AI Promises and Chip Precariousness, Policy Recommendations and a Changing World, Concerns and Counterpoints

Hello, and welcome back to another episode of Sharp Tech. I'm Andrew Sharp, and on the other line, Ben Thompson. Ben, how are you doing?

Feeling—I don't know—

Pregnant. Sheepish.

Oh boy. Yeah, feeling sheepish.

Why?

Because I kind of want to complain—talk about my mental state, feeling tired. And I'm on with a guy with a newborn who is just the wrong person—

I'm here to give you perspective.

—to look for any sort of sympathy from.

Absolutely. Why do you feel the need to complain? Tell me more.

No, not complain. I think it's been very—it's been an interesting week because of the article we're talking about. We're actually launching right into it.

Mm-hmm.

On one hand, I feel very deeply mentally exhausted. And I think part of it is that what I wrote in this article feels so weighty, and there's almost a terror that people actually care and listen to what I have to say about the points in it, which is very mentally taxing.

Taxing, yeah.

That's balanced by the fact that I try to be careful not to get too full of myself and think that what I do matters that much in the grand scheme of things.

Ben Thompson has spoken.

I have to touch some grass. I'm just a blogger, right? At the end of the day. And so I have these conflicting impulses going on, and then I also don't want to tell you that I feel tired.

Indeed.

Well, if you need perspective, we have a newborn. Everyone in my house is horribly sleep-deprived, and last night at 5:00 a.m., every smoke alarm in my house went off because one smoke alarm was low on batteries, and that somehow set off every other smoke alarm. So we were all sleeping peacefully for once, and then the smoke alarms intervened. There's the perspective you were looking for. But I do sympathize with your mentally taxed week here. And that's a good note to get into the article.

1. The Chip Ban Must End

We're going to start with ZX. This is a 2-part email. I'm going to read part 1 from ZX here. He says, “I loved Ben's latest piece, and I particularly enjoyed seeing how his views evolved as he articulated bits of it throughout different podcast episodes, culminating in his final tour de force. The best part was that he actually bit the bullet and put out recommendations, though I can't see decision-makers following through on them.”

I, too, enjoyed watching the evolution of this argument and being part of that evolution over the last month or 2. And I'm happy to revisit this discussion for one more episode, in part because this is arguably one of the most important issues in the world, and not just in tech.

But let's start with you biting that bullet, Ben, and providing recommendations for what chip policy should look like going forward. Per your article, we should be lifting the chip ban in its entirety: 2 nanometers for everybody. Here we go. Tell me why. Tell me how you got there.

Well, I don't know how long ZX has been a reader, but in many respects, it goes back to the very beginning of Stratechery, when I was writing about Intel right off the bat. I wasn't, in that context, necessarily thinking about or referring to Taiwan, but maybe it goes back to that. It goes back to—

Mm-hmm.

—my first flight to Taiwan in 2003. I'm sitting in the back of a China Airlines flight in a middle seat, traveling from L.A. to Taipei, talking to the girl next to me, and we're asking, “Is this a Chinese airline or a Taiwanese airline?” You pull out the in-flight magazine, and it shows the route map, and there are 0 routes to China. That was before there were any sort of links. You had to go through Hong Kong to get to China. And actually, at that time, Taipei to Hong Kong was the busiest air route in the world because there were so many Taiwanese going into China via Hong Kong.

It's the only way in. Yep.

Yeah. “Well, guess it's a Taiwanese airline.” So maybe, initially, that makes all my takes about Taiwan and China suspect because I was that ignorant in 2003. Then again, 2003, scarily enough, was 22 years ago.

I was going to say. Got a couple—

And—

—decades under your belt since then.

And I've been thinking about the Taiwan issue pretty much nonstop since then. I came to Taiwan thinking I'd come for a year. Immediately, within 2 weeks, I remember calling my parents and saying, “Yeah, I'm going to be here for a few years. This is great.” And 22 years later, here I am sitting here.

That, too, is a reason to both listen to and dismiss my argument. It's hard to imagine a geographically related issue that I would struggle to be completely objective about. I have a life and a family here, and also a reason to pay attention because I care a lot.

You're on the ground.

And—

Yep.

My biases would certainly be to be a Taiwanese absolutist. I've been to China. I admire China and what they've done. I have 0 desire to live in China, and I like Taiwan just the way it is and would—

Mm-hmm.

—I would prefer that things don't change. But looking at that route map back in 2003, there's just a really fundamental geographic reality beyond the fact that this island was ethnically Chinese, 70 miles—

Mm-hmm.

—off the coast of China, and had 0 flights. It's the fact that China has one coastline, and there is an island 70 miles off it that is—if you separate everything, you take away Taiwan's democracy, you take away chips, you take away geopolitical considerations—you can just look at a map and realize there is a structural issue here where a very large country has a very small country with which it is avowedly—I think technically—still at war. That's a structural problem.

That was in 2003, when China was not the China of today. It was on its way, I think by and large because of decisions that were made in the 1990s in particular, but it was not yet what it is today, where it dominates global manufacturing. It has a rapidly modernizing military.

Military, yeah.

It has a huge fleet of missiles bristling along the coastline, as it were. I think a strength I've had in my analysis is an ability and willingness to put myself in someone else's shoes. While not agreeing with or endorsing what they think, I try to think, “If I were in this situation, what would I think?”

Mm-hmm.

If the US were in this situation and we had a large landmass, a large economy, but 1 coastline, and a geopolitical rival controlled an island in the middle of that coastline, we would have a problem with that.

We would have a problem.

That's not to say the US controls Taiwan per se, but this sounds kind of trivial: an aspect of this discussion, at the end of the day, comes down to geography.

And to this point, with ZX being a longtime reader, I don't remember when I wrote the article “Chips and Geopolitics.” I think it was probably in the late 2010s, basically saying, “Look, there's a big problem brewing here, and it comes down to geography.” And that's basically where we're at. That is the genesis of this entire discussion.

Well, along those lines, when you wrote “Chips and Geopolitics,” you didn't come out and say, “We shouldn't be banning chips.” You were sort of hinting at it here and there, and then this article was the most explicit you've been in prescribing a different course than the US has chosen over the last 4 or 5 years.

For me, I'll be honest, you kind of lost me with some of the earlier discussions when we were talking about the chip ban over the last month or 2, and I really liked the way you articulated your argument in this piece. So explain more about why you think we should be lifting the ban on chips. You also advocate for strengthening the ban on chip-making equipment, which I think is a very important piece of what's put forth here.

2. How The Chip Ban Started

Okay. “Chips and Geopolitics” I wrote in 2020. The issues with the chip ban really started under Trump with ZTE, even though most of it was codified and put into a larger initiative under the Biden administration.

ZTE was a Chinese phone maker that violated a prohibition against shipping to Iran, which is definitely true. I mean, it's not like—

There are a lot of these discussions, particularly when it gets to international politics, and this has been a framework for a lot of dispute, I think, over the last few years. On this podcast, we're going to request that our users understand that not every time am I going to come out and say, “Yeah, this is morally wrong.” We're—

This isn't a discussion about morals. It's a discussion about reality, and ideally those two are not incompatible, but if there's one arena where they often are, it is international relations. So, just sort of that caveat up front, ZTE could have been—let's assume they're utterly and completely guilty.

Right.

The problem is the U.S. did have the capability to basically kill ZTE: “You're not getting chips, and, oh, by the way, because we control essential technology that goes into chipmaking around the world, we're also banning anyone who makes chips from selling them to you, and we're the United States and we can do that.”

Mm-hmm.

Why can the United States do that? The United States can do that because we basically control the world militarily, and we control the world economically via our consumer market. And so it was a valid and credible threat that basically killed ZTE for kind of a year. I don't know to what extent—and I wrote this at the time—but this is going to set in motion a response that in the very long run is going to be problematic for lots of reasons.

China's always had a drive to build its own chip industry, since the '50s or '60s. Honestly, you could make a case that it was a mistake all along, whereas, say, Japan—of course chips are important, so they plugged into the worldwide chip industry and carved out a huge niche.

Right.

Much of that was sort of taken by South Korea, but Taiwan is a similar thing. They plugged into the global supply chain, and the whole reason we're in the situation we are is because of that. China didn't do that. China's like, “No, we need to do the whole thing ourselves.”

We do it all. Yeah.

And they failed miserably for lots of understandable reasons. It's a very hard thing to do. You need volume, you need customers. You go back to Bob Noyce at what I think was still Fairchild Semiconductor. He brings out a chip that cost them a dollar to make, and he sells it for 25 cents. Why did he do that? Because he realized that the actual cost of chipmaking is all upfront.

The way to get to profitability is not to sell a chip at a profit; it's to drive volume such that the profitability on a per-chip basis—the bar to clear—drops precipitously to our marginal cost. The marginal cost is the actual materials in the chip, which are negligible; it's the electricity and the workforce that actually puts it together, which Fairchild had exported to Asia.

Mm-hmm.

The semiconductor industry was the tip of the spear when it comes to globalization. They wanted those workers—

Cheap labor, yeah.

—in Hong Kong; that was the first place they went. It quickly spread to Taiwan and Singapore, to be out there literally hunched over on a bench connecting wires. That's how assembly of chips actually worked. And guess what?

Mm-hmm.

Brutal work. When you had a component you wanted to sell for 25 cents, you didn't want to be paying an American worker $2.50 to do it. And so, from day one, this has been built into the system of how semiconductors work.

And guess what? It turned out that—and this goes back to our trade discussion from last year—this is another example of where, in the long run, cheap labor was exported first, but then everything else sort of followed it in over time. And so we started out with Taiwanese and Singaporeans and Hong Kongers assembling the chips, and we ended up where actually the whole thing just happens in Taiwan.

Right.

China was not a part of that, and they would never have been a part of it as long as they could keep buying the chips that they need.

Mm-hmm.

It's so hard. And this is a critical point: You don't—just because the government wishes it—make it happen. You need—and, by the way, this is a lesson that the U.S., I think, is going to learn or is in the process of learning. This is the Intel lesson. Just because you—

Yeah.

—you can make 18A looks like it's going to be a great process. That's not enough. You actually need the customers. What goes into getting customers is very complex. You need the volume. And China was never going to have that, and ZTE was the first step in creating the demand.

You pull a semiconductor industry into existence. It has to be demand-driven. ZTE set the conditions for not the government pushing a semiconductor industry into existence, but industry pulling a semiconductor industry into existence, and that's how it has to happen.

Mm-hmm. I mean, it had also focused the government on this area and motivated them to start pouring billions and billions of dollars into it every year.

I know, but I think—

It's funny, I was laughing about it.

I think that most people have that perspective from you. I don't really share that perspective. And I know Bill and I sort of argued about that a couple weeks ago. I think the role of government and subsidies in pushing—and, by the way, this is a very valid critique of any ideas I have about Intel in this article.

The symmetry between the U.S. government and Intel and what China is trying to do with its chipmaking industry does make me smile, because both countries are sort of fumbling around in the darkness trying to create alternatives to TSMC. But you're saying private demand is essential if you actually want to move forward and make progress.

I'm saying semiconductors are so difficult, it's so hard, and when you can just go elsewhere and buy what you need, and you can buy it at basically the marginal cost of production if necessary—which is practically zero—I just don't think it's viable to build a second semiconductor supply chain in any sort of open market.

That's fair, but from what I understand, the ZTE ban triggered a lot of purchases of, let's say, chipmaking equipment in anticipation of future sanctions and export controls on all of this. And so it accelerated it in that respect.

Right.

That's all I mean.

No, for sure. That absolutely happened, and then the Huawei thing happened, I believe, in 2020. I think that might have been what triggered Chips in Politics, actually: the same ban being extended, or a similar ban being applied to Huawei.

Mm-hmm.

Huawei—now you're dealing with the big fish, like the most important company in China. Or maybe the most important company; I don't know, it might be the most important company in China.

It is, in many respects, the technological arm of the Chinese government. It is a private company. The ties, I think—all the U.S. concerns about Huawei, I think, are totally legitimate. And also, Huawei deserves tremendous respect and is probably still underrated by technologists, by people in the U.S.

They've done incredible stuff. Did they completely wholesale rip off Cisco routers? Yes, they did. But morals don't matter. Reality on the ground matters. And in this—don't clip that: “Morals don't matter.” Of course they matter, but this is why I gave the caveat up front.

We're getting grace from the audience, yes.

ZTE maybe could have been a one-off. Once you went after Huawei, we're locked in. It's not—

Mm-hmm.

—just that the government—it's not about just pushing and dumping dollars and the government wanting this to happen. It's a very deep-rooted, deep-felt belief that we on the demand side are going to make the sacrifices necessary to pull this into existence, not because we're patriots—and this isn't to deny the patriotism of Huawei or anyone else—but because our very existence depends on it.

Turns out that's a pretty good incentive to make things happen.

Right.

And so you have the Huawei example, and then you have the Biden administration basically formalizing all this and making it broadly applicable.

And tethering it directly to AI, because Huawei had transgressed in various ways leading up to those sanctions, and then ZTE, you mentioned, was violating the sanctions on Iran. But the Biden controls were passed, A, before ChatGPT, and B, targeted on the AI sector, recognizing that this was going to be critical to the future of geopolitics and national security.

That's right, that's right. And, by the way, all legitimate points. I think that's a key part of this discussion and why it was so hard to write. I can talk a lot. I can pontificate. I'm happy to give life advice about trade-offs and your priority stack and what goes on top. Actually facing them in the face, looking them in the face, and saying, “Yes, your arguments are legitimate, and I am trading them away in my advice,” is hard to do.

And you're wide open to attack, because it's like, “Oh, this, X, Y, Z.” I'm like, “Yeah, you're right. It is true.”

Yeah.

The implication of giving China access to leading-edge chips is that they will have AI systems on par with ours and very well may have better ones. Do you want that to happen? No.

Indeed.

I don't want it to happen.

Well.

But that's why there's this broader context. And what I would just say about this journey that ZTE is talking about is, I think I actually quite clearly articulated the core takeaway of the article I wrote this week when the Huawei ban happened. The ZTE one, I'm like, look, we're creating the conditions for a competitive industry that I don't think would have existed otherwise. That was half of it.

The second half, which I did write in that update after Huawei—turns out it's worth reading the daily updates—sometimes I don't actually put it in a free article until 5 years later, is that we are significantly diminishing the deterrent factor of TSMC with this action.

Yeah.

3. TSMC Is The Deterrent

Right now, we are in a world where the entire foundations of society, for both the US and China, are dependent on Taiwan. And, by the way, this was radical in 2020, and 2020 actually validated the point. Because you had the COVID crisis. You had all the automakers, which are probably the easiest example. What do they do? They're used to ruling the roost, right? They're the integrators. They make the engine. That's the most important component, at least before EVs.

There's a whole supply chain that serves them. So COVID happens, and they call up all their suppliers and say, “Freeze all deliveries for the next 2 months,” or whatever it might be.

Mm-hmm.

And they just have to take it. What else are they going to do? The entity at the nexus of the supply chain gets to call the shots. And that's what the automakers did. Among those suppliers, they maybe go down the priority list of suppliers. How do they organize the priority list of suppliers? Well, how much does this component cost?

They probably had a deep discussion with their transmission supplier. It's a very expensive component. It's difficult to do. “Yeah, I understand it's going to be tough for you. Let's figure this out.” When you get down to the parts that don't cost very much, you just send out an email. They can deal with it.

Mm-hmm.

The problem is that those parts that don't cost very much included these $1 chips for doing power windows, right? Who cares? It's a dollar chip. The problem is, that dollar chip's existence is downstream of massive investments made years and years ago. It's built on a fab that was cutting edge in 2002, and there's no competition for it because you would never build another old fab.

But it continues to run because all the cost was just building the fab. The marginal cost of that chip is practically 0, so that's why you only have to pay $1 for it. But there's this real fundamental mismatch that is hard for manufacturers like the auto companies to really internalize: this chip may be shockingly cheap, but what goes into that chip existing is incredibly complex and has not just a significant financial component in its creation, but a significant temporal element, in that it had to be created at a particular moment in time, which was far in the past.

And guess what? When you just mindlessly shut off all your suppliers—

They go find other customers.

No problem. It's a commodity. We'll sell it to someone else.

Yeah.

And then you start up your factories 2 months later, or a month later, or whenever it was, and demand is skyrocketing—

And they're back of the line, buddy.

Because suddenly—okay, we'll put you back in line, at the back of the line. And now you thought you were king of the hill, and actually this little $1 chip has completely destroyed your business for weeks or months on end.

Right.

And that's just being at the back of the line. If TSMC were destroyed, it would wreak havoc throughout the US economy and the global economy, in part because of these trailing-edge issues, and obviously the leading edge is its own story there. But yes, all of this was highlighted in 2020 with COVID, and as AI becomes more central to the story of the future of the economy, the leading edge is just as important.

Well, it's not just that; it's all of it.

Yeah.

The lesson of 2020 should have been that our economy is, to a shockingly greater degree than anyone realized, utterly dependent on the continued existence of TSMC, and by extension, the continued independence of Taiwan—de facto independence.

Mm-hmm.

So I have to be careful with my words here.

Sure.

And the alternative is that we have a total dependency on China for these components. At the end of the day, the level of trust that's entailed in computing is very large. This gets back to the UK and Apple bit that we talked about earlier in the week. We had this whole discussion about it.

At the end of the day, you can talk about encryption all you want; you are placing total faith in the maker of your device. There is a level of trust entailed in computing. Someone has to be root, and what we are shaping ourselves up for is making China root of everything.

Making China root of everything in the trailing edge is the concern in the short and medium term. Is that right?

Well, and in the long term, for the leading edge, this gets to—

Right.

This gets to this whole point, which is: we've set China on the path, going back to ZTE and Huawei. First, we created the clear need to build a second supply chain. Then, number 2, we created the clear sort of demand and incentives.

Along the way, it's not just that there was, “We could kill you”; it's that as long as China's dependence on chips is just as large as ours is, and they ideally get all those chips from Taiwan, there was a real mutually assured destruction aspect to Taiwan's existence. Who knows, maybe after 2020, the US threat wouldn't be credible, but: “You want to invade Taiwan? We're going to bomb those fabs on the way out.” And you're not going to have access to all these chips either, because we're not going to have them either. Honestly, maybe it's fair to wonder if that threat was credible, but at least it was possible, right? No way, no one's going to get access.

Mm-hmm.

And we're not there yet, but we've set ourselves on a path for China to build self-sufficiency. And to your point, they got a head start. The immediate response to ZTE and Huawei was to buy as much equipment as they could. This sort of clouded the whole discussion: these companies are lobbying the Commerce Department, saying, “Oh, look, this massive percentage of our revenue comes from China.” Well, yeah, because they're buying equipment that's going straight into storage.

Mm-hmm.

Because they know it's looking like they might not get access in the long run, so what's the actual real number? It's sort of hard to tell. But we also have to figure out how to make all these components on our own, which is an insane thing. It's insane economically. There's no reason for an Apple or an NVIDIA to go to Intel. None. It doesn't make any sense. It doesn't make any sense for any company to build trailing-edge semiconductor capability that can support the car industry. No economic sense. Why? You can just go to TSMC and buy it.

You go to TSMC and get your chips fabbed. You only do that for national security reasons, but even then, it's almost impossible. The one way to make it happen is: I literally don't have chips. I'm Huawei; I don't have chips. My existence as a company is dependent on this happening. Suddenly, you have a possibility—

So SMIC, let's talk.

4. China Builds Chip Independence

And so the US created those conditions. They created the conditions for China to become self-sufficient. Again, they're not self-sufficient yet. Even these Ascend chips that SMIC is making—we're getting reports now, and, you know, we talked about it, so it shouldn't have been a surprise—they're doing 7 nanometers. It was almost certainly unprofitable.

Now we're 3 years on, or something, and it sounds like they may be profitable now, which is totally plausible. I think that's believable. They've had 3 years to figure it out, to do this crazy quad patterning. And, by the way, they lost almost certainly billions and billions of dollars along the way in a way that Intel understandably could not have borne with their 10-nanometer problems.

If they kept at it long enough, they probably would have gotten it, but they would have gone bankrupt in the process. Well, guess what? A way to not go bankrupt is to have no choice, and here the government funding does make sense. You’re going to be backed up.

Yeah.

You’re going to figure this out. And now they probably are making 7-nanometer chips in very evolved versions of them that are probably much faster than the old 7-nanometer chips. They’re probably doing it profitably.

And Huawei’s yields are apparently improving as well. They had 20% yields that are apparently up to 40% now. Who can say how real those numbers are? But it’s reasonable to assume that there’s real progress being made, and it’s reflected in some of the products that Huawei’s been bringing to market.

But we’re now 30 minutes into the podcast, and your take is that we should lift the chip ban and try to recreate the conditions that existed in 2017 and bind China to TSMC once again going forward. Is that right?

5. The Ban Must Be Reversed

Yes. So the Trump administration, again, I’m not sure the ZTE and Huawei bans were fully thought through in their implications, so that’s a negative check mark.

Mm-hmm.

On the positive side, the one really important thing they got done, and the single most important chip regulation—or convincing or jawboning—was EUV. There’s not a single EUV machine in China. Now, this does not mean China will never make an EUV machine. It’s been invented once; it can be invented again.

At some point, they will figure out how to make an EUV machine. But I think it is still reasonable, even with SMIC’s impressive progress with DUV machines and quad patterning and all these sorts of pieces, and their willingness to not worry about yields, to believe that at some point there’s just a hard wall.

You need to figure out how to get that beam of light that does the actual etching small enough to go down to these small dimensions. The light is still wider than the actual size of these chips, but DUV is a bridge too far.

Chips are infinitesimally small, so yeah.

Yeah. And so I think EUV is a wall, and China does not have an EUV machine. If they know an EUV machine exists, they know in broad strokes how it works. At some point, they will make one. It would be a lot easier to make one if they had even one in the country because you could actually see what it is.

Mm-hmm.

I mean, reverse engineering is a specialty of the Chinese economy, generally speaking.

Yeah. And again, this is where, in the empathetic mindset, we reverse-engineered a lot of stuff from England in the 1800s and 1900s, right? This is what it is. We’re in the real world here.

Yeah, I was not being derogatory there. They’re unbelievable at what they’re able to accomplish.

Right. And so you have this sort of wall, and I think that wall is still meaningful. We went to 5 nanometers. Then we’ve gone to 3 nanometers. This year, we’re coming out with 2 nanometers. In a few years, it’ll be basically 1-nanometer equivalent. They’re starting to change the names.

There’s a bit where all this is marketing speak, but these are meaningful shifts. The chips become exponentially smaller, which means you either get way more efficiency or way more speed, or some combination of the two. They run cooler. There are just lots of unalloyed goods and positives, and the entire computing industry and history is based on this process continuing.

We’re still talking about a pretty massive gap that is set, in the long run, to continue to expand for quite a while. And so all the talk about what China has done to have good-enough chips, as far as these Ascend AI accelerators go, the reality is they’re still pretty substantially behind what comes out of TSMC, and there’s a real, significant wall between them and overcoming that.

Again, I think they will get over that wall. It’s still going to take some amount of time, and in that amount of time, the gap is going to continue to widen.

Mm-hmm.

And so despite all these factors, despite the fact that every company in China knows they need their own supply chain, despite the fact that this is a priority for the government and they’ll funnel all this money, I still think this wall is so significant and the benefits of super-powerful chips are so massive.

All the reasons why these chip regulations exist—because we don’t want China to have them for fear of what they will do with them—is exactly why I think that if the chip ban went away, Huawei can talk all they want about the great progress SMIC has made. They’ll be first in line for TSMC’s leading edge tomorrow.

Yeah.

And ByteDance and Alibaba and DeepSeek and everyone else, they’re buying NVIDIA chips tomorrow—the top-of-the-line ones, if they can. And again, the response here is, “Yes, but they will also then be building their own.” Yes, they will, but I think it matters whether your life depends on it versus, “Oh, yes, we should also be doing this thing in parallel.”

Semiconductors are so hard and so difficult that this will meaningfully blunt China’s progress in actual semiconductors and in semiconductor manufacturing equipment, because so much energy and money and focus will go to, “Oh, thank God, we finally have advanced chips. By the way, we need to build AI. Let’s buy from TSMC.”

Yeah.

Is this an ideal situation? If my recommendation is one that would have been better in 2015, 2016, or 2017, maybe I should have come out more full-throated. I should have been yelling from the rooftops then to say, “We are going down the wrong road. This is where it plays out.” I think that’s a totally fair critique.

Mm-hmm.

But this is the reason I felt I had to write this now. And, by coincidence, that Bloomberg report came out about the Trump administration probably strengthening chip controls. You saw this was going to be the natural reaction to DeepSeek. You saw, “Oh, they’re catching up as it is. We need to crack down harder. We need to do it more.”

And my takeaway is, look, no, this is our last chance. Honestly, all the opening part of the article about AI was probably a distraction from getting to the main point. Part of it was that I’ve been trying to write this article for so long, it was hard to get me off the mat to do it, and that was an easy way to get started.

Then it was 2:00 in the morning. I’m like, “I can’t rewrite this and edit it. It’s just going out the way it is.”

You voiced that complaint earlier, and I disagree. I actually think it was an important way to frame the discussion of TSMC and its importance to the American economy, because it’s easy to overlook that, for all the hype around everything that AI is going to do, everything that AI makes possible today, and what it might make possible tomorrow, all of it is still reliant on NVIDIA, which is still reliant on TSMC.

If you take TSMC off the map in some sort of invasion scenario, everyone is screwed and all this progress grinds to a halt. I thought it was important to make that point on the way into the discussion of why it’s important to protect TSMC, think about incentives here, and think about the most effective ways to deter China from invading Taiwan. So I liked the intro more than you did.

6. AI Competition Changes The Tradeoff

I appreciate it. Thank you. That’s exactly it. Honestly, I think the AI industry, just from a broad perspective, is a phenomenal endorsement of capitalism. People want to complain about, “Oh, we need top-down state control,” and so on. This is actually an amazing example of how private industry can do things more effectively and efficiently than a centralized, top-down system can.

We’re in what I called the Goldilocks moment. We sit here on this side and ask, “Is it actually economically sustainable to be building these leading-edge models that people can distill and quickly catch up to?” It’s a fair question, but it’s an open enough question that it’s happening, and it’s happening on multiple fronts.

Anthropic is getting investors and investing. OpenAI is getting investors and investing. Google is investing. We have multiple companies. We have xAI coming out of nowhere. And by the way, I love Grok. It’s making me question some of my assumptions.

On one hand, my point about OpenAI controlling the consumer interface and mindshare is really important. I think that still holds.

Mm-hmm.

For me, I was going on about how I could see why people think Anthropic is a little better than OpenAI, but I like the OpenAI app, I’m used to it, and I’m sticking with it.

You’re a Grok guy, huh?

Deep Research is still amazing, but Grok for just your general chatbot is so much better, in my estimation.

And so, yeah, but which is great. There's massive competition happening here. It's not enough competition to make people stop investing, so we're getting the best of both worlds. Meanwhile, we have prices being driven down to nothing. We have these open-source alternatives that are coming along.

We have all these hyperscalers that are making these decisions. We have the neoclouds that are coming up and absorbing capacity, if you want. I just saw something on Twitter about how H100 utilization is back to being sky-high after DeepSeek, which we saw in China as well.

And you have companies that are always wary of getting locked in to these companies. At worst, there's an open-source option if we always need to go in that direction. Llama obviously has been a player in this regard, too. They have a big challenge in front of them. We're not going to see anything, I think, until this LlamaCon. But what they have from a reasoning model—and can they—

LlamaCon. My God.

Really get back in the game.

Okay.

Again, terrifying if you're worried about AI doom, and so I'm just acknowledging that. We'll get people emailing about it. We don't have room to get into that right now, other than to say the doomers have to adjust to a reality that DeepSeek exists. I called it the—

Mm-hmm.

The minimum open model—that is the level that's just available to anyone. Being precious about anything at that level or below is, by definition, crazy, right? You go back to GPT-2 not being released because it was too dangerous. Fast-forward a very short amount of time, and here we are.

And it's all working.

And there's a single dependency, which is even more than NVIDIA, which I think could be overcome. It is TSMC. And TSMC—that's why I started with the island. It's why I started with that in-flight magazine.

The other problem is that you have this mismatch of interests between the US and China. The US, of course, wants to support Taiwanese democracy and is economically dependent on TSMC. China doesn't care at the end of the day about—

Mm-hmm.

Either. They want Taiwan, and if it's bombed to smithereens, that's fine. They'll still take it and be happy about it. And that's a mismatch of desire and motivations that I think is very problematic in the long run.

Long term.

And it just worries me to accentuate that. China already wants it more, and if they take away incentives not to do it, that just makes the situation worse.

Yeah.

And so, yeah. I don't know. So do we give them a—

Well, no. That's what I like about your argument. I'm not sure I agree wholeheartedly, but I thought it was a compelling case. And to your point on China's willingness to buy from whoever is fabbing the best chips, we've had indications over the past couple of years. It occurred to me while you were talking: TSMC has been accused of fabbing for Huawei subsidiaries over the past few years.

So even as Huawei is working with SMIC and having all this success with the Ascend chips, clearly or allegedly, there's still some appeal to working with TSMC. Another feature of the Chinese market is that it is unbelievably competitive. A company like ByteDance, a company like Huawei—everybody is going to be vying for the best chips—

Yeah, no one is—

If only—

No one is compromising their products—

To stay ahead.

That's right. That's right.

Exactly. And that's important to keep in mind: it may be a communist country, but it is very, very capitalist and cutthroat in a lot of respects.

7. Equipment Controls Need Tightening

In terms of some of the previous debates that we had, I think one of the problems I had talking through it with you is that we were talking about the chip ban and chip loopholes in general terms. What the chip ban has entailed over the last several years is really 2 things. On one hand, we're banning actual chips, and on the other hand, we're banning chip-making equipment.

Some of the loopholes that exist on the equipment side have been pretty maddening to me, just as a matter of coherence. If you're worried about protecting this advantage and protecting America's long-term technological dominance—which had been one of your stated concerns when we were talking a couple of weeks ago—the fact that you would ban chips and highlight the importance of this area, and then not ban some of the critical chip-making equipment and make sure you cut out the knees of the PRC chipmakers, just strikes me as frankly insane and incompetent.

That was the piece of your article that also resonated with me and is important to consider in tandem with the idea of lifting any of the bans. We're also talking about closing some of the massive loopholes on chip-making equipment, and that is crucial.

Yeah, you put your finger on what I think most people who object to or get their hackles raised by this article are offended or bothered by: the trade-off I'm implying from AI—that basically we're giving China the tools necessary to build their own AI systems.

And yes, that is wise up front, and it's a painful trade-off. It's also one that you can go back to—I mean, there are long debates about nuclear—

Mm-hmm.

Clearly, there were people in the US nuclear program who basically just gave it to Russia for whatever reason, whether they were a communist or whatever it might be. You can also look back in the fullness of time and make an argument that those people were traitors and also did more for world peace than almost anyone.

Hm.

Because—and this is hard; again, I almost feel my skin crawling making this point because they were traitors.

It's a real common-room bit of logic, but yeah, you could certainly make the argument in retrospect.

Mutually assured destruction worked.

Mm-hmm.

It really did. And is there a level of arrogance in not thinking through what it actually means for 1 country to control what some people would argue is analogous to nuclear weapons in terms of being a superpower and having the capability to dominate other entities when it comes to AI?

Is it actually, in some respects, more sustainable in the long run that AI is diffused generally? And by the way, that's kind of been my argument about the whole safety issue. It's not that I dismiss the safety issue. It's that I don't think this is containable in the way that there's a lot of hand-waving about—that we're going to actually limit this and control this—and I think DeepSeek really hammered this point home.

AI is coming. We need not just AI; we need good AI to defend ourselves against bad AI. This is that same argument writ large: a balance of powers may be the only long-term sustainable way about this. So anyhow, I haven't even responded to your point yet. So that's—

Mm-hmm.

That's part number 1. That's trade-off number 1. Trade-off number 2 is arguably more painful and does, in some respects, go against the argument that you said I'd made previously.

Saying we're going to block all US semiconductor equipment from China is putting the kiss of death on our semiconductor manufacturing industry in the long run, or at least our dominance of it, because China will go down that path. Some number of allies may or may not decide, “Hey, if the US is sacrificing this, we'll just sell to China.” That's actually going to be bigger and larger in the long run, and we can take over these areas that the US has dominated.

You're setting yourself up for 30 or 40 years—or maybe a shorter 10 or 20 years—where—

Mm-hmm.

It's not just that China has its own semiconductor manufacturing equipment, but also now they're selling outside of China to the world and undercutting us, basically doing to our semiconductor equipment manufacturers what China has done to every other industry. We've seen how that has gone.

And that, in many respects, is the grace that you would give to the Commerce Department for all these loopholes, which is: you know what? China making 7-nanometer work with almost exclusively American—or Western, I should say—equipment, yeah, that sucks, but it's also a good thing because that means they're not making their own semiconductor equipment and they're not moving down that learning curve to do it.

This is the other big trade-off that is in this article, and this is the consequence of writing this article in 2025 versus writing it in 2017—

Mm-hmm.

—or whenever it might have been written. I think that the value and importance of maximizing Chinese dependence on Taiwan is significant enough that, at this point, we've waited so long that it's not just that we're sacrificing a world where we dominate AI and no one else does.

And by the way, I think this is where DeepSeek should change some minds in that regard. But also, we're sacrificing the long-run dominance of our semiconductor equipment manufacturing industry.

And that, at least for me, is actually an even more painful pill to swallow. The AI one, it's not that it's a big deal; it's that I've already resigned myself to that fact mentally.

Mm-hmm.

The second one, for me personally, is in many respects a more difficult trade-off to make.

Yeah. The state of the export controls now is full of holes, and SemiAnalysis last year wrote:

“Current sanctions are strict downstream on AI accelerator chips and similar. They are looser upstream on equipment to produce these chips and nonexistent for components of that equipment. This strongly incentivizes and enables China to develop an indigenous capability in manufacturing equipment, eventually rendering sanctions toothless. Instead, restrictions should be stricter upstream, tight on equipment and tighter on its key components.

“For example, current controls prohibit exports of EUV scanners into China, yet they do not restrict EUV optics from Zeiss. This makes no sense. ASML and any chipmaker using EUV can tell you that projection optics are the number 1 critical component. While it is not practical to control every nut and bolt, certain items obviously must be. Where to draw the line? There are fewer than 100 critical components from specific suppliers that should be restricted, including EUV projection optics, precision mask stages, and radio-frequency plasma generators.”

So having read that piece, when you were talking about lifting the chip ban and relaxing the chip ban a month ago because you were concerned about the long-term industry, that's where I thought: Are we still going to be selling all the components that allow them to work on developing their own chipmaking industry? It is a lot more coherent if we're closing those loopholes and then saying, “Keep buying from TSMC for the next 10 to 15 years, and then let's see where we are.”

The other aspect of the argument that is tricky is, in order to actually control the flow of chipmaking equipment, America would have to work with Japan and Tokyo Electron, the Netherlands, and ASML—

And Germany.

—and Germany and Zeiss. Whether the US can do that effectively is an open question. That's what the Bloomberg story this week was focused on: Trump officials meeting with Japanese and Dutch counterparts to discuss tightening these export controls. The Biden administration tried that, and they were unsuccessful at convincing ASML and Tokyo Electron to get on board. But we'll wait and see how effective the Trump administration can be and what sort of carrots and sticks might play into that process. I think that's one of the biggest questions of 2025 as they try to craft their policy in this area.

8. Pax Americana Is Ending

Yeah. The biggest question, zoomed out, is really: What is the structure of the world in 20 years—or shorter or longer, or what it might be? And this is where I think, who am I to be making pronouncements about this? I'm a tech blogger. My concern is that I'm just thinking way too much about things and spouting my opinions. But it's sort of inescapable at this point and with these geographic considerations.

We've operated, and decisions have been downstream from, 25 years as the US global hegemon. Before that, it was the West versus the communist world, and the West was dominated by the US. So the US has basically dominated its economic and technological sphere since the end of World War II.

I've been thinking, just big picture: What went wrong? What are the issues that we ran into? And I think, big picture, all this worked for the US. The US was like, “We're going to leverage our consumer market to rebuild Europe, to rebuild Japan, and yes, it's going to cost us some degree of manufacturing.” But there's only so much of that that Japan or Germany can actually absorb because they're just smaller countries compared to the US. The US is by far the biggest country in this coalition, and so there was a lot of latitude to be generous in a way that Pax Americana was deeply generous to the rest of the world. In exchange, we're going to be policemen and we're going to guarantee peace. That worked.

The problem with China is China's just too big. That's really the issue. It's not that we can't get into the specifics of the debate—that China overly subsidizes its export industry and insufficiently pushes its consumer market, and that's the real problem. All those can be right in various ways in discussions, but at some point you zoom out far enough, and it actually goes back to the map. China is a very large country.

By virtue of that, when and if they ever got their shit together, they were going to be in a structure where the dominant US dollar heavily favors external production of goods and importing to the US. This is a whole TSMC bit. We've talked about currency and all the bits that go into it. The dollar's not just strong; it's also particularly strong against the Taiwanese dollar in a way that maybe doesn't make total economic sense, which is a big advantage for TSMC.

But China enters into this system and it just got too expensive. It's too expensive, and the US—it's not an accident. We could point to politicians and DOGE and all this stuff, but it's not an accident that the debt exploded in conjunction with Chinese growth and the takeover of this. It's not an accident that where US manufacturing really went away and where all these center towns got devastated was in the context of China.

It's just so big that a system predicated on one country being generous just wasn't sustainable when that generosity was flowing to an entity that wasn't just of equivalent size geographically, but was actually much larger in population. And so there's—this is why structure matters. Is this actually sustainable in the long run?

Again, it's always hard to know exactly what Trump is doing. But there's an Occam's razor explanation for a lot of the weirdness, where the weirdness makes sense in the context of Pax Americana and the US being at the center. But if you're actually thinking about a world where that's over, then a lot more stuff makes a lot more sense.

Things like tariffs that we've talked about. Things like really being aggressive within your continued sphere of influence, which is the Americas. Things like really just not caring what the Europeans say or do and telling them the way it's going to be. If you're living in the Pax Americana world, that seems absolutely bizarre. If you're thinking, actually, this whole thing is going away, do you want that to happen to you, or do you want to aggressively go there on your own?

And this also ties into the China-Taiwan thing. The US stepping back from Europe, one way to think about it is: Pax Americana can't control the whole world, but it's going to continue to exist. We're just shifting from Europe to Asia, and so we're going to double down on defending Taiwan and defending the first island chain—the Philippines, Taiwan, and Japan.

You think about the first island chain, and that gets back to my point. China sees it as a chain as well, a chain they don't like very much, right? But maybe that's one solution. In that solution, Taiwan is a real flashpoint because that's the most important link in the chain.

The second one is depressing, and I'm not necessarily saying this is a good thing, but it's a potentially realistic view: China is becoming so large and so powerful that it's inevitable they're going to be the Asian hegemon, and that by definition has to include Taiwan. We can debate these points, but the implication in terms of TSMC is exactly the same. There's just no world where this isn't the critical flashpoint going forward.

Inevitable problem, yeah. And so the goal in that scenario is to—

It's buying time.

—create economic incentives that help you buy time. Yeah.

It's buying time. And that's really it: There is no more disgusting trade-off to make than to say, “I'm going to trade away AI, or trade away our semiconductor equipment manufacturing, just for 10 more years, or just for 15 more years.”

Mm-hmm.

But that's the problem of waiting until 2025. I don't feel great about it. That's why I'm tired. It's tiring. It's not that it's physically tiring; it's tiring to force yourself to consider and make trade-offs that feel terrible.

Well, I'm tired too. I'm right there beside you, so it's all good.

For a much more legitimate reason.

ZX, part 2. I just enjoy this email.

“Implicit in the essay is Ben's skepticism that we'll get to AGI in the near term, and in the long term, Taiwan will become a part of China. It's understandable why he doesn't say that explicitly. In that world, it's rational to empower Chinese AI in exchange for long-term US semiconductor manufacturing competitiveness. But if one takes seriously the sincerely held belief in big AI labs that we will get to AGI in the next few years, which will result in a step change, then allowing Chinese AI to access the best chips is insane.”

From an AI safety perspective, this will intensify the AI race even more and increase the probability of doom, which he writes as p(doom). I'll just note for the record that the worst part of me working in tech now is that all you nerds use message-board terminology with a straight face on a regular basis. You even dropped p(doom) in your article this week. It's horrible stuff.

But ZX continues: “Maybe the Straussian reading is Ben is advocating for policies that will likely fail, as the structural forces against U.S. semiconductor manufacturing are just too great, but in the process will bring us closer to the pre-2018 world where both the U.S. and China depend on Taiwan, nobody is incentivized to go to war, and Taiwan can be Taiwan, and Ben can continue to stay in Taipei.”

I'm glad I got a smile out of you there. So, Ben, do you have a comment on that Straussian reading? “Straussian reading” is another big tech-guy phrase these days. In the real world, we just call that subtext. But do you have thoughts on what ZX had to say there?

Well, I think the Straussian reading was popularized by Tyler Cowen, who is quite influential in Washington, D.C., so I think you've been swimming in those waters—

There you go.

A local guy.

George Mason's finest.

Absolutely. Shout-out Tyler.

—longer than you might think. I don't think he originated it, but he's very closely associated with it. Maybe he is; I don't remember. It goes back to Leo Strauss, anyhow.

Leo Strauss is the originator, yep.

No, but I can't remember if Leo Strauss invented “Straussian reading” as a phrase, or—

Ah.

—I think it might have been Tyler referring to Leo Strauss. Whatever. We can come back to it. Maybe someone will email us the etymology of “Straussian reading.”

Yeah, we'll settle it on the mailbag episode. Don't worry.

Yeah. Straussian reading is basically what is actually being said here. And that's why I want to be upfront: I'm clearly not objective here. The challenge with Straussian reading is that sometimes these things are intentional and sometimes they're unintentional, right?

Mm-hmm.

To what extent am I motivated by wanting Taiwan to stay exactly the way it is? Very fair question. But with the Straussian reading, I think I'm pretty explicit. Well, no, the Straussian bit is having the whole way to revive Intel. That's where this article arguably went off the rails. There are a lot of crazy proposals I put in there, in part because it's such a hard problem.

There was an aspect of me making a choice: Let's expand the Overton window very wide and just throw stuff out there, knowing not all of it—

The U.S. is going to build a data center with Gaudi AI accelerators made by Broadcom and Apple. That sounds stupid, right, in isolation.

The reason it's in there—let's indemnify for copyright. The reason it's in there is just to emphasize—

And everything's open source.

—we have to pull out all the stops here. Everything needs—

Yeah.

—to be tried. And implicit in the absurdity of some of the recommendations, the unlikelihood of them happening, is perhaps the deep-rooted concern that none of it is going to work, that we are just structurally dependent on Taiwan, and outside of a war, that's not going to change. So let's do everything possible to avoid that war and preserve the status quo as long as possible.

Mm-hmm.

It's a fair point. That might be right. And then I talk about the Trump administration and this shift. Is this an ongoing shift away from Pax Americana to a multipolar world? It may be, and it very well may fail, just because these shifts don't happen because you want them to. Just like you—

Yeah.

—you don't create a semiconductor industry because you want it to. It has to be pulled into existence. Do new world orders happen because you want them to, or do they get pulled into existence by breakages of the old one?

Right. On the Intel point, we didn't really talk about that because, on the other hand, I said I'm happy to discuss the chip ban because it is one of the most important issues in the world. I could not stomach another hour-long discussion of how to revive Intel.

My takeaway from that section of the article was that subsidizing demand in various ways is what's going to be critical to onshoring leading-edge production. Absent a war scenario that removes TSMC from the equation entirely and creates that demand organically—and very painfully, I might add—in any peacetime scenario, if the U.S. government is serious about creating onshore leading-edge manufacturing, the government is going to have to be both creative and really, really aggressive—

Yep.

—if we're serious about enhancing Intel's capabilities.

Yeah, the problem with the CHIPS Act is twofold. Number one is the one we've talked about, which is—and I agree with Trump—it's subsidizing supply, not demand. It's on the—

Mm-hmm.

—wrong side of the equation. The second problem is that it's dramatically too small. I'm sorry, pat yourself on the back over $50 billion. We're talking about a trillion-dollar initiative.

Well, and that's what I appreciate. In terms of the prescriptions that you offer for U.S. chips, at the risk of not doing a Straussian reading here, I think the chip-making effort in this country may fail because the government is unlikely to be reading Stratechery. But the recommendations themselves reflect an understanding of how the chip industry actually operates and how companies like NVIDIA would actually behave. I feel like that particular understanding has eluded U.S. bureaucrats over the last several years as they nominally take this problem seriously, but also haven't really committed full-bore.

Now you're causing problems for me, Andrew. I'm trying to come on here and say that I'm just a blogger.

What's that?

The way it comes from—I know this is being read. The terror is that people do actually follow my advice.

Well, then I hope it's heeded.

And then what verdict do I hear? But I think the broader point—you're right. This isn't the conversation about Intel.

Yeah.

That's in there because it needs to be in there. Some people got fixated on that, like, “Oh, this doesn't make sense,” and so on. I'm like, yeah, that's my whole point here: to impress upon people what a hole we're in. There are no good solutions right now.

And you also wrote about trying to subsidize demand three years ago, and the government still hasn't really taken action on that front. That's where people need to sense the urgency and act.

And that's the whole Intel problem. Everyone's fixated on Intel having lost its lead on the leading edge. The reason that's problematic is because that was the reason for them to get demand whenever they built a foundry.

If they built—if they went the foundry route meaningfully—if they did the Pat Gelsinger plan in 2013, or even better in 2010, or even better when they realized they missed out on mobile, the reason why, say, Apple would go there and endure all the pain and figure it out is because they wanted the fastest possible chips, and Intel had the capability of making the fastest possible chips.

In a world today where they have nothing, there's no reason to go to Intel. And even if 18A is marginally better than TSMC's 2-nanometer process, the gap isn't nearly significant enough in the way it was a decade ago. That's the cost. The cost isn't just that Intel's chips being inferior to AMD's chips has been a big problem for them. The other cost is that they lost their bridge to being a foundry when they lost their process edge.

Another thing to address here: everyone is like, “No, we just need to get TSMC to build more in Taiwan—build more in America.” From TSMC's pure corporate perspective, and from a responsibility-to-shareholders perspective, that's exactly what they should do. TSMC should be hightailing it out of Taiwan to a certain extent.

The problem is that the Taiwanese government isn't stupid, and there's a reason why the fastest chips in the U.S. right now are 5-nanometer. They call them 4-nanometer, but it's an evolution of 5-nanometer. They're building 3-nanometer. The 3-nanometer process will come online once TSMC is down to 1-nanometer, or I think they're calling it A16, or whatever it might be.

Basically, the Taiwanese government has said you cannot export the leading edge and the N-minus-one node. N-minus-two, you can do that. Why are they doing that? Because the number one reason the U.S. will come to Taiwan's defense is for that reason.

TSMC, yeah.

It—and so, of all the things Trump says when he complains about the Taiwanese government trying to hold the U.S. hostage: Yes, they are.

That's exactly what they are doing. They're not stupid. It is the number 1 reason why the US comes to Taiwan's defense.

Now, as part of this, you mentioned the trailing edge. I think it's a huge problem. The trailing-edge situation is tough, especially now because China is doing it. This is a terrible market.

Mm-hmm.

Prices are going to zero. It's not just that building a trailing-edge fab would never make its money back; it's probably going to lose money on an ongoing basis. The economic realities of this are pretty dire.

Well, especially once China develops more trailing-edge capabilities and makes it less economical for anyone else to operate in that market.

It's already happening. It's already happening.

Yeah.

There are just total price crashes in this area, and that's going to continue. So, again, it's the same thing: almost all of that is in Taiwan.

Mm-hmm.

TSMC's been building that for ages, and we don't get any of that from Intel. The way Intel worked is that, as soon as they moved on to the next process, they shut down the old one and repurposed as much of that equipment as they could and sold off the rest, because they made their money by earning high margins on the front end. TSMC, interestingly, has shifted to that.

Yeah.

One thing that's been really interesting is that they've taken 5 nanometers and converted some number of those lines to 3 nanometers. The problem is that, when it comes to your window in your car, you don't need a 3-nanometer chip. It's not worth the cost. You can have an old chip; it's fine.

And so it's a challenge filling them, especially because there's so much demand for leading edge. They've built so many 3- and 5-nanometer fabs. Those are not going to stay filled. They're actually adopting a much more Intel approach to what they're doing out of necessity. That's just the price of being on the leading edge, and it's been a really interesting transition for TSMC.

I think one of the big mistakes TSMC made was underpricing 5 nanometers and underpricing 3 nanometers—not just because they could, since they were the only players, but also because you're not going to make as much money in the long run. You're not going to make all this cash flow off depreciated assets. They're not going to be filled. You're going to have to repurpose the equipment, so you have to charge more money up front. Just from a pure analyst perspective, that's been a really interesting thing.

But if TSMC takes over Intel's assets, you don't just make them work. You have to completely—

Mm-hmm.

Restart the whole thing. An Intel transistor is different from a TSMC transistor all the way down to the molecular level. The chemicals are different. The sizes are different. The density is different. Everything is different. It's not a trivial thing.

So you rebuild it, you get the equipment or whatever it might be, but all the R&D is still in Taiwan. All the government limitations and regulations are still there. And why would the Taiwanese government agree to letting the US—

Export that.

Yeah, the US is like—

Yeah.

“We need your chip industry so that we don't need to defend you anymore.” It's like, wait, what? Why?

So I admire TSMC greatly—

I mean, it strikes me as—

But, like—

—a good idea that you're just sort of spitballing with in a conference room. But yes, thinking through—

Everyone else gets a vote.

—the logistics of how that would work.

Everyone else gets a vote.

Yeah.

The enemy gets a vote, right? In this case, Taiwan's not the enemy, but they get a vote, and China gets a vote. We don't just get to make stuff happen because we demand that it happen. Maybe we could 15 years ago, but we don't get to do that now.

Yeah. Well, speaking of everyone else getting a vote, I had various counterpoints and concerns in the wake of reading your article.

Oh, wait, we've gone long.

I will say—

Do you want me to give you the—

No, no.

I'll give you a 30-second answer to each of them.

No, I don't want to go through each one. But I do think, as far as Pax Americana is concerned, the one thing that occurs to me as far as your proposal binding TSMC to the Chinese economy, I'm not sure how much that would actually sway Xi Jinping in the medium term here.

Providing China with unlimited access to TSMC probably won't be enough to dissuade the PRC from doing everything it can to indigenize its chip supply chain. And then, TSMC or no TSMC, taking Taiwan by force is going to entail tremendous costs to the PRC. But from Xi's perspective, it's possible and probably likely that at some point he will decide those costs are worth it.

That's why, if we're talking about the chip ban, it makes sense in the meantime to use every tool that we have to weaken—

To diminish their power.

—China in every way that we can.

Yep.

And the other thing, as far as Pax Americana is concerned, what worries me about the idea that China has its geopolitical sphere, the US has its political sphere, Monroe Doctrine 2.0—the problem with that thinking, in my eyes, based on consuming news about the US and China for the last 2.5 years, and this is also why I don't ultimately think that Trump is going to concede anything on Taiwan, is that the PRC's expansionist impulses won't end with Taiwan if they do ultimately reunify with Taiwan.

I mentioned this at the end of Sharp China yesterday. Palmer Luckey made a similar point on a podcast recently. But if China takes Taiwan, nothing from the past 10 years suggests that that's where it ends. Japan is right there. They were conducting drills around Australia and New Zealand this past week, and I just feel like the next step will be further harassing other allies in the Pacific, and that will ultimately be felt in the US.

And so it's not like there's some stable equilibrium that would be reached if China does reunify with Taiwan, and it's in the American interest to defend Taiwan because of that. That's at least the way I have understood the problem, and I don't know whether we're in a position to credibly deter them from a military perspective. I also don't know whether the PLA is in a position to actually take Taiwan now, and that is a source of ongoing speculation in the foreign policy community.

I'm not really responding to your point, but the point that President Trump is potentially looking at the world that way strikes me as naive.

Maybe. It may be naive to think that if we just hold our ground in Taiwan, that will be enough. There are shades of this in the idea that if Russia invades Ukraine, they're going to invade all of Europe.

I actually didn't know this. Historically, the only times Russia has invaded Europe are when Europe invaded Russia first. So, is there some aspect of self-projection from the European, sort of, Western mindset onto these pieces? Obviously, there's the Cold War era, when the USSR was clearly pushing abroad. So, who knows? You might be right—

Right.

—but it also might not matter when you get to the actual cold, hard calculations of war. War is about manufacturing. Right now, we're not competitive. I think in a skirmish, in a battle, I would heavily favor the US.

The question is what happens when it gets past that initial skirmish and battle into a war of attrition? Then I'm—

Yeah.

Then I'm a little more pessimistic, I think would be a fair way to put it.

The other thing I would say is, I think you mentioned this, but is AGI about to happen? Are we getting to the takeoff point where AI creates AI, and it's going to create these capabilities—military capabilities, whatever it might be—that become dominant? Sure, Ben, all your arguments make sense, just not right now in 2025. The next 5 years are the single most critical time.

Mm-hmm.

So of all the times for us to control the chips, even if it means sacrificing everything else, we just need to control it right now. I don't buy that argument, and not because I don't—

Even if you grant that that's going to happen, we're still dependent on Taiwanese chips. That's what undergirds the AGI, right? So China could say, “Nice AGI you have there. Here are some missiles in Hsinchu.” Right? Like—

What a shame.

So that was fun while it lasted. There you go. So that's objection number 1.

Objection number 2 is—and this is probably the biggest pushback I have on the Biden administration's view—I think it's hard to unpack the Biden administration's approach to chip controls and their AI regulations in general.

I think all of this was influenced by the AI safetyist contingent, which is also this weirdly accelerationist contingent: AI is really dangerous, it's going to kill us all, therefore we must build AI as quickly as possible. It never fully made sense, to be totally honest with you, but I think that had a very heavy influence on the administration. Therefore, we're going to simultaneously regulate AI and impose this chip ban. And yes, there are long-term costs, but it's really important for the next few years.

Mm-hmm.

The issue with that is, from a Chinese perspective, there is the meme “nothing ever happens,” and that meme exists. The thermostatic effect—just things staying the way they are—is extremely powerful. All my concerns, all my worries—you might say, “Ben, you're just being like the people who thought about the worst possible outcomes of COVID, and actually it was just a respiratory virus that, if we had just treated it normally and not thought up the worst-case scenario, we actually would have been in a much better place.” I think that's a very powerful response, and the meme “nothing ever happens” exists for a reason.

China and the U.S. are deeply enmeshed. China and Taiwan are deeply enmeshed. It's not worth it to go to war. The other reason it's not worth it to go to war from a Chinese perspective is that every year that goes on, their relative advantage over Taiwan and the U.S. gets stronger and stronger. Again, Taiwan is off the coast of China. They have home-court advantage, right?

Mm-hmm.

Just as they become more capable, the balance shifts. If we actually believe that AI takeoff is happening—that this AI is coming that's going to be so amazing, it's going to make all these other AIs, it's going to control the world—that changes that calculus. Suddenly it's like, “Oh, it's not that we're getting stronger every year, so we might as well wait. It's, ‘Oh, if we wait too long, the U.S. is back, baby, with the killer AGI,’” and all these sorts of things.

No, that accelerates the timeline. Absolutely.

Right.

The advantages of the CHIPS Act are that you're locking in a technological advantage that will compound over the next 10 years. But to your point about the U.S. not being the only actor, China could potentially watch that happening and say, “Well, it's time for us to take a boat trip—

Now or never.

—that's 70 miles.”

Yep.

You're right. And I think that's the unstated risk inherent to all of this. If we further close the loopholes and comprehensively lock in an advantage and handicap China's economy and technological progress, then we're accelerating a timeline that jeopardizes the future of TSMC, America's entire economy, the safety and freedoms of Taiwanese people, and just the stability of the world order and the entire world.

That's what I took away from your piece, and I think it's just important to keep in mind: that is a real risk. It's often overlooked in the conversations around what the U.S. should be doing and why everybody should tighten the restrictions. There's all this saber-rattling, particularly in the wake of DeepSeek, and it's like, maybe.

Is that the totally wrong takeaway from DeepSeek? Isn't the DeepSeek takeaway that it's not going to work?

Right.

We're not going to limit China?

Well.

You lose me there, too. DeepSeek was working on NVIDIA chips and could just as easily be evidence of how important American chips can be, and a reminder that they're still central to this story—to every success story in AI. We have not comprehensively controlled the flow of those chips over the last several years. So it can kind of go either direction on that one.

Yeah. And to your point, though, that's why I felt like I had to write this article. It's hard to write an article where you know you might be wrong, and if you're wrong, it's sort of catastrophic because, “Oh, no, China has AI. Oops. It's better than us. Sorry about that. Oh, they're invading Taiwan trivially because they got to AGI first,” right? All of these things could be on the opposite side, and I really screwed that one up, Ben.

Number 1, I'm just a blogger. I take no responsibility. But number 2, it was driving me a little batty that I've been talking about this deterrence issue with TSMC and Taiwan and moving from a world where both were dependent on Taiwan to one where we're really dependent on Taiwan and China is dependent on Taiwan for years.

Yeah.

And it feels like no one else brings this up. Maybe I'm wrong, but it's hard to make those recommendations that entail such painful trade-offs. At least I said my part, and now it's to the powers that be to figure out what to do with it. So now I want to go take a nap.

Indeed.

Yes. Have a great weekend, powers that be. Have a great weekend, Ben. I hope no smoke alarms go off during the middle of your nap in the next couple of hours.

Um, but we are coming back on Monday. Lot of great mail coming in the last week or so. People have celebrated my return by writing a lot of great emails, so thank you to the audience. Email at sharptech.fm if others out there have takes, and maybe we'll address them next week. Any final thoughts, Ben, before we sign off here?

No. I almost apologize. I feel like this was a meandering podcast. I think my daily updates came out super late this week. I was working on this on Monday, and then at 6:00 p.m. I'm like, “I gotta—this isn't going to get done.” So I shifted to writing about Apple, which came out super late. And then this came out late.

This is what I hate when bloggers do: no one cares about you. I tell this to writers. You're out there selling a product. I don't like appeals to the audience: “Oh, well, can you please support me? I do this.” No. You do a great product, you put it for sale, and no one cares about you. Everyone has their own life, right?

Yeah.

All that said, here I am talking about myself.

Well.

It's been really—

There's a difference between writing and podcasting. A lot of navel-gazing is—

No, I turn it out every week.

Navel-gazing is—

I turn it out every week. It's not a problem. I love thinking about this. It's great talking to you. I'm so glad you're back. I missed you.

It's just been really interesting, actually. It's been really interesting. There was something about this that took so much out of me, that it's— And maybe that's a reason to—

Well, it's deeply personal.

Yeah, maybe that's why. I don't know. It just—I don't have a take other than it's been interesting to self-observe, to have been writing for so long and doing this for so long, and to almost not have had even this physiological response to doing some writing. I don't know. It's kind of interesting.

Yes. Well, I appreciated the writing. It did build on a lot of the conversations we'd had, and it felt like a couple of the different Sharp Tech episodes—some of the debates with me, some of what you were talking about with Bill last week—it was like the first drafts of what came together on Stratechery this week.

Well, thank you.

And I think you've earned a nap, you know?

Yeah.

You're clearly tortured over there. But I will talk to you on Monday.

A little too full of myself. Need to go touch grass.

You did remind me that people in the government read Stratechery. I was under no illusions that they don't. But they hadn't listened to you to date on the chip industry. Enjoy your weekend, Ben, and we will circle back next week with plenty of mailbag questions and keep it rolling from there.

Talk to you later.