyone who has been through civics 102 in the US knows that there's lots of ways to get your passport taken away...

But that's irrelevant unless one of those ways is to "possess [ITAR-controlled and requires a security clearance] information," as one of the ancestor comments stated. If that's actually the case, someone should be able to cite law/regulation or provide examples of the practice, but no one has.


Folks with security clearances need to inform their employers about any travel plans. If the plans are suspicious, they can lose their clearance.

But they won’t be stopped at the border. They might be fired when they return though.


I doubt it is that easy thing to do. What would motivate some key researcher to move from USA/Europe to China to work on stuff that they have already done instead of exploring new frontiers? Hardly any money can buy that. Ego? Maybe few people from the wishlist will take a new job title, but you need the entire list.


On top of that, doesn’t one of the leading companies have foundries in China already? Presumably that would be a source for poaching talent.


TSMC has some older tech plants but planning more advanced plant it seems https://focustaiwan.tw/business/202107300008


Leaders in their field usually do not want to live under a totalitarian communist government.


When the dude from Alibaba got disappeared for a few weeks, I said to myself, "Well, I won't be starting my next trillion dollar business there."


Shit, I won't even go there to study kung fu! A few years back I was considering studying Bajiquan ( https://www.studymartialarts.org/school/an-wushu-internation... ), and even then, back before they started detaining random Western businessmen as spies, I figured being an ex-military officer would probably be an unnecessary risk.


If you pay them enough most people don’t care. Chinese companies routinely poach Japanese talent in car manufacturing and Taiwanese talent for chips.


What about doing it the old fashion way and just stealing said information?


At least for semiconductors a lot of the magic is distributed tribal knowledge among engineers and the sophisticated equipment used to fabricate the devices (all major suppliers are in the US, Europe, and Japan). So theft is not really a complete solution, although it can help.


So if there were a war and the existing equipment got bombed, the world would basically be set 10 years behind on computing?


> basically be set 10 years behind on computing

That probably overstates the case a little bit since capacity is distributed around the world, but in some sense yes.

It's not particularly hypothetical. If China (PRC) was to invade Taiwan, that would potentially knock out a frightening proportion of the world's leading-edge capacity.


China launched more rockets to orbit than any other country last year. They make not take short cuts. But they do catch up quickly in terms of development.


Rockets are pretty simple compared to jet turbines, rocket production doesn't have as thick of a moat as jet turbine production.

> Developing jet engine is not an easy game unlike rocket engines as it's comparatively very compact in size & it requires advance engineering to develop it's crystal blades technology & maintain it's thrust to weight ratio & that's what make jet engine technology one of the most complicated technology.

From https://defenceview.in/how-difficult-is-it-to-build-an-jet-e...


Rocket engines are to jet engines basically what paper utensils are to porcelain. But with transition to reusable rocket stages, chances are that the "disposable" designs will get more complex if that extra complexity gets justified by not throwing away the whole unit after several minutes of operation.

> Developing jet engine is not an easy game unlike rocket engines as it's comparatively very compact in size & it requires advance engineering to develop it's crystal blades technology & maintain it's thrust to weight ratio & that's what make jet engine technology one of the most complicated technology.

That's a terribly worded and nonsensical claim (and almost reads like it was written by a ten year old). How could someone claim that rocket engines are "not very compact in size"? The RD-180 turbopump for example generates 170 MW of mechanical power, much more than any jet engine in operation, and in a smaller package to boot. Likewise, the HPFTP on the RS-25 generates around 50 MW of mechanical power, and does so in a package sized 0.55m by 1.1m or something like that.


why are they impossible to reverse engineer?

especially in case of jet turbines, due to their much larger size than chips?


I expect it is because jet turbine blades need to be precisely shaped to maximize proper airflow; spin around so fast that the shear forces destroy most materials; operate in an environment so hot that it melts most materials (which does not help with the shear forces, I expect); and be as lightweight as possible. Some blades have channels on the inside to run cool air through them to cool them down so they don't melt. I wonder if the egress of the air may be used to help keep the flow laminar, I thought I read something about that in experimental airplane wings. So you have fluid dynamics for the shape and a serious limitation in available materials. I know fluid dynamics is a bit of a black art, and I expect that metallurgy is somewhat similar. Especially at 2000 deg C and 3500 - 25,000 RPM!

Even once you have the blade to copy, you don't know why the features are the way they are. What bits are important? Which areas must have high tolerance? After that, there is the question of how to make it. You probably cannot just mass spectrometer it, put those ingredients in a vat, heat it up, and pour it into a mold. Is the order of ingredients important? The exact ratios? How quickly do you cool it? At what temperature? And how do you get those holes in it? How do you do it economically? You can't just 3D print these, and even if you did, it'd be way too slow.


I think they are also formed from single crystals


> but it might be a few decades

They could just hire the experts from the incumbents


The article said that’s exactly what they did from CEO to engineers and they still failed.

HSMC attracted a former top TSMC executive as chief executive. QXIC recruited dozens of experienced engineers from Taiwan, including from TSMC, with relatively big pay packages, according to former employees.


I wonder if they were playing both sides, ie getting a ton of money from China and TSMC to just lead them astray :)


Assuming they'll agree to work for them. I imagine many of the people who hold the secret sauce (especially from TSMC) wouldn't work for a Chinese company. I was surprised they managed to hire a former TSMC executive to head up one of the (now failed) companies.

Then again, there may be no "secret sauce" in the way we usually think about it. You might have to hire hundreds of people from an incumbent to pull in enough institutional knowledge to get things going. And even then it's not going to happen overnight.

I feel like the right people to hire, at any rate, aren't the executives. You need the right engineers, scientists, and researchers.


The process/material knowledge is actually pretty easy to keep under lock and key. No one knows the formula, or the key is in the equipment that whose production is strictly controlled. They need to get access to the all of the knowledge all at once, not just bits and pieces of it.


> They have to put in a lot of time and resources to rediscover closely guarded secrets that are almost impossible to reverse engineer

yes, they have to repeat all that work, they have to solve all those solved problems again because companies that solved the problems won't share their solutions.


>they have to solve all those solved problems again because companies that solved the problems won't share their solutions.

...because those solutions are the company's competitive advantage that China has tried to steal for years on end now? Not to mention in many instances those 'solutions' are owned by or access controlled by the parent country of said company, many of which view China as at the least a bad faith actor, and at the worst an inevitable opponent?

Pretending China hasn't engaged in the world's largest industrial espionage campaign over the last two decades, and then victim blaming the companies involved for not "sharing their solutions" with China is a perverse form of logic.


Why would a company share its trade secrets? It's one of the few differentiators left to most companies. Not to mention the national security arguments.


> Why would a company share its trade secrets?

indeed, I wish I could answer this.

> Not to mention the national security arguments.

it is an inteserting excercise to consider this from the rival's point of view


The rival's point of view is utterly irrelevant.


that's a terrible tactic. consider as an example playing chess while disregarding what your opponent may do.


Not just companies, but the Russians were pissed after China copied some of their jet military tech. They were angry enough that they stopped supplying them with military jet turbine parts for awhile, which really tripped things up for the Chinese because they can't produce their own at the same level (they can make fast, they can make economical, but they can't make both yet).


Modern semiconductors seems a lot like the F-1 engine on the Saturn V. Even if you had the blueprints for it, you couldn't build it without doing a lot of work to rediscover all of the missing details that were in the minds of the engineers and tradesmen who built it.

Also, (I didn't read the original article), but there's been some suggestion that some of these companies, like HSMC, were really just investor scams. They raised a whole bunch of money, promised a lot, but didn't actually know how to do what they said they did. Kinda like a Chinese Silicon version of Theranos. I think HSMC even had an ASML EUV machine, but never got to use it.


I have a theory. Not a good one but it's a theory

Any one organisation can only be efficient if the whole organisation can fit in one persons head. It must be possible to reason about changes in one area affecting another. This anecdotally applies to every software system i have worked on - once you don't understand how another part works decisions you make become divorced from reality

As such organisations must either simplify, so they are reason-able or they must split into markets (which seem to be very good at getting efficiency out of complexity) - or hire people with bigger heads ! (this does not mean a niche for AI - at least not yet)

I think what I am trying to get at is that someone at Intel must understand how it all fits together - and someone at Boeing etc


I think this is largely correct, although it helps to have many people understand it. When I worked at Micron, the training on the “Traveller Process” was mandatory and you could go very deep in it. There were a lot of people with at least a passing understanding of what everybody else was doing and in what order. The depth was not always there, especially about what could go wrong, and what that would look like. I started in CVD and then gradually moved my work to parametrics to try and figure out how problems were connected to measurements at each manufacturing step. I got into that role because I was an intern and generally not great at being told what to do and what not to do. I think if I had a normal role I’d have been pigeon-holed in one place with little visibility or impact on anything else. I imagine that I would have learned a lot more over, say, 20 years with full exposure to the entire process, but, I don’t know how someone could possibly develop that kind of knowledge being placed in a specialist role. I think the only people that knew the systems were the people that grew up with the complexity, as in, started early on as a specialist in a relatively simple process, and then that same role became a generalist role over the same function, which grew into a massively complex process that now consumes 100 different types of specialist that will never understand each other. Like, at Google, there will never be another Sanjay and Jeff. Nobody today will ever grow into the technical breadth required to understand that system. The parts will fracture and eventually stop working well with each other, until someone starts from scratch with more modern assumptions, but of course it will come from the outside because Google would never pay anybody to do that. To bring it back to the fab, I think the entire industry is close to this point if not beyond it already. It would be impossible to copy the incumbents, but I would guess that starting from scratch would yield a significant advantage, if only that more people in that team would necessarily have to understand more of the system.


It feels that an open source system (not too sure how you precisely define that) at least forces the "replication" problem. If no-one else can build your code there is a problem. If no-one else can contribute to your open-fab then there isa complexity boundary passed?


Not surprised: $2.3B is chump change in semiconductor manufacturing especially if you're starting from scratch. Double that, add a zero, and wait 5-7 more years.


So 50B? That's nowhere near what you need. Last year both TSMC and Intel announced they are spending a hundred billion each on new fabs (TSMC in April, Intel in July) and they have all the R&D already and good relationship with ASML (and both own a few percent of it, too).

But the US banned ASML from selling to China so you need to catch up with ASML and Intel/TSMC at the same time. Now, I don't know when ASML started researching EUV but they were certainly working on it in 1998 -- which doesn't mean it'll take twenty years to replicate it because surely some of it is public but still, it's a big, big project. https://go.gale.com/ps/i.do?id=GALE%7CA56970339&sid=googleSc...


Looks like ASML started planting the seeds of what eventually became EUV in the early 1990s.

https://news.ycombinator.com/item?id=28266708


Is Intel's & TSMC's spends comparable, though? They are both already executing at very high scale, so further investments need to be pretty huge to be worth it. A new upstart could likely start smaller. Not too small such that you can't develop any economies of scale, but significantly smaller than either Intel's or TSMC's planned operations.


why haven't alternatives emerged to ASML to help drive down the cost of building new fabs?

do you know what the cost breakdown is when building fabs?


Because the EUV machines that ASML build are the most complex machines ever built by humans and getting all the bits to work properly takes decades of R+D

e.g. watch this about the light source, its completely bonkers https://www.youtube.com/watch?v=5Ge2RcvDlgw


And note also that all the software that is needed to engineer those things is also not allowed to be sold to China, so that is also another thing they have to replicate. And by the time you catch up on the decades of R&D and trade secrets needed to catch up with where the software is at now, everything has evolved even further ahead. Keeping pace with semiconductor technology is a never ending treadmill of constant innovation.

Source: my company is one of those not allowed to sell this type of software to China


Wasn't the atomic bomb the most complex machine ever built? Hanford alone had more than 50 000 people working towards it, at Oak Ridge the K-25 building was 1,640,000 square feet...

Although we could mention the B-29 which actually cost 50% more than the entire Manhattan Project -- or, to be closer to present day, the F-35.


Interesting question, maybe relative to the tech at the time it was. But in absolute terms I don't think anything comes close to 'hitting 25 micron molten drops of tin that are moving at 70 meters per second with two co-ordinated lasers in a vacuum 50,000 times a second'. And that's just the very first part of what the machine does.


The Youtube channel Asianometry has a great 18min video "China’s ASML is Years and Years Behind" analyzing the challenges of building fabs and lithography tools: https://youtu.be/DtOyW-JpJjM

The channel's older videos have a lot lot of information about fab economics, though I dunno if there's one that has a general cost breakdown.


I don't know about the whole fabs, but the machines that ASML makes are incredibly precise and powerful. They are built and stay in vacuum, so even shipping them costs millions. And the laser that powers them is room sized; it would probably make a decent weapon for a battleship too.

So I think the cost is largely that every part must be engineered to perfection.


Where is Mill CPU? Been almost 10 years since they started talking about architecture, but nothing tangible emerged yet.


They're still going, but seeing them semi-complain/semi-get-cold-feet about VC funding on their forum makes me think they might not be totally in tune with reality.

From Ivan on the forum "We expected to go out for a funding round (and convert from bootstrap to salary-paying) last spring but, well, 2020.

Right now we’re trying to decide when to make a second try at it. The financial market for the likes of us isn’t back yet, and the virus rates are turning up again, which argues for waiting. But the economy and the future market is real iffy, which argues for doing it now. "

https://news.crunchbase.com/news/global-vc-funding-unicorns-... this suggests there has been a dramatic increase in VC money to be acquired by the right startups.

If I were the mill team, I would get their patents nailed down (which I think they have mostly done), then get some token open source projects out there to show what they actually have in their trousers. Then start giving some talks, public presence, things like that. Show what their toolchain and simulator can actually do.


I asked myself this about 2 weeks ago and went looking. As far as I can tell they're quietly working still.


I always appreciate updates on Mill. Their videos are great.


I have met so many people that think their semiconductor manufacturing needs are exempt from the supply chain issue or backlog.

Typically they all say "We are using this other nanometer size, and so thats not where the backlog is"

is there any truth to that statement?


Yes.

The "size" you are talking about is the "process node". It's basically a shorthand for how large the transistors are. These nodes are not fungible, once a chip is designed and a mask set is taped out it must be fabricated on that specific process node. Even if you have a chip designed on an older/larger node, it cannot be fabricated on a newer/smaller node without significant expense and delay.

Older nodes are generally in lower demand, although strictly speaking they are a tighter supply bottleneck, since nobody is incentivized to build fabs at older process nodes.

But the picture is even more complicated than this. Virtually all semiconductor fabs specialize in certain kinds of chips. Even at roughly the same process node, a DRAM fab cannot easily retool to make NAND, a NAND fab cannot easily retool to make logic (CPU, GPU, FPGA...), a logic fab cannot easily retool to make MEMS or image sensors, and so on.

In fact, not only are individual fabs tooled to make particular kinds of chips, but individual fab operators will generally have their own "design rules" at a certain node. So suppose that you have an ASIC you've been making at TSMC 22nm; just because GlobalFoundries has capacity at 22nm doesn't necessarily mean you can just have them start making your chips.


Your reply looks carefully written but "is there any truth to that statement?" "Yes." confuses me. Maybe the parent's question is not well formed. Would the best direct answer be?

I have met so many people that think their semiconductor manufacturing needs are exempt from the supply chain issue or backlog. [because they are not on a leading node]

Their semiconductor manufacturing needs are not exempt from the supply chain issue or backlog. [for the reasons you explained]


I meant "Yes, there is truth to the statement" in the sense that some market participants are "exempt from", or more accurately "not constrained by", supply chain issues or backlog problems because their specific process nodes are not capacity impacted, and the fungibility problem acts to ensure it stays that way. I actually did not mean to say that leading nodes are the only ones impacted; that's not the case.

The landscape can be surprising at a glance because there's no obvious rhyme or reason to who and what is capacity impacted. Before fabs get greenlit for construction, the beancounters do demand forecasts many years out into the future (getting the first chips out of a fab is a labor of many years). If those forecasts were high, there is overcapacity and prices will settle relatively low for that node. If they were low, then there will be capacity issues.

Over time, trailing nodes will get "arbitraged" because chip designers for new devices that are cost-sensitive or not performance-critical will preferentially select cheaper nodes, but acute changes in demand like we've seen from COVID can still cause problems.


I would not be the least bit surprised if it cost $150B to $250B to capitalize a new 'state of the art' semiconductor fab.

I'm a bit surprised they didn't shoot a bit lower and just try to replicate ASML's machines. Those are one of the unobtainium pieces that cost so much in a new fab.

In my experience the only way to compete with a long established player is to learn the business from the ground up and expect a loooong learning curve. What that might look like with semiconductors is building an ingot plant that can sell ingots that the established player can use. Then start slicing them and selling passivated ingots for a bit more margin, then use "marginal" ingots to make large node (like 140nm or 90nm) chips in the good areas to develop training around the workflow. Ideally those chips have some nominal market value and sales can act as a 'discount' on your investment (say it costs you $1B to make the chips and you sell them for $500M, you've basically invested $500M in learning the work flow instead of a full $1B).

Expect this to take 10 - 15 years minimum, 25 years more typically. Invest in university programs to train process engineers, physicists, and circuit designers. "Seed" several universities with small 300mm fabs to give students real experience rather than just "book learning." Aggressively weed out people who cheat to get through school.

Its a really really huge deal to take on this task. It was insanely complicated in the 90's and now it is even more so. The only thing China has going for it is that it doesn't need to repeat the mistakes made in the last 30 years that other semiconductor companies made and so their path is a bit more "straight line." But if experience is any guide, there is no shortcut to doing the work, no matter how much money you spend.


> Invest in university programs to train process engineers, physicists, and circuit designers. "Seed" several universities with small 300mm fabs to give students real experience rather than just "book learning." Aggressively weed out people who cheat to get through school.

This requires a deep cultural change. Because of high pressure cramming exams that favors rote, Chinese students are incentivized to just cheat from a very young age.

Also, keep in mind that no matter how much you invest, studying at an American university opens the door to working in America and getting a western passport, something more and more Chinese are desperately trying to do.


>Also, keep in mind that no matter how much you invest, studying at an American university opens the door to working in America and getting a western passport, something more and more Chinese are desperately trying to do.

I'd heard the opposite; that in recent years fewer Chinese studying in the US are staying, and that more who had stayed were returning to China, because of greater opportunities.


For a somewhat biased take on the situation, Asianometry on Youtube talked about Hongxin last year, this guy's choice of topic coverage is all over the map, but focuses on Asia, Tech, and Politics:

https://www.youtube.com/watch?v=OZSvDYDfd78


Really good video and channel!


Making EUV chips is hard:

https://semiliterate.substack.com/p/why-cant-china-just-reve.... (Note, “BLUF” at the start means “Bottom Line Up Front.”)


Not even Apple can do it.


Feel if ppl think completely out of the box, they may get enlightened and get good at it much sooner and cheaper …




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: