(预览) 泡沫经济学的隐藏收益、Microsoft 与 OpenAI 达成交易、Taylor Sheridan、Sora 与 Nexperia 随笔
Substrate 是一场半导体豪赌，创始人 James Proud 曾称其成功概率只有 1%，但潜在回报已大到足以吸引融资。 公司称，粒子加速器驱动的 X 射线光刻可以将制造成本削减一半，取消部分多重图形化，并制造出分辨率可与头部晶圆厂媲美的层。Ben Thompson 仍明确表示不确定——「最好能看到更多证据」（“It’d be nice to see more evidence”）——但他强调 Proud 的姿态：尽管胜算极低，仍以「100% 的信念」（“100% conviction”）推进。
同时挑战 ASML 和 TSMC 听起来荒谬，但行业的系统性整合让这套策略在内部逻辑上成立。 Thompson 认为，光刻是整个制造流程围绕其优化的核心环节，因此一种彻底不同的工具，大概率不能直接塞进现有晶圆厂。Substrate 可以采购传统的刻蚀、涂胶和掺杂设备，但可能必须「从第一性原理」（“from first principles”）重建这些环节的协同方式。
ASML 与 TSMC 的共同演进，构筑了强大的在位者锁定。 如今光刻可能占制造成本的 30%-35%；Thompson 粗略估算，现有设备接近 2.5亿美元，下一代设备则接近 5亿美元，同时强调这些数字是他「从自己屁股里掏出来的」（“pulling these numbers out of my rear end”）。一座约 300亿美元、却没有填满产能并运转起来的晶圆厂，会让所有者陷入严重困境；这使 TSMC 「几乎成了 AI 泡沫的刹车」（“almost the brake on the AI bubble”），因为它承担着巨额前置设备风险。
今天 AI 泡沫最有力的辩护，不是 GPU 能否长期保值，而是许多看似不可能的项目正在协同发明。 Big Tech 的支出经由 NVIDIA 流向 TSMC 和 ASML，同时也让 Substrate 这样的局外人获得融资条件：「很多人同时在解决很多很难的问题」（“A lot of people working on a lot of hard problems all at the same time.”）。Thompson 预计会出现过剩和失败，但认为未来取决于「从废墟中浮现出来的东西」（“emerges from the rubble”），以及那些没有泡沫就根本不可能启动的尝试。
一款理论上可将图像生成效率提升 10,000 倍的热力学芯片，是本期节目中最能体现泡沫催生实验的案例。 它的设想是按照某种概率分布激发芯片，再利用系统的物理衰变计算概率输出。这个想法听起来「完全疯了」（“totally insane”），但 Thompson 认为，概率式 AI 与概率式硬件之间存在概念上的契合，或许能让计算方式更贴合概率输出，而不是继续依赖今天的二进制、确定性芯片。
半导体管制、意识形态式抱负与 AI 狂热，正在共同对抗产业停滞。 Thompson 表示，一家新的美国晶圆厂如果独自爬过学习曲线，本来会「破产 100 次」（“100 times over”）；芯片管制反而促使中国承担不具经济性的开发成本，而没有管制的话，TSMC 可能「永远称王」（“rule the roost forever”）。要打破这种均衡，可能需要「宗教般的狂热」（“religious fervor”）：Substrate 坚信美国必须胜出，而 AI 公司行事时仿佛在「创造上帝」（“inventing God”）。
1. Substrate 的 X 射线豪赌，足以审视，但还不足以背书
Andrew Sharp 准确概括了 Substrate 的主张：让光线从粒子加速器穿过一台汽车大小的设备，打印出可与头部晶圆厂媲美的芯片层，并将半导体制造成本砍半。Thompson 对“它到底有没有戏”的回答刻意直接：「我不知道」（“I don’t know.”）。
X 射线光刻本身并不新鲜。EUV 已将光的波长压到 13纳米，但仍需要极其复杂的反射镜，有时还要进行多重图形化；每增加一道曝光，就会增加复杂度和成本。X 射线理论上可以更短，从而有望去掉重复曝光并降低成本。
真正决定成败的问题仍未解决：Substrate 是否已经实现稳定的 X 射线光源，辐射是否会干扰掺杂等其他制造步骤，以及 1980年代和 1990年代曾困扰这一方案的问题是否真的已经被攻克。一个 40年前的想法在今天变得可行，「完全有可能」（“totally plausible”），但还谈不上已经成立。
据报道，Proud 曾告诉投资人项目成功概率只有 1%，但他会以 100% 的信念推进；随着突破出现，他还表示这个概率有所上升。Sharp 认为，Proud 直面胜算、坦承不确定性，反而让自己显得更可信。
Thompson 承认有人批评他在采访 Proud 时技术追问不够，但他表示，真正关键的细节本来也不可能被披露：「只要其中确实有真东西，它显然就是一个价值极高的秘密」（“To the extent there’s something real, it is obviously a massively valuable secret.”）。
2. ASML 的胜出，靠的是优化整个系统，而不只是卖出更好的设备
Thompson 将 ASML 对 Canon 和 Nikon 的崛起，追溯到它与 TSMC 的合作，并将这段关系与 Nikon 对 Intel 的绑定作对比。「TSMC 的崛起就是 ASML 的崛起，反过来也一样」（“The rise of TSMC is the rise of ASML, and vice versa.”）。
从 200毫米晶圆转向 300毫米晶圆，最能说明两者的差异。Intel 凭借垄断利润可以容忍对吞吐量和良率关注不足，因为利润足以覆盖低效；ASML 和 TSMC 则共同重做了流程，让更大的晶圆能够以足够快的速度流转，真正兑现规模带来的产量提升。
Thompson 认为，光刻如今约占成本的 30%-35%，晶圆厂投资已在 200亿美元左右，下一代可能达到 300亿-500亿美元。他粗略估算，现有设备接近 2.5亿美元，下一代设备接近 5亿美元，同时提醒说这些数字是他「从自己屁股里掏出来的」（“pulling these numbers out of my rear end”）。一座 300亿美元、却没有填满产能并运行起来的晶圆厂，会让所有者陷入严重困境；这也是他称 TSMC 「几乎成了 AI 泡沫的刹车」（“almost the brake on the AI bubble”）的原因。
一场台湾地震——Thompson 认为发生在 2001年——成为另一个转折点：ASML 帮助 TSMC 在几周内恢复设备运行，而日本设备则需要数月。此后，ASML 完成了 EUV 艰难的商业化，成为唯一能够制造这类设备的供应商。
3. 替换光刻，可能意味着围绕新技术重建整座晶圆厂
Sharp 的反驳很直接：Proud 试图同时颠覆 ASML 和 TSMC，尽管过去 20年里挑战其中任何一家公司的尝试都以失败告终。「这显然野心太大了」（“That seems very ambitious.”）。
Thompson 的回应立足于架构。光刻是制造流程中最昂贵、最复杂、对精度要求最高的环节，周围的每一步都已围绕它调校完成——这很像 Windows 与 Intel 的合作关系。Substrate 不必重新发明刻蚀或掺杂设备，但只要改变中心环节，就意味着要重新整合围绕它的一切。
在位者的强大，反而可能制造突破口。TSMC 和 Intel 对现有方法拥有巨量专业知识，却已经「押上了全部筹码」（“pot committed”）；Thompson 认为，在满足当前需求的同时，TSMC 几乎不可能认真推进一套彻底替代方案。正因为它占据支配地位，全新的方法反而更有机会——「不会有回应」（“there’s not gonna be a response”）。
4. 泡沫经济学，会协调那些正常市场不会接受的发明
Thompson 过去主要将泡沫与浪费及声誉风险联系在一起。类似 Pets.com 的项目今天依然存在，但底层模式花了 20年、经历一场衰退，以及科技行业的「寒霜冰河期」（“cold frost ice age in tech”）才最终浮现。
传统上，生产性泡沫的回报是搁浅基础设施：破产的铁路建设者仍会留下铁轨，失败的电信公司会留下闲置光纤，电力投资则可能留下持续运行 50年或100年的电厂。GPU 缺乏这种寿命，因此今天的泡沫最终会留下什么持久遗产，反而更难回答。
Thompson 借鉴 Byrne Hobart 的 Boom，给出了新的框架：「协调式创造、协调式创新、协调式发明」（“coordinated creation, coordinated innovation, coordinated invention”）。科技公司的数百亿甚至上千亿美元流入 NVIDIA，再流向 TSMC 和 ASML；机会的规模则为大量相互依赖的实验提供资本，其中少数成功或许足以证明大量失败值得发生。
热力学芯片的案例正好说明了这一点：按照选定的概率分布激发芯片，让它衰减至中性能量，再利用这一物理过程进行计算。其理论结果——图像生成效率提升 10,000 倍——可能失败，但「在正常时期，没人会」（“in a normal period, no one”）为这种尝试提供资金。
5. 概率式技术周期，需要概率式分析
Thompson 预计这场狂热终将破裂：「总有一天我们会说，‘哇，那时候我们有点疯了’」（“At some point we’re gonna say, ‘Wow, we got a little crazy then.’”）。但他仍认为，泡沫资助的这些项目几乎肯定会从废墟中催生出有价值的技术。
这也改变了他自己的分析姿态。Stratechery 的第一个 10年提供的是相对确定的链条——聚合理论、稳定的 Big Tech 赛道、A 到 B 再到 C 的推演；而如今，他试图进行的是「建立在概率结果之上的确定性分析」（“deterministic analysis on top of probabilistic outcomes”）。Proud 能否成功无法得知；但有人正在尝试一件重要且有价值的事情，这一点要清晰得多。
地缘政治层面的同一机制，就是芯片管制：它促使中国承担普通经济学会拒绝的开发成本。Substrate 宣称自己是「一家意识形态公司」（“an ideological company”），认为美国应该获胜，这自然会招致外界要求其拿出技术证明；但 Thompson 认为，艰难的突破有时需要同样的「宗教般的狂热」（“religious fervor”），正如 AI 公司花钱时仿佛在「创造上帝」（“inventing God”）。
Hello, and welcome to a free preview of Sharp Tech. 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?
I’m doing fantastic. You look great, Andrew. Why do you look so good today?
Coming in nice and crisp today. A lot of good things happening in the Sharp Tech universe. You want to tell the people?
I have fiber. We have functional internet.
No, no, no. I think the thing to emphasize is that Ben didn’t have fiber internet for the last month or two, and so in some of the recording sessions, we were dealing with internet from 10 or 15 years ago. But now we’re coming in high-def, with no interruptions to any of the podcasts. I’m very excited about the future.
No, you understate how dire the situation was. I acquired—I bought the location that I’m at because one of my best friends literally lives next door. This, actually, I learned totally by accident. I moved into a place in Taiwan to be closer to the kids’ school.
Mm-hmm.
I moved into a place in Taiwan. Everything is a condo, right? You’re all in sort of high-rises. We’re in a 15-story building. The building we’re in has two units per floor. Literally next door ends up being this other guy from Wisconsin, who ends up becoming one of my best friends. We still hang out all the time. You’ve had the pleasure—
I’ve met him.
Yeah.
He’s great. It’s amazing that you ran into each other like that.
It is.
Yeah.
And I realize, actually, a real fundamental key to happiness is, in fact, location, location, location.
Mm-hmm.
Now, usually location is about weather, and I think weather does contribute to happiness. You should ask for an update on my weather feelings in 2 to 3 months.
Yeah. We’ll check back in February to see where your head’s at there.
But the other one is being close to friends, which is fantastic. So I got a house that is literally next door to one of my best friends. I did not think about or look into the internet situation at this house.
Mm.
This particular neighborhood was built 30 years ago or so, which means most of the people here are older at this point and maybe don’t appreciate what their expectations should be as far as internet goes. It’s a little bit remote, and their only option is the local cable company. The local cable company knows they’re the only option. There is fiber sort of across the street. I heard they built that out. Very few people took it up, I think, because of the dynamics that I’m talking about, so they just stopped building.
So I’ve basically had nonfunctional internet. I talked to technicians. I’ve had multiple technicians come over. They’re like, “Yeah, your neighborhood’s totally saturated.”
Impressively dysfunctional internet. Not just slow internet, but internet that would shut off inexplicably—
No. It goes down multiple times a day, all the time.
And just be down for 45 minutes—
Usually only for a few seconds.
Yeah.
Going down for a couple of seconds is not great for recording a podcast, okay?
Yeah.
So it’s not a few months—it’s been a while since I’ve been here. You actually had to remind me, “Make sure you switch to Starlink before every podcast.”
Yep.
Because at least it’s reliable, even if theoretically the speeds are not. So, first up, I love Starlink. God bless Elon Musk.
Thank you, Elon.
Yeah.
Yes.
No, it’s amazing. I have one on my house. I have a portable one that I’ve talked about before, using it around. But yeah, it’s absolutely, again, another example to get me back on the sort of anti-monopoly train, albeit a real sort of monopoly with wires in the ground—
Wires in the ground.
—and when there’s one option, they don’t need to fix it and make it better. So I’ve really spent an unfortunate amount of money. The way you can get fiber, you’re not going to get it from AT&T or whoever. However, there are business fiber companies that will build out to various locations if you’re willing to pay for it.
Mm-hmm.
Yeah. Well, you’re running a business over there—
I am.
—in addition to living next to your best friend.
Yes.
So it makes sense, and here we are.
It’s funny, though. I did appreciate and love every time that our podcast cut out for the last 3 months, because it just reaffirmed that this ungodly amount of money that I was spending was going to be worth it.
It’s funny because I would try to tell you, “It’s not that bad,” and you were like, “Don’t tell me that.” Then I would have the freedom to be honest and be like, “Ben, I didn’t hear what you were saying for about 4 minutes of that podcast.”
Well, maybe this is going to be a big problem for you. You’re going to have to have some accountability to actually listen to my ramblings when I’m going on a monologue.
There you go. A lot of pressure going forward.
But yes, back in the world of fiber, sincere thanks to Starlink. It has basically made Stratechery and Sharp Tech possible for the last 5 to 6 months. As good as it is, it does not compare to the clarity that we have right now. It’s amazing.
Mm-hmm.
Yes.
It’s great.
Yes.
One other question on your house: This is your first Halloween back in America. Are you prepped for the neighborhood? I feel like you’ve got a couple [?].
Full-size candy bars, baby. Full-size candy bars.
You go full-size?
Absolutely.
Wow.
You’ve got to make a good first impression.
Good.
Yes.
Good to hear. I love it. I’ll be trick-or-treating more than handing out candy with Charles, who’s about 2 and a half years old. I’m very excited for his first real trick-or-treating adventures.
My son was asking, “I’ve never really gone trick-or-treating.” I mean, they had it in Taiwan. The building always had it, so you’d go up and down the floors and get it. Then there’s one in the neighborhood where we’re at called Tianmu.
Okay.
All the stores will have stuff. It’s not quite the same as sort of walking down a busy street and going to a convenience store and getting some piece of Hi-Chew or something like that.
Yep.
The AIT, the sort of American embassy, had a neighborhood up on the top of Yangmingshan with traditional houses, and you could—
Ah.
—actually go door to door. But again, it was very much a case of, “This is a thing you’re going to do.” The idea of just going door to door in your neighborhood and it being sort of a thing—yeah, it was remarkable. You’re like, “I’ve never done that before.” It was kind of interesting. It’ll be cool.
So, do you do the Christmas light decoration thing, by the way?
What’s that?
Do you decorate your house with Christmas lights?
Oh, well, yes. I was wondering whether there was some special Christmas light thing. Yes, I do put Christmas lights on the house.
Okay. No, just thinking about planning. It’s very funny where all this stuff that is very normal, I’m sort of considering for the first time. So yeah.
A whole new world here in the United States of America.
1. Substrate Challenges Chip Manufacturing
Speaking of the United States of America, we’re going to begin the show with a note from Greg, who wrote in Wednesday responding to a digression of yours on the last episode as we were discussing American resilience. Greg writes, “It’s time for some American ingenuity, huh? Ask and you shall receive.” Then he links to Substrate.com before asking, “Does this have any legs?”
Ben, as fate should have it, you interviewed the CEO of Substrate this week. Delightful conversation. We’ll put a link in the show notes for people. And for anybody who hasn’t been following the news, Substrate is a company that claims to have developed an alternative manufacturing process for semiconductors that would cut the manufacturing cost in half by channeling light from a giant instrument known as a particle accelerator through a tool the size of a car. That technique would allow Substrate to print a high-resolution microchip layer comparable to images produced by the world’s leading foundries. I’m cribbing from The New York Times for that synopsis, but more importantly, Substrate is an American company. What do you think? Does it have legs? Does this idea have legs?
I don’t know.
Responsible answer.
Yeah. It’s really interesting because I got some pushback on this interview saying, “You didn’t push on any technical details. There are so many things that are going to go into this,” all of which are fair and appropriate. Part of the reason I’m not getting too much into technical details is that he wasn’t going to share any of them.
He couldn't, right. Yeah.
Well, yeah. To the extent there's something real, it is obviously a massively valuable secret, whatever they've figured out.
Mm-hmm.
2. Lithography Drives Chip Costs
The concept of X-ray lithography is not new. One of the weird things is that we talk about 1-nanometer, 2-nanometer, and 3-nanometer chips, which is mostly a marketing term. Basically, the idea is that the width of the lines you're drawing and the distance between the lines are getting smaller and smaller all the time.
Yeah.
The big breakthrough was EUV, which reduced the width of light to 13 nanometers. It's like, wait, actually, the light is wider than the lines? Yes, that is true, in the way that it's happening. But what happens is they do this crazy focusing with all these mirrors, and even when they get smaller, they sometimes have to go over a chip multiple times, called multi-patterning. Every time we do that, it's more complex, it's more expensive, and there are all these sorts of challenges.
Lithography has come to be by far the most difficult and most expensive step in the process. It's the number-one driver of why foundries are getting more and more expensive. We're in the $20 billion range. The next generation, we're going to be in the $30 billion to $50 billion range. I think, as you know, it might be in the $30 billion range right now.
And when we say lithography, we're talking about ASML, effectively, right?
Yes. They dominate the space. The story of how they came to dominate it is that it used to be Canon and Nikon. I think particularly Nikon was big in this space. They were very closely aligned with Intel.
There was a period when ASML was the upstart, and ASML and TSMC have always been sort of hand in hand. The rise of TSMC is the rise of ASML, and vice versa. A big factor—I think I might have talked about this on the podcast previously—was the shift from 200-millimeter wafers to 300-millimeter wafers.
Mm-hmm.
The physics of it changed. They're larger, right? Intel basically didn't care. This actually came up on the Intel earnings call, where the CFO said, "To be honest, we didn't really care about yields before."
Intel was always so far ahead and had such a monopoly and could charge such high prices that they didn't really need to care about throughput and efficiency. As long as it was good enough and they got the levels they needed, they could charge high prices and it would sort of pay for itself.
Mm-hmm.
One of the huge challenges they have in becoming a foundry is that you have to maximize throughput. You have to get as much stuff through as possible. Your yields have to be high so that you can be competitive on pricing. You're not internalizing everything; you have to charge a competitive price relative to TSMC or whatever it might be.
Right.
Their throughput was never actually that high—how many chips were going through per hour or per day, or whatever it might be. This is going back about 20 years. When you had this shift, the Nikons of the world—I can't remember whether it was Canon or Nikon that was the big one at the time—said, "Sure, we'll do bigger wafers." But they didn't rework the entire process.
If you're moving a larger wafer, you just do it more slowly because of the centrifugal forces and things like that. ASML actually completely reworked how their machines worked, hand in hand with TSMC. It's a new wafer, a new size; you have to rethink the whole thing so it can go through faster, so you can actually achieve the benefits of having a much larger wafer, which you can put more chips on.
If that larger wafer goes through much more slowly, which is basically what was going on at Intel, you're not getting any gains. You're just doing a bigger wafer more slowly. ASML and TSMC completely reworked this process to get the volume and the gains from going through it. There are other sorts of stories here.
More volume and better yields. Yeah.
Yeah.
It's kind of unbelievable.
There was a big earthquake in Taiwan in 2001, I believe it was. ASML got in there with TSMC and got the machines up and running in a matter of weeks. The Japanese machines took months. That was another sort of moment of this shift in balance and power.
All this culminated with ASML buying some of the critical technology for EUV way back when. We already discussed this. The idea came from Japan, actually. A lot of it was propagated by Intel and the U.S. government. ASML did the hard work to bring it to market, and now it's totally dominant. No one else can make an EUV machine, but they're extremely expensive.
I was going to ask: Ballpark, are we talking about $200 million?
I was like $500 million for the high... I think the current ones are around $250 million, but the next generation are closer to $500 million. I'm pulling these numbers out of my rear end, so—
And the foundry needs, like, 10 of those lithography machines?
Yeah, you need a bunch of them, right? And so—
Yeah.
So that's where the cost is in all of this.
Well, everything's expensive, to be clear, but the biggest center of cost, for sure, is lithography. I think it's around 30% to 35%, and that's much higher than it used to be.
The idea here is, what if we could do this instead of using blowing up tin, creating this artificial light, focusing it through these crazy mirrors, and drawing these lines? What if we could use X-rays to do this instead?
Mm-hmm.
It's actually even thinner, so theoretically you don't need to do multi-patterning, and it can just be cheaper overall. Now, a lot of the details are unclear, like I said, in terms of how it's going to work. The big question for X-rays has been a consistent light source, or a consistent source of X-rays, which they claim they're getting.
Okay.
They're using a particle accelerator to do it, which sounds absolutely crazy. There are questions about the X-rays breaking through and disrupting things. Does it interrupt the doping process, where you're actually infusing these transistors to make them more performant? There are all these open questions about X-rays that were faced in the '80s and '90s, and it was decided that ultraviolet light was actually a better solution with these mirrors and these optics.
That's where the industry went. Now, this theory—that it was an idea that didn't work 40 years ago but might work today—I think in principle is totally plausible.
Yeah.
I also think it's totally plausible that no one in the current semiconductor chain is going to go in this direction because they're all pot-committed to the way things work today, and they're under pressure as it is to make more and do more.
TSMC, as I wrote about this week, is kind of almost the brake on the AI bubble. Yes, they're building a ton and trying to support this. If this whole thing were to pop, they're the ones holding the bag because they're paying all this money for all this equipment.
A lot of upfront investment based on the ecosystem exactly the way it is today.
Right. It just—I mean, if it costs $30 billion to build a fab and that fab is not filled and not running—
Yeah.
—you are in big trouble. So this idea, number one, that an old idea could be possible today when it wasn't possible previously makes sense to me.
Mm-hmm.
The idea, number two, that no one in the current chain is going to do it also makes sense to me. And this idea that if you're going to do it, actually you need to not just make a lithography machine but be a foundry also makes sense because it—
So you say that makes sense. I'm very curious, because that's where I found myself raising an eyebrow. James Proud explained it to you, but Substrate—they're not only—
James Proud, yes.
James Proud, excuse me. They are not only trying to disrupt ASML but also TSMC. That seems very ambitious, and everybody—
It does.
—who's tried to displace either one of those companies has failed over the last 20 years. So why? Why are they building a foundry in addition to this lithography tool?
3. Substrate Must Rebuild The Foundry
Well, the key step—this is where I go back to that TSMC and ASML hand-in-hand partnership—is that the entire process has been optimized around this central step of lithography. It's the most expensive piece. It's the most complex piece. It requires the most precision, and everything sort of fits in around that.
If you're going to replace that, it makes sense that you actually have to rework the whole process, even if most of the steps are still the same. Those steps—you can buy the equipment.
It’s not like they’re going to reinvent etching machines, masking machines, or doping machines, or all these sorts of bits and pieces. But we already know the makers of these machines are happy enough to sell to whoever. The whole point is they’re selling to China. You’re using a lot of the same machines for trailing-edge chips as you are for leading-edge chips. Etching is going to happen at 20 nanometers, 14 nanometers, or 2 nanometers. Same thing—
Yeah.
—with a lot of these chemical coatings, all these parts. So it’s not that they’re going to reinvent every step, but if the key step is completely new, the way all that other stuff integrates into the process, it makes sense it would need to be reworked. And TSMC and ASML, it’s kind of like Windows and Intel, to an extent. It’s this hand-in-hand partnership that makes this all work, and if you’re tearing out one piece, you probably can’t just drop it into TSMC. And if you’re Intel—
Mm-hmm.
Even if you’re behind, all your expertise is still on the old way of doing things. Are you going to completely rework this? It makes sense you sort of have to start from first principles.
Yeah.
It’s also totally insane, right? And the chances of success seem—number 1, does the X-ray lithography even actually work? Have they actually overcome all these problems? It’d be nice to see more evidence than we got, just to be totally honest.
Mm-hmm.
And number 2, are you biting off more than you can chew, right? On one hand, it makes sense that a Thiel Fellow who’s most well known for creating a sleep tracker would take someone totally on the outside—
You would need fresh eyes.
—that’s kind of crazy to do this.
Yeah.
On the other hand, this is the guy who’s going to do this, right? So—
Well, I will say for the record, James Proud was pretty honest about it. He recounted conversations with investors and said, “There’s a 1% chance this will succeed, but I’m going to operate with 100% conviction that it can’t fail.” He said maybe the percentages have gone up a little bit since then as some of the technology breakthroughs have come through. But by the same token, I found him more credible by his frank acknowledgment of how long the odds are for any of this to work.
For sure. So why talk to him now? I mean, there’s an element of press—why are you discussing it? He’s like, “Well, this is not going to be a secret for much longer anyway.”
Mm-hmm.
There’s also a bit where I’m sure they’re going to be raising money soon. I’m sure this is tied to it. Was Chetecri basically used as a vehicle to help them raise money? Probably. But number 1, I’m certainly cheering for him.
Yeah.
That’s, you know, I want to see that succeed. But—
Not only that, but on the last episode, when you were talking about potential breakthroughs in the way we make chips, I recall thinking, on this end of the line, I mean, yeah, it’s great to want things. I don’t know why you would ever mention that, because for me, it’s a good reminder that the technology we have today is not fixed. It’s easy to forget that all of this can evolve, and has, when you look at the history of technology over the last 40 or 50 years.
It does make sense that the tech will be in a different place 10 or 15 years from now. Or—
It—
—at least could be.
4. Bubbles Coordinate Innovation
It does. And a topic we’ve come back to a couple of times over the last few months is that it’s so fascinating to be in a bubble. I was aware of the bubble in the ’90s. I wasn’t working in tech; I was consuming lots of stuff that was coming from the bubble, often to my financial benefit, because all these people were insane. But to actually feel the dynamics internally—
Mm-hmm.
Bubbles, we intrinsically, I think, attribute a negative connotation to them. You feel this as an analyst: “Should I write about this knowing that it could pop tomorrow, and then you look like a dummy because you supported stuff and then it blew up?” Even if you know intellectually this will pay off in the long run, right? I was just—
Yeah.
—I was just talking to John on Dithering: Pets.com truly exists today. It also took 20 years, a massive recession, and a total, total dead, cold-frost ice age in tech. Bubbles are viewed negatively. And there is the theoretical construct—the Carlota Perez construct—of productive bubbles.
Mm-hmm.
We’ve also talked about this. What is going to be the productive payoff from this bubble? Ideally, it’s power generation. We actually spend tons of money on power generation. All the investment that goes into it ends up being kaput, and people go bankrupt, but hey, we have power plants that run for 50 years or 100 years, or whatever it might be, and that is going to be a positive effect, right? The railroad tracks were a huge benefit, even though everyone building the railroad tracks went bankrupt.
My beautiful fiber probably at some point is touching some dark fiber that was laid in the late 1990s. It’s amazing, and it’s great, and we have a great video podcast, right? I appreciate WorldCom and everyone else that went bankrupt back in the day.
Yep.
What is that going to be today? You have GPUs that aren’t going to last for 50 years, right?
Mm-hmm.
And how are we going to get the productive angle? This is where I keep coming back to Byrne Hobart’s book, which was also about bubbles. He called it Boom. We discussed it. I had an interview with him.
What I got from the book originally, but didn’t feel until really recently, is that one of his points is the function of a bubble isn’t just this infrastructure investment. It’s coordinated creation, coordinated innovation, coordinated invention.
Yeah.
Basically, you have a lot of people working on a lot of hard problems all at the same time. A lot of the ones that succeed are going to pay off the other ones. Everyone’s really dependent on everyone else also inventing things at the same time.
And you have, this week, this discussion of this creation. You also have all these tech company earnings where they’re like, “Yeah, you thought we were spending a lot last year?” “Full speed ahead. We’re spending even more.” And that spending—a lot of it’s going to NVIDIA. NVIDIA—some of that money is going to TSMC. A lot of that money to TSMC is going to ASML.
And there’s so much money coming in that it is suddenly making someone like Proud and Substrate seem—
Yeah.
—why not? Yeah, it might be a 1% chance, but the payoff of that 1% is getting so astronomically large that here’s all the money you need to do it.
You have this new chip that was just announced this week. There’s this thermodynamic chip where you excite the chip according to a probability distribution that you want. As it decays and goes back to a neutral energy level, it’s actually doing computation according to a probability curve that you set it to, and you can do things like image generation, at least in theory, 10,000 times more efficiently than you can on chips.
And you think about it: We’ve talked about the idea that our chips today, even GPUs, are 1s and 0s. They’re binary computers, and it’s kind of weird we’re using binary deterministic computing for probabilistic outputs, which is what AI is. The idea, in theory, that you’ll have a probabilistic chip that actually uses natural forces, uses—
Mm-hmm.
—thermodynamics to do computation, of course that’s going to be dramatically more efficient. Of course that could align with probabilistic computing to have a probabilistic decay in this chip that is actually doing your computation. It also sounds totally insane. Who’s—
Yes.
—going to actually spend the money to do this? In a normal period, no one is.
Yeah.
In a bubble, when companies are spending tens of billions, hundreds of billions of dollars, you have all this crazy going wrong. Suddenly, all this insanity seems more viable. And yes—
So it pulls the investments forward; it makes the—
It coordinates—
—the opportunities clear to everybody—
—the investments.
Yeah.
You have all this stuff happening at the same time. And man, Byrne really nailed it with this insight. And we embrace the bubble, right?
Mm-hmm.
This is why I don’t feel bad talking to James, asking, “Is it going to work out? Have I pushed him sufficiently on the tech, or should he share more?” Probably all this is true, but is this all going to blow up, where at some point we’re going to say, “Wow, we got a little crazy then”? Yes. Is the future going to depend on what emerges from the rubble that would never have emerged otherwise if that hadn’t happened? Also, almost certainly yes.
Probably yes. Yeah.
And it’s really hard to write about and to analyze because as an analyst, you want to be sure. You want to have—
Mm-hmm.
—clarity. You want to be able to follow logic.
And it’s kind of like—sorry, dude. I hope you’re doing your job listening in my very—
I’m listening.
—high clarity—
Don’t worry.
—as I’m rambling.
I’m here.
I’ve just been thinking so much about this. It’s kind of like doing probabilistic computing on top of deterministic chips. It’s kind of weird and—
Yeah.
—a little uncomfortable. There is a bit where I want to be a deterministic analyst who follows a logic chain: A to B to C to D to E.
Mm-hmm.
I had the pleasure of doing that for the first decade of Stratechery. Actually, a lot of the job felt very easy because I felt I could see very clearly what was happening. I just needed to articulate it.
I could do a daily update about Microsoft earnings or Google earnings dropping and very clearly tie it to the dynamics of Aggregation Theory and all these sorts of pieces.
Yeah.
Welcome to the nonlinear future, Ben.
Here we are.
Yes. No, exactly.
Who knows what direction it’ll go?
I feel like I’m trying to do deterministic analysis on top of probabilistic outcomes. We’re in a probabilistic world, not just in terms of AI, but in terms of the expected value of what’s going to happen.
Mm-hmm.
Is James Proud going to figure this out? I don’t know. Is someone else—
Yeah.
—going to recreate the chip fabrication process? I don’t know. Is it important and valuable that someone’s trying to do it? Absolutely. Would the funding and the desire to do it exist without the mania that we’re in? Probably not.
Yeah.
That’s super important and valuable, and it’s—
I mean, it’s unbelievable.
—pretty crazy to be a part of.
5. Chip Controls Force Innovation
A premise of our chip conversations for the last 2½ years, specifically the Intel conversations, is that it’s just not economical for anybody in the US to try starting a foundry today. It would be insane. It’s not—
Yeah.
—going to happen.
Yeah. It takes so long to move down the learning curve that, in the meantime, you’re going to go bankrupt 100 times over because China’s making chips. TSMC is making chips. You have old GlobalFoundries running; you can’t just jump to 2 nanometers. We talked about this with China, right?
Mm-hmm.
They’re going to move down the curve because the chip controls give them a motivation to be super uneconomical. In a world without chip controls, nothing happens. TSMC just kind of rules the roost forever.
On the leading edge.
Right.
Yeah.
Right. And because, again, why would you buy from anywhere else? That’s why, number one, the chip controls are actually a big forcing function, and China probably is working on stuff like this as well. I’m sure they are looking at whatever—
They’re working tirelessly.
—Substrate is looking at—
Yeah.
—with a high degree of interest, to say the least. I’m like, “What’s your security situation like?” It better be very high.
Yep.
But leaving aside the geopolitical points and the global competition aspects, the fact of the matter is we have way more investment going into this space than we would normally. In this case, it’s because of political considerations.
The point of the bubble is that, because of the mania, because there’s so much money at stake, and because everyone is pushing in the same direction, it’s also insane. I appreciated James Proud being like, “We’re an ideological company. We think the US should win.” And you look at that, it’s like—
Mm-hmm.
“Could we have a technical reason why you’re going to win? Why this should be a thing?” But that doesn’t sound that much different from the AI companies thinking they’re inventing God.
Yeah.
You guys are religious about this. And it’s like, yes, they’re religious about this. That’s why they’re spending crazy amounts of money with the belief that there’s going to be an AI takeoff. There is going to be AI that dominates the world. And you—
Mm-hmm.
—need religious fervor to break through, to break out.
To push forward.
Stratechery was easy for the first decade because the tech industry was in stasis, because we were just coalescing around these Big Five tech companies who were in their lane, did what they did, and it got a little boring, right?
I thought in 2020 and 2021, “When’s the end?” just because I don’t want to write about antitrust—
What are we doing here?
—lawsuits for the rest of my life, right?
Yeah, totally.
So I’m getting what I wanted, good and hard. It makes my job much more difficult. But it is—
Your brain is melting on a weekly basis, but it is very, very exciting. Stasis is a good word for what the ecosystem can look like when you look around at trillion-dollar companies that are firmly entrenched, and then you remember that there’s always more to the story, and the future tends to be nonlinear and more complicated than one might expect.
Right. It’s precisely because TSMC is so dominant that you feel like a company like Substrate has a chance, just because you can see how TSMC is not going to respond to this. I’m sure they’re looking on with interest, and if they ever get asked about it, they’ll mumble something or whatever it might be.
Mm-hmm.
There’s no way they can even explore or think about this. They’re so large, so pot-committed to the current thing that’s going on, and so dominant in the space that it actually makes it more viable for something completely new to come along because there’s not going to be a response.
Yeah.
So, again, I’m not making any pronouncements on the viability, on how it’s going to happen, or on what’s going to work. I think it’s exciting that it’s happening, and all this stuff ties together.
Right.
The big tech company spending, the cost of these machines, and these crazy moonshots—again, at a 1% expected return, that overall number, if it succeeds, has to be really, really large. And we’re getting to the point of those numbers being really, really large. All right, and that is the end of the free preview. If you'd like to hear more from Ben and I, there are links to subscribe in the show notes, or you can also go to sharptech.fm. Either option will get you access to a personalized feed that has all the shows we do every week, plus lots more great content from Stratechery and the Stratechery Plus bundle. Check it out, and if you've got feedback, please email us at email@sharptech.fm.