（预告）给 TSMC AI 客户的行动倡议、华尔街的 Netflix 焦虑、关于科技界 Cignetti、OpenAI、Starbucks 的问答
- TSMC 的定价克制，与其说是议价能力之谜，不如说是塑造其统治地位的“客户优先”模式留下的遗产。 它靠保证产能赢得信任，不与客户竞争、不拿走客户的 IP，并让资本成本已经收回的晶圆厂运行数十年；Thompson 认为，这种习惯使 5 nm 和 3 nm 在先进制程晶圆厂寿命缩短、成本达到“300多亿美元”之际仍定价过低。
- 更大的投资者问题在于产能：TSMC 今年计划资本开支520亿–560亿美元，2025年为410亿美元，此前几年约为300亿美元，但成本通胀意味着供给增幅将低于表面数字的增幅。 今天的投资决策决定着2028–29年的产出；TSMC 合理地担心2028年 AI 泡沫破裂后留下闲置晶圆厂，但它的克制也可能让客户供给不足。
- TSMC 并没有消除风险，而是将其转移到下游，表现为 NVIDIA、Microsoft、Google 及其他 AI 买方永久放弃的收入。 Thompson 所说的“TSMC 刹车”就是 CEO 们所说“如果产能更多，我们本可以卖得更多”背后的供给上限：选择最稳妥的晶圆代工厂，反而可能把更大的晶圆厂风险压到客户身上。
- 行动倡议是让超大规模云厂商和模型实验室现在就把 Intel 和 Samsung 变成可信的第二供应源，尽管制程迁移需要3至4年，近期也有执行风险。 竞争会让 TSMC 更担心失去黏性强、周期长达数年的客户，而不是担心遥远的产能过剩；只是飞到台湾请求它“请多投资一点”，只会让这道刹车继续生效。
- 对 AI 买方而言，理想终局是有意让各家晶圆代工厂过度扩产，因为下行风险只是产能过剩和“超级便宜的芯片”，而上行收益是拥有足够供给来兑现需求。 Thompson 称当前买方是逃避 Intel 和 Samsung 的“胆小鬼”；如果它们继续等待，产能不足本身可能在2028–29年帮助刺破 AI 泡沫。
- 供应链集中既是地缘政治问题，也是财务问题：Sharp 指出，Dario Amodei 把流向中国的 AI 芯片比作核武器，而节目甚至没有提到台湾。 如果这座“核武器工厂”坐落在距中国海岸约80英里的地方，寻找替代晶圆代工厂既能对冲供给不足，也能对冲 Sharp 提出的战争情景。
1. TSMC 的中立晶圆代工承诺如何变成地缘政治咽喉点
Sharp 开场便猛烈抨击一处遗漏：每当 Dario Amodei 把向中国供货比作把核武器交给朝鲜，Sharp 都会搜索“台湾”，却一无所获。如果 AI 芯片具有如此高的战略意义，他问道，那么它们关键的“核武器工厂”距离中国约80英里，究竟意味着什么？
大约40年前，TSMC 凭借 Philips 提供的制程技术起步，除此之外几乎一无所有。当时它或许落后先进制程5代，因此它的卖点是可靠性：客户拿到的是有保障的产能，而不是每当 Intel 或 Texas Instruments 这样的 IDM 需要自有工厂时，就被挤出去。
中立性构成了第二道护城河。由于 TSMC 不生产与客户竞争的芯片，它可以承诺：“我们不会挤压你们的空间”，也不会据为己有客户的 IP；Sharp 指出，Samsung 可以将晶圆代工与产品业务分开，但 Apple 通过一家智能手机竞争对手生产芯片时，仍会面临根本性的利益冲突。
Sharp 想知道，如果台湾发生中断，是否只会让消费者继续使用 iPhone 13 级别的芯片。Thompson 反转了这一前提：市场注意力集中在先进制程处理器，但更难替代的是大量成熟制程芯片，而美国除了 GlobalFoundries 可能提供的产能外，几乎没有这类产能。
2. 先进制程经济学打破“永远运营”的定价模型
TSMC 历来可能只按约5年为晶圆厂计提折旧，却让工厂运行远超这一期限；一些20世纪90年代末建成的设施至今仍在生产古老、标准化的零部件，资本成本早已收回，产品价格低至几分钱。这种长寿命与“以客户为中心、客户优先的思维”共同塑造了根深蒂固的低成本文化。
芯片变得更昂贵的转折点大约出现在2014或2015年。部分买方停留在28 nm——Thompson 说中国在这一节点投入了大量资金——因为性能提升已不足以抵消溢价；7 nm 在讨论中被称为首个 EUV 节点，随着先进制程客户继续向前迁移，也在一定程度上被搁置。
TSMC 将部分5 nm产能改造为3 nm，说明先进晶圆厂不仅成本更高，寿命也更短。公司必须以积极意义上的“像 Intel 一样”运营：前置获取价值，而不是依赖数十年的后续收入。Thompson 认为，5 nm 和 3 nm 恰恰在“没有替代方案”的时候定价过低。
Thompson 提出的“Ben conjecture”明确“不是报道”：定价不足或许有助于解释前 CEO 兼董事长 Mark Liu 的突然退休，其中可能还有 Morris Chang 的影子。相比之下，Chang 在2008年回归时正值经济衰退和行业收缩，他称低迷期是 TSMC 最大的机会，并将投资押注在智能手机时代。
3. 当下资本开支为2028–29年 AI 供给设定硬上限
ChatGPT 在大约2022年出现后，TSMC 必须估算2025年和2026年的需求。Thompson 认为，TSMC 在2022–24年的保守支出决策，正是如今产能远远落后于 AI 需求的原因：半导体投资从承诺落地为可用供给需要数年。
今年计划资本开支从2025年的410亿美元、以及此前几年约300亿美元，升至520亿–560亿美元。Sharp 起初将其理解为接近翻倍；Thompson 纠正说，按同比计算约为25%，同时强调设备价格上涨意味着投入金额与新增产能之间的关系比线性增长更差。
这笔支出主要决定的是2028年和2029年，而不是2026年。TSMC 的担忧是合理的：如果 AI 泡沫在2028年破裂，数年的投资可能留下没有客户的昂贵设备。由于这项业务资本密集度极高，Thompson 说，结果可能从惊人的盈利一路走到极端情况下的“你会破产”。
4. TSMC 的谨慎将晶圆代工风险转化为客户收入损失
Thompson 的核心判断是：“风险不会从系统中消失。” TSMC 通过限制自身资产负债表上的过度扩产风险，制造了 NVIDIA、Microsoft、Google 等公司的潜在收入损失；当 CEO 们说“如果产能更多，我们本可以卖得更多”时，那些永久失去的销售额正是晶圆代工产能不足的下游结果。
这就是“TSMC 刹车”（TSMC brake）：一家供应商卡住 AI 基础设施的扩张，压低潜在泡沫的形成条件。客户以为选择最好的制程和最可靠的服务是在降低风险，但产出受限意味着它们实际上“把晶圆厂风险压到了自己身上”。
Sharp 的表述得到 Thompson 认可：相比此前增加167%，TSMC 更愿意选择增加约67%的支出，以牺牲潜在利润来限制巨大的下行风险。Thompson 的限定语至关重要：“TSMC 做的是对 TSMC 正确的事”，而它之所以能这样做，是因为缺乏竞争。
当被问到依赖 NVIDIA、Apple 和 AMD 是否构成客户集中风险时，Thompson 说，眼下的问题恰恰相反：TSMC 客户太多、产能不足。它既不能把价格抬到足以促使客户离开的程度，又必须提供足够的先进制程芯片，避免客户因为拿不到货而被迫离开。
5. AI 买方必须制造 TSMC 缺失的竞争
NVIDIA 此前通过拆分 Ampere 一代的产能保持议价能力：游戏芯片交给 Samsung，服务器芯片交给 TSMC。如今 TSMC 仍是最佳选择；Intel 14A“可能”颇具潜力，但现在承诺未必能在3至4年内产出芯片，而将设计迁移到另一种制程从来都不是小事。
一个可信的 Intel 或 Samsung，会把 TSMC 的核心担忧从5年后的产能闲置，改为今天失去客户。晶圆代工切换意味着数年期承诺，而且很难逆转，因此潜在客户流失会迫使 TSMC 增加投资，同时也会推动新供应商扩充产能。
Thompson 说，飞到台湾请求 TSMC 多投资一点的做法“完全错误”；他认为 Sam Altman 应该研究 Intel。供给有限让 TSMC 可以挑选赢家，客户则害怕得罪它，但哀求只会让这道刹车继续生效。这也是他对 OpenAI 反复抱怨“我们没有足够算力”的回应：帮助打造缺失的替代方案。
买方应该希望每一家晶圆代工厂都过度扩产：下行风险是“超级便宜的芯片”、更低的资本成本和充足的供货量。Thompson 说，自称资本家的买方必须“拿出点硬气来”；否则，2028–29年的机会可能会被产能本身扼杀。Sharp 补充说，战争也是推动供应链多元化的理由；Thompson 则称，长期硬件规划是硅谷需要练出的一项“新肌肉”。
Hello, and welcome to a free preview of Sharp Tech. Hello, and welcome back to another episode of Sharp Tech. I'm Andrew Sharp. Joining me in the same room today is Ben Thompson. Ben, how are you doing?
Pretty flustered. Apparently, I’d forgotten—I don’t know that I ever knew—that your podcast voice is a good 15 decibels louder than your regular voice.
A little bit. I was just lectured as I did my hello. I’m trying to speak in hushed tones here as we record in person because I was allegedly too loud. I listened back; it sounded normal, but here we are, sharing your basement together.
Yeah, I don’t know what to say. It’s not quieter at all.
He really is flustered here.
I know. It’s been a day of fluster, to be totally honest. Hilariously, we planned for you and some other folks in the Stratechery universe to come visit this weekend, which is going to be the coldest weekend in literally 45 years or something like that.
Yeah.
I’m trying to talk loud on the podcast today because this is actually going to be the last podcast that I’m ever able to record. Twenty below is the low tomorrow in Madison. I’m glad that I’m actually experiencing real Wisconsin winter, though, with my first trip.
I don’t think you’re glad, because you showed up in loafers and no socks.
Hey.
It was one of the most unbelievable sights I’ve ever seen in my life: Andrew Sharp coming down the airport stairs utterly unprepared for what he was about to encounter.
And immediately being roasted by Ben for about 10 minutes straight for the loafers. But look, one of my rules for adulthood is you have to dress well when you’re traveling, so I had a nice pair of loafers, nice pair of slacks, a nice little Henley here. Casual. Don’t do too much when you’re flying. But I felt good. Look good, feel good, feel good, play good.
You say that now.
Or podcast good.
Yeah, I asked to hold your hand walking to the car to make sure you didn’t slip. But—
Yeah, loafers with no socks. I’m going to have frostbitten ankles. That’s my best-case scenario for tomorrow’s weather.
In any event, here we are. We have a lot to get through today, and we’re going to begin with one of the most important tech companies of the century, and a company that we actually don’t talk about very much on the podcast. So I’m going to read a note from Sam.
That’s a you problem, not a me problem.
I was thinking about it. I was like, it’s probably my fault that we never discuss TSMC. You certainly write about TSMC plenty, but—
Well, before it was cool, for the record.
That’s true. You were early. Early on everything. That’s why people subscribe to Stratechery.
Sam says, “Ben and Andrew, what is stopping TSMC from charging more insane margins? They have clearly increased their margins after the AI boom, but they’re the only option in town, and especially during the duration of this bubble, nobody can go anywhere else. Don’t they have such insane leverage in the current moment that Apple and Nvidia would pay basically whatever they charge? Why have their margins not increased more?”
So, Ben, we get this question every couple of months. We actually answered it about two and a half years ago, but the same question came to mind for me. I was reading about the H200s and Nvidia having to go to TSMC and negotiate for more capacity in order to serve the H200 orders, and it was just a reminder that basically anyone who’s doing anything in AI is ultimately going through TSMC. You wrote about them earlier this week. Do you have an answer for Sam? Why aren’t they able to just extract crazier and crazier margins from everybody?
Well, before we get to that, I do want to compliment you, especially since I revealed your sartorial choices to the world and how inappropriate they were. At least you think about TSMC and the fact that all the AI chips come from there. That is a big improvement over everyone.
Everyone.
Well, you had Dario Amadei at Davos this week talking about how giving chips to China is like giving nuclear weapons to North Korea.
Mm-hmm.
And what did I do? As I do with every article or every comment from him about this issue, I do a Control-F. I search the document that I’m reading. I type in “Taiwan,” and there are no results.
No mention.
Can we actually—
Mm-hmm.
If these are the same as giving nuclear weapons to North Korea, what does it mean if the nuclear weapons plant is 80 miles—
A stone’s throw.
—off the coast of China?
Yeah.
Can we think through the totality of this issue?
Mm-hmm.
Anyhow, that rant aside: TSMC margins. This is actually an interesting time to revisit this question. In broad strokes, TSMC—I don’t know, how long do you want me to go on this? There really is a cultural aspect to this. You have to remember TSMC starts 40-some years ago.
Mm-hmm.
And they have nothing. They get some process technology from Philips. There’s actually an interesting universe of Philips spinoffs. ASML is also a Philips spinoff. TSMC is one. There are several other ones in this ecosystem.
But what they can offer is the fact that they have nothing to offer. And what I mean is, if you’re someone who has an idea to make a chip, you can go to them. Number 1, you can get guaranteed capacity—you’re not going to get crowded out. Because at that point, the alternative is you go to Texas Instruments, you go to Intel—maybe Intel—you beg for extra capacity, which, by the way, if they suddenly have more sales than they need, you get booted out.
Right.
And by the way—
So when you say what they can offer, you’re referring to TSMC back when they were trying to take market share in the beginning, right?
This is literally all they had: Number 1, we can guarantee you capacity; we’re not going to crowd out your order. And number 2, because we’re not making our own chips like Texas Instruments—
We’ll work with you—
—or Intel—
—to make whatever you want to make.
No, we won’t take your IP.
Oh, yeah.
That’s right.
Which is a very real and legitimate concern. There are aspects of Intel that are still not trusted in that regard.
Samsung ran into that issue as well, right?
It’s a little more complicated because, obviously, Apple used to be on Samsung. I think Samsung’s done pretty well in terms of having a wall between their foundry business and the rest of their business, but there is a fundamental conflict of interest. If you’re Apple making your chips at Samsung and Samsung is competing with you, that—
Right. I recall reading something about that tension in the Apple in China book, but I don’t know exactly what the circumstances were. But either way, in broad strokes, the conflict of interest just does not exist at TSMC, which makes it easier for everybody.
And that was literally their selling point.
Yeah.
It’s like, “We’re not going to crowd you out because, by the way, our process is like 5 processes behind the leading edge.”
Mm-hmm.
So you start out and you’re making your basic chips, right? The things that we don’t even think about—we didn’t think about until COVID, when we suddenly ran out of these super-basic chips and nothing could be made.
So they come, and then the other thing is, it’s going to be super cheap, as cheap as we can offer it. What we’re going to do is build these fabs, and they will depreciate over 5 years or whatever the number is.
Yeah.
But we’re going to run those fabs forever. Now, their earliest fabs have long since been closed down, but they still have fabs from the late 1990s. 1998 or 1999, I think, is maybe the oldest fab that they still have in operation, which is making these ancient chips that are fine for what they are. And there are certain products, long-lived products, that have just been standardized on this chip for ages and ages.
Sure.
And they sell these chips for pennies.
Money on top, though, I mean—
Right. But it’s—
They’ve already invested in the fab.
It’s already paid for.
Yeah.
Yep, exactly. And so you have, in general, this very customer-centric, customer-first mindset because that was literally the only thing they could sell, combined with this low-cost mentality—that’s what we have to offer.
We don’t have the fastest chips. We don’t have the leading-edge processes, but we’re a reliable partner for you.
So culturally, is it a struggle, then, to be extracting crazy margins because of the way they began?
Yes.
Okay.
That's basically it. So it's been a really difficult transition from that to being the leading edge.
The premium provider.
And this is one of the things that I've written about a fair bit. I think it would have made a lot of great podcast material, but unfortunately you didn't care.
One thing that was really interesting a few years ago was that they have this model: You build a fab once, and you run it forever.
They incurred extra costs, I think this was 2 or 3 years ago, because they said, “We're going to basically rework some of our 5-nanometer fabs to be 3-nanometer.”
What's interesting about that is that one of the challenges they have is these 7-nanometer fabs that are still included in their advanced-fab numbers. But they're a little bit stranded because there was a turning point somewhere around 2014 or 2015 when chips just definitely got more expensive. If you didn't need the speed or the efficiency, you could just stop.
Mm-hmm.
Actually, the biggest stopping point was probably 28 nanometers. There's just a lot of demand there, and that's where China has really invested a ton. The chips are good enough, and it's not worth the price premium to go to the next step down.
I wonder about that on a more general basis when people talk about Taiwan invasion scenarios. If everybody had to run on iPhone 13-level chips, how many people would actually notice the difference between what we have now and what we had then?
Well, no, the problem is actually the opposite. I wrote about this a few years ago, when I think the chip ban went down. My point there was that people are thinking too much about the leading edge.
Right.
Dylan Patel
The real issue is—
The legacy chips.
There are all these legacy chips that the U.S. has no capacity for. I mean, GlobalFoundries might have something or other. But the reason they don't have it—and why Intel doesn't have it—goes back to the 5-nanometer or 7-nanometer story. Intel was always on the cutting edge.
Mm-hmm.
That was their differentiation. They internalized all the gains from that by thinking, “We could sell the fastest chips because we have the best manufacturing.” But when they went to the next generation, they would dismantle the old generation, or they would try to reuse as much stuff as they could going forward. They didn't keep fabs going on forever.
Mm.
What that meant was that, if you had this 5-year depreciation, you needed to pay for the fab in those 5 years.
Yeah.
Dylan Patel
But they could do that because they could charge very high margins.
So they're taking margin, no question.
Yes, they would charge very high prices and pay for it. What happened to TSMC is that you had this overall market issue where the 28-nanometer fab you built back in the day was the fastest of its time, but then it just became a great line that you could run for 40 years.
Mm-hmm.
That old model sort of worked. When you got down to this point, you really saw it happen with 7 nanometers. It was expensive because that was the first one using EUV, and it just wasn't worth it for—
For most customers.
Dylan Patel
For most customers.
Yeah.
So it became a little bit of a stranded node, to an extent. It's still used, but—
That makes sense because all the people who would pay to be on 7 nanometers would then pay to be on—
Dylan Patel
They moved on.
—5 nanometers.
Dylan Patel
They'd already moved on.
Or 3 nanometers.
Dylan Patel
Right.
Yeah.
What was so interesting about this announcement—“Oh, we're going to incur more costs to transition”—was that they had to become like Intel.
Mm-hmm.
They had to start thinking that these incredibly expensive fabs don't just cost way more; they're also shorter-lived.
Yeah.
We're not necessarily going to be able to run them forever and get the money back, which means we have to raise prices. It was almost more of a bottom-up realization that they needed to raise prices and make more margin.
Otherwise, we're going to be underwater on this 7-nanometer investment, or 5-nanometer investment.
Right. They're not underwater, to be clear.
Right.
But the problem is that money you don't make, you don't get back. You don't get to go back to Apple and say, “3 years ago, we probably should have charged you more.”
Well, these are just tremendously expensive investments.
Oh, yeah. The latest ones are well into the $30 billions. Those 7-nanometer ones were probably $15 billion or $20 billion or something like that. I'm just pulling that out of thin air. I'm not—
So if you're putting $30 million, or $30 billion, down—
Yeah, was I saying millions? Yeah.
A million would be great.
Big difference between millions and billions. What's interesting about this is that this is also combined with the overall situation. A huge strength of TSMC is that it's very customer-centric.
Mm-hmm.
They are a customer-service organization. This is by far the hardest thing for Intel, and it will continue to be the hardest thing for them to figure out: being customer-first. But as part of that, TSMC has had a natural disinclination to jack up prices.
Mm-hmm.
Of course, they would talk about this: “Yes, of course, we need to raise prices, but we also are cognizant of a long-term relationship,” and so on. There's a bit about this where that's actually a good thing because they keep loyal customers. You don't just decide, “I'm going to go with Intel this time.” It's a multiyear commitment even to go to another foundry.
Okay.
That sort of locks them in even more. But what happened was—and this part is all, just to be clear, what’s the word I'm looking for?
Ben theory?
Ben theory. Maybe a tiny bit of scuttlebutt infuses this. What happened with this 7-nanometer bit of stranding, and then when 3 nanometers launched, is that you had this period where TSMC was by far in the lead.
Mm-hmm.
There were no alternatives. Their old model wasn't quite working anymore. They needed to switch to this model, and they did not raise prices nearly enough.
Mm.
I think that might be what happened to the previous, previous CEO, who became the chairman of the board and then suddenly retired.
Who was that?
Dylan Patel
Mark Liu.
Okay.
Again, this is just my theory. It was a weird transition and a weird thing that happened. What I think happened is that TSMC insufficiently raised prices. It cost them a lot of money that's sort of gone forever, and it came from this inherent customer-centric, conservative, bottom-up mindset. This bottom-up, cost-plus sort of thinking is very endemic to Taiwanese business culture.
Yeah.
Whereas what TSMC needs to do, to this emailer's point, is become more like Intel.
Mm-hmm.
They were forced into that in terms of how long they can depreciate their assets and having to learn how to reuse stuff, but that means they need to do it from a pricing perspective, too. People hated Intel not just because it was arrogant, but because Intel knew it was the best—
Putting the screws to them.
Dylan Patel
—and absolutely captured its value.
Well, I'm glad that we don't have to lay the failure to raise prices at Morris Chang's feet, because I love—I consider the story of him coming out of retirement to lead the smartphone era of TSMC one of my favorite stories. I consider him the Michael Jordan—
That is part of the story.
—of tech CEOs.
I'm glad you brought that up, because what happened there? TSMC was in an increasingly strong position—not fully caught up to Intel, but very caught up to Samsung in terms of making ARM chips and being the third-party foundry. Then the global financial crisis happened. The world went into recession, and you had this conservative instinct to pull back.
Right.
Dylan Patel
Morris Chang comes in—
That's my guy, Morris.
Dylan Patel
—fires everyone, and says, “This is the biggest opportunity this company has ever seen. We're investing in this.”
Wearing number 45.
Dylan Patel
“This is the biggest opportunity this company has ever seen.”
Yep.
Dylan Patel
“We're investing in this.” And that's how this whole last 15 years happened.
Now, is Morris Chang involved? I feel like if the former CEO, then chairman of the board, unceremoniously retires when I don't think he was ready to retire, that might have Morris Chang's fingerprints on it.
May have been a little bit involved. Yes. Again, this is pure conjecture. That's what we're looking for.
Yes. Okay.
This is Ben conjecture.
Ben conjecture.
A tiny bit of scuttlebutt. There might be something here, but this is not reporting. This is conjecture. Regardless, around this timeframe, I was hammering on this in the Daily Update.
Yeah.
I'm like, "TSMC is screwing up its pricing," and it's a problem for all of them. It's not just that, to the emailer's point, they have the opportunity. It's that their fundamental structure of their business is changing. They're becoming like Intel, whether they want to be or not—Intel in a positive sense, where they're on the leading edge.
They have to capture a margin much more upfront. They can't count on this trailing in the back end. They need to capture their value, and they need to capture their value then. The reason why it was so damaging with those 5-nanometer and 3-nanometer nodes—I think those 2 nodes were underpriced—was that at the time there was no alternative.
Right.
So, fast-forward to today—
That was going to be my question, because when we had a conversation along these lines 2½ years ago, you talked about NVIDIA wanting to second-source, or at least have another player that they could pit against TSMC.
Which they did, which they always did. Their previous generation, or a few generations ago, was the Ampere generation.
Mm-hmm.
They did the gaming chips with Samsung, and then the server chips with TSMC.
And that's how NVIDIA was able to keep TSMC from extracting too much margin in the value chain. Is there anybody who can play that role today, or is it basically just TSMC that's able to serve the AI demand right now?
Well, this is where it gets really interesting, and this is what I was writing about this week. The issue is TSMC is the best.
Mm-hmm.
Still the best, probably will continue to be the best. There's certainly rumblings about Intel, about 14A maybe being good. But even there, if you commit to Intel today, you're not going to have chips coming off the line for 3 or 4 years.
Yeah.
It's a maybe. Now, again, people might have already committed. It's not announced, so that could happen sooner. Or the same thing with Samsung. It's not a trivial thing to map a chip onto a new process.
If TSMC is there and willing, and the best, why would you want to go somewhere else?
And take that risk—
Take that risk.
—3 or 4 years out.
Exactly. The issue, however, is that I think it's fair to say it's clear now—
Mm-hmm.
—that the conservatism that Morris Chang fired everyone for in 2008—
Yeah.
—and the conservatism that led to them underpricing 4 or 5 years ago—
Yeah.
—it has manifested itself in their CapEx.
Okay.
What I mean is, after ChatGPT happened, there was a choice to make: How much demand do we think there's going to be in 2025 and 2026? So this was—what? Wait, 2022?
Mm-hmm.
TSMC was relatively conservative, and the net result is that, because these decisions are made years ahead of time—this is what's so hard about semiconductors—the issue is that there just isn't nearly enough capacity at TSMC for all the demand.
Mm-hmm.
Everyone's stuck, because to go somewhere else is going to take a few years. But at the same time, TSMC isn't there, and the other ones aren't good enough. So you had this whole thing last year of Sam Altman visiting Taiwan, and Jensen's here, and everyone's like… They're basically coming to TSMC saying, "Please invest more."
Mm-hmm.
But the tricky thing is, the "invest more" isn't about 2026. The "invest more" is about 2028 and 2029.
Yeah.
TSMC comes out—the reason why this is a big deal for earnings is they announced their CapEx plans, and they announced, I think, that they would spend between $52 and $56 billion on CapEx this year. A lot of money.
Mm-hmm.
This year, they spent $41 billion. I think last year, they spent $30 billion, which was way too low. The last 3 years were really—or probably the 2022 to 2024 era was—particularly too low.
This raises a really interesting question for all of the TSMC customers: Is $52 to $56 billion, in a context of prices being up in general and all the equipment being more expensive, enough? That's not a linear increase in capacity. The curve bends in the wrong way because just stuff in general is more expensive. Is that enough for—
Is it enough?
—the demand in 2028 and 2029?
Even doubling it, is that enough? I mean—
Well, but they're not doubling it. It's like a 25% increase from last year.
Okay.
So—
I thought it was $30 billion, and now it's $52 billion.
No, it was $41 billion.
Okay.
$41 billion in 2025. Around the $30 billion mark in the years before that.
Mm-hmm.
So even from there, it's like a 67% increase.
Yeah.
And the reason—the problem is that TSMC is nervous. Think about this timeline.
It's reasonable to be nervous.
Right.
Yeah.
If the bubble bursts—say, 2029 is 3 years away. I keep reading this as 2026. The bubble bursts in 2028.
Mm-hmm.
You've spent all that money, and you spend money the next year and after that, and suddenly you have fabs with all this equipment you've spent money on, and no one needs or wants the chips.
Yeah.
That's the risk of their business—
That's the risk of their business: it's so capital-intensive that if demand ebbs—
If you're—
—along the way.
You're either extremely profitable or you're just not profitable at all. It's a big—
That's what I meant—
No, you're going bankrupt.
—with even being underwater. You could really struggle—
Exactly.
—in dollar years.
But the issue is that TSMC is trying to reduce that risk.
Mm-hmm.
Risk doesn't disappear from the system. That risk is being offloaded to TSMC's customers.
How so?
Whether that be NVIDIA or Microsoft or Google or whoever it might be. What does every single CEO say on their earnings call?
I don't know. I don't listen to nearly as many earnings calls as you do.
They say, "If we had more capacity, we would've sold more."
Mm-hmm.
The risk that TSMC is offsetting is foregone revenue for all these companies. All these companies are realizing that risk right now. On every earnings call, when a CEO's on there saying, "Demand vastly exceeds supply," what that is is foregone revenue. That revenue's gone forever, and that revenue is downstream from TSMC not having enough capacity.
Right. Well, and so, is TSMC, by themselves, gating the AI infrastructure build-out—the bubbling conditions?
Yes.
Mm-hmm.
This is why I called it the TSMC brake. I came up with that last year. I should've written a big article. Maybe I'll write a big article. I think they are the brake on an AI bubble, on AI generally.
Right.
Well, no, when you wrote it last year, it didn't fully register with me. Then I read it earlier this week, and I was like, "Oh, he's making a play on accelerationists and TSMC. They're the ones that are the brakes in the middle of all this." It really is fascinating when you step back and look at the ecosystem and think about how much crazier the numbers could be if TSMC could serve the capacity in the middle of all of it.
That's exactly right. So the issue is that all these folks think they're de-risking by sticking with TSMC.
Mm-hmm.
Because it's the known entity, they have good customer service, and they have confidence it's going to work.
But there's a ceiling on how much they can produce—
They have been—
And a ceiling on how much you can make.
They actually have been loading fab risk onto themselves. That risk they're loading onto themselves is foregone revenue.
Mm-hmm.
That foregone revenue is being realized right now. You could imagine that if AI progresses like people think it will, in 2028 and 2029, this mismatch between supply and demand means $52 billion is not nearly enough.
Interesting. Yeah.
And so you're actually loading up on risk.
I'm glad you said that because, reading your update earlier in the week—I was reading it early in the morning—I was like, "It seems like Ben thinks that they should be investing a lot more in CapEx build-out than they actually are." It sounds like that was an accurate reading of the subtext of your analysis.
I think that TSMC is doing what is right for TSMC.
Okay.
And they're able to do what is right for TSMC because they don't have any competition.
There's no competition, yeah.
What I think behooves everyone in AI to do is, they have to get Samsung and Intel on board.
Mm. So Intel's not just a charity case in this scenario. Okay.
Well, I mean, I think this makes it clearer than ever. I should've made this a big blog article. Maybe I'll write another one. What happens if Intel is a credible alternative, or Samsung is a credible alternative? What becomes TSMC's greater fear? Is it that 5 years from now we might have this overhang, or is it that we're losing business to Intel and Samsung? And we know that once someone switches, it's hard to get them back because of this long sort of cycle.
And right now the risk to TSMC of underinvesting in CapEx build-out is, "Okay, we're going to make less profit in 3 or 4 years than we might have otherwise, but we're offsetting that risk against tremendous downside risks if the bubble bursts. So we're happy splitting the difference with a 67% increase as opposed to a 167% increase." That's their logic and the way they view this. Is that right?
That's right.
But that calculus changes if—
What they're doing is offloading risk onto their customers, right?
And there's a much bigger risk if they get 5 or 6 years down the line and half of their customer base is working with other competitors.
Well, what you want if you're Google, Microsoft, Amazon, OpenAI, or Anthropic is cheap chips. The way you get cheap chips is by there being too much capacity and them having to sell it at very low prices. The way you get more capacity is by having everyone overbuild. You want overbuilding. You want TSMC facing—
That's why Sam flew to Taiwan.
No, Sam's totally wrong.
Oh.
I interviewed Sam in the fall and pushed him on this: "You need to be exploring Intel." And he doesn't want to do that because everyone's a little scared of TSMC—because TSMC has limited supply.
They have limited capacity—
They get to choose—
... who... That's right.
Yeah.
That's right.
But the problem is that you're enabling the brake.
Mm.
It's the brake—the slowdown.
Yeah.
The issue is that Intel and Samsung aren't going to get there without customers actually going for it with them. They need this customer base. The reason to do it is that it shifts the risk back to the foundries. You want the foundries taking risk. You want them building for a huge explosion. The worst-case scenario is that you get super cheap chips because they built too much capacity. That's a good situation.
Mm-hmm.
And you're not going to get there. All these companies need to sack up and stop playing—
And the clear-eyed recognition would be that TSMC is just not going to go that direction—
—unless they're forced to.
It's kind of pathetic. You're flying to TSMC and begging them to put their business a little bit more at risk for your sake. No. What you need to do is go out and empower and enable an actual competitor for TSMC. That's how you get more TSMC volume. And, by the way, when that competitor comes online, there's more volume for everyone. That means lower prices. That means you can actually start creating chips, your capital costs are lower, and you can avoid these insane bills that all these companies have.
This is great stuff. You really should have made that update public.
Yes.
Bad job by you.
Yeah. I'll let it go on Monday.
Well, I have 1 question on TSMC, big picture, and potential risks that they incur in the midst of all this. How much of their revenue comes from 3 customers at this point? It's NVIDIA, Apple, and—
AMD.
Okay, AMD, and I mean in the AI space.
Well, everyone in the world has that at TSMC. Their issue right now is not that they have limited customers; it's that there are too many customers, and people are, by necessity, having to at least consider alternatives.
Mm-hmm.
Because there's not enough capacity. There are 2 parts to this. Number 1, they don't want to raise prices too high because people could leave, but they also need to have enough capacity so that people don't leave because they can't get leading-edge chips.
Get what they need.
Right.
Yeah.
The issue is that all these folks who claim to be big capitalists are being chickens. They're not taking the risk of trying to get an Intel fab going, to get a Samsung fab going, because they're thinking about their short-term risk: What if it's expensive? What if it doesn't work? What if there's a delay? All these sorts of things.
And what if TSMC gets upset and starts playing with our supply for the next 3 years?
What they're not thinking about is the long term. There is a scenario where no one does this. We get to 2028 or 2029, and what actually ends up bursting the bubble is that there's not enough chip capacity. This entire opportunity is killed off because everyone was being chickens in 2025.
Yeah. Or there's a war. That's—
Another reason—
Another omnipresent risk—
Another reason to build up alternatives, yes.
—that's been disregarded.
Well, any final thoughts on TSMC before we shift to Netflix here?
If you can't get Morris Chang to give them a kick in the rear end and take the risks—
Mm-hmm.
—you get sort of Mr. Market.
Yeah.
That's the answer. And we'll see. But this is the time now. The time now is to make decisions for 2028 and 2029, and I think tech companies are uncomfortable thinking that far in the future.
Mm-hmm.
Even though software obviously takes a long time to build, things like this—meaningful CapEx investments—it's almost like a new muscle for Silicon Valley. You've always been able to assume the hardware is there.
Right.
And especially with the rise of the cloud and being able to rent. But there needs to be this development and this increased appetite for risk because you're actually risking more by not thinking about it than you might realize.
You're risking foregoing profits. It's an interesting corner of the space, though, because even for TSMC, it's hard to chart a course for where all of this is going to be and whether the bubble will or won't burst, and whether the demand will or won't be there—
Right, but you want to put that risk on the foundries. The way you force TSMC—
—in 4 years.
Well, if you're 1 of the AI companies, of course you do. But it's a fascinating aspect of the whole discussion because they're the ones that really do have to think about the end of the decade as they're making decisions today—TSMC, that is, in addition to—
But if you're an AI company, you need to think about the end of the decade too.
Mm-hmm.
Because you're going to show up at the end of the decade and not be making nearly as much revenue as—
you could because there weren't enough chips.
Sam Altman in his interview with you is making all these investments today for what demand—
He likes—
…will look like in 2028.
OpenAI is whining over and over again: “We don’t have enough compute. We don’t have enough compute. We don’t have enough compute.” Well, then—
Or we want to be ready when there’s even more demand than there is today.
Right. And—
They’re not taking the steps—
Half a hand—
…they need to take.
Short-sighted.
Yeah.
Yeah.
Yeah.
Well, it, it's... Yeah.
All right. Well, on that note, we can shift gears and go to Netflix because you interviewed Netflix co-CEO Greg Peters this week. 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.