TSMC Risk

  Rather, the only thing that will truly motivate TSMC to take on more risk is competition. 

If big tech wants TSMC to increase capacity drastically without TSMC having to take all the risk of CapEx, then they can pre-pay for wafers from TSMC.

They can each give TSMC $10b now in cash and guarantee themselves wafers in 2-3 years that it takes to bring a new fab online.

TSMC is rightfully conservative. If they commit to spending an extra $30b on a fab now that won't make a single wafer until 2029, without any guarantees from big tech, they're stupid. Who knows if the demand will still be there (my guess is yes, but who knows?).

In my opinion, I think it's getting close to this. Nvidia will surpass Apple as TSMC's biggest customer this year. This will start a war for TSMC wafers in 2026 in my opinion. When you have that much demand, customers will be forced to pay well in advance.

There is already a war for memory, silver, copper, energy. No reason why chip production won't be next.

The more complex the process becomes, the harder it is to have equivalent competition so you're bound to have issues where a single company's investment decisions have widespread impacts.

My perspective on the China risk differs some, though. China wouldn't benefit much from attacking TSMC. This is the first time I've heard anyone suggest that they might. At best they'd like to have it in-tact if they do take Taiwan, but there have been talks about machines being rigged to explode to deny them from China, or the US striking them in that scenario.

If neither we nor China get to work with TSMC, then we're still ahead in relative terms. If China did attack TSMC, they set the norm that the fabs (including their own) are now a fair target which would be a larger disadvantage for them than it would be for us since China's physical power projection remains pretty regional outside of Chinese nationals abroad engaging in sabotage.

That is one of their biggest weaknesses. Yes they have a lot of manufacturing capacity and a large population with many talented people, but in a way we have lent them the power to scale up to see what they'll do. We are already putting some pressure on that scale now that they've shown who they are, but if it came to war it would be very doable to start reversing their scale and their capacity to do the same to us would degrade as ours increases.

Even if all the AI in the world was destroyed, that's how it would play out. The problem is that Taiwan remains in close proximity to China so similar to Ukraine it would likely come down to how long they're willing to throw everything at it.

If Russia and China wanted to be powerful, it's just idiocy to show the existing superpower that you cannot be trusted with the power you have. If they fancy a merit based society, they forgot that merit isn't omnipotence and you still need the right ideas to be at the top to accompany the merit. For China maybe they need AI for that alone, but western societies at least have ways for the right ideas to make it to the top without the strict need of AI.

> Anthropic Chief Executive Officer Dario Amodei said selling advanced artificial intelligence chips to China is a blunder with “incredible national security implications” [...] “I think this is crazy. It’s a bit like selling nuclear weapons to North Korea.”

This is all a smoke screen. He knows very well that China can and will develop their own hardware to train AI models (and in fact, they are successfully doing just that; e.g. the recently released GLM-Image was trained on their own silicon). His only objective here is to slow them down enough so that they don't eat Anthropic/Claude's lunch releasing open-weight models that are increasingly competitive. But he can't just openly say "hey, we don't like that they release open weight models for free", so he's engaging in the AI version of the "think of the children" argument.

Anthropic's whole modus operandi was always pretty much "we should control this technology, no one else". It's not a coincidence they're the only major lab which has not released any open weight models, they don't publish any useful research (for training models) and they actively lobby the lawmakers to restrict people's access to open weight models. It's incredibly ironic that Dario is worried about (I quote) "1984 scenarios" while that's exactly what his company is aiming towards (e.g. giving Palantir access to those models is not "unsafe", but an average Joe having unrestricted local access is an immediate 1984-style dystopia).

You hedge this risk by migrating non AI workloads to new architectures that need 80% less energy. You don’t need TSMC chips. AI arch itself is also weak right now. It can be refactored to work distributed and with much more reasoning power on much less hardware.

Basically the current stack is primitive, so we waste 80% of the juice and then worry about how to power our inefficient mess. That’s the wrong set of questions.

No one will listen to me until I build it myself because everyone thinks they know everything.

I really don't understand why companies are ignoring intel's foundry services... for the first time since probably the 2000's, intel's 18A nodes are significantly ahead of what TSMC is offering. Apparently they have capacity and are demonstrating wafer production with their own chips.

It seems wholly illogical that Apple would get refused wafer volume by TSMC and still refuse to give volume to intel foundry services. When you layer on geopolitical factors and national security implications + the fact that Apple is a US company - what reason could they possibly have to turn the shoulder to intel's foundries?

If Taiwan ends up imploding in any of the numerous ways we are aware of today - and which this article adds to - I think there are exactly zero reasons to feel like this couldn't have been avoided.

"AI has a physical dependency in Taiwan that can be easily destroyed by Chinese missiles, even without an invasion" ?

Arguably false. Why do you think the US has encouraged TSMC foundries, now inside Arizona ? It's obviously to protect against the scenario that China takes Taiwan. In that case, give it 6 months or less for US TSMC foundries to produce the finest. China taking Taiwan will likely not result in the CCP getting any technology, certainly Taiwanese have "contingency plans" to vaporize all tech in the event they are invaded.

> “I would ban the sale of semiconductor manufacturing equipment to Chinese fabs”.

Ahem, ASML, who makes the manufacturing equipment for TSMC is a Dutch company, not a US one.

> Rather, the only thing that will truly motivate TSMC to take on more risk is competition.

Maybe I missed something, but if Google, Microsoft, Apple, OpenAI, etc want more capex on fabs, they can front that money themselves?

From a NatSec perspective, TSMC isn't really a bottleneck - most weapon systems use SoCs and microcontrollers that can be fabbed on "legacy nodes" (ie. 28/40/60/90nm) or 14/20/22nm nodes, and compound semiconductors.

The ability to mass produce a Pascal or Volta comparable GPU or Apple A11 comparable SoC is all you need for more cutting edge systems.

Power Electronics and Compound Semiconductors (GaN, SiC) have historically been the biggest bottleneck.

The bigger risk for the TSMC-China aspect is TSMC's planned exit of GaN foundry production by 2027. Most Chinese manufacturers still depend on TSMC-produced GaNs wafers instead of domestically produced GaN vendors due to reliability concerns. China will probably end up matching TSMC's specs for compound semiconductors in 4-7 years, but that implies that the Sullivan Doctrine still holds and is a loss for China.

Every other country with compound semiconductor production capabilities at scale (US, Germany, France, South Korea, Japan, Israel, Russia, India) either limits their exports or cannot export them to China without facing sanctions from other buyers (primarily Russia as India does not allowing commingling of SKUs for defense vendors who sell to Pakistan/China as well, and Russian vendors are members of India's EW and DEW program).

If a US-China War was to arise, worst case we would probably see a mass reversion back to 2018-22 level technology, which isn't the end of the world.

Additonally, the China-Taiwan situation is orthogonal to semiconductor dependency.

There are some near ready foundries in the US and in EU, not to mention South Korea. It would take a few years to catch up of course.

What I worry more about is the full lock-in of TSMC production capacity by nvidia/apple/amd/etc for their chips on their latest and greatest silicon process (aka the best in the world). There is 'no space' for performant large RISC-V implementations or other alternative (and it will require several iterations and mistakes will be made)

Perhaps there is an unpleasant but more expedient geopolitical, temporary solution: Taiwan needs its own nukes. Strength seems to be all that PRC respects.

I remember Elon saying in an interview recently that the only piece of the vertical stack he doesn't own is chips.

I strongly suspect some sort of fab built by Elon associated companies will be announced soon. Almost all supply can be bought by Tesla and xAI.

It makes sense, IF he can get the tech to work at the bleeding edge. But he seems to be quite good at this.

Im always surprised people forget that Intel exists and still has high performance nodes (just release panther lake on their newest node). They even have a plant in Ireland.

> AI has a physical dependency in Taiwan that can be easily destroyed by Chinese missiles, even without an invasion

Taiwan has missiles with the range and warheads to strike the three gorges dam.

An attack by China would end very poorly for everybody. There are millions of people living in the inundation zone.

Honest question, from non-american: what is up with all the China scare? I just can't understand it. Is it because China is socialist and we want to see capitalism winning over?

Anyone who is unironically saying China attacking Taiwan is a real threat (eg The Anthropic CEO quoted in the article) is either simply echoing the administration's painting of China as a geopolitical bogeyman or they're just ignorant of geopolitics, likely because they're projecting American economic imperialism onto China.

I'm glad the article dismissed this as a threat because it isn't one. The official policy of the US is the One China policy. You'll see this described as "strategic ambiguity". That's another way of saying that the official policy is simply to lie about supporting Taiwan's independence.

China can only hurt their position by taking military action against Taiwan. Also, it's highly debatable if they even have the military capability to invade Taiwan. Naval blockade? Sure. But to what end?

China is going to make their own chips. They'll just hire the right people to replicate EUV lithography. The article brought up nuclear weapons. It's a good analogy. At the end of WW2 the thinking of the US military was that the USSR would take 20+ years to get the bomb if they ever got it. It took 4 years. The gap with the hydrogen bomb was even less.

Western chauvinism in policy circles completely underestimates China's capacity to catch up in lithography. Not selling the best chips to China created a captive market for Chinese chipmakers.

I also think TSMC is being understandably cautious in not expanding their CapEx. AI companies really should focus on an economic use case for AI more than worry that foundry capacity will somehow limit a theoretical future AI use case.

TSMC just needs to change its name to USMC and the stock will easily go up 20%