Hacker News Reader

Comment by lumost | Hacker News Reader

Company that solves hard problems in key industry gets high profit margins, news at 11.

The longer I've worked in tech, the more it becomes obvious why traditional business school approaches do not work in innovative fields. Between 2000 and 2016 Semi firms focused on outsourcing the expensive and high risk portion of their product to firms like TSMC, starting around 2016 a lot of companies looked at the risks and costs of further node shrinks in what they believed was a commodity business - shrugged their shoulders and decided they wouldn't chase after it.

In 2022, TSMC is effectively the only game in town for the latest generation of chips - and isn't letting up. Anyone who wants to "catch-up" is now stuck with the daunting prospect of spending ~100 billion dollars to re-create a commodity business.

View on HN

>Company that solves hard problems in key industry gets high profit margins, news at 11.

And yet this has to be repeated time and time again on HN. It is actually quite tiring. People used to look at Intel spending $10B in CapEx and felt the number is unsustainable. ( Part of the reason why Intel lost ), while TSMC is on track to spend ~$50B in 2023.

In the sci-fi novel Dune there is the Space Guild that became so advanced with shipbuilding that no other entity can possibly compete with them, and they internalized all the knowledge. This made all the noble houses have to deal with them.

Perhaps some technologies are just so advanced and difficult, with so many people and specialized knowledge, that firms like TSMC are unbeatable. No one can compete and their global power is assured.

There are many examples of such phenomenon throughout history. However, there are always economic incentives for those with knowledge to strike out on their own to increase their personal gain. That’s generally why such knowledge-monopolies end. An example of this is how textile machines were built in the US by memorizing plans from Britain.

Making a textile factory is a lot less capital intensive than a semiconductor fab.

Think in terms of available capital at the time, and availability of the requisite skilled labour and technical knowledge.

That lacks something of the "so what" follow up. TSMC is in the usual economic vice.

1) If they charge a lot and don't innovate, then there will be a huge reward waiting for whoever invests capital to compete with them.

2) If they are innovative and cheap, then they can fend off competition and nobody cares because they are innovative and cheap.

> Perhaps some technologies are just so advanced and difficult, with so many people and specialized knowledge, that firms like TSMC are unbeatable.

Except that modern chipmaking is so complicated that even one company can't contain all the specialized knowledge. My understanding is that the Dutch company ASML is the only producer of EUV lithography machines in the world, which the chipmakers depend on to make their most advanced chips.

> Except that modern chipmaking is so complicated that even one company can't contain all the specialized knowledge.

(And capital intensive). But then again, these words could have been said about wannabe private rocket companies (SpaceX) or squeezing inside car market (Tesla), that actually turned out to be possible. So I wonder if any startup/billionaire will start to serve some niche chip market, be it capital intensive, but with some clever tactics...

Falcon 1 vs currently flying F9 rocket and potential Starship is a huge difference after all.

Thing to not is that both examples wouldn't have been possible or where they are today without notable government projects/tax credits/etc.

With TSMC i think we'll see government churning up the market the moment the US expects it to be at risk from China or the like.

> No one can compete and their global power is assured.

until they get disrupted. For example, a wholly new approach to CPU design using lazers and light, instead of silicon, or quantum computing perhaps.

Or say a next door global superpower decides to get testy

Over a long enough time horizon eventually every dominant technology gets disrupted. But I'll bet that we'll all be retired before a credible alternative to semiconductor chips comes along for general purpose computing.

Even if quantum computers can be made to work at scale they'll always remain niche devices for running specialized algorithms. No one is ever going to calculate their income taxes on a quantum computer.

Yes, likely they will become something like the modern GPU card.

Because these semiconductors are approaching fundamental limits of physics, it is unlikely that TSMC can achieve this status with anything like the current technology.

I think that Sophie Wilson said that cost-per-part is rising since we surpassed 22nm, heat is more of a problem because finfet can't dissipate as well as planar, and reliability is declining at smaller nodes. If I needed a 20-year service life, I would not choose 7nm.

Sophie addresses the cost rise at 22:00.

https://m.youtube.com/watch?v=zX4ZNfvw1cw

It's much more complicated then that. TSMC relies on a company called ASML which itself relies on a company called Carl Ziess.

ASML makes the EUV etchers, and carl ziess makes the optical components needed to focus the light... both companies hold exclusive technological monopolies on their respective products.

In addition to this, Samsung and Intel are not that far behind the technology curve for making these chips... with Intel quickly shifting to include the capability of being a global foundry for making chips outside of their own designs.

Historically the landscape in the chip industry changes very quickly. I'm sure your analogy fits some technology in the world, but semiconductors is not it.

ASML has 4750 suppliers. The supply chain in the chip world is probably the most complex system in existence. Both astounding and frightening.

> the most complex system in existence

And by definition, anything that uses a chip will depend on this, I think? Cars, airplanes, game consoles, mobile phones... Heck, we have chips in pregnancy pee sticks these days.

So, I think all supply chains for "modern" things are insanely complex and effectively incomprehensible.

And this is not new, see the classic "I, Pencil"[0].

> Since only God can make a tree, I insist that only God could make me. Man can no more direct these millions of know-hows to bring me into being than he can put molecules together to create a tree.

[0] https://en.wikipedia.org/wiki/I,_Pencil

It would be highly interesting to see a "tree of suppliers" for a single product - including how many people each part relied on, history etc.

I would imagine a smartphone for example having an unimaginably large diagram.

Who makes the acid used to bleach the recycled cardboard for the box?

Seriously if you include the box, a phone touches so much of our economy it’s astounding.

Outside of specific fields/circumstances and/or a generous interpretation of just what counts as part of the supply chain, it’s basically impossible. The moment even the simplest items, like surplus vegetables grown and harvested by a slightly above subsistence level farmer on traditionally owned land, get carried in a truck, or even a bicycle, your supply chain explodes into a vast interconnected web of circular links. Which trucks trucked the parts to make the trucks to truck the parts to make the trucks that trucked the parts to make the trucks that… and so on.

I presume you’re referencing this?

https://twitter.com/foone/status/1302820468819288066?s=21

Almost. This was first:

https://news.ycombinator.com/item?id=24371703

The complexity of modern supply chains is in my opinion the strongest case for government regulations and against a lot of libertarian economic ideas.

It is impossible for the end-consumer, and even a lot of producers to know the provenance of every component in the supply chain. Even if every American consumer is against the destruction of rainforests, connecting the dots to palm oil cultivation and understanding the scale of ubiquity of palm oil in supermarket products is impossible. The only way out of this is government regulations.

I think that Eli Goldratt would not agree with you; number of connections does not determine the complexity of the system, it can be less complicated than system with smaller amount of connections. I believe these suppliers are pretty specialized, and that among those 4k+ there are "backup" ones too. Disclosure: I know nothing about chip making :)

Saying that TSMC relies on ASML and Zeiss is a bit reductive. If those were the ingredients of semiconductor success, then Intel wouldn't be in the position it's in.

Historically, the landscape of the chip industry has had many leading node players.

Now, it has ~1.75. TSMC & Samsung. With Intel and SMIC behind them.

The reduction to those 4 players, coupled with demand from mobile, GPUs, and ML accelerators, is pretty unique since the earliest days of the industry... when pretty obscene profits were also made for a long time (Intel, late 80s to 2000).

I'd call a decade of dominance a pretty stable industry.

>Saying that TSMC relies on ASML and Zeiss is a bit reductive. If those were the ingredients of semiconductor success

I never said this at all.

All I am saying is that the supply chain is complex and that to say TSMC is the sole monopoly holder of semiconductors is very very inaccurate.

ASML supplies the machines, TSMC builds the infrastructure around them to maximize yields/make the investment profitable. Both of them have built complementary IP that is nearly impossible to replicate without violating multiple patents and spending 10s of billions of dollars.

In case any readers are (like I was) torn on the question of whether such a patent based legal monopoly (or duopoly or oligopoly...) is justifiable compensation for the resources these companies undoubtedly invested, check out this short intro to a highly recommended book that deals with this issue:

http://www.dklevine.com/papers/imbookfinal01.pdf

Still even without the patents. The knowledge and know how is incredibly hard to replicate. China doesn't even respect patents and they can't fully replicate the associated tech.

That’s a great reference. A number of mainland countries were only able to catch up to Britain in industrialisation by suspending patent law for a while. Looking back I think it’s obvious that the result of this was hugely positive.

This also makes me somewhat sympathetic towards the lack of ip protections in China and other developing economies. Too often the patent system is used to kick away the ladder once one party has gained an advantage.

> Historically the landscape in the chip industry changes very quickly. I'm sure your analogy fits some technology in the world, but semiconductors is not it.

Semiconductors is as close as we've ever come to a natural monopoly. Aircraft are second. Boeing and Airbus are the only games in town.,

for the big aircraft, yes. there's bombardier in the smaller segment and it's doing quite well.

And don't forget about Embraer.

i totally forgot! yes, absolutely.

TSMC obviously relies on ASML. But a foundry is a complex beast. You don't simply place an order with ASML, set up a clean room and when the machine arrives press the on-button. There are A LOT of other processing steps that have to work properly (lithography is only one of them, albeit arguably the most complex one). What TSMC does is HARD. Just as TSMC cannot build their on EUV machine, ASML wouldn't be able to run their own fab.

Of course. Never stated otherwise. However what TSMC does is arguably more easily replicate-able by another foundry (say Samsung or Intel) than the EUV machine.

That being said, my main point wasn't to say that ASML has an exclusive monopoly over EUVs and therefore all chips... far from it. I was only illustrating the complexity of the supply chain. Many different entities have monopolies and the landscape constantly shifts.

As ASML CEO once said "Dont ask me about how well our EUV TwinScan is working, ask TSMC, they know more about our machine than we do."

TSMC made the long term investments in R&D and expertise and are reaping it's rewards.

Just like the short scene spoken by the Bene Gesserit in the 2021 film:

B.G: "..our plans are measured in centuries; we have other prospects if he fails his promise." .. "On Arrakis we have done all we can for you. The path has been laid, let's hope he doesn't squander it."

One thing that gets overlooked in pop semiconductor talk, and casual tech discussion in general, is that not everybody needs the latest and greatest. There may be one or multiple companies making the futuristic chips, but there is also market room for smaller firms to produce last gen products and clones of older tech. As long as it is cheap and commoditized and available, there will be consumers. Think of Small Soldiers, when surplus missile cpu's find there way into toys because they are technically available and nobody wants them. At some point, things are good enough, small enough, and cheap enough to get mass produced by shoddier and shoddier participants. The market can support more than TSMC.

The thing with semiconductor tech, though, is that not only are new nodes faster, they also use less power and (eventually) are cheaper as well. Don't invest in new nodes, and it'll cost more money for the same number of transistors.

It's not as simple as that. You wouldn't make chip with less than a 100 million transistors on TSMC N5. Chips have not only maximal size, but minimal size as well. And you shouldn't forget about horrendous cost of masks. If you need a lot of very small chips for something simple you are better off using something like 28nm or older node. At the end total cost per chip is what matters for such customers, even if cost per transistor is higher. And, IIRC, cost for transistor is currently lower on N6 than on N5.

Power usage gets passed onto the consumer. People buying $25 tablets probably aren’t evaluating yearly power costs.

Or are you saying less power in manufacturing? Like I said, at some point cheap enough is cheap enough.

Battery life. Thermal constraints.

That is not going to negate the existence of cheap chips. Those are luxury features of higher end components.

In Dune, the Spacing Guild held a monopoly on space travel because their navigators knew how to use the Spice to fold spacetime. Later in Frank Herbert's books, the Ixians replicate this ability technologically, ending their monopoly and the guild as an entity.

The lesson from Dune is that even the widest moat can be crossed.

The Dune saga ends with Paul gaining the throne. No one knows what happened after that.

Herbert spent years improving Dune since no one wanted to publish it. He shined it to a perfect diamond.

When Dune was published, it was a good enough success to warrant sequels. Herbert wrote those sequels quickly, and they did not get any polish. They are not diamonds, and no one ought to extract lessons from them.

You have reinvented the economic theory of the natural monopoly. Notice that it's still possible to come (close) to the technological frontier, just not economically rational, and that if it costs under a billion dollars it's definitely not a natural monopoly.

That is one of the purposes of actual medieval guilds. To keep specialised manufacturing knowledge secret so the guild members can charge more.

Intel isn't ballasted so much by chip fabrication for phone handsets. TSMC is. The few years of stagnation in the client computing segment coincided with early delays in the 10nm transition. It's always easier to lay out big money for long-term projects when your business is strong, but that just wasn't the case for Intel around 2015 or so. Handsets were booming, as was parallelized GPU computing. TSMC was a natural beneficiary due to its customer base.

Intel could have gotten into higher-end GPUs or AI processors; they have the demand, a GPU team, their management just wouldn't give them the physical space in any products to add it.

They actually developed a discrete GPU starting around 2008, even demoed it a bit, but ended up canning the project

https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)

IMO this was plagued from the start, including the HPC accelerators (Xeon Phi or whatever those were called). Intel put too much emphasis on x86 compatibility and even then couldn't keep up with Nvidia, neither in performance nor software support, except for a few edge cases. Even back then, if you want to build a compute coprocessor, it needs to be a thoroughbred optimized for throughput at various floating point precisions depending on target market. Nowadays the entry bar is set at being able to reuse your ALUs at multiple bit widths to boost both NN and classic HPC applications with the same silicon.

> Intel could have gotten into higher-end GPUs

They're supposedly working on it with Xe-HPG DG2. But knowing Intel and their shenanigans, I'll believe it when I see independent coverage and benchmarks.

But notably, it’s being manufactured by TSMC...

I could never understand why Raja Koduri gets all the credit and press. ( Even in his ATI / AMD days ) And now coming close to 5 years at Intel. He has yet ship a single Discreet GPU.

We will see how Xe GPU will turn out.

> TSMC is on track to spend ~$50B in 2023

Yet, if history is any guide, then at some point a tech disruption will turn everything on its head and there is no guarantee that the incumbent players survive.

Quantum / optical / bio chips? Nothing on the horizon anywhere close to market, but there is no fundamental reason to think silicon is the end game.

I think this is perfectly fine for TSMC if they invest into these moonshot projects. They certainly have the means!

It's not like someone will invent these techniques in their basement, like you can with software. You need serious investments.

Samsung Foundry is still in the game, and Intel hasn't dropped out yet.

They're both trailing TSMC at the moment, but not by a whole lot more than in TSMC has trailed Intel in the (somewhat) recent past. Samsung allegedly will have 3nm in production early this year while TSMC 3nm has slipped to late this year. Remains to be seen how that actually ramps, but barring a significant problem for either, they will be close.

TSMC I think has tended to have an edge at the same node size name recently, although that can flip. Samsung will have GAA while TSMC will be on FinFET for example, so the 3nm battle will be interesting.

I don't think TSMC has quite emerged as the sole survivor just yet. Although it's also not clear that the industry trend for consolidation and competition gradually dropping out has changed to a more stable equilibrium.

> Samsung allegedly will have 3nm in production early this year while TSMC 3nm has slipped to late this year.

You're making the mistake of taking their process names at face value.

Samsung 3nm is less dense than TSMC's 5nm which has been shipping since 2020. They are more than 2 years behind, as they still haven't matched TSMC 5nm even when their new 3nm launches.

When did this whole “process node isn’t actually the size” nonsense start?

afaik 45nm->32nm, so 2007->2010, because that hit the limit of "simple" scaling.

> You're making the mistake of taking their process names at face value.

You're missing where I said in the next paragraph you snipped off that TSMC has been ahead at the "node size name".

> Samsung 3nm is less dense than TSMC's 5nm which has been shipping since 2020. They are more than 2 years behind, as they still haven't matched TSMC 5nm even when their new 3nm launches.

For high performance logic, drive currents, leakage, switching power, metal resistance, etc can all be equally or more important than density.

The only thing I'd add is "key industries" would be highly capitalized industries. Chip making is the largest scale, most highly capitalize industry in human history. Chip plants are the largest buildings in the world, suppliers for chip manufacturers are themselves large and highly capitalized.

True, other good recent examples include Moderna, and Biontech.

How does it actually compare to other industries, namely defense, aerospace and oil. These are all capital intensive, and if you want to be generous re the term "building", some of those refineries and shipyards are gigantic.

Communist states managed aerospace, defence and oil. But they failed to make modern computer chips even after trying, instead they started importing computers from capitalist countries.

Even worse: There was an export embargo for computers by the West, so the soviet union had to reverse engineer western chips and manufacture their own.

Inspired by an Apple II https://en.wikipedia.org/wiki/Agat_(computer)

clone of the Intel 8086 CPU https://en.wikipedia.org/wiki/K1810VM86

> How does it actually compare to other industries, namely defense, aerospace and oil.

The $12 billion given as the cost for TSMC's new Phoenix fab would buy you approximately 1 new Ford-class American aircraft carrier or a ~10% share in Canada's oil extraction industry.

Chip manufacturing really is the industrial big leagues.

> 1 new Ford-class American aircraft

And that's the bullshit, $30,000-toilet seat, parts-manufactured-in-every-congressional-district, Defense Contractor Special™ price. It should really buy you three or four aircraft carriers.

Yeah the Ford-class is obnoxiously overpriced. There is no way the key metrics of an aircraft carrier (sustained sortie generation rate, survivability, long-term maintenance) improved by the leaps-and-bounds necessary to justify tripling the price compared to a late-model Nimitz-class.

> tripling the price

The last Nimitz-class carrier, CVN-77, cost $6.2 billion and was ordered in 2001, while the Ford cost $13 billion and was ordered in 2008. After adjustment for inflation in that period, it's only about 1.5x as expensive.

Two comparable industries are aerospace and nuclear energy (nuclear facilities cost 2-3x of a fab and nuclear weapons are on a different level of expense).

Nuclear weapons are a lot easier than cutting edge semiconductors.

Maybe the fact that TSMC files around 1000 patents a year has also something to do with the lack of competition.

https://www.taipeitimes.com/News/biz/archives/2021/03/15/200....

> Anyone who wants to "catch-up" is now stuck with the daunting prospect of spending ~100 billion dollars to re-create a commodity business.

Is this cause for concern given the risk of China taking Taiwan?

No, $100B is a tiny, single digit fraction of NATO military spending. I’m more concerned about the time and knowledge to spin up more manufacturing capacity.

We can't spin it up until we build more fabs. And we need infrastructure for energy, water, and logistics before we can simply build a bunch of fabs.

> No, $100B is a tiny, single digit fraction of NATO military spending

Are you sure about that math? Total NATO spending is 1.2T.

$100B is ~8% of $1.2T. Single digit, yes. Tiny - guess that's more subjective than I thought.

I’ve heard it’s less of a concern because by the time troops got to the fabs they’d be useless to the Chinese.

sure but you still get the people while denying/disrupting your war enemy since the US and China may likely be in a shooting war if that were to happen

The problem is that they would also be useless to NATO countries.

I think it's safe to assume there's a plan to exfiltrate the critical information to reproduce what TSMC has accomplished in Taiwan

i doubt information is the only critical component in reproducing the plants.

Critical equipment, which are difficult to manufacture, is a huge part of the equation.

An interesting amount of that knowledge is inside various folks' heads.

There are methods to exfiltrate that knowledge too.

Are there? The US could not even exfiltrate the knowledge of where Osama Bin Laden was.

"Company that solves hard problems in key industry gets high profit margins, news at 11."

Well ... its' the company that develops power in the value chain, gets the big profit margins!

In this case, probably the 'hardest problems' are solved by ASML, in the Netherlands the sole provider of the key equipment TSMC uses to make their gear.

TSMC is the operating entity that uses their various bits of operational advantage, know-how and positioning to great effect.

It's almost like TSMC and ASML should be both talked about in every article, and that both Europe and US should be contemplating strategic initiatives in the operational part of leveraging the fact that their is a lot of money being left on the table in their home turf.

Edit: one of the hard parts about saying 'let's make the chips here' ... is that the 'stuff' that uses the chips are often made in China and Taiwan! Not always, but often.

You can see the effects of China 'creeping up the value chain' - it's hard to displace certain things if they control 'all the other moving parts, even the not very profitable one's' ...

Gee it sounds alot like strategy to manage risk doesn't pay out in a world where you have to innovate and get paid for it.

It's rational for chip making to consolidate as the capex grows. 5nm would probably not exist in volume today without a successful pure play foundry.

As a consumer, I just hope we stabilize with two or three leading competitors, such as TSMC, Intel, and Samsung.

ASML & EUV is a stronger example of a high risk gambit in an innovative field locking down the market.

Several of the TSMC top executives were educated at traditional business schools so at least some of those approaches seem to be working.

https://www.tsmc.com/english/aboutTSMC/executives

Having diverse experiences and educational background on the team is usually helpful in avoiding groupthink and strategic errors.

The fact that so many materials science PhDs are on their board is very interesting.

I'd go as far as to argue most chips TSMC produce, especially mobile, are no longer a commodity. In consumer devices at least, people want TSMC vs say, Samsung. Their last couple efforts for Qualcomm have been quite horrendous. So much so that they're cutting a TSMC 8 gen 1 release soon.

Makes one wonder when "listening to business school majors" will be taken into economic school books as "value destructing behavior".

Or convincing a government to bankroll you

That was true in the early years of TSMC. We're so far from government subsidies now that I think it's a moot point.

I mean for a new player.

Likely this will come from multilateral government forced break up of TSMC if they get too big.

I believe the government of Taiwan seen TSMC as existentially important -- the company is so valuable and unique that attacking them is essentially an attack on the world economy.

What about a slow seemingly unarmed takeover? I think it's unlikely that the west would respond militarily, and that's probably what the Chinese govt counts on.

I mean, I guess if China convinces Taiwan to peacefully rejoin them, that's their right, and it'd really be none of our business, right? It seems like it would be quite a stretch, though, given recent events in the area. In any case, such a change would be slow enough that we could switch over gracefully.

Intel and Samsung are only a couple nodes behind, it would be expensive and annoying to change over but we'd manage it. I mean, go back 5 years, right? Computers have not massively changed since then.

You mean like Intel? https://www.intel.com/content/www/us/en/newsroom/news/intel-...

Do we really need so many cutting edge chips?

The world wouldn't come to an end or anything, but if we didn't have so many cutting edge chips, we'd generally be spending more money to use lesser chips to perform fewer and lesser tasks while consuming more resources to do so.

They reduce power consumption and thermal management needs. That means more compute per a given battery constraint or thermal envelope. High-end mobile phones and laptops benefit.

Thankfully web developers took care of that.

Is consumer computing really the major demand now and going forward. Or is it senors, automation, networking, and cloud compute?

And I can almost garuentee it's better for the environment if I continue using my device or an old one rather than upgrading to the new iphone soon.

Reply on news.ycombinator.com