Dylan Patel — The single biggest bottleneck to scaling AI compute

频道: Dwarkesh Patel
视频: https://www.youtube.com/watch?v=mDG_Hx3BSUE
原文语言: en
统计: 共 189 轮 · Dwarkesh 28 · Dylan 128

[0:00] Dwarkesh

All right, this is the episode where  my roommate teaches me semiconductors. It's also the send off for this current set. It is. After you use it, I'm like, "I can't use this again. I gotta get out of here." No sloppy seconds for Dwarkesh. Dylan is the CEO of SemiAnalysis. Dylan, here’s the burning question I have for you. If you add up the big four—Amazon, Meta, Google, Microsoft—their combined forecasted CapEx this  year that you published recently is $600 billion. Given yearly prices of renting that compute,  that would be close to 50 gigawatts. Obviously, we're not putting  on 50 gigawatts this year, so presumably that's paying for compute that is  going to be coming online over the coming years.

好，这一期是我室友给我讲半导体的一集。这也是现在这套布景的告别演出。没错。你用过之后我就觉得：“这套我没法再用了，我得赶紧撤。”Dwarkesh 可不捡别人用过的。Dylan 是 SemiAnalysis 的 CEO。Dylan，我有个特别想问你的问题。把四大巨头——Amazon、Meta、Google、Microsoft——加起来，按你们最近发布的数据，它们今年合计的 CapEx 预测是 6000 亿美元。按每年租用这些算力的价格折算，那差不多相当于 50 吉瓦。显然我们今年不可能新增 50 吉瓦，所以这笔钱想必是在为未来几年才会上线的算力买单。

[0:46] Dwarkesh

How should we think about the timeline around  when that CapEx comes online? Similar question for the labs. OpenAI just announced they  raised $110 billion, and Anthropic just announced they raised $30 billion. If you look at the compute they have coming online this year—you should  tell me how much it is, but is it on the order of another four gigawatts total? The cost to rent the compute that OpenAI and Anthropic will have this year to sustain their  compute spend is $10 to $13 billion a gigawatt. Those individual raises alone are enough  to cover their compute spend for the year. And this is not even including the revenue  that they're going to earn this year.

我们该怎么理解这些 CapEx 真正落地上线的时间线？对各家实验室也是同样的问题。OpenAI 刚宣布融了 1100 亿美元，Anthropic 刚宣布融了 300 亿美元。看它们今年即将上线的算力——具体多少你来告诉我，但总共是不是也就再多 4 吉瓦的量级？OpenAI 和 Anthropic 今年要维持算力开支，租用这些算力的成本是每吉瓦 100 亿到 130 亿美元。光这几轮融资本身就足够覆盖它们今年的算力开支了，这还没算上它们今年要挣的收入。

[1:26] Dwarkesh

So help me understand: first,  what is the timescale at which the Big Tech CapEx actually comes online? And second, what are the labs raising all this money for if the yearly price of a  one-gigawatt data center is $13 billion? So when you talk about the CapEx of these  hyperscalers being on the order of $600 billion, and you look across the rest of the supply chain,  it gets you to the order of a trillion dollars. A portion of this is immediately for compute  going online this year: the chips and the other parts of CapEx that get paid this year. But there's a lot of setup CapEx as well. When we're talking about 20 gigawatts of  incremental added capacity this year in America, a portion of this is not spent this year. A portion of that CapEx was actually spent the prior year. When you look at Google having $180 billion, a big chunk of that is  spent on turbine deposits for '28 and '29.

所以帮我捋一捋：第一，大厂的 CapEx 到底是按什么时间尺度真正落地上线的？第二，如果一座 1 吉瓦数据中心的年租价是 130 亿美元，那这些实验室融这么多钱是要干嘛？——当你说这些 hyperscaler 的 CapEx 在 6000 亿美元量级时，再把供应链其余环节都算上，就到了一万亿美元的量级。其中一部分是马上用在今年上线的算力上：芯片，以及其他今年就要付掉的 CapEx。但还有大量是前期铺垫性的 CapEx。我们说今年美国新增 20 吉瓦的增量产能，其中一部分钱并不是今年花的，有一部分 CapEx 其实是前一年就花出去了。你看 Google 那 1800 亿美元，其中很大一块是花在 2028、2029 年用的燃气轮机订金上。

[2:25] Dylan

A chunk of that is spent on data  center construction for '27. A chunk of that is spent on power purchasing  agreements, down payments, and all these other things they're doing further out into the future  so they can set up this super fast scaling. This applies to all the hyperscalers  and other people in the supply chain. So with roughly 20 gigawatts deployed this year,  a big chunk is hyperscalers, and a chunk is not. For all of these companies, their biggest  customers are Anthropic and OpenAI. Anthropic and OpenAI are at roughly  two to two-and-a-half gigawatts right now, and they're trying to scale much larger. If you look at what Anthropic has done over the last few months, with $4 billion  or $6 billion in revenue added, we can just draw a straight line and say they'll  add another $6 billion of revenue a month.

还有一块花在 2027 年的数据中心建设上。还有一块花在购电协议、各类预付定金，以及他们为更远的将来做的各种布局上，好为这种超高速扩张打好地基。这适用于所有 hyperscaler 和供应链里的其他玩家。所以今年部署的大约 20 吉瓦里，很大一块是 hyperscaler 的，也有一块不是。对所有这些公司来说，最大的客户就是 Anthropic 和 OpenAI。Anthropic 和 OpenAI 现在大概在 2 到 2.5 吉瓦的水平，而且还想扩到大得多的规模。看看 Anthropic 过去几个月的表现，新增了 40 亿或 60 亿美元的收入，我们干脆画条直线外推，说他们每个月还能再新增 60 亿美元收入。

[3:16] Dylan

People would argue that’s bearish,  and that they should go faster. What that implies is they're going to add $60  billion of revenue across the next ten months. At the current gross margins Anthropic had,  as last reported by media, that would imply they have roughly $40 billion of compute spend for  that inference, for that $60 billion of revenue. That $40 billion of compute, at roughly  $10 billion a gigawatt in rental costs, means they need to add four gigawatts of  inference capacity just to grow revenue. That’s assuming their research and  development training fleet stays flat. In a sense, Anthropic needs to get to well  above five gigawatts by the end of this year.

有人会说这还是太保守了，他们应该跑得更快。这意味着接下来十个月他们要新增 600 亿美元收入。按媒体最近报道的 Anthropic 毛利率算，这意味着为了支撑那 600 亿美元收入的推理，他们大概要有 400 亿美元的算力开支。这 400 亿美元的算力，按每吉瓦约 100 亿美元的租用成本，意味着光是为了收入增长，他们就得新增 4 吉瓦的推理容量。这还是假设他们研发训练用的机群规模保持不变。从某种意义上说，Anthropic 到今年年底得远超 5 吉瓦才行。

[3:59] Dylan

It's going to be really tough for  them to get there, but it's possible. Can I ask a question about that? If Anthropic was not on track to have five gigawatts by the end of this year, but it  needs that to serve both the revenue that's gone crazier than expected—and maybe it's going to be  even more than that—plus the research and training to make sure its models are good enough for next  year: Where is that capacity going to come from? Dario, when he was on your  podcast, was very conservative. He said, "I'm not going to go crazy on  compute because if my revenue inflects at a different rate, at a different  point… I don't want to go bankrupt.

他们要做到这一点会非常艰难，但并非不可能。我能就这点问个问题吗？如果 Anthropic 并没有在年底达到 5 吉瓦的轨道上，但它又需要这个量——既要支撑涨得比预期还疯的收入（说不定还会更疯），又要搞研究和训练、确保明年的模型足够强：那这些容量要从哪来？Dario 上你播客的时候非常保守。他说：“我不会在算力上梭哈，因为如果我的收入在不同的时间点、以不同的速度拐头……我可不想破产。”

[4:31] Dwarkesh

I want to make sure that we're being  responsible with this scaling." But in reality, he's screwed the pooch  compared to OpenAI, whose approach was, "Let's just sign these crazy fucking deals." OpenAI has got way more access to compute than Anthropic by the end of the year. What does Anthropic have to do to get the compute? They have to go to lower-quality providers  that they would not have gone to before. Anthropic historically had the best  quality providers, like Google and Amazon, the biggest companies in the world. Now Microsoft is expanding across the supply chain, and they're going to other newer players. OpenAI has been a bit more aggressive on going to many players. Yes, they have tons of capacity from Microsoft, Google, and Amazon, but they also  have tons with CoreWeave and Oracle.

“我要确保我们的扩张是负责任的。”但现实是，跟 OpenAI 比，他这步棋走砸了——OpenAI 的路子是“咱们就把这些疯狂的合同统统签了”。到年底，OpenAI 能拿到的算力比 Anthropic 多得多。那 Anthropic 要拿到算力得怎么办？他们只能去找那些以前根本不会考虑的、质量更低的供应商。Anthropic 历来用的都是最优质的供应商，比如 Google 和 Amazon，全世界最大的公司。现在 Microsoft 在供应链各环节扩张，他们也开始去找一些更新的玩家。OpenAI 在广撒网这件事上一直更激进。没错，他们从 Microsoft、Google、Amazon 拿了海量容量，但在 CoreWeave 和 Oracle 那儿也有一大堆。

[5:20] Dylan

They've gone to random companies, or companies  one would think are random, like SoftBank Energy, who has never built a data center in their life  but is building data centers now for OpenAI. They've gone to many others,  like NScale, to get capacity. There's this conundrum for Anthropic because  they were so conservative on compute, because they didn't want to go crazy. In some sense, a lot of the financial freakouts in the second half of last year  were because, "OpenAI signed all these deals but they didn't have the money to pay  for them…" Okay, Oracle's stock is going to tank, CoreWeave's stock is going to tank. All these companies' stocks tanked, and credit markets went crazy because people  thought the end buyer couldn't pay for this.

他们还找过一些很“随机”的公司——或者说外人看来很随机的公司——比如 SoftBank Energy，这辈子从没建过数据中心，现在却在给 OpenAI 建数据中心。他们还找了很多别家，比如 NScale，来拿容量。Anthropic 的困境就在这儿：他们当年在算力上太保守了，因为不想冒进。某种意义上，去年下半年那一波金融恐慌，很大程度就是因为“OpenAI 签了这么多合同，但他们没钱付……”好，Oracle 的股票要崩，CoreWeave 的股票要崩。这些公司股价全跌了，信贷市场也炸了锅，因为大家觉得最终买家付不起这笔钱。

[6:02] Dylan

Now it's like, "Oh wait,  they raised a ton of money. Okay, fine, they can pay for it." Anthropic was a lot more conservative. They were like, "We'll sign  contracts, but we'll be principled. We'll purposely undershoot what we think we  can possibly do and be conservative because we don't want to potentially go bankrupt." The thing I want to understand is, what does it mean to have to acquire compute in a pinch? Is it that you have to go with neoclouds? Do they have worse compute? In what way is it worse? Did you have to pay gross margins to a cloud provider that you wouldn't have otherwise had to  pay because they're coming in at the last minute?

现在又变成了“哦等等，他们融了一大笔钱。行吧，付得起。”Anthropic 当时保守得多。他们的态度是：“合同我们会签，但要讲原则。我们会故意把目标定得低于自己实际可能做到的水平，保守一点，因为我们不想有破产的风险。”我想搞明白的是，“临时抱佛脚去抢算力”到底意味着什么？是不是就得去找 neocloud？它们的算力更差吗？差在哪儿？是不是因为最后一刻才进场，就得给云服务商付一笔本来不用付的毛利？

[6:35] Dwarkesh

Who built the spare capacity such  that it's available for Anthropic and OpenAI to get last minute? What is the concrete advantage that OpenAI has gotten if they end up  at similar compute numbers by 2027? Are they just going to end this  year with different gigawatts? If so, how many gigawatts are Anthropic and  OpenAI going to have by the end of this year? To acquire excess compute, yes,  there is capacity at hyperscalers. Not all contracts for compute  are long-term, five-year deals. There's compute from 2023 or 2024, or H100s  from 2025, that were signed at shorter terms. The vast majority of OpenAI's compute is signed  on five-year deals, but there were many other customers that had one-year, two-year,  three-year, or six-month deals, on demand.

是谁建了这些闲置产能，让 Anthropic 和 OpenAI 能在最后关头拿到？如果到 2027 年两家的算力规模差不多，那 OpenAI 具体到底赚到了什么优势？还是说他们只是今年年底的吉瓦数不一样？如果是，那到今年年底，Anthropic 和 OpenAI 各自会有多少吉瓦？要获取额外算力——是的，hyperscaler 那里是有容量的。并不是所有算力合同都是五年期长约。有 2023、2024 年的算力，或者 2025 年的 H100，签的是更短的期限。OpenAI 的算力绝大部分签的是五年期合同，但还有很多其他客户签的是一年、两年、三年、六个月的合同，甚至是按需的。

[7:20] Dylan

As these contracts roll off,  who is the participant in the market most willing to pay price? In this sense, we've seen H100 prices inflect a lot and go up. People are willing to sign long-term deals for above $2 even. I've seen deals where certain AI labs—I'm being a little bit vague here for a reason—have signed at  as high as $2.40 for two to three years for H100s. If you think about the margin, it costs  $1.40 to build Hopper, across five years. Now, two years in, you're signing deals for two  to three years at $2.40? Those margins are way higher. Now you can crowd out all of these other  suppliers, whether Amazon had these, or CoreWeave, or Together AI, or Nebius, or whoever it is. These neoclouds are the firms that had a higher percentage of Hopper in general  because they were more aggressive on it.

这些合同陆续到期时，市场上最愿意出价的参与者是谁？从这个角度看，我们已经看到 H100 价格明显拐头上涨。甚至有人愿意以 2 美元以上的价格签长约。我见过一些交易——我在这儿故意说得模糊一点是有原因的——某些 AI 实验室签 H100 两到三年的约，价格高达每小时 2.40 美元。想想这利润：Hopper 按五年摊下来，成本是每小时 1.40 美元。现在都过去两年了，你还能以 2.40 美元签两到三年的约？这利润率高太多了。这样你就能把其他所有这些供应商手里的货都抢过来，不管是 Amazon 手里的，还是 CoreWeave、Together AI、Nebius，随便谁的。这些 neocloud 普遍 Hopper 占比更高，因为他们当年在这上面更激进。

[8:23] Dylan

They also tended to sign shorter-term  deals, not CoreWeave but the others. So if I want Hopper, there  is some capacity out there. Also, while most of the capacity at an Oracle  or a CoreWeave is signed for a long-term deal in terms of Blackwell, anything that's  going online this quarter is already sold. In some cases, they're not even hitting all the  numbers they promised they would sell because there are some data center delays, not just those  two, but Nebius, Microsoft, Amazon, and Google. But there are a lot of neoclouds, as well as some  of the hyperscalers, who have capacity they're building that they haven't sold yet, or capacity  they were going to allocate to some internal use that is not necessarily super AGI-focused,  that they may now turn around and sell.

而且他们签的往往也是更短期的合同——CoreWeave 除外，其他家是这样。所以如果我想要 Hopper，市面上还是有一些容量的。另外，虽然 Oracle 或 CoreWeave 那里的容量在 Blackwell 上大多签了长约，本季度要上线的基本已经全卖光了。有些情况下，他们承诺要卖的数字甚至都没能全部兑现，因为数据中心有延期——不只这两家，Nebius、Microsoft、Amazon、Google 也一样。但还有不少 neocloud，以及部分 hyperscaler，手里有在建但还没卖出去的容量，或者原本打算留给某些内部用途——不一定是那种全力奔 AGI 的用途——现在可能掉头拿出来卖。

[9:06] Dylan

Or in the case of Anthropic, they don't  have to have all the compute directly. Amazon can have the compute and serve Bedrock,  or Google can have the compute and serve Vertex, or Microsoft can have the compute  and serve Foundry, and then do a revenue share with Anthropic, or vice versa. Basically, you're saying Anthropic is having to pay either this 50% markup in the sense of the  revenue share, or in the sense of last-minute spot compute that they wouldn't have otherwise  had to pay had they bought the compute early. Right, there's a trade-off  there. But at the same time, for a solid four months, everyone was saying to  OpenAI, "We're not going to sign deals with you."

或者对 Anthropic 来说，他们不必把所有算力都直接握在自己手里。可以是 Amazon 拿着算力去跑 Bedrock，或者 Google 拿着算力去跑 Vertex，或者 Microsoft 拿着算力去跑 Foundry，然后和 Anthropic 做收入分成，或者反过来。你的意思基本上是，Anthropic 要么以收入分成的形式付这 50% 的加价，要么就是买最后一刻的现货算力多花钱——如果早点把算力买下来，这些钱本来是不用付的。对，这里面有取舍。但与此同时，有整整四个月，所有人都对 OpenAI 说：“我们不会跟你签约。”

[9:43] Dylan

That sounds crazy, but it was  because, "you don’t have the money." Now everyone's saying, "OpenAI,  we believed you the whole time. We can sign any deal because  you've raised all this money." Anthropic is constrained in that sense. There are not that many incremental buyers of compute yet, because Anthropic hit the capability  tier first where their revenue is mooning. That's interesting. Otherwise you  might think having the best model is an extremely depreciating asset, because three  months later you don't have the best model. But the reason it's important is that you  can sign these deals, lock in the compute in advance, and get better prices. Maybe this is an obvious point.

听起来很离谱，但原因就是“你没钱”。现在所有人又都说：“OpenAI，我们一直都信你。你融了这么多钱，什么合同都能签。”Anthropic 在这个意义上是受限的。目前市场上增量的算力买家还没那么多，因为 Anthropic 率先迈进了那个能力档位，收入一飞冲天。这挺有意思的。否则你可能会觉得“拥有最强模型”是个折旧极快的资产，因为三个月后你手里的就不是最强模型了。但它之所以重要，是因为你可以靠它签下这些合同，提前锁定算力，还能拿到更好的价格。也许这是个显而易见的点。

[10:22] Dwarkesh

But at least until recently, people had made this  huge point about the depreciation cycle of a GPU. The bears, the Michael Burrys or  whoever, have said, "Look, people are saying four or five years for these GPUs. Maybe it's because the technology is improving so fast, but it in fact makes sense to have  two-year depreciation cycles for these GPUs," which increases the reported amortized CapEx  in a given year and makes it financially less lucrative to build all these clouds. But in fact you’re pointing out that maybe the depreciation cycle is even longer than five years. If we're using Hoppers—especially if AI really takes off and in 2030 we’re saying, "We have  to get the seven-nanometer fabs up, we have to go back and turn on the A100s again"—then the  depreciation cycle is actually incredibly long.

但至少直到最近，大家还在大谈特谈 GPU 的折旧周期。那些看空的人，Michael Burry 之类的，说：“你看，大家都说这些 GPU 能用四五年。可能正因为技术迭代太快，其实按两年折旧才合理。”——这样一来，每年报表上的摊销 CapEx 就变高了，建这些云的生意在财务上就没那么划算了。但你其实是在指出，折旧周期说不定比五年还要长。如果我们还在用 Hopper——尤其是如果 AI 真的起飞了，到 2030 年我们说“得把 7 纳米的 fab 开起来，得回头把 A100 重新点亮”——那折旧周期其实长得不可思议。

[11:18] Dwarkesh

I feel like that's an interesting financial  implication of what you're saying. There's a few strings to pull on there. One is, what happens to depreciation of GPUs? I guess I didn't answer your prior question,  which is that I think Anthropic will be able to get to five gigawatts-ish, maybe a little  bit more by the end of the year through themselves as well as their product being  served through Bedrock, Vertex, or Foundry. I think they'll be able to get to five or six  gigawatts, which is way above their initial plans. OpenAI will be roughly the same, actually  a little bit higher based on our numbers.

我觉得这是你这番话里一个很有意思的财务含义。这里有好几条线可以往下捋。其一是，GPU 的折旧到底会怎样？另外我好像还没回答你前面那个问题——我认为 Anthropic 到年底能做到 5 吉瓦左右，也许再多一点，一部分靠自己，一部分靠产品通过 Bedrock、Vertex 或 Foundry 对外提供服务。我觉得他们能到 5 到 6 吉瓦，远超他们最初的计划。OpenAI 大致相当，按我们的数据其实还会略高一些。

[11:59] Dylan

But anyway, the depreciation cycle of a GPU. Michael Burry was saying it's three years or less. That’s sort of his argument.  There are two lenses to look at this. Mechanically, there's a TCO model, total cost of  ownership of a GPU, where we project pricing out for GPUs and build up the total cost of a cluster. There are a number of costs: your data center cost, your networking cost, your smart hands and  people in the data center swapping stuff out. There's your spare parts, your  actual chip cost, your server cost. All these various costs get lumped together. There's some depreciation cycles on it, certain credit costs on it. You build up to, "Hey, an H100 costs $1.40/hour to deploy at volume across five  years if your depreciation is five years."

话说回来，GPU 的折旧周期。Michael Burry 说是三年甚至更短，这大概就是他的论点。看这个问题有两个视角。从机制上讲，有一个 TCO 模型，也就是 GPU 的总拥有成本：我们对 GPU 的价格做长期预测，然后把一个集群的总成本一层层搭出来。成本项有很多：数据中心成本、网络成本、数据中心里负责换件的现场运维人员（smart hands），还有备件、芯片本身的成本、服务器成本。所有这些成本加在一起，再叠加折旧周期和一定的信贷成本。最后你得出：“一台 H100 按五年折旧、规模化部署的话，每小时成本是 1.40 美元。”

[12:48] Dylan

If you sign a deal at $2/hour for those five  years, your gross margin is roughly 35%. It's a little bit above that. If you sign it for $1.90, it's 35% roughly. Then you assume at that fifth year,  the GPU falls off a bus and is dead. In some cases, the argument people are making is  if you didn't sign a long-term deal, because every two years NVIDIA is tripling or quadrupling the  performance while only 2X-ing or 50% increasing the price… Then the price of an H100… Sure  maybe the value in the market was $2 at 35% gross margins in 2024, but in 2026, when Blackwell  is in super high volume and deploying millions a year, you’re actually now worth $1/hour. And when Rubin in '27 is in super high volume—even though it starts shipping this year, it’s super  high volume next year—doing millions of chips a year deployed into clouds, you've got another  3X in performance, another 50% or 2X in price,

如果你以每小时 2 美元签下那五年的合约，毛利率大概是 35%，比 35% 略高一点。签 1.90 美元的话，也差不多是 35%。然后你假设到第五年，那块 GPU 就像从车上摔下来一样彻底报废。某些情况下，人们的论点是：如果你没签长约——因为 NVIDIA 每两年就把性能提高三四倍，而价格只翻一倍或涨 50%……那么一块 H100 的价格……好吧，也许 2024 年它在市场上的价值是每小时 2 美元、35% 毛利，但到了 2026 年，当 Blackwell 大规模量产、一年部署几百万块的时候，你实际上只值每小时 1 美元了。而当 Rubin 在 27 年进入超大批量时——虽然它今年就开始出货，但明年才是真正的大批量——每年几百万颗芯片部署进各家云，性能又是 3 倍提升，价格又涨 50% 或翻倍，

[13:44] Dylan

then the Hopper is only worth $0.70/hour. So the price of a GPU would continue to fall. That's one lens. The other lens is,  what is the utility you get out of the chip? If you could build infinite Rubin  or infinite of the newest chip, then yes, that's exactly what would happen. The price of a Hopper would fall at a spot or short-term contract rate as the new chips  come out and the price per performance goes up. But because you are so limited on semiconductors  and deployment timelines, what actually prices these chips is not the comparative thing I  can buy today, but rather what is the value I can derive out of this chip today. In that sense, let's take GPT-5.4.

那 Hopper 就只值每小时 0.70 美元了。所以 GPU 的价格会持续下跌。这是一种视角。另一种视角是：你能从这块芯片里获得多少效用？如果你能无限量地造 Rubin、无限量地造最新的芯片，那没错，事情就会那样发展：随着新芯片问世、每单位性能的价格不断改善，Hopper 的现货价或短约价会一路下滑。但正因为半导体产能和部署周期的限制这么大，真正给这些芯片定价的，不是「我今天能买到的替代品」这种比较，而是「我今天能从这块芯片里榨出多少价值」。从这个意义上说，我们拿 GPT-5.4 来举例。

[14:31] Dylan

GPT-5.4 is both way cheaper to run than  GPT-4 and has fewer active parameters. It's much smaller, in that sense  of active parameter, because it's a sparser MoE versus GPT-4 being a coarser MoE. There's also been so many other advancements in training, RL, model architecture, and data  qualities that have made GPT-5.4 way better than GPT-4. And it's cheaper to serve. When you look  at an H100, it can serve more tokens per GPU of 5.4 than if you had ran GPT-4 on it. So it's producing more tokens of a model that is of higher quality. What is the maximum TAM for GPT-4 tokens? Maybe it was a few billion dollars, maybe  it was tens of billions of dollars. Adoption takes time. For GPT-5.4, that number  is probably north of a hundred billion.

GPT-5.4 的运行成本比 GPT-4 低得多，激活参数也更少。就激活参数而言它小得多，因为它是一个更稀疏的 MoE，而 GPT-4 是一个更粗粒度的 MoE。再加上训练、RL、模型架构、数据质量上的种种进步，GPT-5.4 比 GPT-4 好太多了，而且服务成本更低。看一块 H100：跑 5.4 时，每块 GPU 能输出的 token 数比跑 GPT-4 更多。所以它是在以更高的产量，输出一个质量更高的模型的 token。GPT-4 的 token 市场规模（TAM）上限是多少？也许是几十亿美元，也许是几百亿美元。普及需要时间。而 GPT-5.4 的这个数字，很可能超过一千亿美元。

[15:23] Dylan

But there's an adoption lag, there's  competition, and there's the constant improvements that everyone else is having. If improvements stopped here, the value of an H100 is now predicated on the value  that GPT-5.4 can get out of it instead of the value that GPT-4 can get out of it. These labs are in a competitive environment, so their margins can't go to infinity. You sort of have this dynamic that is quite interesting in that an H100 is worth  more today than it was three years ago. That's crazy. It's also interesting from  the perspective of just taking that forward. If we had actual AGI models developed, if  we had a genuine human on a server… These are such hand wave-y numbers about  how many flops the brain can do.

但普及有滞后，有竞争，而且其他所有人也在不断进步。如果进步就此停住，那一块 H100 的价值现在就取决于 GPT-5.4 能从它身上榨出的价值，而不是 GPT-4 能榨出的价值。这些实验室处在竞争环境中，所以它们的利润率不可能无限膨胀。于是你就看到一个非常有意思的动态：一块 H100 今天比三年前更值钱了。这太疯狂了。把这个逻辑继续往前推也很有意思。如果我们真造出了 AGI 模型，如果服务器上真有一个货真价实的「人」……关于人脑能做多少 flops，这些数字都是非常拍脑袋的估算。

[16:08]

But on a flop basis, an H100 is estimated to  do 1e15, which is how much some people estimate the human brain does in flops. Obviously, in terms of memory, the human brain has way more. An H100 is 80 gigabytes, and the brain might have petabytes. Oh, yeah, you've got petabytes? Name a petabyte of ones and zeros, bro. Name me a string. Well, this is actually the point. No, we’ve just got the best  sparse attention techniques ever. Genuinely though. In the amount of information  that is compressed, it might be petabytes. The brain is an extremely sparse MoE. But anyways, imagine a human knowledge worker can produce six figures a year of value. If an H100 can produce something close to that, if we had actual humans on a server, the  value of an H100 is such that it can repay itself in the course of a couple of months. So when I interviewed Dario, the point I was trying to make is not that I think the singularity  is two years away and therefore Dario desperately needs to buy more compute, although the revenue is  certainly there that he needs to buy more compute.

但按 flops 算，一块 H100 估计能做到 1e15，这也正是有些人估计的人脑的 flops 水平。当然，论内存，人脑要多得多。一块 H100 是 80GB，而人脑可能有 PB 级。「哦是吗，你有好几个 PB？兄弟，那你给我背出一个 PB 的 0 和 1 来，报一段字符串听听。」「呃，这其实正是重点所在。」「不不，我们只是掌握了史上最强的稀疏注意力技术。」不过说正经的，就压缩进去的信息量而言，可能确实是 PB 级。人脑就是一个极度稀疏的 MoE。总之，想象一个人类知识工作者一年能创造六位数（美元）的价值。如果一块 H100 能创造接近这个数的价值——如果服务器上真有「人」——那 H100 值钱到几个月就能回本。所以我采访 Dario 的时候，我想表达的并不是我认为奇点还有两年就到、所以 Dario 急需买更多算力——尽管从营收上看，他确实需要买更多算力。

[18:12] Dwarkesh

The point I was trying to make is that given what  Dario seems to be saying—given his statements that we're two years away from a data center of  geniuses, and certainly not more than five years away, and a data center of geniuses should  be earning trillions upon trillions of dollars of revenue—it just does not make sense why  he keeps making these statements about being more conservative on compute or, to your point,  being less aggressive than OpenAI on compute. I guess that point got lost because then people  were roasting me, saying, "Oh, this podcaster is trying to convince this multi-hundred  billion dollar company CEO to YOLO it, bro."

我想说的是：按 Dario 自己的说法——他说我们距离「装满天才的数据中心」只有两年，而且肯定不超过五年，而一个装满天才的数据中心理应赚到数万亿又数万亿美元的收入——那他一直强调要在算力上更保守、或者像你说的在算力上不如 OpenAI 激进，这就完全说不通。我猜这个点没传达出去，因为后来大家都在喷我，说「哦，这个播客主持人想说服一家市值几千亿美元公司的 CEO 去梭哈，兄弟。」

[18:44]

I was just trying to say that internally,  his statements are inconsistent. Anyway, it's good to iron it out. I think going back to the earlier view that if the models are so powerful, the  value of a GPU goes up over time, right now only OpenAI and Anthropic have that viewpoint. But as we approach further out, everyone is going to be able to see that value skyrocket per GPU. So in that sense, you should commit now to compute. Interestingly, in Anthropic fashion, there's a bit of a meme that they have  commitment issues and are sort of polyamorous. Not Dario, but this is a bit of a meme. Explains everything. By the way, there's this interesting economic effect called  Alchian-Allen, which is the idea that if you increase the fixed cost of different goods,  one of which is higher quality and one which is lower quality, that will make people choose  the higher quality good, on the margin.

我只是想说，他的表态在内部逻辑上是自相矛盾的。总之，把这事说开了挺好。回到前面那个观点：如果模型真的那么强，GPU 的价值会随时间上涨——现在只有 OpenAI 和 Anthropic 持这种观点。但随着时间往后走，所有人都会看到每块 GPU 的价值飙升。从这个意义上说，你现在就该锁定算力。有意思的是，很有 Anthropic 风格的一点：圈子里有个梗，说他们有「承诺恐惧症」，有点「多边恋」。不是说 Dario 本人，但这确实是个梗。这就全说通了。对了，有个很有意思的经济学效应叫 Alchian-Allen 效应，大意是：如果给两种商品——一种质量高、一种质量低——同时加上一笔固定成本，那在边际上，人们会更倾向于选质量高的那个。

[19:59] Dwarkesh

To give a specific example, suppose the  better-tasting apple costs two dollars and the shittier apple costs one dollar. Now suppose you put an import tariff on them. Now it's $3 versus $2 for a great  apple versus a medium apple. Is that because they both increased by a  dollar, or should it be a 50% increase? No, because they both increased by $1. The whole effect is that if there's a fixed cost that is applied to both. Then the price difference between them, the ratio, changes. Previously, the more expensive one was 2X more expensive. Now it's just 1.5X more expensive. So I wonder if applied to AI that would mean  that, if GPUs are going to get more expensive, there will be a fixed cost  increase in the price of compute.

举个具体例子：假设更好吃的苹果卖 2 美元，差一点的苹果卖 1 美元。现在对它们加征进口关税，变成好苹果 3 美元、普通苹果 2 美元。这是因为两者都涨了 1 美元，还是应该按 50% 的比例涨？不，是因为两者都涨了 1 美元。整个效应的关键就在于：当一笔固定成本同时加在两者头上，它们之间的价格差——那个比值——就变了。原来贵的那个要贵一倍（2 倍），现在只贵 50%（1.5 倍）。所以我在想，把这套用到 AI 上是不是意味着：如果 GPU 会越来越贵，算力价格就会有一笔固定成本式的上涨。

[20:43] Dwarkesh

As a result, that will push people to be willing  to pay higher margins for slightly better models. Because the calculus is, I'm going to be  paying all this money for the compute anyway. I might as well just pay slightly more to  make sure it's the very best model rather than a model that's slightly worse. So the Hopper went from $2 to $3. If a Hopper can make a million tokens of Opus  and it can make two million tokens of Sonnet, the price differential between Opus and  Sonnet has decreased because the price of the GPU has increased by a dollar from $2 to $3.  Interesting. I think that makes a ton of sense.

结果就是，这会推动人们愿意为稍好一点的模型付出更高的溢价。因为算盘是这么打的：反正算力这笔钱我横竖都要花，不如再多花一点点，确保用的是最顶尖的模型，而不是稍差一点的那个。比如 Hopper 从每小时 2 美元涨到 3 美元。如果一块 Hopper 能产出一百万 token 的 Opus，或者两百万 token 的 Sonnet，那由于 GPU 价格从 2 美元涨了 1 美元到 3 美元，Opus 和 Sonnet 之间的价差（比值）就缩小了。「有意思。我觉得这非常有道理。」

[21:22] Dylan

We just see all of the volumes  are on the best models today, all the revenue is on the best models today. In a compute-limited world, two things happen. One, companies that don't have commitment issues  and have these five-year contracts for compute have locked in a humongous margin advantage. They've locked in compute for five years at the price it transacted at  two, three, or five years ago. Whereas if you're three years into that  five-year contract and someone else's two-year or three-year contract rolled off, and  now they're trying to buy that at modern pricing, when it's priced to the value of models,  the price is going to be up a lot more.

我们现在看到的就是：所有的用量都在最好的模型上，所有的收入也都在最好的模型上。在一个算力受限的世界里，会发生两件事。第一，那些没有「承诺恐惧症」、签了五年期算力合约的公司，已经锁定了巨大的利润率优势。他们以两年、三年甚至五年前成交的价格锁定了五年的算力。而如果你的五年合约刚走到第三年，别人的两年或三年合约到期了，现在要按当下的价格去买——当算力是按模型价值来定价的时候——价格会高出一大截。

[22:05] Dylan

So the person who committed early  has better margins in general. The percentage of the market that is in long-term  contracts is much larger than the percentage of the market in short-term contracts that can be  this flex capacity you add at the last second. At the same time, where does the margin go? Because models get more valuable, how much can the cloud players flex their pricing? If you look at CoreWeave, their average  term duration is over three years right now. For ninety-eight percent plus of  their compute, it's over three years. They end up with this conundrum  where they can't actually flex price.

所以总体上，早早锁定的人利润率更好。市场上签长约的比例，远大于签短约的比例——短约就是你最后一刻才临时加的那种弹性容量。与此同时，利润流向了哪里？既然模型越来越值钱，云厂商能把价格上调多少？看 CoreWeave：他们目前的平均合约期限超过三年。98% 以上的算力，合约期都在三年以上。于是他们陷入一个两难：其实根本没法灵活调价。

[22:46] Dylan

But every year they're adding incrementally  way more capacity than they had previously. This year alone, Meta's adding as much capacity as  they had in their entire fleet of compute and data centers for all purposes for serving WhatsApp,  Instagram, and Facebook in 2022, and doing AI. They're adding that alone this year. In the same sense, you talk about Meta doing that, CoreWeave, Google, and Amazon, all these companies  are adding insane amounts of compute year on year. That new compute gets transacted at the new price. In a sense, yes, you've locked in, as long as we're in a takeoff. "Oh, OpenAI went from six  hundred megawatts to two gigawatts last year, and from two gigawatts to six plus this  year, and six to twelve next year."

但他们每年新增的容量都比以前多得多。光是今年，Meta 新增的容量就相当于他们 2022 年整个算力和数据中心机群的总和——那可是当年支撑 WhatsApp、Instagram、Facebook 全部业务外加做 AI 的全部家当。他们今年一年就要新增这么多。同样地，你说 Meta 在这么干，CoreWeave、Google、Amazon，所有这些公司每年都在疯狂新增算力。而新增的算力是按新价格成交的。某种意义上，没错，只要我们还处在起飞阶段，你锁定了就是赚的。「哦，OpenAI 去年从 600 兆瓦涨到 2 吉瓦，今年从 2 吉瓦涨到 6 吉瓦以上，明年再从 6 涨到 12。」

[23:29] Dylan

The incremental added compute is where all the  cost is, not the prior long-term contracts. Then who holds the cards is the  infra providers for charging margin. Now the cloud players, the neoclouds, or  the hyperscalers can charge the margin. They can to some extent, but then as you go  upstream to who has access to all the memory and logic capacity, it's Nvidia for the most part. They've signed a lot of long-term contracts. They've got ninety billion dollars of long-term  contracts today, and they're negotiating three-year deals today with the memory vendors. You've got Amazon and Google through Broadcom, Amazon directly, and AMD. These companies hold all the cards because they've secured the capacity. TSMC is not raising prices, but memory vendors are, to some extent, raising a lot of price. They're going to double or triple price again, but then they're also signing these long-term deals. Who is able to accrue all the margin dollars is potentially the cloud, potentially the  chip vendors, and the memory vendors,

所有的成本都在新增的那部分算力上，而不是之前的长约上。那么，握着「收利润」这副牌的是谁？是基础设施提供商。云厂商——新兴云（neocloud）或超大规模云厂商（hyperscaler）——可以收这个利润。他们能收一部分，但再往上游走，谁掌握着所有的内存和逻辑芯片产能？大部分是 Nvidia。Nvidia 签了大量长约，今天手里有 900 亿美元的长期合约，而且现在正跟内存厂商谈三年期的协议。还有 Amazon 和 Google 通过 Broadcom、Amazon 自己直接谈，以及 AMD。这些公司握着所有的牌，因为他们锁定了产能。TSMC 没在涨价，但内存厂商在某种程度上大幅涨价。他们还会再涨一倍甚至两倍，但同时也在签这些长约。最终能把利润赚走的，可能是云厂商，可能是芯片厂商，还有内存厂商，

[24:28]

until TSMC or ASML break out and say,  "No, we're going to charge a lot more." But at the same time, do the model  vendors get to charge crazy margins? At least this year, we're going to see  margins for the model vendors go up a lot. Because they're so capacity constrained,  they have to destroy demand. There's no way Anthropic can continue at  the current pace without destroying demand. Let's get into logic and memory. How specifically has Nvidia been able to lock up so much of both? I think according to your numbers, by '27, Nvidia is going to have +70% of  N3 wafer capacity, or around that area.

直到 TSMC 或 ASML 跳出来说：「不行，我们要大幅提价。」但与此同时，模型厂商能收取离谱的高利润吗？至少今年，我们会看到模型厂商的利润率大幅上升。因为产能太紧张，他们不得不去「摧毁需求」。Anthropic 不可能在不压制需求的情况下维持现在的增速。「我们来聊聊逻辑芯片和内存吧。具体来说，Nvidia 是怎么把这两样都锁走这么多的？按你们的数据，到 27 年，Nvidia 会拿下 N3 晶圆产能的 70% 以上，大概是这个量级。」

[25:12] Dwarkesh

I forget what the numbers were for memory  at SK Hynix and Samsung and so forth. Think about how the neocloud business  works and how Nvidia works with that, or how the RL environment business  works and how Anthropic works with that. In both those cases, Nvidia is purposely trying to  fracture the complementary industry to make sure that they have as much leverage as possible. They're giving allocation to random neoclouds to make sure that there's not one  person that has all the compute. Similarly, Anthropic or OpenAI, when they're  working with the data providers, they say, "No, we're going to just seed a huge industry  of these things so that we're not locked into any one supplier for data environments." And I wonder why on the 3 nm process—that's going to be Trainium 3, that's going to be  TPU v7, other accelerators potentially—why is TSMC just giving it all up to Nvidia  rather than trying to fracture the market?

我不记得 SK Hynix 和 Samsung 那边内存的具体数字了。想想 neocloud 这门生意是怎么运作的、Nvidia 是怎么跟它们打交道的；或者 RL 环境这门生意是怎么运作的、Anthropic 是怎么跟它们合作的。这两种情况下，Nvidia 都在刻意让互补产业碎片化，确保自己握有尽可能大的话语权。他们把配额分给各种不知名的 neocloud，确保没有哪一家独占全部算力。类似地，Anthropic 或 OpenAI 在跟数据供应商合作时会说：「不，我们要直接扶植出一整个行业，这样就不会被任何一家数据环境供应商锁死。」那我就纳闷了：在 3 纳米制程上——上面会有 Trainium 3、TPU v7，可能还有其他加速器——为什么 TSMC 就把产能全让给了 Nvidia，而不是想办法让市场碎片化？

[26:09] Dylan

There are a couple points here. On 3 nm, if we go back to last year, the vast majority of 3 nm was Apple. Apple is being moved to 2 nm. Memory prices are going up, so  Apple's volumes may go down. As memory prices go up, either  they cut margin or they move on. There's some time lag because they have  long-term contracts, but Apple likely reduces demand or moves to 2 nm faster, where  2 nm is only capable of mobile chips today. In the future, AI chips will move there.  So Apple has that. Apple is also talking to third-party vendors because they're  getting squeezed out of TSMC a little bit. TSMC's margins on high-performance computing—HPC,  AI chips, et cetera—are higher than they are for mobile, because they have a bigger  advantage in HPC than they do in mobile.

这里有几个点。关于 3 纳米：回到去年，3 纳米的绝大部分产能是 Apple 的。Apple 正在被迁往 2 纳米。内存价格在涨，所以 Apple 的出货量可能下滑。内存越贵，他们要么压缩利润，要么转走。因为有长约，这里有时间差，但 Apple 很可能会减少需求，或者更快转向 2 纳米——而 2 纳米今天只做得了移动芯片，未来 AI 芯片才会迁过去。所以 Apple 是一个因素。Apple 还在跟第三方代工厂接触，因为他们在 TSMC 那边有点被挤出去了。TSMC 在高性能计算——HPC、AI 芯片等等——上的利润率比移动芯片高，因为他们在 HPC 上的领先优势比在移动端更大。

[27:00] Dylan

When you look at TSMC’s running calculus  here, they're actually providing really good allocations to companies that are doing CPUs. When you think about Amazon having Trainium and Graviton, both of those are on 3 nm, Graviton  being their CPU, Trainium being their AI chip. TSMC is much more excited to give  allocation to Graviton than they are to Trainium because they view the CPU  business as more stable, long-term growth. As a company that is conservative and doesn't  want to ride cycles of growth too hard, you actually want to allocate to the market that  is more stable with a lower growth rate first before you allocate all the incremental capacity  to the fast growth rate market. That is the case generally. Same for AMD. The allocations they  get on their CPUs, TSMC is much more excited about those than they are for GPUs. Likewise  for Amazon. Nvidia is a bit unique because yes, they have CPUs, they make switches, they make  networking, NVLink, InfiniBand, Ethernet, NICs.

从 TSMC 的算盘来看，他们其实把很不错的配额给了做 CPU 的公司。想想 Amazon 同时有 Trainium 和 Graviton，两者都在 3 纳米上——Graviton 是他们的 CPU，Trainium 是 AI 芯片。TSMC 更乐意把配额给 Graviton 而不是 Trainium，因为在他们眼里，CPU 业务是更稳定的长期增长。作为一家风格保守、不想过度追逐增长周期的公司，你会先把产能分给增速较低但更稳定的市场，然后才把增量产能全给高增速市场。大体上都是如此。AMD 也一样：他们 CPU 拿到的配额，TSMC 比对 GPU 的更上心。Amazon 同理。Nvidia 有点特殊，因为没错，他们有 CPU，还做交换机、做网络——NVLink、InfiniBand、以太网、NIC 网卡。

[28:11] Dylan

By and large, most of these things will  be on 3 nm by the end of this year with the Rubin launch and all the chips in that  family, the GPU being the most important one. Yet Nvidia is getting the majority of supply. Part of this is because you look at the market and TSMC and others forecast market demand in  many ways, but it's also the market signal. The market signaled, "Hey, we need this much  capacity next year. We need this much. We'll sign non-cancelable, non-returnable. We  may even pay deposits." Nvidia just did it way earlier than Google or Amazon. In some cases, Google and Amazon had stumbling blocks. One of the chips got delayed slightly by a couple quarters. Trainium and all these sorts of things happened.

总体来说，到今年年底，随着 Rubin 发布，这些东西大部分都会用上 3nm——整个产品家族的所有芯片，其中 GPU 是最重要的一颗。而 NVIDIA 拿到了大部分产能。一部分原因是，你看这个市场，TSMC 和其他厂商会用很多方式预测市场需求，但另一部分靠的是市场信号。市场发出的信号是：「嘿，我们明年需要这么多产能，需要这么多。我们愿意签不可取消、不可退货的订单，甚至可以付定金。」NVIDIA 只是比 Google 或 Amazon 早得多就这么干了。有些情况下，Google 和 Amazon 自己遇到了绊脚石，有一颗芯片延期了几个季度，Trainium 之类的事情都发生过。

[29:01] Dylan

In that case, there was a huge sort  of, "Well, these guys are delaying, but Nvidia is wanting more, more, more, more. And we are checking with the rest of the supply chain, is there enough capacity?" They're going to all the PCB vendors and saying, "Is there enough PCB?" Victory Giant is one of the largest suppliers of PCBs to Nvidia, and they're a Chinese company. All the PCBs come from China, or many of them. They're like, "Do you have enough PCB capacity?  Great. Hey memory vendors, who has all the memory capacity? Okay, Nvidia does. Great." When you  look at who is AGI-pilled enough to buy compute on long timelines at levels that seem ridiculous  to people who aren't AGI-pilled—but nonetheless, they're willing to pay a pretty good margin and  sign it now because they view in the future that ratio is screwed up—the same thing happens  with the supply chain for semiconductors.

在那种情况下，就出现了一个巨大的反差：「这几家在往后拖，而 NVIDIA 在喊还要、还要、还要。我们正在跟供应链其他环节确认，产能够不够？」他们跑去问所有 PCB 厂商：「PCB 够不够？」Victory Giant（胜宏科技）是 NVIDIA 最大的 PCB 供应商之一，它是一家中国公司。PCB 全部或者说很大一部分来自中国。他们就问：「你们 PCB 产能够吗？很好。嘿，内存厂商们，内存产能都在谁手里？哦，都在 NVIDIA 手里。很好。」你去看谁足够信 AGI（AGI-pilled）、愿意在很长的时间线上、以在不信 AGI 的人看来离谱的量级去买算力——但他们照样愿意付相当不错的溢价、现在就签，因为他们判断未来那个供需比会彻底失衡——半导体供应链上发生的是同一件事。

[29:54]

I don't think Nvidia is quite AGI-pilled. Jensen doesn't believe software is going to be fully automated and all these things. Accelerated computing, not AI chips, right? It's AI chips. But that's what he calls it, right? Yeah. I think it's a broader term, AI is within  that, but also physics modeling and simulations. But it's like he's not  embracing the main use case. I think he's embracing it, but I just don't  think he's AGI-pilled like Dario or Sam. But he's still way, way more AGI-pilled than  Google was in Q3 of last year, or Amazon was in Q3 of last year, and he saw way more demand.  The reason is pretty simple. You can see all the data center construction. He's like, "Okay, I want to have this market share." We have all the data centers tracked, and there's a lot of data centers  that could be one or the other.

我不觉得 NVIDIA 算得上真正的 AGI 信徒。Jensen 并不相信软件会被完全自动化这些说法。——他讲的是加速计算（accelerated computing），不是 AI 芯片，对吧？——其实就是 AI 芯片，但他就是这么叫的，对吧？——对。我觉得那是个更宽的词，AI 包含在里面，但也包括物理建模和仿真。——可这等于他没有拥抱最主要的用例。——我觉得他是拥抱的，只是他不像 Dario 或 Sam 那样是 AGI 信徒。但他还是远比去年三季度的 Google、去年三季度的 Amazon 信得多，而且他看到的需求也多得多。原因很简单：所有数据中心的建设你都看得见。他就想：「好，我要拿下这个市场份额。」我们把所有数据中心都跟踪着，有很多数据中心用谁的芯片都有可能。

[30:44]

To some extent, Google and Amazon, Google  especially, even though their TPU is just better for them to deploy, they have to  deploy a crap load of GPUs because they don't have enough TPUs to fill up their  data centers. They can't get them fabbed. I have a question about that. Google sold a million, was it the v7s? Yes. —the Ironwoods to Anthropic, and you're saying the  big bottleneck right now, this year or next year, I guess going forward forever now,  is going to be the logic and memory, the stuff it takes to build these chips. Google has DeepMind, the third prominent AI lab. If this is the big bottleneck, why would they  sell it rather than just giving it to DeepMind?

某种程度上，Google 和 Amazon——尤其是 Google——尽管自家 TPU 对他们来说明明更划算，也不得不部署海量 GPU，因为他们的 TPU 不够填满自己的数据中心，芯片根本排不上产（fab）。——我有个问题。Google 卖了一百万颗，是 v7 吧？——对。——就是卖给 Anthropic 的 Ironwood。你又说现在最大的瓶颈，今年也好明年也好，估计从今往后一直是逻辑芯片和内存，也就是造这些芯片所需的东西。Google 自己有 DeepMind，第三大 AI 实验室。既然这才是大瓶颈，他们为什么要把芯片卖掉，而不是直接给 DeepMind 用？

[31:24] Dylan

This is again a problem of… DeepMind people  were like, "This is insane. Why did we do this?" But Google Cloud people and Google  executives saw a different thought process. You and I know the compute team at Anthropic. Both of the main people came from Google. They saw this dislocation, they negotiated  a deal, and they were able to get access to this compute before Google realized. The chain of events, at least from our data that we found, was in early Q3, over  the course of six weeks, we saw capacity on TPUs go up by a significant amount. It went up multiple times in those six weeks. There were multiple requests. Google  even had to go to TSMC and explain to them why they needed this increase in  capacity because it was so sudden.

这又是一个……DeepMind 的人当时就觉得：「这太疯狂了，我们为什么要这么干？」但 Google Cloud 的人和 Google 高管的思路不一样。你我都认识 Anthropic 的算力团队，两位核心人物都是从 Google 出来的。他们看到了这个错位，谈成了一笔交易，赶在 Google 反应过来之前拿到了这批算力。事件链条——至少从我们找到的数据看——是在三季度初，六周之内，我们看到 TPU 的产能上调了一大截，那六周里上调了好几次，有多次追加请求。Google 甚至不得不去 TSMC 解释为什么突然需要这么大的产能增量，因为来得太突然了。

[32:21] Dylan

A lot of that capacity increase  was for selling to Anthropic. Because Anthropic saw it before Google. And then Google had Nano Banana and Gemini 3 which caused their user metrics to skyrocket. Then leadership at Google was like, "Oh." Then they started making the statement that  we have to double compute every six months, or whatever the exact number was. They really woke up a lot more, and then they went to TSMC and said, "We want more. We want  more." TSMC replied, "Sorry guys, we're sold out. We can maybe get 5-10% more for 2026,  but really we're going to work on 2027." There was this information asymmetry among  the labs, in my mind. I don't know exactly.

那次产能增量里很大一部分就是为了卖给 Anthropic。因为 Anthropic 比 Google 自己更早看到了这件事。后来 Google 出了 Nano Banana 和 Gemini 3，用户指标一飞冲天，Google 领导层才反应过来：「哦。」然后他们开始放话说我们的算力必须每六个月翻一倍——具体数字记不清了。他们真正醒过来了很多，于是跑去 TSMC 说：「我们还要更多、还要更多。」TSMC 回答：「抱歉各位，卖光了。2026 年也许还能多挤出 5-10%，但真正能给你们做的是 2027 年。」在我看来，各家实验室之间存在这种信息不对称。具体我也不完全确定。

[32:59] Dylan

It's the narrative I've spun myself from  seeing all the data in the supply chain on wafer orders and what's going on with the data  centers that Anthropic and Fluidstack signed. It's pretty clear to me that Google screwed up. You can see this from Google's Gemini ARR. They had next to nothing in Q1 to Q3—in Q3  a little bit once they started inflecting. But in Q4 they reached $5 billion  in revenue on an ARR basis. It's clear Google didn't see  revenue skyrocket initially. In a sense, Anthropic had a little bit of  commitment issues before their ARR exploded, even though they had far more information  asymmetry and saw what was coming down the pipe.

这是我自己从供应链的所有数据里——晶圆订单、Anthropic 和 Fluidstack 签的数据中心的动向——拼出来的叙事。在我看来 Google 搞砸了这一点很清楚。从 Google 的 Gemini ARR 就能看出来：一季度到三季度几乎没有收入，三季度开始拐头才有一点点，但到四季度他们按 ARR 口径达到了 50 亿美元。很明显 Google 一开始没有预见到收入会暴涨。某种意义上，Anthropic 在自家 ARR 爆发之前也有点下不了决心，尽管他们掌握的信息不对称优势大得多，早就看到了后面要来的东西。

[33:44] Dylan

Google is going to be more conservative  than Anthropic and Google had even less ARR. So they were just not willing to do it,  and then they realized they should do it. Since then, Google has gotten absurdly  AGI-pilled in terms of what they're doing. They bought an energy company. They're  putting deposits down for turbines. They're buying a ridiculous  percentage of powered land. They're going to utilities and  negotiating long-term agreements. They're doing this on the data center  and power side very aggressively. I think Google woke up towards the end  of last year, but it took them some time.

Google 只会比 Anthropic 更保守，而当时 Google 的 ARR 还更少。所以他们就是不愿意下注，后来才意识到应该下。从那以后，Google 在动作上变得离谱地「AGI 信徒化」了：他们买了一家能源公司；给燃气轮机下定金；买下了高得离谱的比例的已通电土地（powered land）；跑去跟电力公司谈长期协议。在数据中心和电力这一侧，他们做得非常激进。我认为 Google 是去年年底才醒过来的，但花了他们一些时间。

[34:26]

How many gigawatts do you think Google  will have by the end of next year? Buy my data. You charge for that kind of information. Yes, yes. I feel like every year the  bottleneck for what is preventing us from scaling AI compute keeps changing. A couple years ago it was CoWoS. Last year it was power. You'll tell me  what the bottleneck is this year. But I want to understand five years  out, what will be the thing that is constraining us from deploying the singularity? The biggest bottleneck is compute. For that, the longest lead time supply chains  are not power or data centers. They're actually the semiconductor  supply chains themselves.

你觉得到明年年底 Google 会有多少吉瓦？——买我们的数据吧。——这种信息你们是收费的。——对，对。——我感觉每一年，卡住 AI 算力扩张的瓶颈都在换。前几年是 CoWoS，去年是电力，今年是什么瓶颈你待会儿告诉我。但我想搞清楚的是五年之后，到底是什么东西会卡住我们部署「奇点」？——最大的瓶颈是算力本身。而在这件事上，交付周期最长的供应链不是电力也不是数据中心，恰恰是半导体供应链自己。

[35:01] Dylan

It switches back from power and data  centers as a major bottleneck to chips. In the chip supply chain, there's  a number of different bottlenecks. There's memory, logic wafers from  TSMC, and the fabs themselves. Construction of the fabs takes two to three years,  versus a data center which takes less than a year. We've seen Amazon build data  centers in as fast as eight months. There's a big difference in lead  times because of the complexity of building the fab that actually makes the chips. The tools also have really long lead times. The bottlenecks, as we've scaled, have shifted based on what the supply  chain is currently not able to do.

瓶颈会从电力和数据中心切回到芯片。芯片供应链里有好几个不同的瓶颈：内存、TSMC 的逻辑晶圆，还有晶圆厂（fab）本身。建一座 fab 要两到三年，而数据中心一年都用不了——我们见过 Amazon 最快八个月建成一座数据中心。交付周期差距这么大，是因为真正造芯片的 fab 复杂度完全不同。设备的交付周期也非常长。随着规模扩张，瓶颈一直在随着供应链当下「做不到的那件事」而转移。

[35:44] Dylan

It was CoWoS, power, and data centers, but  those were all shorter lead time items. CoWoS is a much simpler process  of packaging chips together. Power and data centers are ultimately way simpler  than the actual manufacturing of the chips. There's been some sliding of capacity  across mobile or PC to data center chips, which has been somewhat fungible. Whereas CoWoS, power, and data centers have had to start anew as supply chains. But now there's no more capacity for the mobile and PC industries—which used to be the majority  of the semiconductor industry—to shift over to AI. Nvidia is now the largest customer at TSMC  and SK Hynix, the largest memory manufacturer.

之前的瓶颈是 CoWoS、电力、数据中心，但那些都是交付周期较短的环节。CoWoS 是把芯片封装到一起，工艺简单得多；电力和数据中心归根结底也远比芯片制造本身简单。过去还能把手机、PC 的产能挪一部分给数据中心芯片，这块是有一定弹性的；而 CoWoS、电力、数据中心则得当作全新的供应链从头建。但现在，手机和 PC 行业——它们曾经占半导体行业的大头——已经没有更多产能可以再挪给 AI 了。NVIDIA 现在是 TSMC 的最大客户，也是最大内存厂商 SK Hynix 的最大客户。

[36:33]

It's sort of impossible for the  sliding of resources away from the common person's PCs and smartphones  to shift any more towards the AI chips. So now the question is how do  we scale AI chip production? That's the biggest bottleneck as we go to 2030. It would be very interesting if there's an absolute gigawatt ceiling that you can project  out to 2030 based just on "We can't produce more than this many EUV machines." To scale compute further, there are different bottlenecks this year and  next year, but ultimately by 2028 or 2029, the bottleneck falls to the lowest rung  on the supply chain, which is ASML.

从普通人的 PC 和智能手机那边继续往 AI 芯片挪资源，基本已经不可能再挪了。所以现在的问题是：AI 芯片产量怎么扩上去？这是我们走向 2030 年最大的瓶颈。——那就很有意思了：是不是可以纯粹基于「EUV 光刻机的产量就这么多」推出一个到 2030 年的吉瓦绝对天花板？——要继续扩算力，今年和明年各有各的瓶颈，但最终到 2028 或 2029 年，瓶颈会落到供应链最底层的那一环，也就是 ASML。

[37:16] Dylan

ASML makes the world's most  complicated machine: an EUV tool. The selling price for those is $300-400 million. Currently, they can make about 70. Next year, they'll get to 80. Even under very aggressive supply chain expansion, they only get to a little  bit over 100 by the end of the decade. What does that mean? They can make a hundred of these  tools by the end of the decade, and 70 right now. How does that actually translate to AI compute? We see all these numbers from Sam Altman and many others across the supply chain:  gigawatts, gigawatts, gigawatts. How many gigawatts are we adding? We see Elon saying a hundred gigawatts in space.

ASML 造的是全世界最复杂的机器：EUV 光刻机。一台的售价是 3 到 4 亿美元。目前他们一年能造 70 台左右，明年能到 80 台。即便供应链非常激进地扩张，到这个十年结束也只能做到 100 台出头。这意味着什么？十年末他们一年能造一百台这种机器，现在是 70 台。这怎么换算成 AI 算力？我们看到 Sam Altman 和供应链上上下下很多人都在报数字：吉瓦、吉瓦、吉瓦。我们到底在新增多少吉瓦？Elon 还说要在太空里搞一百吉瓦。

[37:55]

A year. A year. The problem with any of these numbers, or the challenge to these numbers,  is actually not the power or the data center. We can dive into that, but  it's manufacturing the chips. Take a gigawatt of Nvidia's Rubin chips. Rubin is announced at GTC, I believe the week this podcast goes live. To make a gigawatt worth of data center capacity of Nvidia's latest chip that they're  releasing towards the end of this year, you need a few different wafer technologies. You need about 55,000 wafers of 3 nm. You need about 6,000 wafers of 5 nm, and then  you need about 170,000 wafers of DRAM memory.

——一年一百吉瓦。——对，一年。这些数字的问题，或者说对这些数字的真正挑战，其实不在电力也不在数据中心——那个我们可以展开聊——而在芯片制造。以一吉瓦的 NVIDIA Rubin 芯片为例。Rubin 会在 GTC 上发布，我记得就是这期播客上线那一周。要用 NVIDIA 今年年底推出的最新芯片建一吉瓦的数据中心容量，你需要几种不同的晶圆工艺：大约 55,000 片 3nm 晶圆、大约 6,000 片 5nm 晶圆，然后还需要大约 170,000 片 DRAM 内存晶圆。

[38:41] Dylan

Across these three different buckets,  each requires different amounts of EUV. When you manufacture a wafer, there are thousands  and thousands of process steps where you're depositing material and removing them. But the key critical step—which at least in advanced logic is 30% of the  cost of the chip—is something that doesn't actually put anything on the wafer. You take the wafer, you deposit photoresist, which is a chemical that chemically  changes when you expose it to light. Then you stick it into the EUV tool, which  shines light at it in a certain way. It patterns it. There's what's called a mask,  which is effectively a stencil for the design.

这三大类晶圆各自需要的 EUV 用量不同。制造一片晶圆有成千上万道工序，不断地沉积材料、再去除材料。但最关键的一步——在先进逻辑芯片里占到芯片成本的 30%——其实并不往晶圆上添加任何东西。你拿来晶圆，涂上光刻胶（photoresist），这是一种见光会发生化学变化的化学品；然后把它送进 EUV 光刻机，机器以特定方式向它照射光线，进行图形化。里面有个叫掩膜版（mask）的东西，本质上就是电路设计的「模板」。

[39:16] Dylan

When you look at a leading-edge 3 nm wafer, it has  70 or so masks, 70 or so layers of lithography, but 20 of them are the most advanced EUV. If you need 55,000 wafers for a gigawatt, and you do 20 EUV passes per wafer, you can do the math. That's 1.1 million passes of EUV for a single gigawatt. It's pretty simple. Once you add the  rest of the stuff, it ends up being 2 million, across 5 nm and all the memory. You're at roughly 2 million EUV passes for a single gigawatt. These tools  are very complicated. When you think about what it's doing across a wafer, it's taking  the wafer and scanning and stepping across.

看一片最先进的 3nm 晶圆，它大约有 70 层掩膜、70 层光刻，其中 20 层用的是最先进的 EUV。如果一吉瓦需要 55,000 片晶圆，每片要过 20 道 EUV，你算一下：单单一吉瓦就是 110 万次 EUV 曝光。很简单。再把其他部分加上——5nm 和所有内存——总数就到了 200 万次。也就是说一吉瓦大约要 200 万次 EUV 曝光。这些机器非常复杂。想想它在晶圆上干的事：拿着晶圆，一边扫描一边步进地移过去。

[40:03] Dylan

It does this dozens of times  across the whole wafer. When you're talking about how  many EUV passes, that’s the entire wafer being exposed at a certain rate. An EUV tool can do roughly 75 wafers per hour, and the tool is up roughly 90% of the time. In the end, you need about three and a half EUV tools to do the 2 million EUV  wafer passes for the gigawatt. So three and a half EUV  tools satisfies a gigawatt. It's funny to think about the numbers. What does  a gigawatt cost? It costs roughly $50 billion. Whereas what do three and a half EUV tools cost?  That's $1.2 billion. It's actually quite a lower number, which is interesting to think about. Fifty gigawatts of economic CapEx in the data center, and what gets built on top of  that in terms of tokens is even larger.

它要在整片晶圆上重复几十次这个动作。我们说多少次 EUV 曝光，指的是整片晶圆按某个速率被完整曝光一遍。一台 EUV 光刻机每小时大约能处理 75 片晶圆，设备正常运转时间大约 90%。算下来，要完成一吉瓦所需的 200 万次 EUV 晶圆曝光，大约需要三台半 EUV 光刻机。也就是说三台半 EUV 机器就能撑起一吉瓦。这些数字想想挺有意思：一吉瓦要花多少钱？大约 500 亿美元。而三台半 EUV 机器多少钱？12 亿美元。其实是个低得多的数字，这一点很值得琢磨。数据中心里 50 吉瓦的经济性 CapEx，再算上在这之上用 token 产出的东西，规模还要更大。

[40:56] Dylan

It might be $100 billion worth of  AI value into the supply chain, three years, TSMC has done $100 billion of CapEx.  So it's $30/$30/$40 billion. A small fraction of that is being used by Nvidia for the 3 nm, or  previously 4 nm, that it's using for its chips. What were its earnings last  quarter? It was $40 billion. So $40 billion times four is $160 billion. Nvidia alone is turning some small fraction of $100 billion in CapEx, which is going to  be depreciated over many years and not just this one year, into $160 billion in a single year. That gets even more intense when you go down the supply chain to ASML, which is taking a billion  dollars' worth of machines to produce a gigawatt.

可能是 1000 亿美元的 AI 价值进入供应链，而三年里 TSMC 做了 1000 亿美元的资本开支，大概按 300 亿/300 亿/400 亿这样分。其中只有一小部分被 NVIDIA 用在它芯片所用的 3 纳米（之前是 4 纳米）产能上。NVIDIA 上个季度的营收是多少？400 亿美元。400 亿乘以四就是 1600 亿。光 NVIDIA 一家，就把这 1000 亿资本开支里的一小部分——而且这些开支要折旧很多年，不是只摊在这一年——变成了一年 1600 亿美元。往供应链下游走到 ASML，这个倍数就更夸张了：价值 10 亿美元的机器就能产出一吉瓦（的 AI 芯片）。

[41:54]

Of course, those machines last for more  than a year so it’s doing more than that. Now I want to understand, how many  such machines will there be by 2030, if you include not just the ones that are sold  that year, but have been compiling over the previous years? What does that imply? Sam Altman  says he wants to do a gigawatt a week in 2030. When you add up those numbers,  is it compatible with that? That's completely compatible,  if you think about it. TSMC and the entire ecosystem have  something like 250 to 300 EUV tools already. Then you stack on 70 this year, 80  next year, growing to 100 by 2030.

当然，这些机器能用不止一年，所以它实际产出的还不止这个数。那我想搞清楚：到 2030 年会有多少台这样的机器——不只算当年卖出的，还要算前几年累积下来的？这意味着什么？Sam Altman 说他想在 2030 年做到每周一吉瓦。把这些数字加起来，跟他这个目标对得上吗？——你仔细算一下，完全对得上。TSMC 和整个生态现在已经有大约 250 到 300 台 EUV 光刻机了。然后今年再加 70 台，明年 80 台，到 2030 年增长到每年 100 台。

[42:30] Dylan

You're at 700 EUV tools by the end of the decade.  700 EUV tools, at three and a half tools per gigawatt—assuming it's all allocated to AI, which  it's not—gets you to 200 gigawatts worth of AI chips for the data centers to deploy. Sam wants 52 gigawatts a year. He's only taking 25% share then. Obviously, there's some share given to mobile and PC, assuming we're even allowed to have consumer  goods still and we don't get priced out of them. But roughly, he's saying 25% market  share of the total chips fabbed. That's very reasonable given that this  year alone, I think he's going to have access to 25% of the Blackwell GPUs  that are deployed. It's not that crazy.

到这个十年结束时你就有 700 台 EUV 光刻机了。700 台 EUV，按每吉瓦三台半算——假设全都分配给 AI，实际上并不是——能给数据中心提供 200 吉瓦的 AI 芯片。Sam 想要一年 52 吉瓦，那他也只占 25% 的份额。当然，还得分一些给手机和 PC——前提是我们还被允许拥有消费电子产品、没被价格挤出去。但粗略来说，他要的就是全部代工芯片里 25% 的市场份额。这其实很合理，因为光是今年，我估计他就能拿到已部署 Blackwell GPU 里 25% 的量。所以没那么疯狂。

[43:23]

When did ASML start shipping  EUV tools, when 7 nm started? I don't know when that was exactly. You're saying in 2030, they're going to be using machines that initially were shipped in 2020. So for ten years, you're using the same most important machine in this  most technologically advanced industry in the world? I find that surprising. ASML's been shipping EUV tools now for roughly a decade, but it only entered mass volume production  around 2020. The tool's not the same. Back then, the tools were even lower throughput. There are various specifications around them called overlay. I was mentioning you're stacking layers on top of each other. You'll do some EUV, you'll do a bunch of different process steps—depositing stuff,  etching stuff, cleaning the wafer—dozens of those steps before you do another EUV layer. There's a spec called overlay, which is: you did all this work, you drew these lines  on the wafer, now I want to draw these dots.

ASML 是什么时候开始出货 EUV 光刻机的？是 7 纳米量产开始的时候吗？我不清楚具体是哪年。你的意思是，到 2030 年他们还在用最早 2020 年出货的机器。也就是说在这个全世界技术最先进的行业里，最关键的那台机器一用就是十年？我觉得这挺让人意外的。——ASML 出货 EUV 光刻机到现在差不多有十年了，但真正进入大规模量产是在 2020 年前后。而且机器早就不是同一台机器了。当年的机器吞吐量还要低得多。这些机器有各种规格指标，其中一个叫套刻精度（overlay）。我前面提过，你是在一层一层往上叠：先做一道 EUV，然后做一堆不同的工艺步骤——沉积材料、刻蚀、清洗晶圆——几十道工序之后才做下一层 EUV。有个指标叫 overlay，意思是：你做完了前面所有工作，在晶圆上画好了这些线，现在我要往上画这些点。

[44:17] Dylan

Let's say I want to draw these dots to  connect these lines of metal to holes, and then the next layer up is another set of lines  going perpendicular, so now you're connecting wires going perpendicular to each other. You have to be able to land them on top of each other. It's called overlay. Overlay is  a spec that's been improved rapidly by ASML. Wafer throughput has been  improved rapidly by ASML. The price of the tool has gone up, but not  as much as the capabilities of the tool. Initially, the EUV tools were $150 million. Over time, they're now $400 million as I look out to 2028. But the capabilities of the tools have more than doubled as well, especially  on throughput and overlay accuracy, which is the ability to accurately align the subsequent  passes on top of each other even though you do tons of steps between. ASML is improving super  rapidly. It's also noteworthy to say that ASML is maybe one of the most generous companies  in the world. They have this linchpin thing.

比方说，我要画这些点，把这些金属线连到通孔上，再上一层又是一组垂直方向的线，这样就把互相垂直的导线连了起来。你必须让它们精确地落在彼此正上方，这就叫套刻（overlay）。overlay 这个指标被 ASML 快速改进，晶圆吞吐量也被 ASML 快速提升。机器的价格是涨了，但涨幅比不上机器能力的提升。最早 EUV 光刻机是 1.5 亿美元一台；往 2028 年看，现在已经到 4 亿美元了。但机器的能力也提升了不止一倍，尤其是吞吐量和套刻精度——也就是中间隔着大量工序，仍能把后续每次曝光精确对准前一层的能力。ASML 进步非常快。还有一点值得一提：ASML 可能是全世界最慷慨的公司之一。他们是那种『咽喉要道（linchpin）』式的存在。

[45:19] Dylan

No one has anything competitive. Maybe China will  have some EUV by the end of the decade, but no one else has anything even close to EUV, and yet they  haven't taken price and margins up like crazy. You go ask some other folks that we  talk to all the time, like Leopold, and they're like, "Let's have the price go  up." Because they can. The margin is there. You can take the margin. Nvidia takes the  margin. Memory players are taking the margin. But ASML has never raised the price more than  they've increased the capability of the tool. In a sense, they've always provided  net benefit to their customers.

没有任何人拿得出有竞争力的东西。也许中国到这个十年结束时会有一些 EUV，但除此之外没人有任何接近 EUV 的东西——可 ASML 并没有借机疯狂抬价、拉高利润率。你去问我们经常聊的另一些人，比如 Leopold，他们会说：『让价格涨上去啊。』因为他们做得到，利润空间就在那儿，你可以把利润吃掉。NVIDIA 就在吃这个利润，内存厂商也在吃。但 ASML 从来没有让涨价幅度超过机器能力的提升幅度。某种意义上，他们始终在给客户提供净收益。

[45:54]

It's not that the tool is stagnant,  it's just that these tools are old. Yes, you can upgrade them some,  and the new tools are coming. For simplicity's sake, we're  ignoring the advances in overlay or throughput per tool for this podcast. You say we're producing 60 of these machines this year and then 70, 80 over subsequent years. What would happen if ASML just decided  to double its CapEx or triple its CapEx? What is preventing them from  producing more than 100 in 2030? Why are you so confident that  even five years out, you can be relatively sure what their production will be? I think there are a couple factors here.

不是说机器停滞不前，只是这些机器确实旧了。对，你可以做一些升级，新机器也在路上。为了简化，这期播客我们先忽略单台机器在 overlay 和吞吐量上的进步。你说今年生产 60 台这种机器，之后几年是 70 台、80 台。那如果 ASML 干脆把资本开支翻倍、翻三倍会怎样？是什么挡着他们在 2030 年产出超过 100 台？为什么你这么有信心，哪怕往后看五年，也能相对确定他们的产量？——我觉得这里有几个因素。

[46:31]

ASML has not decided to just go YOLO,  let's expand capacity as fast as possible. In general, the semiconductor  supply chain has not. It's lived through the booms and busts,  and we can talk a bit more about it. Basically some players have recently woken  up, but in general no one really sees demand for 200 gigawatts a year of AI chips, or  trillions of dollars of spend a year in the semiconductor supply chain. They're  not AI-pilled. They're not AGI-pilled. We're going to get to a  trillion dollars this year. Yeah, I feel you, but I'm saying no one  really understands this in the supply chain.

ASML 并没有决定 YOLO 一把、以最快速度扩产能。整个半导体供应链总体上都没有。他们经历过一轮轮的繁荣与萧条，这个我们可以再多聊聊。基本上，最近有些玩家醒过来了，但总体上没有人真的预见到一年 200 吉瓦的 AI 芯片需求，或者半导体供应链一年几万亿美元的开支。他们没有被 AI 洗礼（AI-pilled），也没有被 AGI 洗礼。——我们今年就要到一万亿美元了。——是，我懂你的意思，但我想说的是，供应链里没有人真正理解这件事。

[47:11] Dylan

Constantly, we're told our numbers are  way too high, and then when they're right, they're like, "Oh, yeah, but your next  year's numbers are still too high." ASML's tool has four major components. It has the source, which is made by Cymer in San Diego. It has the reticle stage, which is made in Wilmington, Connecticut. It has the wafer  stage. It has the optics, the lenses and such. Those last two are made in Europe. When you look at each of these four, they're tremendously complex supply chains that,  (A) they have not tried to expand massively, and (B) when they try to expand  them, the time lag is quite long.

我们一直被人说数字定得太高，等到被证明是对的了，他们又说：『哦对，但你们明年的数字还是太高。』ASML 的光刻机有四大核心部件：光源，由圣迭戈的 Cymer 制造；掩模台（reticle stage），在康涅狄格州 Wilmington 制造；晶圆台（wafer stage）；还有光学系统，也就是镜头之类。后两样在欧洲制造。你逐个去看这四块，每一块背后都是极其复杂的供应链，而且（A）它们没有尝试过大规模扩张，（B）真要扩的时候，时间滞后也相当长。

[47:55] Dylan

Again, this is the most complicated machine  that humans make, period, at any sort of volume. Let's talk about the source specifically. What  does the source do? It drops these tin droplets. It hits it three subsequent  times with a laser perfectly. The first one hits this tin  droplet, it expands out. It hits it again, so it expands  out to this perfect shape, and then it blasts it at super high power. The tin droplets get excited enough that they release EUV light, 13.5 nanometer, and then  it's in this thing that is collecting all the light and directing it into the lens stack. Then you have the lens stack, which is Carl Zeiss, as you mentioned, and some other folks, but  Zeiss being the most important part of it.

再说一遍，这是人类以任何量产规模制造的机器里最复杂的一台，没有之一。具体说说光源。光源是干什么的？它滴下这些锡液滴，然后用激光连续精准地打它三次。第一发打中锡滴，它膨胀开；再打一发，膨胀成一个完美的形状；然后再用超高功率轰击它。锡滴被激发到足以释放出 EUV 光——13.5 纳米波长——然后有一个装置把所有光收集起来，导入镜组。接下来就是镜组，像你提到的，是 Carl Zeiss 和另外一些厂商做的，其中 Zeiss 是最关键的部分。

[48:36] Dylan

They also have not tried to expand  production capacity because they don't see... They're like, "We're growing a lot because of AI. We're growing from 60 to 100." It's like, "No, no, no. We need to go to a couple hundred, but it's  fine. Whatever." Each of these tools has, I think, 18 of these lenses, effectively. They are multilayer mirrors, which are perfect layers of molybdenum  and ruthenium, if I recall correctly, stacked on top of each other in many layers,  and then the light bounces off of it perfectly. When we think about a lens, it's in  a shape, and it focuses the light. This is like a mirror that's also  a lens, so it's pretty complicated.

他们也一样没有去扩产能，因为他们看不到……他们的想法是：『我们因为 AI 已经增长很多了，从 60 涨到 100。』而实际情况是：『不不不，我们需要的是几百台。不过算了，随便吧。』每台光刻机里大概有 18 组这样的『镜头』。它们其实是多层反射镜——如果我没记错，是钼和钌的完美薄层，一层层叠很多层，光在上面完美地反射。我们通常理解的透镜是有个形状、负责聚焦；而这个东西既是反射镜又是透镜，所以相当复杂。

[49:16] Dylan

Any defect in these super thinly  deposited stacks will mess it up. Any curvature issues will mess it up. There are a lot of challenges with scaling the production. It's quite artisanal in this sense because you're not making tens of thousands  of these a year, you're making hundreds, you're making thousands. 60 tools a year, 18 of  these per tool, you’re still in the hundreds, of tools, or you're at the thousand number  roughly for these lenses and projection optics. Then you step forward to the reticle stage,  which is also something really crazy. This thing moves at, I want to say, nine Gs. It will shift nine Gs because as you step across a wafer, the tool will go... The  wafer stage is complementary. It's the wafer part. You line these two things up. You're taking all the light through the lenses that's focused, and here's  the reticle, here's the wafer.

这些极薄的沉积叠层里只要有任何缺陷，就全毁了；曲率有任何问题，也全毁了。扩大生产面临很多挑战。从这个意义上说它相当『手工作坊』，因为你一年做的不是几万件，而是几百件、几千件。一年 60 台机器、每台 18 组镜片：机器还是几百台的量级，这些镜片和投影光学件也就大约一千件的量级。再往前看掩模台，那也是个非常疯狂的东西。它的运动加速度我记得是 9 个 G。它会以 9 个 G 移动，因为在晶圆上逐区曝光时，机器会……晶圆台是与之配套的，负责晶圆那一侧。你要把这两样东西对齐：所有经过镜组聚焦的光，这边是掩模（reticle），那边是晶圆。

[50:16] Dylan

The reticle's moving one direction, the  wafer's moving the other direction as it scans a 26x33 millimeter section  of the wafer, and then it stops. It shifts over to another part  of the wafer and does it again. It does that in just seconds. Each of them is moving at nine Gs in opposite directions. Each of these things is a wonder and marvel of chemistry, fabrication, mechanical engineering,  and optical engineering, because you have to align all these things and make sure they're perfect. All of these things have crazy amounts of metrology because you have  to perfectly test everything. If anything is messed up, the yield goes to  zero, because this is such a finely tuned system.

掩模朝一个方向移动，晶圆朝相反方向移动，同时扫描晶圆上一块 26x33 毫米的区域，然后停下，挪到晶圆的另一块区域再来一遍。整个过程只需要几秒钟，而两个部件各自以 9 个 G 朝相反方向运动。这里每一样东西都是化学、制造工艺、机械工程和光学工程的奇迹，因为你必须把所有这些对齐并确保完美。所有环节还配了数量惊人的量测（metrology）设备，因为一切都必须被完美检测。任何一处出问题，良率就直接归零——这个系统就是精密调校到这种程度。

[50:58]

By the way, it's so large that you're building  it in the factory in Eindhoven, Netherlands, and they're deconstructing it and shipping it on  many planes to the customer site, and then you're reassembling it there and testing it again. That process takes many, many months. There are so many steps in the supply  chain, whether it's Zeiss making their lenses and projection optics or Cymer, which is  an ASML-owned company, making the EUV source. Each of these has its own complex supply chain. ASML has commented that their supply chain has over ten thousand people in it. Like individual suppliers?

顺便说一句，这台机器大到什么程度：你在荷兰 Eindhoven 的工厂里把它造出来，然后拆解、用很多架飞机运到客户现场，再在那边重新组装、重新测试。这个过程要花好几个月。供应链里的环节实在太多——不管是 Zeiss 做镜片和投影光学件，还是 Cymer（ASML 旗下的公司）做 EUV 光源，每一家背后又都有自己复杂的供应链。ASML 说过，他们的供应链里有超过一万个参与方。——你是说一万家独立供应商？

[51:32] Dylan

Yes. It might not be directly. It might  be through Zeiss having so many suppliers and XYZ company having so many suppliers. If you just think about it, you're talking about two physically moving objects that are  the size of a wafer, and it has to be accurate to the level of single-digit nanometers or even  smaller because the entire system, the overlay, the layer-to-layer overlay variation,  has to be on the order of 3 nanometers. If the overlay is 3 nms, that means each  individual part, the accuracy of its physical movement has to be even less than that. It has to be sub-one nanometer in most cases, because the error of these things stack up. There's no way to just snap your fingers and increase production. Things as simple as power.  The US going from zero percent power growth to two percent power growth, even though China's  already at thirty, was so hard for America to do.

对。不一定都是直接供应商，可能是 Zeiss 有一大堆供应商，某某公司又有一大堆供应商。你想想看：两个物理上高速运动、尺寸跟晶圆差不多大的物体，精度却要做到个位数纳米甚至更小，因为整个系统的套刻——层与层之间的 overlay 偏差——必须控制在 3 纳米的量级。如果 overlay 是 3 纳米，那意味着每个单独部件物理运动的精度还得比这更小，大多数情况下要做到 1 纳米以下，因为这些误差是会叠加的。你没法打个响指就把产量提上去。就连电力这种简单的事：美国从电力零增长提到 2% 的增长——而中国已经在 30% 了——对美国来说都那么难。

[52:34]

And that's a really simple supply chain with  very few people in it who make difficult things. There are probably 100,000  electricians and people who work in the electricity supply chain, or more, in the US? When you look at ASML, they employ so few people. Carl Zeiss probably employs less than a  thousand people working on this, and all of those people are super, super specialized. You can't just train random people up for this in the snap of a finger. You can't just get your entire supply chain to get galvanized. Nvidia's had to do a lot to get the entire supply chain to even deliver  the capacity they're going to make this year.

而电力还算是一条相当简单的供应链，里面做高难度工作的人并不多。美国电力供应链里的电工和从业者大概有 10 万人，甚至更多吧？再看 ASML，他们雇的人非常少。Carl Zeiss 做这块的可能不到一千人，而且这些人个个都是超级专业化的。你没法打个响指就把随便什么人训练成这样，也没法让整条供应链一下子动员起来。NVIDIA 光是为了让整个供应链交付出他们今年要的产能，就已经下了很大功夫。

[53:15] Dylan

When you go talk to Anthropic, they're  like, "We're short of TPUs, we're short of training, and we're short of GPUs." When you go talk to OpenAI, they're like, "We're short of these things." OpenAI and Anthropic know they need X. Nvidia is not quite as AGI-pilled.  They're building X - 1. You go down the supply chain, everyone's doing X - 1. In some cases, they're doing X ÷ 2, because they're not AGI-pilled. You end up with this time lag for the whip to react. The AI-pilledness and the desire to increase production takes so long. Once they finally understand that they need to increase production rapidly…  They think they understand.

你去跟 Anthropic 聊，他们说：『我们缺 TPU，缺训练算力，缺 GPU。』你去跟 OpenAI 聊，他们也说缺这缺那。OpenAI 和 Anthropic 知道自己需要 X。NVIDIA 没那么信 AGI（AGI-pilled），他们按 X-1 来建。再往供应链下游走，每一层都在做 X-1，有些环节甚至在做 X÷2，因为他们不信 AGI。结果就是整条『鞭子』的反应存在时间滞后。这种对 AI 的信念、以及扩产的意愿，传导起来太慢了。等他们终于明白需要快速扩产的时候……他们是自以为明白了。

[53:57] Dylan

They think AI means we have to go from 60  to 100, in addition to the tools getting better and faster, the source getting  higher power from 500 watts to 1,000, and all these other aspects of the supply chain  advancing technically and increasing production. They think they're actually  increasing production a lot. But if you flow through the  numbers… What does Elon want? He wants 100 gigawatts a year  in space by 2028 or 2029. Sam Altman wants 52 gigawatts a  year by the end of the decade. Anthropic probably needs the  same, and Google needs that. You go across the supply chain, and it's  like, wait, no, the supply chain can't possibly build enough capacity for everyone  to get what they want on the side of compute.

他们认为 AI 意味着我们必须从 60 提到 100，除此之外还要让设备变得更好更快、光源功率从 500 瓦提到 1000 瓦，供应链的其他各个环节也都要在技术上进步、产能上扩张。他们确实认为自己在大幅扩产。但你把数字一路算下来……Elon 想要什么？他想在 2028 或 2029 年之前做到每年往太空送 100 吉瓦。Sam Altman 想在这个十年结束前做到每年 52 吉瓦。Anthropic 大概也需要同样的量，Google 也需要。你顺着供应链看一圈就会发现：等等，不对，供应链根本不可能建出足够的产能，让每家在算力这边都拿到自己想要的量。

[55:44] Dwarkesh

I feel like in the data center  supply chain for the last few years, people have been making arguments like, "We are  bottlenecked by this specific thing, therefore AI compute can't scale more than X." But as you've written about, if the grid is a bottleneck, then we just do behind the  meter on the site, we do gas turbines, et cetera. If that doesn't work, there are all these  other alternatives that people fall back on. I want to ask whether we can imagine a similar  thing happening in the semiconductor supply chain. If EUV becomes a bottleneck, what if we  just went back to 7 nm and did what China is doing currently, producing 7 nm chips  with multi-patterning with DUV machines?

我感觉过去几年在数据中心供应链里，大家一直在做这样的论证：「我们被某个具体环节卡住了，所以 AI 算力的扩张不可能超过 X。」但正如你写过的，如果电网是瓶颈，那我们就在场地上搞表后供电（behind the meter）、上燃气轮机等等。如果那也不行，还有一堆别的替代方案可以退而求其次。我想问的是，半导体供应链里能不能想象出类似的事情。如果 EUV 成了瓶颈，我们干脆退回 7 纳米、照中国现在的做法来行不行——用 DUV 光刻机加多重曝光（multi-patterning）生产 7 纳米芯片？

[56:31] Dwarkesh

If you look at a 7 nm chip like the  A100, there's been a lot of progress obviously from the A100 to the B100 or B200. How much of that progress is just numerics? If you just hold FP16 constant from A100 to B100. The B100 is a little over one petaflop, and the A100 is like 300 teraflops. Yeah, 312. Holding numerics constant, you have  a 3x improvement from A100 to B100. Some of that is the process improvement, some of  that is just the accelerator design improving, which we could replicate again in the future. It seems there's actually a very small effect from the process improving from 7nm to 4 nm. I don't know the numbers offhand but let's say there's 150k wafers per month of 3 nm  and eventually similar amounts for 2 nm.

你看 A100 这种 7 纳米芯片，从 A100 到 B100 或 B200 显然有很大进步。但这些进步里有多少只是数值格式（numerics）带来的？如果把 FP16 固定不变，从 A100 到 B100——B100 是一 petaflop 出头，A100 大概是 300 teraflops。对，312。数值格式不变的话，从 A100 到 B100 有 3 倍提升。其中一部分来自制程进步，一部分就只是加速器设计本身的改进，而后者我们未来还可以再复制一遍。看起来从 7 纳米到 4 纳米，制程进步的贡献其实非常小。具体数字我记不太清，但假设 3 纳米每月有 15 万片晶圆产能，2 纳米最终也会有差不多的量。

[57:31]

But then there's a similar amount for 7 nm. If you have all those old wafers and there's maybe a 50% haircut because the bits per  wafer area are 50% less or something, it doesn't seem that bad to just bring on 7  nm wafers if that gives you another fifty or hundred gigawatts. Tell me why that's naive. We potentially do go crazy enough that this happens because we just need incremental  compute, and the compute is worth the higher cost and power of these chips. But it's also unlikely to a large extent because some of these are not fair comparisons. For example, from A100, which is 312 teraflops, to Blackwell, which is 1,000 or 2,000 FP16,  and then Rubin is 5,000 or so FP16… It's not a fair comparison because these chips  have vastly different design targets.

但 7 纳米也有差不多的量。如果把那些旧制程晶圆都用上——也许要打个对折，因为单位晶圆面积的算力少了 50% 之类的——但如果这能再给你五十或一百吉瓦，直接把 7 纳米晶圆拉进来似乎也没那么糟。告诉我这个想法为什么太天真。——我们有可能真的疯狂到让这件事发生，因为我们就是需要增量算力，而这些算力值得为这些芯片付出更高的成本和功耗。但这在很大程度上也不太可能，因为其中一些对比并不公平。比如从 312 teraflops 的 A100，到 FP16 下 1000 或 2000 的 Blackwell，再到 FP16 下 5000 左右的 Rubin……这不是公平的对比，因为这些芯片的设计目标截然不同。

[58:38] Dylan

With A100, Nvidia optimized  for FP16 and BF16 numerics. When you look at Hopper, they didn't care  as much about that; they cared about FP8. When you look at Rubin, they don’t  care about FP16 and BF16 so much, they care mostly about FP4 and FP6. Numerics are what they've designed their chip for. Let's say we make a new chip design on 7 nm,  optimized for the numerics of the modern day. The performance difference is  still going to be much larger than the FLOPS difference you mentioned. Often it's easy to boil things down to FLOPS per watt or FLOPS per dollar,  but that's not a fair comparison.

A100 那一代，Nvidia 优化的是 FP16 和 BF16 数值格式。到了 Hopper，他们就没那么在意这个了，他们在意的是 FP8。到了 Rubin，他们对 FP16 和 BF16 更不上心，主要在意 FP4 和 FP6。芯片就是围绕这些数值格式来设计的。假设我们在 7 纳米上做一款新芯片，针对当下的数值格式做优化，性能差距仍然会远大于你刚才说的那个 FLOPS 差距。大家经常习惯把事情简化成每瓦 FLOPS 或每美元 FLOPS，但那不是公平的比较。

[59:32] Dylan

Let's look at Kimi K2.5 and DeepSeek. When you look at those two models and their performance on Hopper versus  Blackwell on very optimized software, you get vastly different performance. Most of this is not attributed to FLOPS or numerics, because those  models are actually eight-bit. So it's not like Blackwells and Hopper are both  optimized for eight-bit, and Blackwell is not really taking advantage of its four-bit there. The performance gulf is actually much larger. Sure it's one thing to shrink process  technology and make the transistor smaller so each chip has X number of FLOPS,  but you forget the big gating factor.

看看 Kimi K2.5 和 DeepSeek。这两个模型在 Hopper 和 Blackwell 上、都用高度优化的软件来跑，性能差距非常悬殊。这里面大部分不能归因于 FLOPS 或数值格式，因为这些模型实际上是 8 比特的。所以并不是说 Blackwell 和 Hopper 都为 8 比特做了优化、Blackwell 只是没发挥出它 4 比特的优势——实际的性能鸿沟要大得多。没错，缩小制程、把晶体管做小，让每颗芯片有 X 数量的 FLOPS，这是一回事，但你忘了真正卡脖子的大因素。

[1:00:18] Dylan

These models don't run on a single chip. They run on hundreds of chips at a time. If you look at DeepSeek's production  deployment, which is well over a year old now, they were running on 160 GPUs. That's what they serve production traffic on. They split the model across 160 GPUs. Every time you cross the barrier from one chip to another, there is an efficiency loss. You have to transmit over high-speed electrical SerDes, which brings  a latency cost and a power cost. There are all these dynamics that hurt. As you shrink and shrink the process node, you've increased the amount of compute in a single chip. Now in-chip movement of data is at least tens of terabytes a second, if not  hundreds of terabytes a second.

这些模型不是跑在单颗芯片上的，而是同时跑在几百颗芯片上。看 DeepSeek 的生产部署——那都是一年多前的事了——他们当时跑在 160 块 GPU 上，生产流量就是用这个规模承接的，模型被切分到 160 块 GPU 上。每当你从一颗芯片跨到另一颗芯片，就有一次效率损失。你必须通过高速电气 SerDes 传输，这带来延迟成本和功耗成本。这里面有一堆拖累性能的因素。随着制程节点不断缩小，你把越来越多的算力塞进单颗芯片，如今片内数据搬运速度至少是每秒几十 TB，甚至几百 TB。

[1:01:04] Dylan

Whereas between chips, you're on  the order of a terabyte a second. Then you have this movement of data between chips  that are super close to each other physically. You can only put so many chips  close to each other physically, so you have to put chips in different racks. The movement of data between racks is on the order of hundreds of gigabits a second, 400 gig or 800  gig a second, so roughly 100 gigabytes a second. So you have this huge ladder: on-chip  communication is super fast, within the rack is an order of magnitude slower, and outside  the rack is an order of magnitude lower than that.

而芯片与芯片之间，只有每秒 1 TB 的量级。而且这还是物理上贴得非常近的芯片之间的数据搬运。物理上能挨在一起的芯片就那么多，所以你不得不把芯片放进不同的机架。机架之间的数据搬运是每秒几百 gigabit 的量级——每秒 400 gig 或 800 gig，也就是大约每秒 100 GB。于是你有了一个巨大的阶梯：片内通信超快，机架内慢一个数量级，机架外再慢一个数量级。

[1:01:39] Dylan

As you break the bounds of chips,  you end up with a performance loss. The reason I explain this is because  when you look at Hopper versus Blackwell, even if both are using a rack's worth of  chips, Hopper is significantly slower. The amount of performance you have leveraged to  the task within each domain—tens of terabytes a second of communication between these processing  elements versus terabytes a second between these processing elements—is much, much higher and  therefore the performance is much higher. When you look at inference at 100 tokens  a second for DeepSeek and Kimi K2.5, the performance difference between Hopper  and Blackwell is on the order of 20x.

一旦突破芯片的边界，你就会有性能损失。我解释这些是因为，你看 Hopper 对比 Blackwell，即使两边都用上一整个机架的芯片，Hopper 也明显更慢。在每个域内你能调动来干活的性能——这边处理单元之间每秒几十 TB 的通信，对比那边处理单元之间每秒 1 TB 的通信——高得多，因此整体性能也高得多。当你看 DeepSeek 和 Kimi K2.5 以每秒 100 token 做推理时，Hopper 和 Blackwell 的性能差距是 20 倍的量级。

[1:02:21] Dylan

It's not 2x or 3x like the FLOPS  performance difference indicates, even though those are on the same process node. There are just differences in networking technologies and what they've worked on. You can translate some of these back, but when you look at what they're doing  on 3 nm with Rubin, some of those things are simply not possible to do all the way back  on A100, even if you make a new chip for 7 nm. There are certain architectural improvements  you can port and certain ones you cannot. The performance difference is not just  going to be the difference in FLOPS. It's in some senses cumulative between  the difference in FLOPS per chip, networking speed between chips, how many FLOPS  are on a chip versus a system, and memory bandwidth on a single chip versus an entire  system. All of these things compound.

不是 FLOPS 性能差距所显示的 2 倍或 3 倍——尽管这两代还在同一个制程节点上。差别就在于网络技术，以及他们各自下功夫的方向。有些东西你可以移植回去，但你看他们在 3 纳米的 Rubin 上做的一些事情，有些就是不可能一路移植回 A100——哪怕你在 7 纳米上做一款全新芯片。有些架构改进可以移植，有些不行。性能差距不会只是 FLOPS 的差距，某种意义上它是多项差距的叠加：单芯片 FLOPS 的差距、芯片间网络速度、单芯片对比整个系统的 FLOPS 数、单芯片对比整个系统的内存带宽。所有这些因素会复利式叠加。

[1:03:03]

Can I ask you a very naive question? The B200 now has two dies on a single chip, so you can get that bandwidth without  having to go through NVLink or InfiniBand. Next year, Rubin Ultra will  have four dies on one chip. What is preventing us from just doing  that with an older… How many dies could you have on a single chip and still  get these tens of terabytes a second? Even within Blackwell, there are differences in performance when you're communicating  on the chip versus across the chips. Those bounds are obviously much smaller than  when you're going out of the entire chip. When you scale the number of chips  up, there is some performance loss.

我能问个很外行的问题吗？B200 现在把两个 die 放在一颗芯片上，这样不用走 NVLink 或 InfiniBand 就能拿到那个带宽。明年 Rubin Ultra 会在一颗芯片上放四个 die。那有什么阻止我们在旧制程上也这么干……一颗芯片上最多能放多少个 die，还能保持每秒几十 TB 的速度？——即便在 Blackwell 内部，片上通信和跨片通信的性能也有差别。当然，这些差距比走出整个芯片封装时小得多。芯片数量往上扩的时候，总会有一些性能损失。

[1:03:50]

It's not perfect, but it is way  better than different entire packages. How large can advanced packaging scale? The way Nvidia is doing it is CoWoS. Google, Broadcom, MediaTek, and  Amazon's Trainium are all doing CoWoS. But actually you can go look back at what Tesla  did with Dojo, which they cancelled and restarted. Dojo was a chip that was  the size of an entire wafer. They had 25 chips on it. There were some  tradeoffs. They couldn't put HBM on it. But the positive side was  that they had 25 chips on it. To date, it is still probably the best chip  for running convolutional neural networks.

它不完美，但比完全独立的封装之间好太多了。——先进封装能扩到多大？——Nvidia 的做法是 CoWoS。Google、Broadcom、MediaTek，还有 Amazon 的 Trainium，用的都是 CoWoS。但你其实可以回头看看 Tesla 的 Dojo——就是他们砍掉又重启的那个项目。Dojo 是一颗整片晶圆大小的芯片，上面放了 25 颗芯片。有一些取舍，比如没法在上面放 HBM。但好处是它上面有 25 颗芯片。直到今天，它可能仍然是跑卷积神经网络最好的芯片。

[1:04:35] Dylan

It's just not great at transformers because the  shape of the chip, the memory, the arithmetic, and all these various specifications are just  not well-suited for transformers. They're well-suited for CNNs. Dojo chips were optimized  around that, and they made a bigger package. But as you make packages bigger and bigger,  you have other constraints: networking speed, memory bandwidth, and cooling capabilities. All of these things start to rear their heads. It's not simple. But yes, you will see a trend  line of more chips on the package, and yes, you're going to be able to do that on 7 nm. In fact, that's what Huawei did with their Ascend 910C or D. They initially put one, and then they did two.

它只是不太适合 transformer，因为这颗芯片的形态、内存、算术单元等各项规格就是不适配 transformer，它们适配的是 CNN。Dojo 芯片是围绕 CNN 优化的，他们做出了更大的封装。但当你把封装越做越大，其他约束就来了：网络速度、内存带宽、散热能力，这些问题都会一个个冒头。这并不简单。但没错，你会看到封装里芯片越放越多的趋势线，而且没错，这在 7 纳米上也能做。事实上 Huawei 的 Ascend 910C 或 D 就是这么干的——一开始放一颗，后来放了两颗。

[1:05:20]

They're focusing on scaling the  packaging up because that is an area where they can advance faster than  process technology where they can't shrink. But at the end of the day, that’s something  you can do on the leading-edge chips too. Anything you do on 7 nm, you can also  probably do on 3 nm in terms of packaging. If we end up in this world in 2030 where the  West has the most advanced process technology but has not ramped it up as much, whereas  China… I don't know if you think by 2030 they would have EUV and 2 nm or whatever. But they are semiconductor-pilled and they are producing in mass quantity. Basically, I'm wondering what the year is where there's a crossover, where our  advantage in process technology has faded enough, and their advantage in scale has increased enough. And also, if their advantage in having one country with the entire supply chain indigenized—rather  than having random suppliers in Germany and the Netherlands—would mean that China would  be ahead in its ability to produce mass flops.

他们专注于把封装规模做大，因为这是他们能比制程技术进步更快的领域——制程上他们没法继续缩小。但说到底，这些事在最先进的芯片上同样能做。你能在 7 纳米上做的封装，大概率也能在 3 纳米上做。——如果 2030 年我们走到这样一个世界：西方拥有最先进的制程技术但没把产能拉起来，而中国……我不知道你是否认为到 2030 年他们会有 EUV 和 2 纳米之类的。但他们是全力押注半导体的，而且在大规模量产。我其实想问：交叉点会出现在哪一年——我们在制程技术上的优势消退到足够小、他们在规模上的优势增长到足够大？还有，他们把整条供应链在一国之内本土化的优势——而不是供应商零散地分布在德国和荷兰——是否意味着中国在量产 FLOPS 的能力上会领先？

[1:06:22] Dylan

To date, China still does not have an entirely  indigenized semiconductor supply chain. But would they in 2030? By 2030, it's possible that they do. But to date, all of China's 7 nm and  14 nm capacity uses ASML DUV tools. The amount that they can  import from ASML is large. But the vast majority of ASML's revenue,  especially on EUV all of it, is outside of China. The scale advantage is still in  the favor of the West plus Taiwan, Japan, and Korea, et cetera. But they're trying to make their own DUV and EUV tools, right? They're trying to do all these things. The question is how fast can they advance  and scale up production as well as quality.

到今天为止，中国还没有一条完全本土化的半导体供应链。——那 2030 年会有吗？——到 2030 年，有可能。但迄今为止，中国所有的 7 纳米和 14 纳米产能用的都是 ASML 的 DUV 设备。他们能从 ASML 进口的量很大。但 ASML 的绝大部分收入——尤其是 EUV 的全部收入——都来自中国以外。规模优势仍然在西方加上台湾、日本、韩国等这一边。——但他们在努力自研 DUV 和 EUV 设备，对吧？他们在尝试做所有这些事。——问题是他们能以多快的速度进步，并把产量和质量一起做上去。

[1:07:08] Dylan

To date, we haven't seen that. Now I'm quite bullish that they're going to be able to do these things  over the next five to ten years. They will really scale up production  and kick it into high gear. They have more engineers working on it and  more desire to throw capital at the problem. So by 2030, will they have fully indigenized DUV? I think for sure. DUV, yes. And fully indigenized EUV by 2030? I think they'll have working tools. I don't think that they'll be  able to manufacture a bunch yet. There's having it work, and  then there's production hell. ASML had EUV working in the  early 2010s at some capacity.

到目前为止我们还没看到。不过我相当看好他们在未来五到十年里能做成这些事。他们会真正把产能拉起来、火力全开。他们有更多工程师在攻关，也更愿意往这个问题上砸钱。那么到 2030 年，他们能完全本土化 DUV 吗？我认为肯定能。DUV，能。那到 2030 年完全本土化 EUV 呢？我认为他们会有能用的设备，但我不认为他们那时已经能大批量制造。「能用」是一回事，「量产地狱」是另一回事。ASML 早在 2010 年代初就有了某种程度上能用的 EUV。

[1:07:49]

The tools were not accurate enough. They were not scaled for high-volume manufacturing or reliable enough. They had to ramp production, and that all took time. Production hell takes  time. That's why it took another five to seven years to get EUV into mass production at  a fab rather than just working in the lab. How many DUV tools do you think  they'll be able to manufacture in 2030? ASML? No, China. That's a great question. It's a bit of a  challenge to look into this supply chain especially. We try really hard. In some instances,  they're buying stuff from Japanese vendors. If they want a fully indigenized supply chain,  they need to not buy these lenses, projection optics, or stages from Japanese vendors. They need to build it internally.

那时候设备精度不够，既没有为大规模量产做好准备，可靠性也不行。他们必须爬产能，而这些都需要时间。量产地狱是要熬时间的。这就是为什么 EUV 从实验室里「能用」到真正在 fab 里大规模量产，又花了五到七年。——你觉得他们 2030 年能造出多少台 DUV 设备？——ASML 吗？——不，是中国。——这是个好问题。这条供应链尤其难看透，我们已经非常努力在查了。在某些环节，他们还在从日本供应商那里买东西。如果他们想要一条完全本土化的供应链，就不能再从日本供应商那里买这些镜头、投影光学系统或工件台，得自己在内部造出来。

[1:08:40] Dylan

It's really tough to say where  they'll be able to get to. I honestly think it's a shot in the dark. But it's probably not unlikely that they'll be able to do on the order of 100 DUV tools  a year, whereas ASML is currently doing hundreds of DUV tools a year. No company has a process node where they make a million wafers a month. Elon says he wants to do it and China is obviously going to do it. TSMC is trying to do that. The memory makers may get to a million wafers  a month as well, but not in a single fab. It's mind-boggling to think of  that scale, and challenging to see the supply chain galvanized for that. I don't want to doubt China's capability to scale.

很难说他们最终能走到哪一步，老实讲这就是瞎猜。但他们做到一年上百台 DUV 光刻机的量级，恐怕并非不可能——而 ASML 现在一年出货的是几百台 DUV。目前没有任何一家公司在单一制程节点上做到每月一百万片晶圆。Elon 说他想干，中国显然也会去干，TSMC 也在朝这个方向努力。存储厂商也许也能到每月一百万片，但不会是在单一一座 fab 里。这个规模想想都令人咋舌，也很难想象整条供应链为此被动员起来。不过，我不想怀疑中国的规模化能力。

[1:09:29] Dwarkesh

I guess this is an interesting question. I think at some point SemiAnalysis will do the deep dive on this. By when would indigenized Chinese production be bigger than the rest of the West combined. And put in the input of your model of when they'll have DUV machines and EUV machines at scale? Because there's this question around if you have long timelines on AI—by long meaning  2035, which is not that long in the grand scheme of things—should you expect a world  where China is dominating in semiconductors? It doesn't get asked enough  because if you're in San Francisco, we're thinking on timescales of weeks. If you're outside of San Francisco, you're not thinking about AGI at all. What if we  have AGI? What if you have this transformational thing that is commanding tens or hundreds  of trillions of dollars of economic growth and token output, but it happens in 2035? What does that imply for the West versus China?

我觉得这是个很有意思的问题。我想 SemiAnalysis 迟早会就此做一次深度研究：到哪一年，中国本土化的产能会超过整个西方世界的总和？再把他们何时能规模化拥有 DUV 和 EUV 设备作为你们模型的输入。因为有这么一个问题：如果你对 AI 持长时间线——这里的『长』指 2035 年，放在大历史尺度里其实并不算长——你是否应该预期一个中国主导半导体的世界？这个问题被问得太少了，因为你要是在旧金山，大家是按『周』来思考的；你要是在旧金山之外，大家压根不去想 AGI。可如果我们真有了 AGI 呢？如果这个能撬动几十万亿、上百万亿美元经济增长和 token 产出的变革性东西，是在 2035 年才发生呢？那对西方对中国意味着什么？

[1:10:33] Dylan

SemiAnalysis has got to write  the definitive model on this. It's really challenging when you  move timescales out that far. What we tend to focus on is tracking every  data center, every fab, and all the tools. We track where they're going, but the time  lags for these things are relatively short. We can only make reasonably accurate estimates  for data center capacity based on land purchasing, permits, and turbine purchasing. We know where all these things are going, that's the data we sell. As you go out to 2035, things are just so radically different. Your error bars get so large it's hard to make an estimate. But at the end of the day, if takeoff or timelines are slow enough, I don't see why  China wouldn't be able to catch up drastically.

SemiAnalysis 确实应该写出这方面的权威模型。但时间尺度拉到那么远之后，事情就非常难做。我们平时专注的是追踪每一座数据中心、每一座 fab、所有的设备——我们追踪它们的去向，但这些事情的时间滞后相对较短。我们只能基于购地、许可证、燃气轮机采购这些信息，对数据中心产能做出比较准确的估计。我们知道所有这些东西流向哪里，这就是我们卖的数据。可一旦推到 2035 年，情况就截然不同了，误差棒大到很难给出估计。但说到底，如果起飞速度或时间线足够慢，我看不出中国为什么不能大幅追上来。

[1:11:28] Dylan

In some sense, we've got this valley where, three  to six months ago, or maybe even now, Chinese models are as competitive as they've ever been. I think Opus 4.6 and GPT 5.4 have really pulled away and made the gap a little bit bigger, but  I'm sure some new Chinese models will come out. As we move from selling tokens where they  provide the entire reasoning chain, to selling automated white-collar work—an automated  software engineer, you send them the request, they give you the result back, and there's a bunch  of thinking on the back end that they don't show you—the ability to distill out of American  models into Chinese models will be harder.

从某种意义上说，我们正处在一个低谷期：三到六个月前，甚至可能就是现在，中国模型的竞争力达到了有史以来最强。我认为 Opus 4.6 和 GPT 5.4 已经真正拉开了身位，把差距扩大了一点，但肯定还会有新的中国模型出来。而当我们从『卖 token、把完整推理链都给你』转向『卖自动化的白领工作』——比如一个自动化软件工程师，你把需求发过去，它把结果还给你，后台有一大堆思考过程是不展示给你的——从美国模型往中国模型里蒸馏就会变得更难。

[1:12:05] Dylan

Second, look at the scale of  the compute the labs have. OpenAI exited the year with  roughly two gigawatts last year. Anthropic will get to  two-plus gigawatts this year. By the end of next year, they'll  both be at ten gigawatts of capacity. China is not scaling their AI  lab compute nearly as fast. At some point, when you can't distill the  learnings from these labs into the Chinese models, plus with this compute race that OpenAI,  Anthropic, Google, and Meta are all racing on, they end up getting to a point where the model  performance should start to diverge more. Then look at all this CapEx  being spent on data centers.

第二，看看这些实验室手里的算力规模。OpenAI 去年年底大约有 2 吉瓦，Anthropic 今年会到 2 吉瓦以上。到明年年底，两家都会达到 10 吉瓦的容量。中国 AI 实验室的算力扩张远没有这么快。到了某个时点，当你没法再把这些实验室的成果蒸馏进中国模型，再叠加 OpenAI、Anthropic、Google、Meta 都在比拼的这场算力竞赛，模型性能应该会开始进一步分化。然后再看看砸在数据中心上的这些资本开支（CapEx）。

[1:12:49] Dylan

Amazon is spending $200  billion, Google $180 billion. All these companies are spending  hundreds of billions of dollars on CapEx. There's nearly a trillion dollars  of CapEx being invested in data centers in America this year, roughly. What's the return on invested capital here? You and I would think the return on invested  capital for data center CapEx is very high. If we look at Anthropic's revenues,  in January they added $4 billion. In February, which was a shorter  month, they added $6 billion. We'll see what they can do in March and April, given that compute constraints are  what's bottlenecking their growth.

Amazon 在花 2000 亿美元，Google 1800 亿。所有这些公司都在往 CapEx 上砸数千亿美元。今年美国投进数据中心的 CapEx 大致接近一万亿美元。那这里的投入资本回报率是多少？你我都会觉得数据中心 CapEx 的回报率非常高。看 Anthropic 的收入：1 月份他们增加了 40 亿美元；2 月份虽然天数更短，却增加了 60 亿。3 月、4 月能做到什么程度我们拭目以待，毕竟算力约束正是卡住他们增长的瓶颈。

[1:13:27] Dylan

The reliability of Claude is quite low  because they're so compute constrained. But if this continues, then the ROIC  on these data centers is super high. At some point, the US economy starts growing  faster and faster over this year and next year because of all this CapEx, all the revenue these  models are generating, and the downstream supply chain. China doesn't have that yet. They  have not built the scale of infrastructure to invest in models, get to the capabilities,  and then deploy these models at such scale. When you look at Anthropic,  they're at $20 billion ARR. The margins are sub-50 percent, at least  as last reported by The Information.

Claude 的可靠性相当低，就是因为他们算力太紧张。但如果这个势头延续下去，这些数据中心的 ROIC 就会非常高。到某个时点，美国经济在今明两年会越长越快——因为这些 CapEx、这些模型创造的收入，以及下游整条供应链。中国目前还没有这些。他们还没有建起这种规模的基础设施去投模型、把能力做上去、再以这种规模部署这些模型。看 Anthropic，他们的 ARR 已经到了 200 亿美元，而毛利率不到 50%——至少按 The Information 最近的报道是这样。

[1:14:09] Dylan

So that's $13 or $14 billion of compute that it's  running on rental cost-wise, which is actually $50 billion worth of CapEx that someone laid out  for Anthropic to generate their current revenue. China has just not done this. If and when Anthropic 10Xs revenue again—and I think our answer would be when, not if—China  doesn't have the compute to deploy at that scale. So there is some sense that  we're in a fast takeoff. It's not like we're talking  about a Dyson sphere by X date, it's more like the revenue is compounding at  such a rate that it does affect economic growth. The resources these labs are  gathering are growing so fast.

也就是说，它跑在按租金成本算约 130 到 140 亿美元的算力上，而这背后其实是有人为 Anthropic 铺下了价值 500 亿美元的 CapEx，才撑起他们现在的收入。中国就是没有做这件事。等 Anthropic 收入再翻十倍的时候——我想我们的答案是『何时』而不是『会不会』——中国没有能以那种规模部署的算力。所以从某种意义上说，我们已经身处一场快速起飞之中。倒不是说到某个日期就造出戴森球那种，而是收入正以这样的速度复利增长，实打实地影响到经济增长。这些实验室聚拢资源的速度快得惊人。

[1:14:51] Dylan

China hasn't done that yet, so in that case,  the US and the West are actually diverging. The flip side is that these infrastructure  investments have middling returns. Maybe they're not as good as hoped. Maybe Google is wrong for wanting to take free cash flow to zero and  spend $300 billion on CapEx next year. Maybe they’re just wrong and people on  Wall Street who are bearish and people who don't understand AI are correct. In that case, the US is building all this capacity but doesn't get great returns. Meanwhile, China is able to build a fully vertical, indigenized supply chain, instead of  the US/Japan/Korea/Taiwan/SE Asia/Europe countries together building this less vertical supply chain. In a sense, at some point China is able to scale past us if AI takes longer to get to certain  capability levels than the vast majority of your guests on this podcast believe. It's fast timelines, the US wins;

中国还没做到这一点，所以在这种情形下，美国和西方实际上是在拉开差距。反过来的另一面是：这些基础设施投资的回报平平，也许没有预期那么好。也许 Google 想把自由现金流打到零、明年在 CapEx 上花 3000 亿美元是错的。也许他们就是错了，华尔街那些看空的人、那些不懂 AI 的人反而是对的。那种情况下，美国建了这么多产能却拿不到好回报；与此同时，中国建成了一条完全垂直、全面本土化的供应链，而不是像美国/日本/韩国/台湾/东南亚/欧洲这些国家和地区一起拼出的这条没那么垂直的供应链。从某种意义上说，如果 AI 达到特定能力水平所需的时间比你这档播客上绝大多数嘉宾相信的更久，中国到某个时点就能反超我们。时间线快，美国赢；

[1:15:50]

long timelines, China wins. Yeah but I don't know what fast timelines means. I don't think you have to believe in AGI  to have the timelines where the US wins. Let's go back to memory. I think people on  Wall Street and people in the industry are understanding how big this is, but maybe generally  people don't understand what a big deal it is. So we've got this memory crunch,  as you were talking about. And earlier I was asking about,  oh, could we solve for the EUV tool shortage by going back to seven nanometers? So let me ask a similar question about memory. HBM is made of DRAM, but has three  to four times fewer bits per wafer area than the DRAM it's made out of. Is it possible that accelerators in the future could just use commodity  DRAM and not HBM, so we can get much more capacity out of the DRAM we have? The reason I think this might be possible is, if we're going to have agents that are  just going off and doing work, and it's not a synchronous chatbot application, then you  don't necessarily need extremely fast latency.

时间线慢，中国赢。是啊，但我不知道『时间线快』到底指什么。我觉得你不必相信 AGI，也能得出美国赢的时间线。我们回到内存这个话题。我觉得华尔街和业内的人正在理解这件事有多大，但普通人可能还不明白这是多大的事。就像你刚才说的，我们碰上了内存紧缺。之前我问过：能不能通过退回七纳米来化解 EUV 设备短缺？那让我就内存问一个类似的问题。HBM 是用 DRAM 做出来的，但单位晶圆面积的比特数比它所用的 DRAM 少三到四倍。未来的加速器有没有可能干脆用普通的商品级 DRAM 而不用 HBM，这样我们手里的 DRAM 就能榨出多得多的容量？我觉得这可能行得通的原因是：如果将来是 agent 自己出去干活，而不是同步的聊天机器人应用，那你并不一定需要极快的响应延迟。

[1:16:57]

Maybe you can have lower bandwidth,  because the reason you stack DRAM into HBM is for higher bandwidth. Is it possible to go to HBM accelerators and basically have the opposite  of Claude Code Fast, like have Claude Slow? At the end of the day, the incremental  purchaser who's willing to pay the highest price for tokens also ends up being  the one that's less price-sensitive. Compute should be allocated, in a capitalistic  society, towards the goods that have the highest value, and the private market  determines this by willingness to pay. To some extent, Anthropic could  actually release a slow mode.

也许带宽低一点也可以，毕竟把 DRAM 堆叠成 HBM 就是为了更高带宽。有没有可能做不用 HBM 的加速器，搞一个 Claude Code Fast 的反面——比如来个 Claude Slow？说到底，愿意为 token 出最高价的那个边际买家，恰恰也是对价格最不敏感的。在资本主义社会里，算力应该被分配到价值最高的用途上，而私人市场是靠支付意愿来决定这一点的。某种程度上，Anthropic 确实可以发布一个慢速模式。

[1:17:39] Dylan

They could release Claude Slow Mode and increase  tokens per dollar by a significant amount. They could probably reduce the price of Opus 4.6  by 4-5x and reduce the speed by maybe just 2x. The curve on inference throughput versus  speed is already there just on HBM. And yet they don't, because no one  actually wants to use a slow model. Furthermore, on these agentic tasks, it's great  that the model can run at a time horizon of hours. But if the model was running slower,  those hours would become a day. Vice versa, if the model is running  faster, those hours become an hour. No one really wants to move to a day-long wait  period, because the highest-value tasks also have some time sensitivity to them. I struggle to see… Yes, you could use regular DRAM. There are a couple of challenges with this.

他们可以发布一个 Claude 慢速模式，把每美元能买到的 token 数大幅提高。他们大概可以把 Opus 4.6 的价格降到四五分之一，而速度也许只降一半。仅靠 HBM，推理吞吐量与速度之间的那条权衡曲线就已经摆在那儿了。可他们没这么做，因为根本没人想用慢模型。再说，在这些 agent 任务上，模型能以小时级的时间跨度连续干活固然很棒，但如果模型跑得更慢，那几个小时就会变成一整天；反过来，如果模型跑得更快，那几个小时就会变成一个小时。没人真想接受一天的等待，因为价值最高的任务往往也有时效性。我很难想象……是的，你可以用普通 DRAM，但这里有几个难点。

[1:18:44] Dylan

One of the core constraints of chips is  that a chip is a certain size, and all of the I/O escapes on the edges. Often, the left and right of the chip are HBM—so the I/O from the chip  to the HBM is on the sides—and then the top and bottom are I/O to other chips. If you were to change from HBM to DDR, all of a sudden this I/O on the edge  would have significantly less bandwidth, but significantly more capacity per chip. But the metric you actually care about is bandwidth per wafer, not bits per wafer. Because the thing that is constraining the FLOPS is just getting in and out the next matrix,  and for that you just need more bandwidth.

芯片的一个核心约束是：一颗芯片就那么大，所有 I/O 都要从边缘引出去。通常芯片的左右两侧是 HBM——也就是芯片到 HBM 的 I/O 走两侧——上下两侧则是通往其他芯片的 I/O。如果你把 HBM 换成 DDR，边缘上这些 I/O 的带宽会一下子小得多，虽然每颗芯片的容量会大得多。但你真正在意的指标是每片晶圆的带宽，而不是每片晶圆的比特数。因为卡住 FLOPS 的，就是把下一个矩阵搬进搬出，而这需要的就是更多带宽。

[1:19:39] Dylan

Yeah, getting out the weights and  getting in and out the KV cache. In many cases, these GPUs are not  running at full memory capacity. It's obviously a system design thing:  model, hardware, and software co-design. You have to figure out how much KV cache  you need, how much you keep on the chip, how much you offload to other chips and  call when you need it for tool calling, and how many chips you parallelize this on. Obviously, the search space for this is very broad, which is why we have InferenceX,  an open-source model that searches all the optimal points on inference for a  variety of different chips and models.

对，把权重读出来，把 KV cache 搬进搬出。很多情况下，这些 GPU 并没有把内存容量用满。这显然是个系统设计问题：模型、硬件、软件要协同设计。你得算清楚需要多少 KV cache、多少留在芯片上、多少卸载到其他芯片上等工具调用要用时再取回来，以及把这一切并行到多少颗芯片上。显然这个搜索空间非常大，所以我们做了 InferenceX 这个开源项目，可以在各种不同芯片和模型的组合上搜索推理的所有最优点。

[1:20:16]

The point is, you're not always  necessarily constrained by memory capacity. You can be constrained by FLOPS, network  bandwidth, memory bandwidth, or memory capacity. If you really simplify it down,  there are four constraints, and each of these can break out into more. If you switch to DDR, yes, you produce four times the bits per DRAM wafer, but all of  a sudden the constraints shift a lot and your system design shifts. You go slower.  Is the market smaller? Maybe. But also, all these FLOPS are wasted because they're  just sitting there waiting for memory. You don't need all that capacity because you can't  really increase batch size because then the KV cache would take even longer to read. Makes sense. What is the bandwidth difference between HBM and normal DRAM? An HBM4 stack—let's talk about the stuff that's in Rubin, because that's what we've been  indexing on—is 2048 bits across, connected in an area that's 13 millimeters wide. It transfers memory at around 10 giga-transfers a second. So a stack of HBM4 is 2048 bits on an area that's roughly 11 to 13 millimeters wide. That's the shoreline you're taking on the chip.

重点是，你并不总是被内存容量卡住。你可能被 FLOPS 卡住、被网络带宽卡住、被内存带宽或内存容量卡住。极度简化地说就是这四个约束，每一个还能再往下细分。如果你换成 DDR，没错，每片 DRAM 晶圆产出的比特数是原来的四倍，但约束会突然大幅漂移，你的系统设计也随之改变——你会变慢。市场会更小吗？也许。而且这些 FLOPS 全浪费了，因为它们只能干等着内存。你也用不上那么大的容量，因为你没法真正加大 batch size——那样 KV cache 的读取会更加耗时。有道理。那 HBM 和普通 DRAM 的带宽差距是多少？一个 HBM4 堆栈——我们就聊 Rubin 里用的这一代，因为我们一直拿它做基准——位宽是 2048 位，接口落在大约 13 毫米宽的区域里，传输速率大约是每秒 10 giga-transfers。所以一个 HBM4 堆栈就是 2048 位、占大约 11 到 13 毫米宽的区域。这就是它在芯片上占用的『岸线』（shoreline）。

[1:21:33] Dylan

In that shoreline, you have 2048 bits  transferring at 10 giga-transfers per second. You multiply those together and divide by eight, bits to a byte, and you're at roughly  2.5 terabytes a second per HBM stack. When you look at DDR, in that same  area, it's maybe 64 or 128 bits wide. That DDR5 is transferring at anywhere from  6.4 to maybe 8,000 giga-transfers a second. So your bandwidth is significantly lower. It's 64 times 8,000 divided by eight, which puts you at 64 gigabytes a second. Even if you take a generous interpretation of 128 times 8 giga-transfers, you're at 128  gigabytes a second for the same shoreline, versus 2.5 terabytes a second. There's an order of magnitude difference in bandwidth per edge area. If your chip is a square, or 26 by 33 millimeters—which is the maximum size for an  individual die—you only have so much edge area.

在这段岸线里，你有 2048 位以每秒 10 giga-transfers 的速率传输。两者相乘再除以八（八个比特一个字节），每个 HBM 堆栈大约是每秒 2.5 TB。再看 DDR，同样的面积里位宽大概只有 64 位或 128 位，而 DDR5 的传输速率大约在每秒 6.4 到 8 giga-transfers 之间。所以带宽低得多：64 乘 8000（MT/s）再除以八，也就是每秒 64 GB。就算按 128 位乘每秒 8 giga-transfers 的慷慨算法，同样的岸线也只有每秒 128 GB，对比每秒 2.5 TB——单位边缘面积的带宽差了一个数量级。如果你的芯片是方形的，比如 26 乘 33 毫米——这是单颗 die 的最大尺寸——你的边缘面积就只有那么多。

[1:22:32]

On the inside of that chip,  you put all your compute. There are things you can do to try and  change that, like more SRAM or more caching. But at the end of the day, you're  very constrained by bandwidth. Then there's the question of where you can  destroy demand to free up enough for AI. I guess the picture is especially bad because,  as you're saying, if it takes four times more wafer area to get the same byte, for HBM you have  to destroy four times as much consumer demand for laptops and phones to free up one byte for AI. What does this imply for the next year or two? Sorry for the run-on question, in your newsletter  you said 30% of Big Tech's CapEx in 2026 is going towards memory? Yes.

芯片内部则全放你的计算单元。你可以做一些事来改善这一点，比如加更多 SRAM、做更多缓存，但归根结底，你被带宽死死卡住。接下来的问题就是：你能到哪里去『消灭』需求，好腾出足够的产能给 AI。我想局面之所以格外糟糕，就是因为像你说的：HBM 拿到同样一个字节要用四倍的晶圆面积，那你就得消灭四倍的笔记本、手机等消费需求，才能为 AI 腾出一个字节。这对未来一两年意味着什么？抱歉问题有点连珠炮——你们的 newsletter 里说，2026 年大型科技公司 30% 的 CapEx 要花在内存上？是的。

[1:23:16] Dwarkesh

That's insane, right? Of the $600 billion  or whatever, 30% is going just to memory. Yes. Obviously, there's some level  of margin stacking that Nvidia does, so you have to separate that out and apply  their margin to the memory and the logic. But at the end of the day, a third  of their CapEx is going to memory. That's crazy. What should we expect over the  next year or two as this memory crunch hits? The memory crunch will continue to get  harder, and prices will continue to go up. This affects different parts  of the market differently. Are people going to hate AI more and more? Yes, because smartphones and PCs are not going to get incrementally better year on year. In fact, they're going to get incrementally worse.

太疯狂了，对吧？6000 亿美元还是多少的开支里，光内存就占了 30%。是的。当然，NVIDIA 存在一定程度的利润叠加，所以你得把这部分拆出来，把他们的利润率分摊到内存和逻辑芯片上。但归根结底，他们资本开支的三分之一花在了内存上。这太夸张了。随着这波内存紧缺来袭，未来一两年我们该预期什么？内存紧缺会越来越严重，价格会继续上涨。这对市场不同部分的影响各不相同。大家会不会越来越讨厌 AI？会的，因为智能手机和 PC 不会再一年比一年更好了。实际上，它们会一年比一年更差。

[1:24:00] Dwarkesh

If you look at the bill of materials for an  iPhone, what fraction of it is the memory? How much more expensive does an iPhone get  if the memory is two times more expensive? I believe an iPhone has 12 gigabytes of memory. Each gig used to cost roughly $3-4, so that's $50. But now the price of memory has tripled. Let's say it's $12 per gig for DDR. Now you're talking about $150 versus $50. That's a $100 increase in cost for Apple. Apple has some margin, they're  not just going to eat the margin. NAND also has the same market dynamics,  so in reality, it's probably a $150 increase on the iPhone. So now that’s a $100 cost increase and that’s just on the DRAM. The NAND also has the same sort of market.

如果看 iPhone 的物料清单（BOM），内存占多大比例？如果内存贵两倍，一部 iPhone 会贵多少？我记得 iPhone 有 12GB 内存。过去每 GB 大概 3-4 美元，也就是 50 美元左右。但现在内存价格涨了三倍。假设 DDR 每 GB 12 美元，那就是 150 美元对比原来的 50 美元。对 Apple 来说是 100 美元的成本增加。Apple 有利润空间，但他们不会就这么白白吞掉。NAND 也是同样的市场态势，所以实际上 iPhone 的成本大概要涨 150 美元。所以这 100 美元的成本增加还只是 DRAM 这一块，NAND 也是类似的市场情况。

[1:24:37] Dylan

So in fact it’s probably a  $150 increase on the iPhone. Apple either has to pass that  on to the consumer or eat it. I don't see Apple reducing their margin  too much, maybe they eat a little bit. But at the end of the day, that means the end  consumer is paying $250 more for an iPhone. Now that’s just on last  year’s pricing versus today’s. There is some lag before Apple feels the heat  because they tend to have long-term contracts for memory that last three months to a year. But at the end of the day, Apple gets hit pretty hard by this. They won't really adjust until the next iPhone release. But that's the high end of the market, which is only a few hundred million phones a year. Apple sells two or three hundred million phones annually. The bulk of the market is mid-range and low-end.

所以实际上 iPhone 的成本大概要涨 150 美元。Apple 要么转嫁给消费者，要么自己吞下。我不觉得 Apple 会大幅压缩自己的利润率，也许会消化一点。但归根结底，这意味着终端消费者买一部 iPhone 要多付 250 美元。而这还只是拿去年的价格跟今天比。Apple 感受到压力会有一定滞后，因为他们的内存往往签的是三个月到一年的长期合同。但最终 Apple 会被狠狠冲击一把。他们要到下一代 iPhone 发布时才会真正调整。不过那只是市场的高端部分，一年也就几亿部手机。Apple 每年卖两三亿部。市场的大头是中端和低端。

[1:25:25] Dylan

It used to be that 1.4 billion  smartphones were sold a year. Now we're at about 1.1 billion. Our projections are that we might drop to 800 million this year, and  down to 500 or 600 million next year. We look at data points out of China  from some of our analysts in Asia, Singapore, Hong Kong, and Taiwan. They've been tracking this, and they see Xiaomi and Oppo cutting low-end  and mid-range smartphone volumes by half. Yes, it’s only a $150 BOM increase on a $1,000  iPhone where Apple has some larger margin. But for smaller phones, the percentage of the BOM  that goes to memory and storage is much larger.

过去全球一年卖 14 亿部智能手机。现在大约是 11 亿部。我们的预测是今年可能跌到 8 亿部，明年跌到 5 到 6 亿部。我们参考了亚洲——新加坡、香港、台湾——的分析师从中国拿到的数据点。他们一直在跟踪，看到 Xiaomi 和 Oppo 正在把低端和中端智能手机的出货量砍半。没错，在一部 1000 美元、Apple 利润空间较大的 iPhone 上，BOM 只涨 150 美元。但对更便宜的手机来说，内存和存储在 BOM 里占的比例要大得多。

[1:26:08] Dylan

And the margins are lower, so there's  less capacity to even eat the margins. And they have also generally tended not  to do long-term agreements on memory. Why this is a big deal is that if smartphone  volumes halve, that drop will happen in the low and mid-range, not the high end. So it’s not like the bits released are halving. Currently, consumer devices account  for more than half of memory demand. Even if you halve smartphone volumes,  because of the shape of the halving, the low end gets cut by more than half, while  the high end gets cut by less than half, because you and I will still buy the high-end  phones that cost north of a thousand dollars.

而且它们利润更薄，更没有余地去消化这部分成本。它们通常也不签内存的长期协议。这件事之所以重要，是因为如果智能手机出货量腰斩，跌掉的会是低端和中端，而不是高端。所以释放出来的内存容量（bits）并不是减半。目前，消费电子设备占内存需求的一半以上。即便智能手机出货量腰斩，由于腰斩的结构，低端砍掉的超过一半、高端砍掉的不到一半，因为你和我还是会买那些 1000 美元以上的高端手机。

[1:26:46] Dylan

We'll buy them even if they get  a little bit more expensive. And Apple's volumes will not go down as  much as a low-end smartphone provider. The same applies to PCs. What this  does to the market is quite drastic. DRAM gets released and goes to AI chips, who are  willing to do longer-term contracts and pay higher margins, because at the end of the day the margin  they extract from the end user is much larger. This probably leads to people hating AI even more. Today, you already see all the memes on PC subreddits and gaming PC Twitter. It's cat dancing videos saying, "This is why memory prices have doubled and  you can't get a new gaming GPU or desktop."

就算贵一点我们也照买。而且 Apple 的销量下滑不会像低端手机厂商那么厉害。PC 也是一样。这对市场的影响相当剧烈。DRAM 释放出来后流向 AI 芯片，因为 AI 买家愿意签更长期的合同、付更高的溢价——毕竟归根结底，他们从终端用户身上榨取的利润要大得多。这大概率会让人们更讨厌 AI。今天你已经能在 PC 相关的 subreddit 和游戏 PC 圈的 Twitter 上看到各种梗图。配着猫咪跳舞视频说：「这就是内存价格翻倍、你买不到新游戏 GPU 和台式机的原因。」

[1:27:33] Dylan

It's going to be even worse when memory  prices double again, especially DRAM. Another interesting dynamic is that  it's not just DRAM, it's also NAND. NAND is also going up in price. Both of these markets have expanded capacity very slowly over the last few years, NAND almost zero. The percentage of NAND that goes to phones and PCs is larger than the percentage  of DRAM that goes to phones and PCs. As you destroy demand, mostly for  DRAM purposes, you unlock more NAND that gets allocated and can go to other markets. The price increases of DRAM will be larger than those of NAND because you've released  more from the consumer, and in fact, you've produced more memory for AI. Sorry, maybe you just explained it and I missed it. Is it because SSDs are being used in large quantities for data centers? They are, but not in as large quantities as DRAM.

等内存价格再翻一倍，尤其是 DRAM，情况会更糟。另一个有意思的动态是：不只是 DRAM，NAND 也在涨价。这两个市场过去几年的产能扩张都非常慢，NAND 几乎是零。NAND 流向手机和 PC 的比例，比 DRAM 流向手机和 PC 的比例还要高。当你摧毁需求——主要是围绕 DRAM 的需求——你就释放出更多 NAND，可以分配到其他市场去。DRAM 的涨幅会比 NAND 更大，因为你从消费端释放出来的更多，而且实际上你为 AI 生产了更多内存。抱歉，也许你刚解释过而我没跟上。是因为数据中心在大量使用 SSD 吗？确实在用，但量没有 DRAM 那么大。

[1:28:27] Dwarkesh

Okay, so they will also increase because  they'll be using some quantity, but there's not as much of a need as there is for HBM. Makes  sense. One thing I didn't appreciate until I was reading some of your newsletters is that the  same constraints preventing logic scaling over the next few years are quite similar to what's  preventing us from producing more memory wafers. In fact, literally the same exact machine,  this EUV tool, is needed for memory. So I guess the question someone could ask right  now is, why can't we just make more memory? The constraints, as I was mentioning earlier,  are not necessarily EUV tools today or next year.

好，所以 SSD 也会涨价，因为数据中心也会用掉一部分，但需求没有 HBM 那么迫切。有道理。有一件事是我读你们的 newsletter 之前没意识到的：未来几年阻碍逻辑芯片扩产的那些约束，跟阻碍我们生产更多内存晶圆的约束非常相似。事实上，内存需要的就是同一台机器——EUV 光刻机。所以现在有人可能会问：为什么我们不能直接多造点内存？就像我前面提到的，眼下和明年的约束其实并不在 EUV 设备上。

[1:29:11] Dylan

They become that as we get to  the latter part of the decade. Currently, the constraints are more that  they physically just haven't built fabs. Over the last three to four years,  these vendors have not built new fabs because memory prices were really low. Their margins were low, and in fact, they were losing money in 2023 on memory. So they decided they weren't building new fabs. The market slowly recovered over time but  never really got amazing until last year. In 2024, we were banging on the drums  that reasoning means long context, which means a large KV cache, which  means you need a lot of memory demand.

要到这个十年的后半段，EUV 才会变成约束。目前的约束更多在于：他们实际上根本没建晶圆厂（fab）。过去三四年，这些厂商都没有建新 fab，因为内存价格实在太低了。利润率很低，实际上 2023 年他们做内存还在亏钱。所以他们决定不建新厂。市场后来慢慢恢复，但直到去年之前都算不上好。2024 年我们就一直在敲鼓呐喊：推理（reasoning）意味着长上下文，长上下文意味着大 KV cache，也就意味着巨大的内存需求。

[1:29:48] Dylan

We've been talking about that  for a year and a half, two years. People who understand AI went  really long on memory then. So you’ve seen that dynamic, but now  it has finally played out in pricing. It took so long for what was  obvious: long context means the KV cache gets bigger, you need more memory. Half the cost of accelerators is memory. Of course they're going to  start going crazy on it. It took a year for that to  actually reflect in memory prices. Once memory prices reflected that, it  took another three to six months for the memory vendors to start building fabs. Those fabs take two years to build.

这个我们已经讲了一年半到两年。真正懂 AI 的人当时就重仓做多内存。这个动态早就能看到，但现在才终于反映到价格上。这么显而易见的事花了那么久：长上下文意味着 KV cache 变大，你需要更多内存。加速器一半的成本就是内存。他们当然会开始疯抢。这件事花了一年才真正反映到内存价格上。价格反映出来之后，内存厂商又花了三到六个月才开始动工建 fab。而这些 fab 要两年才能建好。

[1:30:24] Dylan

So we won't have really meaningful fabs to even  put these tools in until late 2027 or 2028. Instead, you've seen some really  crazy stuff to get capacity. Micron bought a fab from a company in  Taiwan that makes lagging-edge chips. Hynix and Samsung are doing some pretty  crazy things to try and expand capacity at their existing fabs, which also have  large knock-on effects in the economy. So why can't we build more capacity? There's nowhere to put the tools. It's not just EUV; there are other  tools involved in DRAM and logic. In logic, for N3, about 28% of the cost of the final wafer is EUV. When you look at DRAM, it's in the teens.

所以要到 2027 年底或 2028 年，我们才会有真正上规模的 fab 来安放这些设备。与此同时，你能看到一些为了抢产能干出来的相当疯狂的事。Micron 从台湾一家做成熟制程芯片的公司手里买下了一座 fab。Hynix 和 Samsung 也在用各种相当激进的手段扩张现有 fab 的产能，这对经济还有很大的连带效应。那为什么不能建更多产能？因为设备没地方放。不只是 EUV，DRAM 和逻辑芯片还涉及其他设备。在逻辑芯片上，N3 制程最终晶圆成本约 28% 来自 EUV。而在 DRAM 上，这个比例只有百分之十几。

[1:31:19] Dylan

It's going up, but it's a much  smaller percentage of the cost. These other tools are also bottlenecks, although  their supply chains are not as complex as ASML's. You see Applied Materials, Lam  Research, and all these other companies expanding capacity a lot as well. But you don't have anywhere to put the tool, because the most complex buildings people make  are fabs, and fabs take two years to build. I interviewed Elon recently, and his whole plan  is that they're going to build this TeraFab and they're going to build the clean rooms. I won't even ask you about the dirty rooms thing, but let's say they build the clean rooms. I have a couple of questions.

这个比例在上升，但在成本里占比仍小得多。其他设备同样是瓶颈，只不过它们的供应链没有 ASML 那么复杂。你能看到 Applied Materials、Lam Research 等公司也都在大举扩产。但你就是没有地方放设备，因为人类建造的最复杂的建筑就是 fab，而 fab 要两年才能建成。我最近采访了 Elon，他的整套计划是要建一座 TeraFab，他们要自己盖洁净室。「脏房间」那套说法我就不问你了，但假设他们把洁净室盖起来了。我有几个问题。

[1:32:58] Dwarkesh

One, do you think this is the kind of  thing that Elon Co. could build much faster than people conventionally build it? This is not about building the end tools. This is just about building the facility itself. How complicated is it to just build the clean room extremely fast? Is this something that Elon, with his "move fast" approach, could do much faster if that's  what we're bottlenecked on this year or next year? Two, does that even matter if, in two years,  your view is that we're not bottlenecked on clean room space, but on the tooling? As with any complex supply chain, it takes time, and constraints shift over time. Even if something is no longer a constraint, that doesn't mean that market no longer has margin. For example, energy will not be a big bottleneck a couple of years from now, but that  doesn't mean energy isn't growing super fast and there's no margin there. It's just not the key bottleneck.

第一，你觉得这种事是不是 Elon 的班子能比常规做法快得多建起来的？这说的不是造最终的设备，只是建设施本身。极快地盖出洁净室到底有多复杂？如果今明两年我们卡在这上面，Elon 那套「快速推进」的打法能不能快很多？第二，如果按你的判断，两年后瓶颈已经不在洁净室空间而在设备上，那这还重要吗？任何复杂供应链都一样：需要时间，而且约束会随时间转移。即使某样东西不再是约束，也不代表那个市场就没有利润了。比如，几年后能源就不会是大瓶颈了，但这不意味着能源不在高速增长、没有利润空间。它只是不再是关键瓶颈而已。

[1:33:47] Dylan

In the space of fabs, clean rooms are the  biggest bottleneck this year and next year. As we get to 2028, 2029, 2030, there  will still be constraints there. The thing about Elon is he has a tremendous  capability to garner physical resources and really smart people to build things. The way he recruits amazing people is by trying to build the craziest stuff. In the case of AI, that hasn't really worked because everyone's trying to build AGI. Everyone  is very ambitious. But in the case of going to Mars, making rockets that land themselves, fully  autonomous electric cars, or humanoid robots, these are methods of recruiting the people who  think that's the most important problem in the world to work on that problem, because  he's the only one trying really hard.

在 fab 这个领域，洁净室是今明两年最大的瓶颈。到 2028、2029、2030 年，那里仍然会有约束。Elon 的特别之处在于，他有超强的能力去调动物理资源和真正聪明的人来建东西。他招揽顶尖人才的方式，就是去造最疯狂的东西。在 AI 这件事上，这招不太灵，因为所有人都在造 AGI，所有人都野心勃勃。但在去火星、造能自己着陆的火箭、全自动驾驶电动车、人形机器人这些事上，这些项目就是他的招募手段——那些认为这是世界上最重要问题的人会来跟他干这个问题，因为只有他在玩命地做。

[1:34:31] Dylan

In the case of semiconductors, he stated he wants  to make a fab that's a million wafers per month. No one has a fab that big. It's possible that he's able to recruit a lot of really awesome people and get them on this  crazy task of building a million wafers a month. Step one is to build the clean room,  and that I think he probably can do. His mindset around deleting things, that it  can be dirty, it's fine, is probably not right. Actually I think it’s 100% not right. You need the fab to be very clean. All of the air in the fab gets replaced  every three seconds, it’s that fast. There have to be so few particles. But I think he can build the clean room.

在半导体上，他放话要建一座每月产 100 万片晶圆的 fab。没有人有这么大的 fab。他有可能招来一大批非常出色的人，让他们扑到「月产百万片晶圆」这个疯狂任务上。第一步是建洁净室，这一步我觉得他大概能做到。但他那套「能删就删、脏一点没关系」的思路恐怕不对。其实我认为是 100% 不对。fab 必须极其洁净。fab 里的空气每三秒就全部置换一次，就是这么快。颗粒物必须少到那种程度。但洁净室本身，我认为他能建出来。

[1:35:14] Dylan

It'll take a year or two. Initially, it won't be super fast, but over time, he'll get faster at it. The really complex part is actually developing a process technology and building wafers. I don't think he can develop that quickly. That has a lot of built-up knowledge. The most complicated integration of very expensive tools and supply chains  is done by TSMC, Intel, or Samsung. These two other companies aren't even that  great at it, and they're tremendously complex. How surprised would you be if in 2030  there just happened to be some total disruption where we're not using EUV? What if we're using something that has much better effects, is much simpler to produce,  and can be produced in much bigger quantities?

这要花一两年。一开始不会特别快，但随着时间推移他会越来越熟练。真正复杂的部分是开发制程技术并实际造出晶圆。我不认为他能很快搞定这个。这里面有大量长期积累的知识。把极其昂贵的设备和供应链做最复杂整合的，是 TSMC、Intel 或 Samsung。而后两家其实都没那么擅长，可见这有多复杂。如果到 2030 年恰好出现某种彻底颠覆——我们不再用 EUV 了，你会有多惊讶？比如我们用上了某种效果更好、生产更简单、能大批量制造的东西？

[1:35:56] Dwarkesh

I'm sure as an industry insider that  sounds like a totally naive question, but do you see what I'm asking? What probability should we put on something coming totally out of left  field to make all of this irrelevant? Something that's very simple and easy to  scale, I assign a very, very low probability. There are a number of companies  working on effectively particle accelerators or synchrotrons that generate  light that's either 13.5 nanometer, like EUV, or an even narrower wavelength, like X-ray at  7 nanometers, to then use in lithography tools. But those things are massive particle  accelerators generating this light.

我相信在业内人士听来这是个特别外行的问题，但你明白我在问什么吧？某种完全从天而降的东西冒出来，让这一切都变得无关紧要——我们该给这种情况多大概率？如果是那种非常简单、容易规模化的东西，我给的概率非常非常低。确实有一些公司在做本质上是粒子加速器或同步辐射光源的装置，用来产生 13.5 纳米（跟 EUV 一样）甚至更短波长的光，比如 7 纳米的 X 射线，然后用到光刻设备上。但那些东西是靠巨型粒子加速器来产生这种光的。

[1:36:32] Dylan

It's a very complicated thing to build. There are a couple of companies and I think that could be a big disruption  to the industry beyond EUV. But I don't think we're going to  magically build something new that is direct write and super simple, and can  be manufactured at huge volumes, although there are some attempts to do things like this. I ask because if you think about Elon's companies in the past, rocketry was this thing that was  thought to be—and is—incredibly complicated. Look, I'm just a naive yapper compared to Elon.  What have I built? So maybe it's possible. In order to build more memory in the  future, could we build 3D DRAM the way we do 3D NAND and then go back to DUV? That is the hope currently. Everyone's roadmap for 3D DRAM is that you'll still use EUV  because you want to have that tighter overlay.

这东西造起来极其复杂。确实有那么几家公司在做，我也觉得那可能成为 EUV 之外对整个行业的一次大颠覆。但我不认为我们会奇迹般地造出一种全新的、直写式的、超级简单、还能大规模量产的东西——虽然确实有人在做这类尝试。我这么问是因为，回看 Elon 过去的那些公司，火箭当年就被认为是——而且确实是——极其复杂的东西。听着，跟 Elon 比我就是个只会耍嘴皮子的外行，我造出过什么？所以也许真有可能。为了未来造出更多内存，我们能不能像做 3D NAND 那样做 3D DRAM，然后退回去用 DUV？这正是目前大家的指望。不过所有人的 3D DRAM 路线图里仍然要用 EUV，因为你需要更紧的套刻（overlay）精度。

[1:37:24] Dylan

When you're doing these subsequent processing  steps, everything is vertically stacked and you have more layers on top of each other. You want the pitches to be tighter. So generally, people are still  trying to do it with EUV. But what 3D would do is change the calculation  of how many bits a single EUV pass can make. That number would go up drastically if you  go to 3D DRAM. That is the hope. Right now, everyone's roadmap goes from the current 6F cell,  to a 4F cell, and then finally 3D DRAM by the end of the decade or early next decade. There's still a lot of R&D, manufacturing, and integration to be done. I wouldn't call that out of the cards.

在做后续这些工艺步骤时，一切都是垂直堆叠的，层数越叠越多，你会希望间距（pitch）更紧。所以总体上大家还是想用 EUV 来做。但 3D 化会改变'一次 EUV 曝光能做出多少 bit'这笔账——转向 3D DRAM 后，这个数字会大幅上升。这就是大家的指望。目前所有人的路线图都是从现在的 6F 单元走到 4F 单元，最后在这个十年末或下个十年初实现 3D DRAM。研发、制造、整合还有大量工作要做，但我不会说这事没戏。

[1:38:04] Dylan

I think it's very likely going to happen. It's also going to require a huge retooling of fabs. The breakdown of tools in a fab will be very different. The lithography tool is actually the only thing that isn't that different. But the number of them relative to different types of chemical vapor deposition, atomic layer  deposition, dry etch, or different kinds of etch chambers with different chemistries… You have all  these different tools for different process nodes. You can't just convert a logic fab to a  DRAM fab, or vice versa, or a NAND fab to a DRAM fab, in a short amount of time. In the same way, existing DRAM fabs require a lot of retooling just to go from 1-alpha to 1-beta  to 1-gamma process nodes, because they have to add DUV and change the chemistry stacks for when  you’re using EUV in terms of deposition and etch.

我认为它大概率会发生。但这也需要对 fab 做大规模的设备更换，fab 里各类设备的构成会变得非常不同。光刻机其实是唯一变化不大的东西，但它相对其他各类设备的数量比例会变——化学气相沉积、原子层沉积、干法刻蚀，还有用不同化学体系的各种刻蚀腔……不同制程节点要用各种不同的设备。你没法在短时间内把一座逻辑 fab 改成 DRAM fab，反过来也不行，NAND fab 改 DRAM fab 也一样。同理，现有的 DRAM fab 光是从 1-alpha 升级到 1-beta 再到 1-gamma 制程节点，就得大量换装设备，因为要引入 EUV，还要为用 EUV 的环节改换沉积和刻蚀的化学品体系。

[1:38:51] Dylan

And the EUV tool has to be there. Furthermore, when you change to 3D DRAM, there's going to be an even larger shift, so a  lot of retooling of these fabs needs to happen. That would be a big disruption. That would make EUV demand generally lower. But as we've seen across time, lithography demand  as a percentage of wafer cost has trended up. Around the 2014 era, it was 17% of the wafer cost,  and it's gone to 30% over the last fifteen years. For DRAM, it was in the low to mid-teens,  and now it's trended toward the high teens. Before we get to 3D DRAM, it'll  likely cross into the 20% range. But then, if we get to 3D DRAM, the total end  wafer cost as a percentage of EUV tanks again.

而且 EUV 设备必须到位。再往后，转向 3D DRAM 时变化会更大，这些 fab 需要大量重新换装设备。那会是一次大颠覆，会让 EUV 的总体需求下降。但纵观历史，光刻在晶圆成本中的占比一直在往上走：2014 年前后大约占晶圆成本的 17%，过去十五年涨到了 30%。DRAM 那边原来是 13-15% 上下，现在往 17-19% 靠。在走到 3D DRAM 之前，很可能会突破 20%。可一旦真到了 3D DRAM，EUV 占最终晶圆成本的比例又会暴跌。

[1:39:39] Dwarkesh

I guess you care less about the percent of cost  and more about how much it bottlenecks production. Right, but the percentage of cost— It’s a proxy, yeah. If you're Jensen or Sam Altman, or whoever stands to  gain a lot from scaling up AI compute, there are these stories that they'd go to  TSMC and say, "Why can't we access Y and Z?" But I think the point you're  making is that it doesn't really matter what TSMC does in some sense. In fact, even if you have Intel and Samsung building more foundries, in the  long run, you're going to be bottlenecked by ASML and other tool and material makers. First, is that a correct interpretation?

我猜你其实不太在乎成本占比，更在乎它对产能的瓶颈作用有多大。对，但成本占比——算是个替代指标吧。如果你是 Jensen 或者 Sam Altman，或者任何能从 AI 算力扩张里大赚一笔的人，有种说法是他们会跑去 TSMC 问：'为什么我们拿不到 Y 和 Z？'但我理解你的观点是：某种意义上，TSMC 做什么其实无关紧要。事实上，就算 Intel 和 Samsung 建更多晶圆厂，长期看你还是会被 ASML 以及其他设备和材料厂商卡住。第一个问题：这样理解对吗？

[1:40:18] Dwarkesh

Second, should Silicon Valley people be  going to the Netherlands right now to try to pitch ASML to make more tools so that  in 2030 they can have more AI compute? It's a funny dynamic we saw  in 2023, 2024, and 2025. People who saw the energy bottleneck  before others asymmetrically went to Siemens, Mitsubishi, and of course GE  Vernova, and bought up turbine capacity. Now they're able to charge  excess amounts for deploying these turbines in places because of energy. In the same sense, this could be done for EUV, except ASML is not just going to trust any  random bozo who wants to buy EUV tools.

第二个问题：硅谷的人现在是不是该飞去荷兰游说 ASML 多造设备，好让自己在 2030 年拥有更多 AI 算力？2023、2024、2025 年我们看到过一个很有意思的现象：那些比别人更早看到电力瓶颈的人，不对称地跑去 Siemens、Mitsubishi，当然还有 GE Vernova，把燃气轮机产能买断了。现在因为电力紧张，他们在各地部署这些轮机时能收超额溢价。同样的打法理论上也能用在 EUV 上——只不过 ASML 可不会随便信任哪个想买 EUV 设备的无名之辈。

[1:41:00] Dylan

These turbines are much cheaper than EUV  tools, and there's many more of them produced. Especially once you get to industrial gas  turbines, not just combined-cycle but the cheaper, smaller, less efficient ones, people put  down deposits for these. Someone could do this. Someone should go to the Netherlands  and be like, "I'll pay you a billion dollars. You give me the right to purchase ten EUV tools  two years from now, and I'm first in line." Then over those two years, you go around  and wait for everyone to realize, "Oh crap, I don't have enough EUV tools," and you  try to sell your option at some premium.

这些轮机比 EUV 设备便宜得多，产量也大得多。尤其是到了工业燃气轮机这一档——不只是联合循环机组，还有那些更便宜、更小、效率更低的——人们是会为它们交定金的。有人可以这么干。应该有人去荷兰说：'我付你十亿美元，你给我一个两年后购买十台 EUV 设备的权利，而且我排第一位。'然后这两年里你就四处转悠，等着大家反应过来：'糟了，我的 EUV 设备不够用'，这时你就把手里的期权加价卖出去。

[1:41:38] Dylan

All you're effectively doing  is saying, "ASML, you're dumb. You weren't making enough margin on these. I'm going to make a margin." The question is, will ASML even  agree to this? I don't think so. There's a world where they at least get the  demand signal from that to increase production. Potentially. I agree. But it sounds like you're saying they couldn't even increase production  if they wanted to, given the supply chain. Right. But that's exactly the market in  which… If they can't increase production, just like TSMC cannot increase production  that fast, and yet demand is mooning, then the obvious solution is to arbitrage this. You and I know demand is way higher than they're projecting and their capability to build. You arbitrage this by locking up the capacity, doing a forward contract, and then  trying to sell it at a later date once other people realize everything is  fucked and we don't have enough capacity.

你实际上等于在说：'ASML，你们真傻，这些设备的利润你们没赚够，那我来赚这个差价。'问题是，ASML 会答应这种事吗？我觉得不会。也存在一种可能：他们至少能从中拿到需求信号，从而扩产。有可能，我同意。但听你的意思是，就算他们想扩产，受供应链所限也扩不了。对。但恰恰是在这样的市场里……如果他们没法扩产——就像 TSMC 也没法那么快扩产——而需求又在暴涨，那显而易见的解法就是做套利。你我都知道，需求远高于他们的预测，也远超他们的建造能力。套利的方式就是锁定产能、签远期合约，等其他人意识到'全完了、产能根本不够'的时候再高价卖出。

[1:42:26] Dylan

Then you'll have this insane margin that  ASML and TSMC should have been charging. But the thing is, I don't know if  ASML and TSMC will ever agree to this. Let me ask you about power now. It sounds like you think power can be arbitrarily scaled. Not arbitrarily, but yes. But beyond these numbers. If I'm  remembering correctly, your blog post on how AI labs are increasing power implied that  GE Vernova, Mitsubishi, and Siemens could produce 60 gigawatts a year in gas turbines. Then there are these other sources, but they're less significant than the turbines. Only a fraction of that goes to AI, I assume.

到那时你就能吃到这笔本该由 ASML 和 TSMC 收取的疯狂利润。但问题在于，我不知道 ASML 和 TSMC 到底会不会答应。接下来我想问问电力。听起来你认为电力是可以任意扩张的？不能说任意，但差不多。但要超出这些数字的话——如果我没记错，你那篇讲 AI 实验室如何扩充电力的博客文章里提到，GE Vernova、Mitsubishi 和 Siemens 一年能生产 60 吉瓦的燃气轮机。此外还有其他电源，但都不如轮机重要。我猜其中只有一部分流向了 AI。

[1:43:10] Dwarkesh

If in 2030 we have enough logic and memory to  do 200 gigawatts a year, do you just think that these things are on a path to ramp up to more  than 200 gigawatts a year, or what do you see? Right now we're at 20 or 30. This is critical IT capacity, by the way, which is an important thing to mention. When I'm talking about these gigawatts, I'm talking about critical IT capacity. Server plugged in, that's how much power it pulls. But there are losses along the chain. There is loss on transmission, conversion, cooling, et cetera. So you should gross this factor up from 20 gigawatts for this year, or 200  gigawatts by the end of the decade, to some number 20-30% higher. Then you have capacity  factors. Turbines don't run at 100 percent.

如果到 2030 年我们的逻辑芯片和内存足以支撑一年 200 吉瓦，你是认为这些电源正走在爬坡到每年 200 吉瓦以上的路径上，还是你怎么看？我们现在是 20 或 30 吉瓦。顺便说一句，这指的是关键 IT 容量（critical IT capacity），这点很重要。我谈这些吉瓦数时，说的都是关键 IT 容量：服务器插上电，它实际抽走多少功率。但整条链路上是有损耗的——输电、电力变换、散热等等都有损失。所以不管是今年的 20 吉瓦还是这个十年末的 200 吉瓦，你都得在此基础上再上浮 20-30%。然后还有容量系数的问题——轮机不会 100% 满负荷运转。

[1:43:54] Dylan

If you look at PJM, which I think is the largest  grid in America—covering the Midwest and some of the Northeast area—in their models they want  to have roughly 20 percent excess capacity. Within that 20 percent excess capacity,  they're running all the turbines at 90% because they are derated some for  reliability, maintenance, and so on. In reality, the nameplate capacity for energy is  always way higher than the actual end critical IT capacity because of all these factors. But it's  not just turbines. If you were just making power from turbines, that's simple, boring, and easy. Humans and capitalism are far more effective.

看看 PJM——我记得它是美国最大的电网，覆盖中西部和东北部部分地区——他们的模型里希望保有大约 20% 的富余容量。而在这 20% 富余之内，所有轮机也只按 90% 运行，因为出于可靠性、维护之类的原因要做降额。现实中，铭牌容量永远远高于最终的关键 IT 容量，就是因为这一堆因素。但电力不只靠轮机。如果只用轮机发电，那太简单、太无聊、太容易了。人类和资本主义可比这有办法得多。

[1:44:41] Dylan

The whole point of that blog was that, yes, there  are only three people making combined-cycle gas turbines, but there's so much more we can  do. We can do aeroderivatives. We can take airplane engines and turn them into turbines. There are even new entrants in the market, like Boom Supersonic trying to  do that and working with Crusoe. Also there's all the other ones like  that already exist in the market. There are also medium-speed  reciprocating engines: engines that spin in circles, like a diesel engine. There are ten people who make engines that way. I'm from Georgia, and people used  to be like, "Oh man, you got a Cummins engine in there," regarding RAM trucks. Automobile manufacturing is going down, so these companies all have capacity and could scale  and convert that for data center power.

那篇博客的核心观点就是：没错，联合循环燃气轮机只有三家在做，但我们能做的事多得多。可以搞航改燃机（aeroderivative）——把飞机发动机改造成轮机。市场上甚至有新玩家在做这个，比如 Boom Supersonic 就在尝试，还在和 Crusoe 合作。市场上现成的同类产品也不少。还有中速往复式发动机：像柴油机那样转圈的发动机，做这种发动机的厂家有十家。我是 Georgia 人，以前大家聊 RAM 皮卡时总会说'哥们，你这车里装的可是 Cummins 发动机'。汽车制造业在走下坡路，这些公司都有富余产能，完全可以扩产并转产数据中心电力。

[1:45:26] Dylan

You stick all these reciprocating engines in. It's not as clean as combined-cycle, but maybe you can convert them from diesel to gas if you want.  What about ship engines? All of these engines for massive cargo ships are great. Nebius is doing that for a Microsoft data center in New Jersey. They're running ship engines to generate power. Bloom Energy is doing fuel cells. We've been very positive on them for a year and a half now because they have such  a capability to increase their production. Their payback period for a production  increase is very fast, even if the cost is a little bit higher than combined-cycle,  which is the best for cost and efficiency.

你把这些往复式发动机全堆上去。它不如联合循环那么清洁，但想要的话也许还能把它们从烧柴油改成烧天然气。船用发动机呢？那些巨型货轮上的发动机都很棒。Nebius 就在 New Jersey 为一座 Microsoft 数据中心这么干——用船用发动机发电。Bloom Energy 在做燃料电池。我们看好他们一年半了，因为他们的产能扩张能力很强，扩产的回本周期非常快，哪怕成本比联合循环——成本和效率上的最优选——稍高一点。

[1:46:06] Dylan

Then there's solar plus battery, which can come  online as those cost curves continue to come down. There's wind, where you might only expect 15  percent of the maximum power because things oscillate, but you add batteries. There are  all these things. The other thing is that the grid is scaled so we don't cut off power at  peak usage on the hottest day of the summer. But in reality, that's a load spike  that is 10-20% higher than the average. If you just put enough utility-scale  batteries, or peaker plants that only run a small portion of the year—and those could  be gas, industrial gas turbines, combined-cycle, batteries, or any of the other sources  I mentioned—then all of a sudden you've unlocked 20% of the US grid for data centers. Most of the time that capacity is sitting idle.

然后是光伏加电池，随着成本曲线持续下探，它们会陆续上线。还有风电——因为出力波动，你可能只能指望最大功率的 15%，但可以配上电池。这类方案有一大堆。另外一点是：电网的规模是按'盛夏最热那天的用电高峰也不断电'来建的。但实际上，那只是一个比平均值高 10-20% 的负荷尖峰。只要装上足够的电网级储能电池，或者一年只运行一小段时间的调峰电站——可以是燃气、工业燃气轮机、联合循环、电池，或者我提到的任何其他电源——你一下子就为数据中心解锁了美国电网 20% 的容量。这部分容量大多数时候都在闲置。

[1:47:00] Dylan

It's really only there for that peak, which is  just a few hours over a few days of the year. If you have enough capacity  to absorb that peak load, then all of the sudden you’ve transferred it all. Today, data centers are only 3-4% of the power of the US grid, and by 2028 they'll be 10%. But if you can unlock 20% of the US grid like this, it's not that crazy. The US grid is terawatt-level, not hundreds-of-gigawatts-level. So we can add a lot more energy. I'm not saying it's easy. These things are going to be hard. There's a lot of hard engineering, risks people have to take, and new  technologies people have to use.

它的存在基本只为了应付那个尖峰——一年里也就那几天的几个小时。只要有足够的容量去吸收那个峰值负荷，一下子就全腾出来了。今天数据中心只占美国电网用电的 3-4%，到 2028 年会到 10%。但如果能像这样解锁美国电网的 20%，其实并不夸张。美国电网是太瓦级的，不是几百吉瓦级的。所以我们还能新增很多电力。我不是说这容易——这些事都会很难，有大量硬核工程要做，有人得去承担风险，还得用上各种新技术。

[1:47:40] Dylan

But Elon was the first to do this behind-the-meter  gas, and since then we've seen an explosion of different things people are doing to get power. They're not easy, but people are gonna be able to do them. The supply chains are just way simpler than chips. Interesting. He made the point during the  interview that for the specific blade for the specific turbine he was looking at, the lead  times go out beyond 2030. Your point is that— That's great. There are so many other ways to  make energy. Just be inefficient. It's fine. Right now, combined-cycle gas turbines  have CapEx of $1,500 per kilowatt.

但 Elon 是第一个搞'表后'（behind-the-meter）燃气发电的，从那以后我们看到人们为了搞到电力各显神通，方案层出不穷。都不容易，但人们是做得到的。这些供应链比芯片简单太多了。有意思。他在那期访谈里提到，他看中的那款轮机所需的那种叶片，交货周期已经排到 2030 年以后了。你的观点是——那没关系，发电的路子多得是。低效一点就低效一点，没事。目前联合循环燃气轮机的 CapEx 是每千瓦 1500 美元。

[1:48:17] Dwarkesh

Are you saying it would make sense to  have either technologies that are much more expensive than that, or other things are  getting cheap enough to make it competitive? Exactly. It can be as high as $3,500 per kilowatt. It could be twice as much as the cost of combined-cycle, and the total cost of the GPU on  a TCO basis has only gone up a few cents per hour. Because we've been talking about Hopper pricing,  $1.40, let's say the power price doubles. The Hopper that was $1.40 is now $1.50 in cost. I don't care, because the models are improving so fast that the marginal utility of them is worth  way more than that ten-cent increase in energy.

你的意思是：要么用远比这贵的技术也说得通，要么其他方案正变得足够便宜、开始有竞争力了？正是。哪怕高到每千瓦 3500 美元——相当于联合循环成本的两倍——按 TCO 算，GPU 的总成本每小时也只涨几美分。我们一直在聊 Hopper 的价格，1.40 美元；假设电价翻倍，原来成本 1.40 美元的 Hopper 现在变成 1.50 美元。我根本不在乎，因为模型进步太快了，它们的边际效用远远超过那 10 美分的电费涨幅。

[1:49:00] Dwarkesh

So you're saying 20 percent of the grid—the grid  is about one terawatt—can just come online from utility-scale batteries, increasing what  you'd be comfortable putting on the grid. The regulatory mechanism  there is not easy, by the way. But that's 200 gigawatts, if  that hypothetically happens. Just from the different sources of gas generation  you mentioned—the different kinds of engines and turbines—combined, how many gigawatts  could they unlock by the end of the decade? We're tracking this in our data. There are over 16 different manufacturers of power-generating things just from gas alone. Yes, there are only three turbine manufacturers for combined-cycle, but we're  tracking 16 different vendors, and we have all of their orders. It turns out there are hundreds of gigawatts of orders to various data centers. As we get to the end of the decade, we think something like half of the capacity  that's being added will be behind the meter.

所以你的意思是，电网总量大约是 1 太瓦，其中 20% 可以靠公用事业级电池直接上线，从而提高你敢往电网上加的负载量。顺便说一句，那里的监管机制可不简单。但假设这真的发生了，那就是 200 吉瓦。光是你刚才提到的各种燃气发电来源——不同类型的发动机和燃气轮机——加起来，到这个十年结束时能释放出多少吉瓦？我们的数据一直在追踪这个。光是燃气这一块，就有超过 16 家不同的发电设备制造商。没错，联合循环的燃气轮机制造商只有三家，但我们追踪着 16 家不同的供应商，而且掌握了他们所有的订单。结果发现，各个数据中心的订单加起来有几百吉瓦。到这个十年末，我们估计新增容量里大约有一半会是「表后」（behind the meter）的。

[1:49:59] Dylan

Behind the meter is almost always more expensive  than grid-connected, but there are just a lot of problems with getting grid-connected: permits and  interconnection queues and all this sort of stuff. So even though it's more expensive,  people are doing behind the meter. What they're doing behind the meter ranges widely. It could be reciprocating engines, ship engines, or aeroderivatives. It could be combined-cycle, although combined-cycle is not  that great for behind the meter. It could be Bloom Energy fuel  cells, or solar plus battery. It could be any of these things. And you're saying any of these individually could do tens of gigawatts? Any of these individually will do tens of gigawatts, and as a whole, they  will do hundreds of gigawatts.

表后发电几乎总是比并网更贵，但并网的麻烦实在太多：许可证、并网排队等等这一堆事。所以尽管更贵，大家还是在做表后发电。表后用的东西五花八门。可以是往复式发动机、船用发动机，或者航改燃机。可以是联合循环，虽然联合循环其实不太适合表后场景。可以是 Bloom Energy 的燃料电池，或者太阳能加电池。什么都有可能。你是说这里面任何一种单拎出来都能做到几十吉瓦？任何一种单拎出来都能做到几十吉瓦，加在一起能做到几百吉瓦。

[1:50:36] Dwarkesh

Okay. So that alone should more than— Electrician wages will probably double or triple again. There are going to be a lot of new people entering that field, and a ton of people who make money,  but I don't see that as the main bottleneck. Right now in Abilene, at the 1.2-gigawatt data  center that Crusoe is building for OpenAI, I think they have 5,000 people  working there, or at peak they did. If you turn that into 100 gigawatts—and  I'm sure things will get more efficient over time—that would be 400,000 people  it would take to build 100 gigawatts. If you think about the US labor force, and  how many electricians there are and how many construction workers there are… I  guess there are 800,000 electricians.

好。那光这一项就应该绰绰有余——电工的工资大概还会再翻两三倍。会有一大批新人进入这个行业，很多人会赚到钱，但我不认为这是主要瓶颈。现在在 Abilene，Crusoe 给 OpenAI 建的那个 1.2 吉瓦数据中心，我记得有 5,000 人在那里干活，至少高峰期是这样。如果把它换算成 100 吉瓦——当然效率肯定会随时间提高——那建 100 吉瓦就需要 40 万人。想想美国的劳动力总量，有多少电工、多少建筑工人……我记得电工大概有 80 万。

[1:51:24] Dwarkesh

I don't know if they're all  substitutable in this way. There are millions of construction workers. But if we're in a world where we're adding 200 gigawatts a year, are we going to be  crunched on labor eventually, or do you think that is actually not a real constraint? Labor is a big constraint. It's a humongous constraint in this. People have to  be trained. Likewise, we'll probably start importing the highest-skilled labor. It makes sense that a really high-skilled electrician in Europe who was working  on destroying power plants now comes to America and is building high-voltage  electricity moving across a data center.

我不确定他们是否都能这样直接顶上去。建筑工人有几百万。但如果我们进入一个每年新增 200 吉瓦的世界，劳动力最终会不会被挤爆？还是你觉得这其实不是真正的约束？劳动力是个大约束。在这件事里是个巨大的约束。人得经过培训。同样，我们大概会开始引进最高技能的劳动力。这很合理：一个原本在欧洲拆电厂的高技能电工，现在来到美国，在数据中心里架设高压输电线路。

[1:52:03] Dylan

Humanoid robots or robotics at least might start  to help, but the main factor for reducing the number of people is going to be modularizing  things and making them in factories in Asia. Unfortunately for America, places like Korea,  Southeast Asia, and in many ways China as well are going to ship more and more built-out sections  of the data center and those will be shipped in. Today you currently ship servers or a rack in,  and then you plug that into different pieces that you're shipping from different places. But now you'll ship it to a factory and integrate the entire thing. Maybe this is a two-megawatt block, and this block goes from high-voltage AC  power to the DC voltage that you deliver to the rack, or something like this. Or with cooling, you ship a fully integrated unit that has a lot of the  cooling subsystems already put together, because plumbers are also a big constraint here. Furthermore, instead of just a single rack where you have people wiring up all these racks with  electricity, you take a skid and put an entire row of servers on it that is  shipped directly from the factories.

人形机器人，或者至少是机器人技术，可能会开始帮上忙，但真正减少用人数量的主要因素，是把东西模块化、放到亚洲的工厂里生产。对美国来说很不幸，韩国、东南亚，以及在很多方面也包括中国，会把越来越多预制好的数据中心模块整段运过来。今天你是把服务器或整机柜运进来，然后把它接到从各地运来的不同部件上。而以后你会把东西运到工厂里，把整个系统集成好。比如这是一个 2 兆瓦的模块，这个模块负责把高压交流电转换成送到机柜的直流电压，诸如此类。或者在散热方面，你直接运来一个把大量冷却子系统都预装好的完整集成单元，因为管道工在这里也是个大瓶颈。再进一步，与其让工人给一个个机柜逐一接电，不如用一个滑撬（skid）装上一整排服务器，直接从工厂整体发货。

[1:53:25] Dylan

Today, a single rack may be 120 or 140 kilowatts,  but as we get to next-generation Nvidia Kyber and things like that, it's almost a megawatt. In addition, if you do an entire row, it'll have the rack, the networking, the  cooling, and the power all integrated together. Now when you come in, you have much less to cable. There's less networking fiber, fewer power connections, and fewer plumbing things. This can drastically reduce the number of people working in data centers, so our  capability to build them will be much larger. Along the way, some people will move faster  to new things, and some will move slower.

今天单个机柜可能是 120 或 140 千瓦，但到了下一代 NVIDIA Kyber 之类的产品，就快到一兆瓦了。另外，如果做成一整排，机柜、网络、冷却、供电就全都集成在一起了。这样进场施工时，要接的线就少多了。网络光纤更少、电源接头更少、管道活也更少。这能大幅减少数据中心里的用工数量，我们的建设能力就会大得多。在这个过程中，有些人向新方案转得快，有些人转得慢。

[1:54:08] Dylan

Crusoe and Google have been talking  a lot about this modularization, as have companies like Meta and many others. The people who move faster to new things may face delays, while the people who  are slower will face labor problems. There will always be dislocations in the market  because this is a very complex supply chain. At the end of the day, it's still  simple enough that we will be able to solve it through capitalism and human  ingenuity on the timescales required. Speaking of big problems to solve, Elon  Musk is very bullish on space GPUs. If you're right that power is not a constraint  on Earth… I guess the other reason they would make sense is that even if there will be  enough gas turbines or whatever on Earth, Elon's next argument is that you can't get the  permitting to build hundreds of gigawatts on Earth. Do you buy that argument? Land-wise, America is big. Data centers don't actually take up that  much space, so you can solve that.

Crusoe 和 Google 一直在大谈这种模块化，Meta 和很多其他公司也是。转向新方案快的人可能会遇到延误，而转得慢的人则会碰上劳动力问题。市场上永远会有错配，因为这是一条非常复杂的供应链。但归根结底，它还是足够简单，我们能靠资本主义和人类的聪明才智在所需的时间尺度内把它解决掉。说到要解决的大问题，Elon Musk 非常看好太空 GPU。如果你说得对、地球上电力不是约束……我猜他们讲得通的另一个理由是：就算地球上燃气轮机什么的够用，Elon 的下一个论点是，你在地球上拿不到建几百吉瓦所需的审批许可。你买这个论点的账吗？从土地上说，美国很大。数据中心其实占不了多少地方，这个问题解决得了。

[1:55:09] Dylan

Permitting-wise, air pollution permits are  a challenge, but the Trump administration made it much easier. You go to Texas, and you can skip a lot of this red tape. Elon had to deal with a lot of this complex stuff in Memphis, and then building a power  plant across the border for Colossus 1 and 2. But at the end of the day, there's a lot more  you can get away with in the middle of Texas. Given that Elon lives in Texas,  why didn't he just go to Texas? I think it was partially that they over-indexed  on grid power for a temporary period of time. That's just what they thought they needed more of. Because they had an aluminum refinery connected to the grid there. It was actually an idled appliance factory.

从审批上说，空气污染许可是个难题，但 Trump 政府让它容易了很多。你去 Texas，可以跳过一大堆繁文缛节。Elon 在 Memphis 就跟这些复杂的事纠缠了很久，还为 Colossus 1 和 2 在州界那边建了个电厂。但说到底，在 Texas 腹地你能通融的空间大得多。既然 Elon 住在 Texas，他当初为什么不直接去 Texas？我觉得部分原因是，他们在某段时间里过度押注了电网供电。那只是他们当时以为自己更需要的东西。因为那里有个接入电网的铝精炼厂。其实那是一座闲置的家电工厂。

[1:55:50] Dylan

But I think they may have indexed more to  grid power, water access, and gas access. I think they bought that knowing the gas  line was right there and they were going to tap it. Same with water. It was a  whole host of different constraints. It was probably an area where  electricians were easier to find. At the end of the day, I'm not  exactly sure why they chose that site. I bet Elon would've chosen somewhere in  Texas if he could've gone back because of the regulatory challenges he faced. Ultimately, permitting is a challenge, but America is a big place with 50  states, and things will get done.

但我觉得他们当时可能更看重电网供电、水源和天然气供应。我想他们买那块地时就知道燃气管线就在旁边，打算直接接上去。水也是同理。是一整套不同约束的组合。那里大概也是个更容易招到电工的地方。说到底，我也不完全确定他们为什么选那个地址。我打赌，如果能重来，考虑到他遇到的那些监管麻烦，Elon 会选 Texas 的某个地方。归根结底，审批是个难题，但美国地方大、有 50 个州，事情总能办成。

[1:56:27] Dylan

There are a lot of small jurisdictions where  you can just transport in all the workers you need for a temporary period of three to  twelve months, depending on the contractor. You can put them in temporary housing and pay out  the butt, because labor is very cheap relative to the GPUs and the networking, and the end  value of the tokens it's going to produce. So there is plenty of room to  pay for all of these things. Also, people are also diversifying now. Australia, Malaysia, Indonesia, and India are all places where data centers  are going up at a much faster pace. But currently, over 70% of AI  data centers are still in America, and that continues to be the trend. People are figuring out how to build these things.

有很多小的辖区，你可以把需要的工人全部临时运进去，干上三到十二个月，取决于承包商。你可以给他们安排临时住房，工钱砸重金也无所谓，因为相对于 GPU、网络设备，以及最终产出的 token 的价值，劳动力实在太便宜了。所以完全有余地把这些成本都付掉。另外，大家现在也在做分散布局。Australia、Malaysia、Indonesia、India 这些地方的数据中心都在以快得多的速度拔地而起。但目前超过 70% 的 AI 数据中心仍然在美国，而且这个趋势还在延续。大家正在摸索怎么把这些东西建起来。

[1:57:19] Dylan

Ultimately, dealing with permitting and  red tape in middle-of-nowhere Texas, Wyoming, or New Mexico is probably a hell of  a lot easier than sending stuff into space. Other than the economic argument making less sense  once you consider that energy is a small fraction of the total cost of ownership of a data center,  what are the other reasons you're skeptical? Obviously, power is basically free in space. That's the reason to do it. Yeah, that's the reason to do it. But there are all the other counterarguments. Even if power costs double on Earth, it's  still a fraction of the total cost of the GPU.

说到底，在 Texas、Wyoming 或 New Mexico 那种前不着村后不着店的地方对付审批和繁文缛节，大概率比把东西送上太空容易得多。除了经济账上的问题——一旦你考虑到能源只占数据中心总拥有成本的一小部分，这个论点就不太成立了——你持怀疑态度还有哪些别的理由？显然，太空里电力基本上是免费的。这就是做这件事的理由。对，这就是做它的理由。但还有其他所有的反面论证。就算地球上电价翻倍，它仍然只是 GPU 总成本的一个零头。

[1:57:54] Dylan

The main challenge is… We have  ClusterMAX, which rates all the neoclouds. We test over 40 cloud companies,  including the hyperscalers and neoclouds. Outside of software, what differentiates these  clouds the most is their ability to deploy and manage failure. GPUs are horrendously unreliable.  Even today, around 15% of Blackwells that get deployed have to be RMA'd. You have to take them out. Sometimes you just have to plug them  back in, but sometimes you have to take them out and ship them back to Nvidia or  their partners who do the RMAs and such. What do you make of Elon's argument that after an  initial phase, they actually don't fail that much?

主要的挑战在于……我们有 ClusterMAX，它给所有 neocloud 打分。我们测试了 40 多家云公司，包括 hyperscaler 和 neocloud。抛开软件不谈，这些云厂商之间最大的差异，在于它们部署和应对故障的能力。GPU 的可靠性糟糕透顶。即便在今天，部署下去的 Blackwell 大约有 15% 得走 RMA（退回返修）。你得把它们拆下来。有时候只要重新插一下就行，但有时候就得拆下来寄回给 Nvidia 或它们负责 RMA 的合作伙伴。那你怎么看 Elon 的说法——过了最初一个阶段之后，它们其实没那么容易坏？

[1:58:34] Dylan

Sure, but now you've done this, tested them all,  deconstructed them, put them on a spaceship, launched them into space, and then put  them online again. That takes months. If your argument is that a GPU has a useful life of  five years, and this takes six additional months, that is 10% of your cluster's useful life. Because we're so capacity-constrained, that compute is theoretically most valuable  in the first six months you have it. We're more constrained now  than we will be in the future. That compute can contribute to a better  model in the future, or generate revenue today that you can use to raise more money. All these things make now the most important moment, but you've potentially delayed  your compute deployment by six months.

行，但现在你得先做完这一切：全部测完、拆解、装上飞船、发射上太空、再重新上线。这要花好几个月。如果你的论点是 GPU 的使用寿命是五年，而这一套多花了六个月，那就是你整个集群使用寿命的 10%。因为我们现在算力极度紧缺，理论上这些算力在你拿到手的头六个月里价值最高。我们现在比未来任何时候都更紧缺。这些算力可以贡献给未来更好的模型，或者在今天产生收入，让你能去融更多的钱。所有这些都让「现在」成为最重要的时刻，而你却可能把算力部署硬生生推迟了六个月。

[1:59:25] Dylan

What separates these cloud providers is… We  see some clouds taking six months to deploy GPUs right here on Earth. We see clouds that take a lot less than six months. So the question is, where does space get in there? I don't see how you could test them all on Earth,  deconstruct them, and ship them to space without it taking significantly longer than just leaving  them in the facility where you tested them. The question I wanted to ask is about  the topology of space communication. Right now, Starlink satellites talk to  each other at 100 gigabits per second. You could imagine that being much  higher with optical intersatellite laser links optimized for this. That actually ends up being quite close to InfiniBand bandwidth,  which is 400 gigabytes a second.

这些云服务商之间的差距在于……我们看到有些云在地球上部署 GPU 都要花六个月。也有些云用的时间远少于六个月。那么问题来了，太空排在哪个位置？我不明白你怎么可能在地球上把它们全测完、拆解、运上太空，还不比直接把它们留在测试它们的机房里花的时间长得多。我想问的问题是关于太空通信的拓扑结构。现在 Starlink 卫星之间的通信速率是每秒 100 吉比特。可以想象，用专门为此优化的星间光学激光链路，这个数字会高得多。这样算下来其实相当接近 InfiniBand 的带宽，也就是每秒 400 吉比特。

[2:00:09] Dylan

But that's per GPU, not per rack. So multiply  that by 72. Also, that was Hopper. When you go to Blackwell and Rubin, that 2x's and 2x's again. But how much compute is happening per… During inference, are the different scale-ups  still working together, or is inference just happening as a batch within a single scale-up? A lot of models fit within one scale-up domain, but many times you split them  across multiple scale-up domains. As models become more and more sparse,  which is the general trend, you want to ping just a couple of experts per GPU. If leading models today have hundreds, if not a thousand, of experts, then you'd want to  run this across hundreds or thousands of chips, even as we advance into the future. So then you end up with the problem of needing to connect all these satellites  together for communications as well.

但那是每块 GPU 的带宽，不是每个机柜的。所以要乘以 72。而且那还是 Hopper 的数字。到了 Blackwell 和 Rubin，还要再翻一倍、又翻一倍。但每个……的计算量有多少？在推理的时候，不同的 scale-up 域之间还需要协同工作吗，还是说推理就是在单个 scale-up 域内按批次完成的？很多模型能装进一个 scale-up 域，但很多时候你会把它们拆到多个 scale-up 域上。随着模型越来越稀疏——这是大趋势——你希望每块 GPU 只 ping 少数几个专家。如果今天的头部模型有几百个甚至上千个专家，那你会想把它跑在几百甚至几千块芯片上，而且往后发展也是如此。于是你就会碰到一个问题：还得把所有这些卫星在通信上也连起来。

[2:01:09] Dylan

That would be tough. If there's a world where  you could do inference for a batch on a single scale-up, then maybe it's more plausible. But if not, it's a different story. Networking these chips together  is a problem, and you can't just make the satellite infinitely large. There are a lot of physics challenges to making a satellite really big. That's why you need these interconnects between the satellites. Those  interconnects are more expensive. In a cluster, 15-20% of the cost is networking. All of a sudden, you're using space lasers instead of simple lasers that are manufactured in  volumes of millions with pluggable transceivers.

那会很难。如果存在一个世界，一个批次的推理能在单个 scale-up 域里完成，那也许还更说得通一些。但如果不行，那就是另一回事了。把这些芯片联网本身就是个难题，而且你不可能把卫星做到无限大。把卫星做得非常大有一大堆物理上的挑战。所以你才需要卫星之间的这些互连。而这些互连更贵。在一个集群里，15% 到 20% 的成本是网络。突然之间，你用的是太空激光，而不是那种以百万级产量制造、配可插拔光模块的普通激光器。

[2:01:50] Dylan

And those things are very unreliable as well,  more unreliable than the GPUs by the way. Across the life of a cluster, you have  to unplug and clean them all the time. You have to unplug and replug  them just for random reasons. These things are just not as reliable. So you've got that problem as well. You've got a more expensive, complicated  space laser to communicate instead of this pluggable optical transceiver that's  been produced in super high volume. So all in all, what does that  imply for space data centers? Space data centers effectively are  not limited by their energy advantage. They are limited by the same contended resource. We can only make two hundred gigawatts of chips a year by the end of the decade. What are we going to do to get that capacity?

而且那些东西本身也非常不可靠——顺便说一句，比 GPU 还不可靠。在集群的整个生命周期里，你得不停地把它们拔下来清理。有时候纯粹因为莫名其妙的原因就得拔了重插。这些东西就是没那么可靠。所以你还得面对这个问题。你用上了更贵、更复杂的太空激光来通信，而不是那种已经以超大批量生产的可插拔光模块。那么总的来说，这对太空数据中心意味着什么？太空数据中心实际上并不受限于它们的能源优势能不能兑现，而是受限于同一种被争抢的资源。到这个十年结束，我们一年也只能造出两百吉瓦的芯片。我们要怎么拿到这份产能？

[2:02:29] Dylan

It doesn't matter if it's on land or in space. It doesn’t really matter, because you can build that power. Human capabilities and capacity could get to the period where we're adding a terawatt  a year globally of various types of power. At some point, we do cross the chasm where space  data centers make sense, but it's not this decade. It is much further out, once energy  constraints actually become a big bottleneck and land permitting becomes a much bigger  bottleneck as it subsumes more of the economy. And crucially, once chips  are no longer the bottleneck. Right now, chips are the biggest bottleneck. You want them deployed and working on AI the moment they're manufactured. There are a lot of things people are doing to increase that speed faster and faster. They’re modularizing data centers, or even modularizing racks where you put the chip in at  the data center, but only the chip and everything else is already wired up and ready to go. There are things like this people are doing to decrease that time that you cannot do in space. At the end of the day, all that matters in a chip-constrained world is getting  these chips producing tokens ASAP.

在地面还是在太空其实无所谓。真的无所谓，因为电力你是可以建出来的。人类的能力和产能可以发展到全球每年新增一太瓦各类电力的阶段。到某个时点，太空数据中心确实会跨过那道鸿沟、变得有意义，但不是这个十年。那要远得多——得等到能源约束真正成为大瓶颈、土地审批随着 AI 吞掉更大比例的经济而成为更大的瓶颈之后。而且关键是，得等到芯片不再是瓶颈。现在芯片才是最大的瓶颈。你希望芯片一下产线就立刻部署、立刻跑 AI。为了把这个速度越提越快，大家在做很多事：把数据中心模块化，甚至把机架模块化——芯片是到了数据中心才装上去的，但也只装芯片，其他一切都已经布好线、随时可用。诸如此类缩短时间的办法，在太空里是做不了的。说到底，在一个芯片受限的世界里，唯一重要的就是让这些芯片尽快开始产出 token。

[2:03:43] Dylan

Maybe by 2035, the semiconductor industry,  ASML, Zeiss, and suppliers like Lam Research and Applied Materials and other fab  manufacturers will catch up once the pendulum swings and we are able to make enough chips. Then we will be optimizing every dial and it makes sense to optimize the 10-15% of energy costs. As we move to ASICs potentially, and if Nvidia's margins aren't +70%, maybe  that energy cost becomes 30% of the cluster. These are the things to optimize. But Elon doesn't win by doing 20% gains. He never wins that way. Elon wins when he swings for  the fences and does 10X gains. That's what SpaceX is about. That's what Tesla is about. All of his  success has been about that, not chasing the 20%.

也许到 2035 年，半导体行业——ASML、Zeiss，还有 Lam Research、Applied Materials 这些供应商以及其他 fab 设备厂商——会赶上来，钟摆摆回去，我们能造出足够多的芯片。那时我们才会去优化每一个旋钮，去抠那占 10-15% 的能源成本才有意义。等我们可能转向 ASIC，而且如果 Nvidia 的毛利不再是 70% 往上，也许能源成本会占到集群成本的 30%。这些才是那时该优化的东西。但 Elon 从来不靠 20% 的改进取胜。他从不这么赢。Elon 赢的方式是抡圆了打全垒打，做 10 倍的提升。SpaceX 是这么回事，Tesla 也是这么回事。他所有的成功都源于此，而不是去追那 20%。

[2:04:31] Dylan

I think space data centers will eventually  be a 10X gain as Earth's resources get more and more contentious, but that's not this decade. Just to drive some intuition about how much land there is on Earth… Obviously, for the chips  themselves, especially if you move to a world where you have racks that have megawatts— That's the other thing. If manufacturing is the constraint, right now it's roughly one  watt per square millimeter for AI chips. One easy way to improve that is to pump  it to two watts per square millimeter. You may not get 2x the performance,  you may only get 20% more performance, and that requires much more exotic cooling. It requires more complicated cold plates and complex liquid cooling, or maybe  even things like immersion cooling.

我认为随着地球上的资源争夺越来越激烈，太空数据中心最终会成为一个 10 倍级的提升，但不是这个十年。先建立一点直觉，感受下地球上有多少土地……显然，就芯片本身而言，尤其是如果你走向一个机架功率达到兆瓦级的世界——这是另一件事。如果制造才是约束，那么现在 AI 芯片大约是每平方毫米一瓦。一个简单的改进办法就是把它拉到每平方毫米两瓦。你未必能拿到 2 倍性能，可能只多 20% 的性能，而且这需要激进得多的散热方案：更复杂的冷板、更复杂的液冷，甚至可能要用浸没式冷却这类东西。

[2:05:18] Dylan

In space, higher watts per  millimeter is very difficult, whereas on Earth, these are solved problems. One of these things enables you to get a lot more tokens, maybe 20% more tokens per wafer  that's manufactured, and that's a humongous win. Square millimeter, you mean of die area? Yeah, of die area. It would be better for space because more watts  per millimeter means the chip runs hotter. I guess this is a question of computer  chip engineering, but it cools to the fourth power by the Stefan-Boltzmann law. If you can run a very hot chip, it allows a lot of— No, you can't run it hotter. You can only run it denser. The problem is that getting the heat out of that dense area means you have to  move away from standard air and liquid cooling to more exotic forms of liquid cooling, or even  immersion, to get to higher power densities.

在太空里，把每平方毫米瓦数做高非常困难，而在地球上这些都是已解决的问题。这类改进能让你拿到多得多的 token——每片制造出来的晶圆多出差不多 20% 的 token，这是个巨大的胜利。（Dwarkesh：你说的平方毫米，是指 die 面积？）对，die 面积。（Dwarkesh：这样反而对太空更有利吧，因为每平方毫米瓦数越高意味着芯片跑得越热。这大概是个芯片工程问题，但按 Stefan-Boltzmann 定律，辐射散热和温度的四次方成正比。如果你能让芯片跑得很热，就能——）不，你没法让它跑得更热，你只能把它做得更密。问题在于，要把热量从那么密的区域里导出来，你就得放弃标准的风冷和液冷，改用更激进的液冷形式甚至浸没式冷却，才能达到更高的功率密度。

[2:06:05] Dylan

That's more difficult in  space than it is on Earth. Maybe it's worth explaining at this point what  exactly a scale-up is and what it looks like for Nvidia versus Trainium versus TPUs. Earlier I was mentioning how communication within a chip is super fast. Communication within chips that are in the same rack is fast, but not as fast. It's on the order of terabytes. Communication very far away is on  the order of hundreds of gigabytes. As you get further distance, maybe  across the country, the order of magnitude is on the order of gigabytes. A scale-up domain is this tight domain where the chips are communicating  on the order of terabytes a second.

这件事在太空比在地球上难得多。（Dwarkesh：也许这里值得解释一下，scale-up 到底是什么，Nvidia、Trainium、TPU 各自的 scale-up 又长什么样。）我前面提到过，芯片内部的通信是超级快的。同一个机架里芯片之间的通信也很快，但没那么快，量级在每秒太字节（TB）级。距离很远的通信量级是每秒几百吉字节（GB）。距离再拉远，比如跨越整个国家，量级就掉到每秒吉字节级。所谓 scale-up 域，就是芯片之间以每秒太字节量级互相通信的那个紧密域。

[2:06:50] Dylan

For Nvidia, previously this meant  an H100 server had eight GPUs, and those eight GPUs could talk to  each other at terabytes a second. With Blackwell NVL72, they  implemented rack-scale scale-up. That meant all seventy-two GPUs in the rack could  connect to each other at terabytes a second. The speed doubled generation on generation, but  the most important innovation was going from eight to seventy-two in the domain. When we look at Google, their scale-up domain is completely different. It has always been on the order of thousands. With TPU v4, they had pods the  size of four thousand chips.

对 Nvidia 来说，以前这意味着一台 H100 服务器里有八块 GPU，这八块 GPU 之间能以每秒太字节的速度互相通信。到了 Blackwell NVL72，他们实现了机架级的 scale-up：整个机架里全部七十二块 GPU 都能以每秒太字节的速度互联。速度是一代翻一倍，但最重要的创新是把这个域从八块扩到了七十二块。再看 Google，他们的 scale-up 域完全是另一回事，一直都是数千的量级。TPU v4 时代，他们的 pod 就有四千颗芯片的规模。

[2:07:23] Dylan

With v8 or v7, they have pods in  the eight or nine thousand range. What's relevant here is that it's not the  same as Nvidia. It's not like for like. Google has a topology that's a torus. Every chip connects to six neighbors. Nvidia's 72 GPUs connect all-to-all.  They can send terabytes a second to any arbitrary other chip in that pod of scale-up. Whereas Google, you have to bounce through chips. If TPU 1 needs to talk to TPU 76, it has to bounce  through various chips, and there is always some blocking of resources when you do that because  that one TPU is only connected to six other TPUs. So there is a difference  in topology and bandwidth, and there are trade-offs and advantages to both. Google gets to have a massive scale-up domain, but they have the trade-off of bouncing  across chips to get from one to another.

到了 v8 或者 v7，他们的 pod 在八九千颗的范围。这里的关键是，这跟 Nvidia 的不是一回事，不能直接对等比较。Google 用的是环面（torus）拓扑：每颗芯片连接六个邻居。而 Nvidia 的 72 块 GPU 是全互联（all-to-all）的，在那个 scale-up pod 里可以向任意一块芯片以每秒太字节的速度发数据。Google 那边则得在芯片之间跳转：如果 TPU 1 要和 TPU 76 通信，就得经过一串芯片中转，而这么做时总会有资源阻塞，因为那颗 TPU 只连着另外六颗 TPU。所以两者在拓扑和带宽上有差异，各有取舍和优势。Google 拿到了巨大的 scale-up 域，代价是芯片之间要靠逐跳中转。

[2:08:15] Dylan

You can only talk to six direct neighbors. Amazon has mutated their scale-up domain. They're somewhere in between Nvidia and Google. They're trying to make larger scale-up domains. They try to do all-to-all to some extent with  switches, which is what Nvidia does, but they also use torus topologies like Google to some extent. As we advance forward to next generations, all three of them are moving more  towards a dragonfly topology. That means there are some fully connected elements  and some elements that are not fully connected. You can get the scale-up to be hundreds or  thousands of chips, but also have it not contend for resources when bouncing through chips. Related question: I heard somebody make the claim that the reason parameter scaling has been  slow—and only now are we getting bigger models from OpenAI and Anthropic—is that… The original  GPT-4 is over a trillion parameters, and only now are models starting to approach that again. I heard a theory that the reason is that Nvidia's scale-ups have just not  had that much memory capacity.

你只能直接和六个邻居通信。Amazon 则把自己的 scale-up 域做了变种，介于 Nvidia 和 Google 之间。他们在努力做更大的 scale-up 域：一定程度上用交换机做全互联（这是 Nvidia 的路子），但也在一定程度上像 Google 那样用环面拓扑。往后几代演进，三家都在往蜻蜓（dragonfly）拓扑靠：其中一部分是全连接的，一部分不是。这样 scale-up 可以做到几百甚至几千颗芯片，同时在逐跳中转时又不至于争抢资源。（Dwarkesh：一个相关的问题。我听人说过一个说法：参数规模扩展之所以一直很慢——直到现在 OpenAI 和 Anthropic 才又开始出更大的模型——原因是……最初的 GPT-4 参数量超过一万亿，直到现在各家模型才重新逼近那个规模。我听到的一种理论是，原因在于 Nvidia 的 scale-up 域内存容量一直不够大。）

[2:09:37] Dwarkesh

Let's say you have a 5T model running at  FP8, so that's five trillion gigabytes. And then you have the KV cache, let's say it's— Just call it the same size. Okay, let's say it's the same size for one batch. So you need ten terabytes to be able to run… A single forward pass, yeah. And then only with the GB200 and NVL72 do you have an Nvidia scale-up that has twenty  terabytes, and before that they were much smaller. Whereas Google, on the other hand, has had  these huge TPU pods that are not all-to-all, but still have hundreds of terabytes  of capacity in a single scale-up. Does that explain why parameter  scaling has been slow?

假设你有一个 5T 参数的模型跑在 FP8 下，那就是 5 万亿字节（5 TB）。然后还有 KV cache，假设它——（Dylan：就当它一样大。）好，假设单个 batch 的 KV cache 一样大。那你就需要 10 TB 才能跑……（Dylan：跑一次前向传播，对。）而直到 GB200 和 NVL72，Nvidia 的 scale-up 才有了 20 TB，在那之前都小得多。反观 Google，一直有那些巨大的 TPU pod——虽然不是全互联，但单个 scale-up 里就有几百 TB 的容量。这能解释为什么参数规模扩展一直很慢吗？

[2:10:16] Dylan

I think it's partially the capacity and  bandwidth, but also as you build a larger model, the ability to deploy it is slower. In terms of what the inference speed is for the end user, that's kind of irrelevant. What's  really relevant is RL. What we've seen with these models and allocation of compute at a lab… There  are a few main ways you can allocate compute. You can allocate it to inference, i.e. revenue. You can allocate it to development, i.e. making the next model. You can allocate it to research. In development specifically, you  split it between pre-training and RL. When you think about what is happening, the  compute efficiency gains you get from research are so large that you actually want most of your  compute to go to research, not to development.

我觉得容量和带宽是部分原因，但还有一点：模型做得越大，部署它的速度就越慢。对终端用户的推理速度来说，这其实无关紧要；真正要紧的是 RL。从这些模型以及实验室里算力分配的情况看……算力主要有几种分法：可以分给推理，也就是收入；可以分给开发，也就是做下一代模型；可以分给研究。在开发内部，又分成预训练和 RL。你仔细想想就会发现，研究带来的算力效率增益太大了，以至于你其实希望把大部分算力投给研究，而不是开发。

[2:11:04] Dylan

All these researchers are generating new  ideas, trying them out, testing them, and continuing to push the Pareto optimal  curve of scaling laws further and further. Empirically, what we’ve seen is that  model costs get ten times cheaper every year, or even more than that. At the same scale it gets ten times cheaper, and to reach new frontiers it  costs the same amount or more. So you don't want to allocate too  many resources to pre-training and RL. You actually want to allocate most  of your resources to research. In the middle is this development period. If you pre-train a five-trillion-parameter model, how many rollouts do you have to do in RL? Rollouts for a five-trillion-parameter model are five times larger than for  a one-trillion-parameter model.

所有这些研究员都在产生新想法、动手尝试、做测试，不断把 scaling law 的帕累托最优曲线往前推。经验上我们看到的是：模型成本每年降到十分之一，甚至更低。同等规模的模型一年后便宜十倍，而要触及新的前沿，花费则持平或更高。所以你不想往预训练和 RL 上砸太多资源，你其实想把大部分资源投给研究。夹在中间的就是这个开发阶段。如果你预训练了一个五万亿参数的模型，RL 阶段要跑多少次 rollout？五万亿参数模型的 rollout，开销是一万亿参数模型的五倍。

[2:11:54] Dylan

If you wanted to do as many rollouts—maybe  the larger model is two times more sample efficient—now you need 2.5x as much  time of RL to get the model smarter. Or you could RL the smaller model for 2x the time. You'd still have a 25% difference in the big model, which is 2x as sample efficient  and doing X number of rollouts. But the smaller model, which is a  trillion parameters, although its less sample efficient, is doing twice as  many rollouts and is still done faster. You get the model sooner, you've done more RL,  and then you can take that model to help you build the next models, help your engineers  train, and do all these research ideas.

如果你想跑同样多的 rollout——假设大模型的样本效率高两倍——那你现在需要 2.5 倍的 RL 时间才能把模型变得更聪明。或者你可以把小模型 RL 上两倍的时间。大模型样本效率高一倍、跑了 X 次 rollout，两相比较你还是差 25%。但那个一万亿参数的小模型，虽然样本效率低一些，却跑了两倍的 rollout，而且还完成得更快。你更早拿到模型、做了更多 RL，然后就能拿这个模型去帮你构建下一代模型、帮你的工程师训练、去跑所有那些研究想法。

[2:12:33] Dylan

This feedback loop is actually weighed  towards smaller models in every case, no matter what your hardware is. As you look to Google, they do deploy the largest production model of  any of the major labs with Gemini Pro. It's a larger model than GPT-5.4. It's a larger model than Opus. Google does this because they have a unipolar set  of compute. It's almost all TPU. Whereas Anthropic is dealing with H100s, H200s, Blackwell,  Trainiums, and TPUs of various generations. OpenAI is dealing with mostly Nvidia right now,  but going towards having AMD and Trainium as well. The fleets of compute like Google's can  just optimize around a larger model.

这个反馈循环其实在任何情况下都偏向小模型，不管你的硬件是什么。再看 Google，他们确实部署着各大实验室里最大的生产模型——Gemini Pro。它比 GPT-5.4 大，也比 Opus 大。Google 能这么做，是因为他们的算力是单一体系的，几乎全是 TPU。而 Anthropic 要同时应付 H100、H200、Blackwell、各代 Trainium 和各代 TPU。OpenAI 目前主要用 Nvidia，但正在走向同时用 AMD 和 Trainium。像 Google 这样的算力舰队，就可以干脆围绕一个更大的模型去优化。

[2:13:23] Dylan

They can leverage a thousand chips in a scale-up  domain to get the RL time speed much faster so that this feedback loop can be fast. But at the end of the day, in isolation, you almost always want to go with a smaller  model that gets RL'd faster and gets deployed into research and development earlier. You can build the next thing and get more efficiency wins. You have this compounding effect of making a smaller model that can be  deployed into research and development earlier. I spend less compute on the training because I  was able to allocate more compute to the research. This compounding effect of being able  to do research faster and faster is potentially a faster takeoff. That's all these companies want: the fastest takeoff possible. Okay, a spicy question. You've explained that SemiAnalysis sells these spreadsheets. You're always pointing out how six months or a year ago, you warned  people about the memory crunch.

他们可以在一个 scale-up 域里调用上千颗芯片，把 RL 的时间大幅压缩，让这个反馈循环转得飞快。但归根结底，单独来看，你几乎总是想选一个更小的模型：RL 更快、更早部署进研究和开发。你能更快做出下一个东西、拿到更多效率增益。这里有个复利效应——做一个能更早投入研发的小模型。训练上花的算力少了，因为我把更多算力分给了研究。这种让研究越转越快的复利效应，可能意味着更快的起飞（takeoff）。这正是所有这些公司想要的：尽可能快的起飞。（Dwarkesh：好，来个辣一点的问题。你说过 SemiAnalysis 卖的是那些电子表格。你总在强调，半年前、一年前你就提醒过大家内存紧缺。）

[2:14:17] Dwarkesh

Now you're telling people about the cleanroom  crunch, and in the future, the tool crunch. Why is Leopold the only person using your  spreadsheets to make outrageous money? What is everybody else doing? I think there are a lot of people making money in many ways. Leopold jokes that he's the only client of mine who tells me our numbers are too low. Everyone else tells me our numbers are too high, almost ad nauseam. Whether it's a hyperscaler saying, "Hey, that other hyperscaler, their numbers are  too high," and we're like, "Nah, that's it." They're like, "No, no, no, it's  impossible," blah, blah, blah.

现在你又在跟大家讲洁净室（cleanroom）紧缺，将来还有设备紧缺。为什么只有 Leopold 一个人在用你的表格赚离谱的大钱？其他人都在干嘛？（Dylan：我觉得很多人在用很多方式赚钱。Leopold 开玩笑说，他是我唯一一个嫌我们数字太低的客户。其他所有人都嫌我们数字太高，说到你耳朵起茧。比如某家 hyperscaler 说：「嘿，另外那家 hyperscaler 的数字太高了吧。」我们说：「不，就是那么多。」他们就「不不不，不可能」，等等等等。）

[2:14:52] Dylan

You finally have to convince them through all  these facts and data when we're working with hyperscalers or AI labs that in fact, no,  that number isn't too high, that's correct. Eventually, sometimes it takes them  six months to realize, or a year later. Other clients, on the trading  side, also use our data. Roughly 60% of my business is industry. So AI labs, data center companies, hyperscalers, semiconductor companies, the  whole supply chain across AI infrastructure. But 40% of our revenue is hedge funds. I'm not going to comment on who our customers are, but a lot of people use the data. It's just how do you interpret it, and then what do you view as beyond it? I will say Leopold is pretty much the only person who tells me my numbers are too low, always. Sometimes he's too high, sometimes I'm too low.

最后当我们跟 hyperscaler 或 AI 实验室合作时，你得用所有这些事实和数据说服他们：不，那个数字不算高，就是对的。有时候他们要过六个月，甚至一年之后才反应过来。另一些客户，交易那边的，也在用我们的数据。我的业务大约 60% 来自产业界：AI 实验室、数据中心公司、hyperscaler、半导体公司，横跨 AI 基础设施的整条供应链。但 40% 的收入来自对冲基金。我不会评论客户具体是谁，但用这些数据的人很多。关键在于你怎么解读它，以及你在数据之外还看到了什么。我要说的是，Leopold 差不多是唯一一个永远嫌我数字太低的人。有时是他太高了，有时是我太低了。

[2:15:44] Dylan

But in general, I think  other people are doing that. You can look across the space at hedge funds and  look at their 13Fs and see they own, maybe not exactly what Leopold does, because it's always a  question of what is the most constrained thing. What's the thing that's going to  be most outside of expectations? That's what you're really trying to  exploit: inefficiencies in the market. In a sense, our data is making the market  more efficient by making the base data of what's happening more accurate. Many funds do trade on information that is out there… I don't  think Leopold's the only person.

但总体上，我觉得别人也在这么做。你可以扫一遍这个领域的对冲基金，看他们的 13F 持仓，会发现他们持有的——也许不完全是 Leopold 那一套，因为问题永远在于：什么才是最受约束的环节？什么东西会最超出市场预期？你真正要利用的就是这个：市场里的低效。从某种意义上说，我们的数据让「正在发生什么」的基础数据更准确，从而让市场变得更有效。很多基金确实在用公开的信息做交易……我不认为 Leopold 是唯一一个。

[2:16:26]

I think he has the most conviction  about the AGI takeoff, though. Right, but the bets are not  about what happens in 2035. The bets that you're making—that are at least  exemplified by public returns we can see for different funds including Leopold's—are  about what has happened in the last year. The last year stuff could be  predicted using your spreadsheets. It's about buying the next year's spreadsheets. They're not just spreadsheets. There are reports. There's API access to  the data. There's a lot of data. But do you see what I mean? It's not about some crazy singularity thing. It's about, do you buy the memory crunch? You only buy the memory crunch if you believe AI is going to take off in a huge way. The memory crunch, a lot of it was predicated on… At least for people in the Bay Area who  think about infrastructure, it's obvious.

不过我觉得他是对 AGI 起飞最有信念的人。对，但这些押注赌的不是 2035 年会发生什么。你们下的注——至少从各家基金（包括 Leopold 的）公开回报能看到的那些——赌的是过去一年已经发生的事。过去一年的行情，用你们的电子表格就能预测出来。关键是要买到明年的电子表格。那不只是电子表格，还有研究报告、数据的 API 接口，有大量数据。但你明白我的意思吧？这跟什么疯狂的奇点没关系。问题是：你买不买内存紧缺这个逻辑？只有当你相信 AI 会大爆发，你才会去押内存紧缺。内存紧缺这件事，很大程度上是建立在……至少对湾区那些整天琢磨基础设施的人来说，这是显而易见的。

[2:17:17] Dylan

KV cache explodes as context lengths get longer,  so you need more memory. Then you do the math. You also have to have a lot of supply chain  understanding of what fabs are being built, what data centers are being built,  how many chips, and all these things. We track all these different datasets  very tightly, but at the end of the day, it takes someone to fully believe  that this is going to happen. A year ago, if you told someone memory  prices would quadruple and smartphone volumes are going to go down 40% over the  year or two after that, people were like, "You're crazy. That'd never happen." Except a few  people do believe that, and those people did trade memory. And people did. I don't think Leopold  was the only person buying memory companies.

上下文越长，KV cache 就爆炸式膨胀，所以你需要更多内存。然后你去算这笔账。你还得对供应链有很深的理解：哪些 fab 在建、哪些数据中心在建、芯片有多少，诸如此类。我们把所有这些数据集都盯得很紧，但说到底，还是得有人真正相信这件事会发生。一年前，你要是跟人说内存价格会翻两番、智能手机出货量会在之后一两年里跌 40%，人家会说：「你疯了，绝不可能。」但确实有少数人信了，而这些人真的去交易了内存。也确实有人做了。我不认为 Leopold 是唯一在买内存公司的人。

[2:18:00]

He, of course, sized and positioned and did  things in better ways than some, maybe most. I don't want to comment on whose returns  are what, but he certainly did well. Other people also did really well. Wow, you've made me diplomatic for the first time ever. No, no, you're fine.  I think this is hilarious. I'm being a diplomat, whereas usually I'm spicy. Okay, some rapidfire to close out. If you're saying with the memory,  logic, et cetera, the N3 is mostly going to be AI accelerators, but then there's  N2, which is mostly Apple now… In the future, I guess AI would also want to go on N2. Can TSMC kick out Apple if Nvidia and Amazon and Google say, "Hey, we're willing  to pay a lot of money for N2 capacity?"

当然，他在仓位大小和布局上比一些人——也许是大多数人——做得更好。我不想评论谁的回报是多少，但他确实干得不错。其他人也赚得很好。哇，你让我头一次这么外交辞令。不不，你没问题，我觉得这太好笑了。我在当外交官，而我平时可是很毒舌的。好，来几个快问快答收尾。你说内存、逻辑芯片这些，N3 大部分会变成 AI 加速器，但还有 N2，现在主要是 Apple 的……未来我想 AI 应该也想上 N2。如果 NVIDIA、Amazon、Google 说「嘿，我们愿意为 N2 产能砸大钱」，TSMC 能把 Apple 踢出去吗？

[2:18:59] Dylan

I think the challenge with this is chip design  timelines take a long while, so that's more than a year out, and the designs that are  on two nanometer are more than a year out. What would really happen is Nvidia and  all these others will be like, "Hey, we're going to prepay for the capacity  and you're going to expand it for us." Maybe TSMC takes a little  bit of margin, but not a ton. They're not going to kick Apple out entirely. What they're going to do is when Apple orders X, they might say, "Hey, we project you only need  X minus one, and so that's what we're going to give you, X minus one." Then that flex capacity, Apple's kind of screwed on. Traditionally, Apple has always over-ordered by 10% and cut back  by 10% over the course of the year.

我觉得这里的难点在于，芯片设计周期很长，那是一年开外的事，而 2 纳米上的那些设计也都在一年以后。真正会发生的是，NVIDIA 和其他这些公司会说：「嘿，我们预付产能的钱，你替我们扩产。」TSMC 也许多拿一点利润，但不会太多。他们不会把 Apple 整个踢出去。他们会做的是：Apple 下单 X 的时候，他们可能说「嘿，我们预测你只需要 X 减一，所以我们就给你 X 减一」。那部分弹性产能上，Apple 就有点吃亏了。传统上，Apple 一直是先多订 10%，再在一年当中砍掉 10%。

[2:19:38]

Some years they hit the entire 10%. Volumes vary based on the season and macro. I don't think TSMC would kick out Apple. I think Apple will become a smaller and smaller percentage of TSMC's revenue, and therefore be  less relevant for TSMC to cater to their demands. TSMC could eventually start saying, "Hey, you've  got to pre-book your capacity for next year, for two years out, and you have to prepay for  the CapEx," because that's what Nvidia and Amazon and Google are doing. I wonder if it's worth going into specific numbers. I don't have any of them on hand. What percentage of N2 does Apple have its  hands on over the coming years versus AI?

有些年份他们把那 10% 全用上了。出货量随季节和宏观环境波动。我不认为 TSMC 会把 Apple 踢出去。我认为 Apple 在 TSMC 营收里的占比会越来越小，因此 TSMC 也就越来越没必要迁就他们的要求。TSMC 最终可能会开始说：「嘿，你得提前一年、两年预订产能，还得为 CapEx 预付款」——因为 NVIDIA、Amazon、Google 就是这么干的。不知道值不值得聊几个具体数字。我手头没有现成的。未来几年 N2 产能里 Apple 握着多少，AI 又占多少？

[2:20:22] Dylan

This year Apple has the majority of  N2 that's going to get fabricated. There's a little bit from AMD. They are trying to make some AI chips and CPU chips early. There's a little bit, but for the most part, it's Apple. As we go forward to the year after that, Apple still gets closer to half of it as other people  start ramping, but then it falls drastically, just like for N3, where they were half. When I say N2, that includes A16, which is a variant of N2. Over time, those nodes will be the majority. What's also interesting is traditionally,  Apple has been the first to a process node. 2 nm is actually the first time they're  not. Well, that’s besides Huawei. Huawei, back in 2020 and before, was the first with  Apple, but they were both making smartphones.

今年要投产的 N2，大部分在 Apple 手里。AMD 有一小部分，他们想提早做一些 AI 芯片和 CPU 芯片。有一点点，但大体上都是 Apple。到下一年，随着其他家开始放量，Apple 还能占到接近一半，但之后就急剧下滑——就像 N3 一样，他们当年也是占一半。我说的 N2 包括 A16，那是 N2 的一个变体。随着时间推移，这些节点会占到大头。还有个有意思的点：传统上 Apple 一直是第一个用上新制程节点的。2 纳米其实是他们头一次不是第一。呃，华为除外。2020 年及更早，华为和 Apple 是并列第一的，但他们俩当时做的都是智能手机。

[2:21:08] Dylan

Now, with 2 nm, you've got AMD trying  to make a CPU and a GPU chiplet that they use advanced packaging to package  together, in the same timeframe as Apple. This is a big risk for AMD that causes  potential delays because it's a brand-new process technology. It's hard. But at the end  of the day, this is a bet that they want to do to scale faster than Nvidia and try and beat them. As we move forward, when we move to the A16 node, the first customer there is not even  Apple. It's AI. As we move forward, that will become more and more prevalent. Not only will Apple not be the first to a node, they will also not be the majority  of the volume to the new node.

现在到了 2 纳米，AMD 在尝试做一颗 CPU 和一颗 GPU chiplet，用先进封装把它们封在一起，时间点和 Apple 同期。这对 AMD 是很大的风险，可能导致延期，因为这是全新的制程技术，很难。但说到底，这是他们想下的一个赌注：比 NVIDIA 扩张得更快，试着击败他们。再往后，到 A16 节点的时候，第一个客户甚至都不是 Apple 了，是 AI。往后走，这种情况会越来越普遍。Apple 不仅不再是第一个用上新节点的，也不再是新节点上的出货大头了。

[2:21:49]

They'll then just be like any old customer. Because the scale of TSMC's CapEx keeps ballooning, but Apple's business  is not growing at the same pace, they become a less and less relevant customer. They also will just cut their orders because things in the supply chain are  kicking them out, whether it be packaging or materials or DRAM or NAND. These things are increasing in cost. They can't pass on all the cost to customers  likely because the consumer is not that strong. You end up with this conundrum  where they are just not TSMC's best bud like they have been historically. Do you think if Huawei had access to 3 nm, they would have a better accelerator than Rubin? Potentially, yeah. Huawei was the first with a 7 nm AI chip as well. They were the first with a 5 nm mobile chip, but they were the first with a 7 nm AI chip. The Huawei Ascend was two months before the TPU and four months before Nvidia's A100, I think. That's just moving to a process node.

到那时他们就跟普通老客户没两样了。因为 TSMC 的 CapEx 规模在不断膨胀，而 Apple 的生意没有以同样的速度增长，他们就变成一个越来越不重要的客户。而且他们还会主动砍单，因为供应链里的种种东西在把他们挤出去——不管是封装、材料，还是 DRAM、NAND。这些东西的成本都在涨。他们大概率没法把所有成本都转嫁给消费者，因为消费端并不那么强劲。最后你就落到这么一个困局：他们不再是 TSMC 历来那种铁哥们了。你觉得如果华为能用上 3 纳米，他们能做出比 Rubin 更强的加速器吗？有可能，会。华为当年也是第一个做出 7 纳米 AI 芯片的。他们是第一个做 5 纳米手机芯片的，但也是第一个做 7 纳米 AI 芯片的。华为 Ascend 比 TPU 早两个月，比 NVIDIA 的 A100 早四个月，我记得是。不过那只是制程节点上的领先。

[2:22:49] Dylan

That doesn't imply software or hardware  design or all these other things. But Huawei is arguably the only company in the  world that has all the legs. Huawei has cracked software engineers. Huawei has cracked  networking technologies. That's, in fact, their biggest business historically. They have  cracked AI talent. Furthermore, beyond Nvidia, they actually have better AI researchers. Beyond Nvidia, they have their own fabs. And beyond Nvidia, they have their own end  market of selling tokens and things like that. Huawei is able to get the top, top talent. Nvidia is as well, but not with as much concentration, and Huawei  has a bigger pool in China.

这并不代表软件、硬件设计以及其他所有这些东西也领先。但华为可以说是全世界唯一一家几条腿都齐的公司。华为有顶尖的软件工程师；华为吃透了网络技术——事实上那历来就是他们最大的业务；他们也攒下了顶尖的 AI 人才。而且，除 NVIDIA 之外，他们其实有更强的 AI 研究员。比 NVIDIA 多出来的是：他们有自己的 fab。还比 NVIDIA 多出来的是：他们有自己的终端市场，卖 token 之类的。华为能拿到最顶尖的人才。NVIDIA 也能，但集中度没那么高，而且华为在中国有一个更大的人才池。

[2:23:33] Dylan

It's very arguable that Huawei, if they  had TSMC, would be better than Nvidia. There are areas where China has advantages  in areas that Nvidia can't access as easily. Not just scale, but certain optical  technologies China's actually really good at. I think it's very reasonable that if in  2019 Huawei was not banned from using TSMC, Huawei would have already eclipsed  Apple as the biggest TSMC customer. Huawei has huge share in networking,  compute, CPUs, and all these things. They would have kept gaining share, and  they'd likely be TSMC's biggest customer. Wow. That's crazy. I've got a  random final question for you.

很有得争的一点是：如果华为能用 TSMC，他们可能会比 NVIDIA 更强。有些领域中国有优势，而 NVIDIA 不那么容易触及——不只是规模，还有某些光学技术，中国其实非常强。我觉得很合理的一个推演是：如果 2019 年华为没被禁止使用 TSMC，华为可能早就超过 Apple，成为 TSMC 最大的客户了。华为在网络、计算、CPU 这些领域份额巨大。他们会继续抢份额，很可能已经是 TSMC 最大的客户。哇，太疯狂了。我还有个随性的收尾问题问你。

[2:24:18]

The other part of the Elon interview was robots. If humanoids take off faster than people expect, if by 2030 there's millions of humanoids  running around which each need local compute, any thoughts on what that implies? What would be required for that? There's a lot of difficulties with the VLMs  and VLAs that people are deploying on robots. But to some extent, you don't need to  have all the intelligence in the robot. It would be much more efficient to not do that. Because in the cloud, you can batch process and all these things. What you may want to do is have a lot of the planning and longer-horizon tasks  determined by a much more capable model in the cloud that runs at very high batch sizes. Then it pushes those directions to the robots, who interpolate between each subsequent action. Or it is given a command like, "Hey, pick up that cup," and then the model  on the robot can pick up the cup.

Elon 那期访谈的另一部分聊的是机器人。如果人形机器人起飞得比大家预期快，如果到 2030 年满世界跑着几百万台人形机器人，每台都需要本地算力，你觉得这意味着什么？那需要什么条件？现在人们部署在机器人上的 VLM 和 VLA 有很多难题。但某种程度上，你不需要把所有智能都塞进机器人里。不那么做反而效率高得多，因为在云端你可以做批处理之类的。你可能想要的是：规划和更长时程的任务，由云端一个能力强得多、以极高 batch size 运行的模型来决定，然后它把这些指令推给机器人，机器人在每个后续动作之间做插值。或者它收到一条指令，比如「嘿，把那个杯子拿起来」，然后机器人上的模型就能把杯子拿起来。

[2:25:17] Dylan

As it's picking up, things like weight and  force may have to be determined by the model on the robot, but not everything needs to be. It can say, "hey that’s a headphone" and the super model in the cloud can say, "I  know these headphones are Sony XM6s," which is not a Dwarkesh ad spot, but... I’m like, why is this guy's plugging this thing so hard. It's on the table. It's on his  neck when we're interviewing Satya together. Is he getting paid by Sony? Unfortunately not. But anyways, it might say, "Hey, the headband is soft, and  this is the weight of it," and all these things. Then the model on the robot  can be less intelligent, take these inputs, and do the actions. It may get told by the model in the cloud every second, or maybe ten times a second,  depending on the hertz of the action.

在拿的过程中，重量、力度这类东西可能得由机器人上的模型来判断，但不是所有事情都需要。它可以说「嘿，那是个耳机」，然后云端的超级模型说「我知道这是 Sony XM6」——这可不是 Dwarkesh 的广告位啊，不过……我就在想，这哥们怎么把这东西吹得这么起劲。它就摆在桌上；我们一起采访 Satya 的时候它还挂在他脖子上。他是收了 Sony 的钱吗？很遗憾没有。总之，它可能会说「嘿，头梁是软的，重量是多少」等等这些信息。然后机器人上的模型可以笨一点，接收这些输入去执行动作。云端模型可能每秒给它下一次指令，或者每秒十次，取决于动作的频率（赫兹）。

[2:26:09] Dylan

But a lot of that can be offloaded to the cloud. Otherwise, if you do all of the processing on the device, I believe it would be more  expensive because you can't batch. Two, you couldn't have as much intelligence  as you do in the cloud because the models will just be bigger in the cloud. Three, we're in a semiconductor shortage world, and any robot you deploy needs leading-edge  chips because the power is really bad for robots. You need it to be low power and efficient,  and all of a sudden you're taking power and chips that would've been for AI data  centers, and you're putting them in robots.

但其中很大一部分都可以卸载到云端。否则，如果所有处理都在设备上做，我认为会更贵，因为你没法做批处理。第二，你的智能水平赶不上云端，因为云端的模型就是会更大。第三，我们正处在一个半导体短缺的世界，你部署的任何机器人都需要最先进的芯片，因为机器人的功耗约束非常苛刻。你需要它低功耗、高能效，于是突然之间，你就把本来该给 AI 数据中心的电力和芯片，拿去装进了机器人。

[2:26:39]

So now that 200 gigawatts gets lower if  you're deploying millions of humanoids. I think this is very interesting because  something people might not appreciate about the future is how centralized, in  a physical sense, intelligence will be. Right now, there are eight billion humans, and  their compute is in their heads, on their person. In the future, even with robots that are  out physically in the world—obviously, knowledge work will be done in a centralized  way from data centers with hundreds of thousands or maybe millions of instances—the future  you're suggesting is one where there's more centralized thinking and centralized computation  driving millions of robots out in the world.

所以如果你部署几百万台人形机器人，那 200 吉瓦就得往下减。我觉得这非常有意思，因为关于未来，人们可能没意识到的一点是：智能在物理意义上会有多么中心化。现在地球上有 80 亿人，他们的算力长在自己脑袋里、带在自己身上。而在未来，即便机器人实实在在走在物理世界里——显然，知识工作会以中心化的方式在数据中心完成，由几十万甚至上百万个实例来做——你描绘的未来是：更中心化的思考、更中心化的计算，驱动着散布在世界各地的几百万台机器人。

[2:27:25]

That's an interesting fact about the future  that I think people might not appreciate. I think Elon recognizes this, which is why  he's going to different places for his chips. He signed this massive deal with Samsung to make  his robot chips in Texas because I personally think he thinks Taiwan risk is huge. Because of that and the centralization of resources in Taiwan, having his robot  chips in Texas means having a separate supply chain that is not as constrained. No one's really making AI chips on Samsung besides Nvidia's new LPU that they launched. They’re launching it next week, but we're recording this the week before. This episode's coming out Friday.

这是关于未来的一个很有意思的事实，我觉得人们可能没意识到。我认为 Elon 意识到了这一点，所以他的芯片才会另找门路。他跟 Samsung 签了那笔大单，在得州生产他的机器人芯片，因为我个人认为，他觉得台湾风险巨大。正因为这个，加上资源都集中在台湾，把机器人芯片放在得州，就意味着有一条不那么受限的独立供应链。除了 NVIDIA 刚推出的新 LPU，几乎没人在 Samsung 上做 AI 芯片。他们下周就发布，不过我们是提前一周录的。这期节目周五上线。

[2:28:04]

Oh, this episode's coming out before.  Sick. They're launching this new AI chip next week which is built on Samsung, but  that's a recent development from Nvidia. That's the only other AI demand there,  whereas on TSMC, everything is competing. He gets both geopolitical diversification  and supply chain diversity for his robots, and he's not competing as much with the infinite  willingness to pay for the data center geniuses. Final question, on Taiwan. If we believe  that tools are the ultimate bottleneck, how much of Taiwan's place in the AI semiconductor  supply chain could we de-risk simply by having a plan to airlift every single process engineer  at TSMC out if they get blockaded or something?

哦，这期会赶在那之前上线，太好了。他们下周要发布这颗基于 Samsung 制造的新 AI 芯片，不过那是 NVIDIA 最近才有的动作。Samsung 那边的 AI 需求就只有这一家，而在 TSMC 上，所有东西都在抢产能。所以他既拿到了地缘政治上的分散，也为他的机器人拿到了供应链的多元化，而且不用跟那些为数据中心天才们愿意无限掏钱的人抢得那么凶。最后一个问题，关于台湾。如果我们相信设备才是终极瓶颈，那只要有一个预案——一旦台湾被封锁之类的，就把 TSMC 的每一个制程工程师都空运出来——能在多大程度上给台湾在 AI 半导体供应链里的位置去风险？

[2:28:56]

Or do you still need to ship out the EUV  tools, which would be multiple plane loads per single tool and would not be practical? If you ship out all the process engineers and assuming it's hot enough that you destroy the  fabs, no one has all the fabs in Taiwan now, which is a big risk. These tools actually use a lot of semiconductors which are manufactured in Taiwan. It's a snake eating its own tail meme because you can't make the tools without the chips from  Taiwan, which you can't use without the tools in Taiwan. There's obviously some diversification  there. They don't use super advanced chips in lithography tools, but at the end of the  day, there is some dragon eating its tail.

还是说你仍然得把 EUV 设备也运出来？那可是一台设备就要好几架飞机才装得下，根本不现实。如果你把所有制程工程师都撤出来，并且假设局势热到 fab 都被摧毁了，那现在台湾的那些 fab 就谁都没有了，这是个巨大的风险。这些设备本身其实要用到大量在台湾制造的半导体。这就是「蛇吞自己尾巴」的梗：没有台湾的芯片你造不出设备，而没有台湾的设备你又造不出芯片。当然这里面有一些多元化的余地——光刻设备里用的并不是特别先进的芯片——但归根结底，这就是一条咬着自己尾巴的龙。

[2:29:36] Dylan

Just shipping out all the engineers and  blowing up the fabs means China has a stronger semiconductor supply chain than the  rest of the world in terms of verticalization, now that you've removed Taiwan. You've got all the know-how, but you've got to replicate it in,  let's say, Arizona or wherever for TSMC. It's going to take a long time to build all the  capacity that TSMC has built over the years. And so you've drastically  slowed US and global GDP. Not just growth, you've shrunk the GDP  massively, and you've got a lot bigger problems. Your incremental ability to add  compute goes to almost zero.

就算把工程师全都撤出来、把晶圆厂炸掉，也意味着在垂直整合程度上，中国的半导体供应链会比世界其他地区更强——因为台湾已经不在牌桌上了。你手里有全部的技术知识（know-how），但你得在比如亚利桑那或别的什么地方，把 TSMC 的能力复制出来。要把 TSMC 这么多年建起来的产能全部重建，需要非常长的时间。这样一来，你就大幅拖慢了美国乃至全球的 GDP。不只是增长放缓，而是 GDP 被大幅缩水，你会面临一大堆更严重的问题。你新增算力的能力几乎归零。

[2:30:16]

Instead of hundreds of gigawatts  a year by the end of the decade, let's say something happens to Taiwan, now you're  at maybe 10 gigawatts across Intel and Samsung, or 20 gigawatts. It's nothing. Now all of a sudden  you've really caused some crazy dynamics in AI. Of course, you have all the existing capacity,  but that existing capacity pales in comparison to the capacity that's being expanded. Okay. Dylan, that was excellent. Thank you so much for coming on the podcast. Thank you for having me. And see you tonight.

本来到这个十年末每年能新增几百吉瓦，假设台湾出了什么事，现在 Intel 加 Samsung 加起来可能只有 10 吉瓦，或者 20 吉瓦。这根本不算什么。突然之间，你就在 AI 领域引发了一些非常疯狂的态势。当然，现有的产能还在，但和正在扩建的产能相比，现有的这点产能不值一提。好的。Dylan，聊得太精彩了。非常感谢你来上播客。谢谢你邀请我。今晚见。