Cerebras CEO 谈 950 亿美元 IPO、晶圆级芯片与 AI 硬件

欢迎收听 ASI Pill 特别节目，Cerebras CEO Andrew Feldman 刚刚完成 950 亿美元 IPO，与 Alex Wister Gross 对谈。餐盘大小的晶圆级引擎、金字塔般的晶圆厂、轨道上的芯片，以及英特尔为何错过手机？我们开始吧。系好安全带。Cerebras CEO 刚刚为 950 亿美元 IPO 敲钟。感觉如何？

Welcome to the ASI Pill, special drop Cerebras CEO Andrew Feldman fresh off the $95 billion IPO in conversation with Alex Wister Gross. Wafer scale engines the size of a dinner plate, fabs as pyramids, chips in orbit, and why Intel missed the cell phone? Let's go. Strap in. The Cerebras CEO just rang the bell for a $95 billion IPO. How did it feel?

感觉很好。当然，这也是努力换来的。对团队来说超级激动。我们带了一部分员工和他们的家人一起分享。我父母、妻子和继女都来了，我们把它变成了一个家庭活动，真的很特别。

It felt good. Well, and you know, you had to work for it, too. Super exciting for the team. We were able to bring a portion of the organization and their families and share it. I mean, my parents were there, and my wife, and my stepdaughter, and we made it into a family event, and it was really something special.

Andrew 怎么看 Karpathy 加入 Anthropic 这个前沿实验室引力井？

And what does Andrew make of Karpathy joining Anthropic in the frontier lab gravity well?

他是这个领域最重要、最多产的思想家之一，对吧？不仅体现在他构建的东西上，也体现在他如何教导社区。我觉得这很有意思。他认为只有少数前沿实验室足够领先，如果你不在其中，就不在前沿——这个观点可能也适用于硬件。对吧？如果你不从根本上与三大实验室——Google、Anthropic、OpenAI——合作构建硬件，你就看不到他们在想什么。就像模型制造者的想法会漂移一样，你的硬件也会偏离他们的需求。我把这个观点应用到了我们的领域。他在构建有用东西方面的记录非常出色。非常出色。

Look, he is sort of one of the most important and prolific thinkers in the space, right? And not just as reflected in what he's built, but in how he's taught the community. And I think that's interesting. I think his point that there are a small number of frontier labs that are sufficiently far ahead that if you're not with them, you're not on the frontier, probably applies to hardware, too. Right? That if you're not building hardware engaged with at a fundamental level one of the three labs, the three most important labs, Google, Anthropic, OpenAI, you are not seeing what they're thinking. And just like your ideas will drift if you're a model maker, your hardware will drift from what they need as well. And I thought that was really sort of I applied that to our domain. And I think he has just an extraordinary track record for useful stuff he's built. Extraordinary.

Cerebras 创始人怎么看 Elon 和 Sam 的亿万富翁口水战？对 Cerebras 来说，这有关系吗？

And what does the Cerebras founder think of the Elon and Sam billionaire pissing match? For Cerebras, but was it relevant?

我认为这是一个巨大的干扰，亿万富翁的口水战我毫无兴趣。我认为他们是我们这一代最重要的两位思想家。Elon 所构建的东西令人叹为观止。我见过他，一起吃过饭。他是个博学家，才华横溢的思想家。Sam 所构建的，可以说是资本主义历史上增长最快的公司之一。但他的一些想法，比如发明了安全的 Y Combinator，对硅谷的结构意义重大。我认为他们争斗时，所有人都输了。我希望 Elon 建造酷东西，希望 Sam 建造酷东西。我不想浪费时间或阅读分歧。我只希望他们做自己最擅长的事——建造东西。

I think this was a giant distraction and billionaires and pissing matches interest me not at all. I think these are two of the most important thinkers of our generation. I think what Elon has built is breathtaking. I've met him, had dinner together. He's a polymath. He's a brilliant thinker. What Sam has built, sort of one of the fastest growing companies in the history of capitalism. But some of his ideas also in the invention of the safe Y Combinator, these were enormously meaningful in the structure of Silicon Valley. I think everybody loses when they battle. I want Elon building cool stuff. I want Sam building cool stuff. And I don't want to waste time or read about disagreements. I just want these guys doing what they're the best in the world at, which is building stuff.

在 ASI Pill，我们想知道起源故事。你们到底是怎么造出餐盘大小的芯片的？比任何已有芯片大 58 倍。

On the ASI pill, we want the origin story. How do you actually build a chip the size of a dinner plate? 58 times larger than any chip ever made.

我们做到了。创始团队之前都在我上一家创业公司共事。2012 年 AMD 收购了那家公司，到 2015 年我们有点分道扬镳，然后开始碰面，看到了 AI 的曙光。我们知道这种新工作负载会消耗巨大的算力。我们下了两个大赌注。第一个赌注是：就像图形催生了 GPU，移动计算支持了 ARM 处理器的发展一样，这项技术、这项工作会足够庞大，需要专用芯片。第二个赌注是：正确的策略不是构建 GPU 的衍生品，而是从一张白纸开始，做一些根本不同的事情。这些都是非常反共识的赌注，当时如此，后来证明完全正确。在此基础上，我们继续创新思考，认为 AI 需要的是内存带宽——数据从内存到计算的移动速度。在这个维度上创新的方法是使用与其他人不同的内存类型，对吧？我们有两种内存：一种存储量大但速度慢，我们称之为 DRAM 或 HBM；另一种速度快但每平方毫米存储量小。所以我们假设，如果能造出餐盘大小的芯片——比任何已有芯片大 58 倍——就能塞满 SRAM，从而克服其每平方毫米存储量小的弱点，发挥其速度优势。这被证明是一个非常棘手的问题。但当我们解决后，证明是正确的。在任何推理问题上，我们比 GPU 快 15 到 20 倍。过程中的挑战是，在计算机工业 75 年历史中，从未有人造过这么大的芯片。

We did. The founders had all worked together at my previous startup. And AMD bought that in 2012 and by 2015 we'd wandered off a little bit and we started meeting and we saw AI on the horizon. And what we knew was that this new workload would eat through extraordinary amounts of compute. And we made two really big bets. We made a bet that said like graphics produced the GPU and like mobile compute supported the development of the arm processor, that this technology, this work, would be big enough to require dedicated silicon. And the second bet we made was that the right strategy wasn't to build a derivative of the GPU. You needed to start with a clean sheet of paper and you needed to do something fundamentally different. And these were enormously contrarian bets. At the time and both proved to be dead right. And from that foundation we continued the innovative thinking and we said what AI is going to need is memory bandwidth. That's the speed with which you can move data from memory to compute. And the way to innovate on that dimension is to use a different type of memory than everybody else uses, right? We have two types of memory. We have memory that can store a lot that's slow and we call that DRAM or HBM. And we have memory that is fast but can't store very much per square millimeter. And so what we hypothesized was that if we could build a chip the size of a dinner plate, a chip 58 times larger than any chip ever built before, we could stuff it to the gills with SRAM. Thereby overcoming its weakness in not being able to store very much per square millimeter and benefit from its strength. And that proved to be a very difficult problem to tackle. But when we got it, proved to be right. We are somewhere between 15 and 20 times faster than the GPU on any inference problem. And so the challenge along the way was that nobody ever built a chip this big, not once in the 75-year history of the computer industry.

是啊，他们一直把芯片切得越来越薄、越来越小。

Yeah, they got them slicing them thinner and thinner and smaller and smaller.

没错。就连我们行业中的传奇人物，比如 Gene Amdahl，也曾失败过，一败涂地。有趣的是，即使我们解决了这个问题，还有人参观我们的实验室后尝试复制，也失败了。所以这需要多年的坚持和创新，所有功劳都归于工程团队——Gary、Shawn、Michael、JP 和我们的团队。我们失败了好几年。2019 年 8 月，我们宣布解决了这个长期未解的问题。我们以为所有人都会涌来，但世界根本不在乎。为什么？世界完全漠不关心。接下来，第一代产品大概卖了 12 套系统。哇。第二代卖了 300 到 350 套。第三代正在卖出成千上万套。所以情况是：我们解决了问题，但远远领先于市场。直到 2024 年底、2025 年初，模型才变得足够快、足够聪明，人们开始到处使用推理。就这样发生了。我们拥有地球上快几个数量级的最快推理机，突然人们想用 AI 了。而使用 AI 的方式就是推理。我们的需求被压垮了。然后在 2025 年 12 月，我们与 OpenAI 签署了价值超过 200 亿美元、为期数年的协议，这是硅谷有史以来最大的协议之一。

That's right. And even those on the Mount Rushmore of our industry, people like Gene Amdahl, had failed, crashed and burned. And interestingly, even after we solved this problem, we had people come and visit our labs and then try and build it and they also failed. And so what it took was years of perseverance and innovation and all the credit goes to the engineering team, Gary and Shawn and Michael and JP and the team we had. And we failed for years. And in August of 2019, we announced we'd solved this problem that had been unsolved forever. And we thought everybody would rush to our door and the world didn't care one bit. Why? The world was utterly indifferent. And over the next, in the first generation, I think we sold 12 systems. Wow. And in the second generation, we sold 300 to 350. And in the third generation, we're selling many many many thousands. And so what happened was we solved this problem and we were way ahead of the market. And it wasn't till 2024, late 24, early 25 that the models got fast enough and the models got smart enough that people wanted to use inference everywhere. And that's what happened. And so there we were with the fastest inference machines on Earth by orders of magnitude. And suddenly people wanted to use AI. And the way we use AI is with inference. And we were just crushed with demand. And then in December of 2025, we signed a deal with OpenAI north of $20 billion over several years, one of the largest deals ever signed in Silicon Valley.

三月份，我们与 AWS 签署了条款清单，用于在其数据中心部署。此后业务一直不错。

In March we signed a term sheet with AWS for deployment in their data centers. And business has been pretty good since.

安德鲁认为我们哪些赌注大错特错？

Which of all those bets does Andrew think we're dead wrong?

我认为我们有很多赌注都错了。任何 CEO 在回顾一个像我们这样快速发展的十年后，说自己全都做对了，那可能不是你想邀请参加生日派对的人。我们犯了大量错误。但有一件事我们做对了：理解我们用训练创造 AI，用推理使用 AI。如果 AI 要变得智能且有用，你就需要推理业务。我们很早就看到了这一点。2020 到 2024、2025 年间的真正问题是，AI 还不够智能到有用。所以每个人都关注参数数量。现在人们不在乎了。唯一的问题是：它能写出好代码吗？它能给出好答案吗？它能做我想做的事吗？这是因为我们已经进入了一个 AI 有用的阶段。这就是衡量标准。所以我们认识到了这一点。我们非常擅长训练。但现在，对快速推理的需求如此巨大，以至于我们把大量注意力分配给了它。

I think we got many bets wrong. Any CEO who looks back over a decade that moved as quickly as ours and says they got it all right is probably not a guy you want at your birthday party. We got an enormous amount wrong. But one thing we got right was understanding that we make AI with training and use it with inference. If AI is going to be smart and useful, you need an inference business. We saw that early. The real problem between 2020 and 2024, 25 was that it wasn't smart enough to be useful. So everyone focused on number of parameters. Now people don't care. The only question is: does it write good code? Does it give good answers? Can it do things I want done? That's because we've moved into a regime where it's useful. That's how it's measured. So we recognized this. We are really good at training. But right now, there's such overwhelming demand for fast inference that we're allocating a lot of attention to it.

我很好奇，安德鲁。你之前提到了 SRAM。最大的模型，有些达到 10 万亿参数，而晶圆级引擎 3 代可能只有几十 GB 的 SRAM，你怎么看 SRAM 的未来？

I'm curious, Andrew. SRAM, you mentioned SRAM earlier. The largest models, up to 10 trillion parameters, how do you think about SRAM when the Wafer Scale Engine version 3 has maybe tens of gigabytes of SRAM?

40、50 GB 左右。但最大的模型有数万亿参数。我认为这种规模的模型必须被分割，无论使用 GPU、TPU 还是我们的芯片。它们必须被切分并分布在多个芯片上。这种规模的模型在注意力头中有非常大的矩阵乘法。这无法放在一个 GPU 上。你必须切分它并进行张量模型并行。用我们的芯片则不需要这样做。但你确实需要将模型分布在四、六或八个芯片上。你非常小心地分割模型，使得没有一层跨越两个芯片。你将结果从一层传递到下一层。你可以传递它，因为结果向量非常小，通过 100G 以太网传输。这一小跳确实更慢。但占据大部分时间的计算快得多，因此将模型分割成 4、8、16、20 个芯片的代价很小，大约 2%。其他 SRAM 解决方案，比如 Groq（被 Nvidia 收购），由于只有 800 平方毫米，必须将大模型分割到 2 到 3 千个芯片上。每一跳都损害性能。我们只需要几跳，而他们需要数千跳。没有任何大小的内存能容纳所有内容，但我们有一种简单优雅的方法来分割模型并分布在多个芯片上。昨天我们宣布并发布了 Kimiko 2 的数据，这是一个开源社区的万亿参数模型。我们比任何人都快一个数量级。

40, 50, something like that. But the largest models are in the trillions of parameters. I think models that size have to be divided up, whether using GPUs, TPUs, or us. They have to be cut up and spread over multiple chips. Models that size have a very large matrix multiply in the attention head. That doesn't fit on a GPU. You have to cut it up and go tensor model parallel. You don't have to do that with us. But you do have to spread it over four, six, or eight chips. You divide the model very carefully so that no layer runs over two chips. You move results from one layer to the next. You can move it because it's a very small vector, a results vector, over 100 gig Ethernet. It is slower, that little hop. But the calculations that take up such a big portion of the time are so much faster that you pay a very small penalty for breaking it up into 4, 8, 16, 20 chips, on the order of 2%. Other SRAM solutions that are small, like Groq, which Nvidia acquired, have to break it up because they have only 800 square millimeters. They have to break a big model over 2 or 3,000 chips. Each hop hurts their performance. We have to do a few hops; they have to do thousands. There's no way to fit everything on any size of memory, but there is a nice and simple way for us to cleave models and spread them over multiple chips. Yesterday we announced and posted numbers on Kimiko 2, a trillion-parameter model in the open-source community. We were an order of magnitude faster than anybody else.

哦，真的吗？用了多少芯片？多少晶圆？

Oh, really? How many chips? How many wafers on that?

我忘了，我一直很忙。但大约是每秒 1000 个 token，而像 Fireworks 这样非常优秀的 GPU 厂商只能跑到每秒 70 个 token。所以 15 倍，相当不错。

I forget. I've been busy. But it was about 1,000 tokens per second, where a really good GPU shop like Fireworks is running at 70. So 15x, that's pretty good.

彼得和我昨天在 Google I/O 大会上。他们展示了一整架 TPU 协同工作，每秒生成 1400 个 token 来写代码。看到那个，你会想，我明天就需要它。

Peter and I were at Google I/O yesterday. They showed a whole rack of TPUs operating together generating 1,400 tokens per second writing code. You see that and you think, I need that tomorrow.

这里的把戏，戴夫，有时——Nvidia 非常擅长这种障眼法——是不告诉你他们指的是每个用户的每秒 token 数还是总吞吐量。GPU 在生成慢 token 方面极其出色。你可以让 MVL 72 以每秒 35 个 token 的速度生成，这慢得令人痛苦，但可以生成数百万个 token。另一方面，如果你要求它以每个用户每秒 200 个 token 的速度生成，它只能支持一两个用户。那是一个价值 400 万美元的解决方案，只为一个用户服务。所以当你深入分析这些数字时，重要的是：他们说的是总吞吐量吗？这是很多对性能不满意的客户吗？还是他们能够为单个客户提供服务，以及能服务多少客户？

The trick there, Dave, sometimes, and Nvidia has been masterful at this sleight of hand, is not telling you whether they mean tokens per second per user or aggregate throughput. The GPU is an extraordinarily good machine at generating slow tokens. You can generate an MVL 72 at 35 tokens per second, which is painfully slow, but can generate millions of tokens. On the other hand, if you ask it to generate tokens at 200 tokens per second per user, it can support one or two users. That's a $4 million solution working on one user. So it's really important when you dig into these numbers: are they telling you gross throughput? Is this a lot of customers who are unhappy with their performance? Or are they able to serve that to individual customers and how many?

是的。因为安德鲁，你说过……

Yeah. Because Andrew, you said...

那埃隆的 Terafab 呢？实际需要多长时间才能建成？

And what about Elon's Terafab? How long does that actually take to build?

听着，我认为埃隆在多个维度上证明了自己。他证明了自己是一个有远见的人。很多人说你试图在弗里蒙特造车是白痴。我们拥有全国最高、也许是世界最高的劳动力成本。我们有对商业不友好的监管体制。很多人说你不能造火箭公司。很多人不明白他造火箭公司是因为他想要一个卫星公司。他领先所有人很长时间了。但你正在采取一个非常……那是愿景方面。他也能在部分事情上执行，并非全部。酷的地方在于：他尝试做别人做不到的事。现在，我对这个具体问题略知一二。建造晶圆厂非常困难。这与他攻击的其他问题不同，困难的方式不同。我不是说他做不到。我是说它总是会比他说的花更长时间。它会花费多得多的钱。这就是建造非凡事物的挑战。依我拙见，这不是一个 5 年或 10 年的项目。我以前错过，但我认为这是一个 15 到 20 年的项目。我认为这很有趣。美国拥有本土晶圆厂可能对大家都有好处。但我认为有一个原因：即使使用完全相同的 ASML 设备，三星和台积电也不在同一个节点上。台积电领先，他们非常出色。

Look, I think Elon has proven himself on multiple dimensions. He's proven himself to be a visionary. The number of people said you're an idiot to try and build cars in Fremont. We've got the highest labor rates in the country, maybe among the highest in the world. We've got a regulatory regime unfriendly to business. The number of people who said you shouldn't build a rocket company. The number of people who didn't understand he was building a rocket company because he wanted a satellite company. He has been ahead of everybody for a very long time. But you're taking a very... that's the vision side. He's also been able to execute on some of them, not all. That's what's cool: he tries to do things other people can't do. Now, this particular problem I know a little about. Building fabs is very hard. It is hard in a different way than some of the other problems he's attacked. I'm not saying he can't do it. I'm saying it will always take longer than he says. It will cost vastly more money. That's the challenge of building extraordinary things. It is not a 5 or 10-year project in my humble view. I've been wrong before, but I put this at a 15 or 20-year project. I think it's interesting. It's probably good for the US that we have domestic fabs. But I think there is a reason why, even with the exact same equipment from ASML, Samsung and TSMC aren't at the same node. TSMC is ahead and they're extraordinary.

他们从几代人建造的晶圆厂中积累的经验和知识不可低估。但如果有人能做到，埃隆就能做到。那么建造晶圆厂到底难在哪里？

And the amount of received wisdom and learning from the fabs they've built over generations cannot be underestimated. But if anybody can do it, Elon can do it. Then what is so hard about building a fab, really?

这些东西很难建造，对吧？我是说，晶圆厂就是金字塔。它们是我们这个时代的金字塔，而台积电是地球上最伟大的制造公司。挑战在于这些东西需要 5 年、6 年才能建成，花费 400 亿到 500 亿美元，而且是由建造上一个晶圆厂的人来建。没错。无论是台积电、三星还是其他伟大的建造者，都是如此。这些东西建造起来极其困难。这就是为什么它们在美国一直落后于计划。我认为他们遇到了一些未曾预见的挑战。我认为我们有政治上的挑战，因为这些项目耗时太长，跨越了行政任期。对吧？当你的项目无法在 4 年内完成，必须跨越多个行政任期，可能两到三个不同的行政任期，随着时间的推移，当地方条例阻碍建设时，就像三星在德克萨斯的晶圆厂那样。他们因为一条毫无道理的地方消防条例重新设计了晶圆厂。这些都是我们系统尚未找到克服方法的痛苦问题。所以我认为我们必须找到更好的方法，因为我认为晶圆厂的回流——不仅仅是晶圆厂，晶圆厂获得了所有荣耀，还有封装业务——每个人都很重要，而我们在晶圆厂离开时完全失去了这些。

These things are hard to build, right? I mean, fabs are pyramids. They are our pyramids and TSMC is the greatest manufacturing company on Earth. And the challenge is these things take 5 years to build, 6 years to build, and $50 billion, $40 billion from the people who built the last one. Right. And that's true whether it's TSMC or Samsung or any of the great builders here. These are unbelievably difficult to build. And that's why they've been behind schedule in the US. I think they encountered some challenges that were unforeseen. I think we have political challenges in that these things take a long enough time that they cross administration boundaries. Right? When your projects can't be done in 4 years and have to cut across multiple administrations, maybe two or three different administrations, over a period of time, when you have local ordinances that get in the way of building, as happened with Samsung's fab in Texas. They redesigned the fab because of a local fire ordinance that made no sense. These are painful problems that our system hasn't found a way to overcome. And so I think that we have to find a way to do better because I think the reshoring of fab, and not just the fab, fab gets all the glory, but the packaging business. Everybody's important and something we lost entirely when the fabs left.

是的，实际上是个好问题。你知道，当某样东西准备好放入你的数据中心或第三方数据中心时，那片晶圆经历了多少个不同的制造合作伙伴？

Yeah, actually a good question. You know, by the time you get something ready to put into one of your data centers or a third-party data center, how many different manufacturing partners has that wafer been through?

相当多。然后它送到某家公司，比如 ASE，他们在背面沉积 RDL。它被切割、清洗，然后来到我们这里作为第一步。我是说，这是一个漫长的过程。我想，你知道，当我们在 90 年代不再关心晶圆厂，IBM 和 GlobalFoundries 等晶圆厂离开时，我们没做任何事来留住它们。我们失去了这一整套周边专业知识。对吧？当芯片从晶圆厂出来时，它是一块死硅。封装是你如何给它注入能量和生命的方式。如何让 IO 进入，如何让电力进入。这也是一项极具挑战性的技术。它需要材料科学家、制造工程、工艺工程、沉积工程，而我们因为不关心这个行业而放弃了这一切。而这一切都集中在台北和韩国。

A fair number. And then it goes to someone who, ASE, who deposits RDL on the backside. It's diced, it's cleaned, it comes to us first step. I mean, it is a long process. I think, you know, when we stopped caring about fabs in the '90s and IBM sort of left and GlobalFoundries as fabs, we didn't do anything to keep them. We lost this collection of surrounding expertise. Right? When a chip comes off a fab, it's a dead piece of silicon. The package is how you breathe power and life into it. How you get IO into it and how you get power into it. And that's also an enormously challenging technology. It takes material scientists, manufacturing engineering, process engineers, deposition engineering, and we punched all of that by not caring about this industry. And it's all sitting in Taipei and in Korea.

材料是在日本制造的。京瓷是那里的领导者之一。

The materials are manufactured in Japan. Kyocera's one of the leaders there.

我们必须把这一切都带回来。我们必须对这个行业做出长达几十年的承诺。那么 Cerebras 实际使用哪家晶圆厂？他们永远不会碰哪家？你知道，我们已经将 3 纳米设计交给了台积电。所以这会让我们稍微领先一点。我们确实在三星制造一些组件，并且非常尊重三星的晶圆厂能力。我们从未使用过英特尔。你知道，Lip-Bu 是一位杰出的领导者，也是硅谷硬件的长期倡导者。如你所知，Dave，在 2006 年到 2007 年左右到 2015 或 16 年期间，每家 VC 公司都充斥着来自 VMware 的人，他们对硬件一无所知，认为计算是由云中的跳蚤粪便产生的。认为以太中有某种东西在生成计算。我们花了很长时间试图解释，制造更多虚拟计算的方法是从真实计算开始。

And we got to get it all back. And we got to make a decades-long commitment to this industry. And which fab does Cerebras actually use? And which one would they never touch? You know, we've committed our 3 nanometer design to TSMC. So that will take us out a little bit. We do manufacture some components at Samsung and have a great deal of respect for Samsung's fab capabilities. We have never used Intel. You know, Lip-Bu is an extraordinary leader and a long-time advocate for hardware in Silicon Valley. As you know, Dave, there was a period between about 2007, 2006, and 2015 or 16 where every VC firm was filled with somebody from VMware who didn't know anything about hardware, who thought compute was made by flea feces in the cloud. And that there was some sort of thing in the ether that somehow was generating compute. And we tried to explain for a long time that the way you make more virtual compute is to begin with real compute.

好吧，我得告诉你，你激励了现在校园里的很多人，他们渴望成为你夺回这一切的使命的一部分。所以，我会尽可能多地把人引向这座大楼。

Well, I got to tell you, you've inspired so many people that are on campuses right now that are eager to be part of your mission to get that back. So, the more I'm going to route as many as I can through this building.

请便。我是说，像 Andy Bechtolsheim、Lip-Bu 和其他一些人，他们持续投资硬件，Pierre Lamond 也继续这样做，并在我们那段艰难的融资时期支持我们。而且我认为，嗯，我知道 Lip-Bu，他们还有很多工作要做。他到目前为止做了很棒的事情，但在我们转向英特尔之前，我们还有一些工作要做。

Please do. I mean, guys like Andy Bechtolsheim and Lip-Bu and a few others were continuing to put money into hardware, continuing to Pierre Lamond continued to do it and support us as it was tough going to raise money over that time period. And I think, well, I know Lip-Bu, they've got a lot of work to do. And he's done great things so far, but we've got some work to do before we could move to Intel.

Alex 或 Salim，你们有问题想问吗？

Alex or Salim, do you have a question you want to ask?

快速问一下。Cerebras 在 Nvidia 甚至还没遇到之前就悄悄解决了什么问题？制造工艺，你希望它缩小到什么程度？

A quick one. What is the thing Cerebras quietly solved 7 years before Nvidia even hit it? The manufacturing process, what do you hope it shrinks to?

嗯，我说第一个花了 4 年时间，可能 5 亿美元。大约在 4 亿到 5 亿美元之间。这就是为什么我和妻子出去吃饭时带着它。就像一个 10 岁孩子带着他的第一辆越野摩托车。我是说，它会上床。我是说，它在卧室里。它不在外面的车库里。我走到哪儿都带着它。我有一片晶圆。我认为这些发明涵盖了光刻、芯片架构、封装、冷却、电力输送和冷却。它们包括某种编译器发明、算法发明。事实上，我们遇到的一些最困难的问题是封装。你知道，我们在其他人遇到它们之前 7 或 8 年就解决了。所以，B100 或 B200 晚了 18 个月。它延迟是因为他们的 CoWoS 出了问题。那是什么意思？CoWoS 是一个工艺步骤，台积电使用一块 65 纳米的硅片作为主板。然后，他们把 Nvidia 的芯片和内存放在上面。而不是放在绿色电路板上，就像传统主板那样，他们把它放在一块硅片上。而且硅中的导线效率更高，可以更窄。所以这是一个重大发明。但我们知道会有热膨胀系数的问题。我们知道，因为我们在 2018 年就解决了那个问题。所以他们在 2024 年、2025 年还在为一个我们 7 年前就解决的问题而挣扎。这就是做开创性工作的结果。你会遇到问题。你有机会在行业其他人甚至遇到它们并知道它们是问题之前很久就解决它们。所以那是乐趣之一。显然，在我们做的所有工程中都有权衡。缺点是有一些低谷的日子。

Well, I say that the first one took 4 years and maybe half a billion dollars. Somewhere between 400 and 500 million dollars. That's why I take it to dinner when I go with my wife. Like a 10-year-old with a first dirt bike. I mean, it's coming to bed. I mean, it's in the bedroom. It's not outside in the garage. It is being carried around everywhere I go. I got a wafer. I think the inventions cut across lithography, chip architecture, packaging, cooling, power delivery, and cooling. They included sort of compiler inventions, algorithmic inventions. In fact, some of the hardest problems that we encountered were packaging. And you know, we solved them 7 or 8 years before others encountered them. So, the B100 or the B200 was 18 months late. And it was late because they had a problem with their CoWoS. What's that mean? CoWoS is a process step where TSMC uses a 65 nanometer chunk of silicon as a motherboard. And so, they put on that they put on Nvidia's chips and the memory. And instead of putting it on a green board, when you a traditional motherboard, they put it on a piece of silicon. And the wires are more efficient in silicon. They can be narrower. And so this was a big invention. But we knew that there would be a problem with the coefficient of thermal expansion. And we knew that because we'd solved that problem in 2018. And so there they were in 2024, 2025 struggling with a problem that we'd solved 7 years earlier. And that's what happens when you do pioneering work. You encounter problems. You have a chance to solve them long before the rest of the industry even encounters them and knows they're a problem. And so that is one of the joys. Obviously in everything we do in engineering there's a trade-off. The downside is there's some low days.

对吧？有些日子你回家，这些日子会累积起来。

Right? There are some days you go home and some of these days stack up.

我们有大约 18 个月的时间，每个月花 800 万却解决不了问题。每六周开一次董事会，天哪，你来了还是解决不了，还是解决不了。你亏了 1 亿，然后亏了 1.2 亿，再亏了 1.4 亿，还是解决不了。那真是低谷期。

And we had about 18 months where we're spending 8 million a month that we couldn't solve the problem. And when you have board meetings every 6 weeks, oh my god. You come in and you still can't solve it. You still can't solve it. And you're $100 million more in the hole. And then you're $120 million in the hole. Then you're $140 million in the hole and you still can't solve it. These are some low days.

然后你迎来了年度 IPO，那又是高光时刻。

And then you have the IPO of the year and it's a high day.

没错。我觉得，Peter，你想要一个典型的企业家完整故事，创业之旅。

That's right. I think, Peter, you want an archetypal full story of an entrepreneur, the entrepreneurial journey.

是的。而且我认为，Salim，我一路走来学到的一件事——这是我的第五家创业公司——就是如果你不能调节高潮和低谷，这会要了你的命。

Yeah. And I think, Salim, one of the things I've learned along the way, this is my fifth startup, is that this will kill you if you can't modulate the highs and the lows.

是的，会的。对于每个企业家、每个 CEO，我首先告诉他们这是对灵魂的压力测试。其次，作为创业公司的 CEO，你在午饭前被重击多少次还能觉得今天是好日子，这很惊人。你宁愿做别的事吗？不，这是我唯一知道怎么做的。我是与歌利亚战斗的专业大卫。这就是我所知道的。我对做其他事情没有兴趣，也只想和那些想攻克最难问题的人一起工作。

Yeah. It will. And that for every entrepreneur, every CEO, I tell them first that this is a pressure test on your soul. And second, the number of times you can get kicked in the gut before lunchtime and have it still be a good day as a CEO of a startup is amazing. Would you rather be doing anything else? No, this is all I know how to do. I'm a professional David in the battle with Goliath. This is what I know. I have no interest in doing other things and I have no interest in working with people who are other than those who want to attack the hardest problems.

太棒了。Alex，请讲。说到最难的问题，这几乎就在 Cerebras 的名字里。我想，你的第三代晶圆级引擎有 4 万亿晶体管的预算。我想聊聊彩虹的尽头，比如展望 10 年后，当你做到第 n 代模型时。未来是什么样子？看起来像在 WSE8 上运行的大脑上传吗？杀手级应用是什么？10 年后会是什么样？

Amazing. Alex, please. Speaking of the hardest problems and it's almost in the name Cerebras. You have, I think, a 4 trillion transistor budget with your third generation wafer scale engine. I'd love to talk maybe a little bit about what's at the end of the rainbow, projecting out say 10 years when you're on your nth generation model. What does the future look like? Does it look like brain uploads running on WSE8? What's the killer app? Where does this look like in 10 years?

所以 Alex，我认为作为基础设施构建者，有趣的一点是你不需要有那些想法。

So Alex, I think one of the fun things about being an infrastructure builder is you don't have to have those ideas.

不，真的。我在 90 年代中期与团队一起——其中很多人今天也在——帮助降低了网络成本。我们制造了最早、最快的以太网交换机之一。我们不知道 WhatsApp 会出现，并让即使是最贫穷的社会成员也能与家人通话。我 70 年代长大时，听到祖母在电话里说的唯一一句话是“让你哥哥接，太贵了。”我妈妈打电话给在澳大利亚的她妈妈，每分钟 4 美元。她们每周只讲 6 分钟。我听到祖母说的唯一一句话就是我说“你好 Bubba”，她说“让你哥哥接，太贵了。”对吧？我们创办了一家公司叫 Yago，与 Juniper 等许多公司一起，为降低 IP 传输成本添了一块砖，使得别人能发明技术，让每个人无论多穷都能与祖父母通话。这是我们当时不知道的。那不是我和公司当初要解决的问题。作为基础设施构建者，我们解决的问题是修路。你在路上开什么、开多远，那是别人的工作。我们想做的是让人们能在我们的基础设施上做非凡的事。所以当我想到我们赋能了什么，那是 Sam 的工作，是 Ilya 的工作，是别人的工作。我们想做的是创造一个算力平台，让他们的想法得以起飞。我们知道的是你需要更快的计算。

No, really. I was with the team, and many of them are here, in the mid-90s that helped drive down the cost of networking. We built some of the first and fastest Ethernet switches. And we had no idea that WhatsApp would arrive and that it would make possible even for the poorest members of our society to communicate home. And when I grew up in the '70s, the only thing I heard my grandmother say on the phone was "put your brother on, it's expensive." It was $4 a minute for my mother to call Australia where her mother was. They spoke for 6 minutes a week. And the only thing I heard my grandmother say was I'd say hello Bubba, she'd say "put your brother on, it's expensive." Right? And we put in a company called Yago, along with many others with Juniper and others. We put a small brick in the wall that made the cost of IP transport so low that somebody else could invent a technology that made it such that every person can talk to their grandparents no matter how poor they are anywhere in the world. And that's something we didn't know. That's not the problem I set out to solve and our company set out to solve. We set out to solve a problem as an infrastructure builder: we build roads. And what you drive over those roads and how far you take them, that's other people's work. What we're trying to do is allow people to do extraordinary things on our infrastructure. And so when I think about what we're enabling, that's work for Sam. That's work for Ilya. That's work for others. What we're trying to do is make a compute platform on which their ideas can take flight. And what we know is you need faster calculations.

所以我认为我听到你说的是，你刻意不对未来在你的基础设施上运行的工作负载形态持有意见，目前主要依赖前沿实验室来引导未来工作负载的架构。今天的前沿实验室或新的前沿实验室，对吧？

So what I think I heard you say is that you're very deliberately not having opinions as to the future shape of the workloads that will run on your infrastructure and you're primarily at this point deferring to the frontier labs to steer the future architecture of workloads. Today's frontier labs or new frontier labs, right?

我们正在押注世界将继续依赖稀疏线性代数作为所有这些计算的基础。

We are making bets that the world will continue to depend on sparse linear algebras as an underpinning for all these calculations.

那么晶圆级引擎在太空中的杀手级应用是什么？

And what is the killer app for the wafer scale engine in space?

我认为首先我们在太空中有巨大优势。在太空中，最昂贵的部分之一是芯片间通信。对吧？我们在太空中使用芯片已经很长时间了。卫星就是那样。卫星看起来像一块 PC 主板，上面粘着一个大相机，一个大望远镜。对吧？如果你拆开一颗小卫星，每个电脑爱好者都会说：“天哪，这看起来像一块服务器主板，上面粘着一个大望远镜。”然后它被加固了。在构建集群时进行通信实际上要复杂得多，因为你需要做大量通信工作。将数据从地面传到集群是我们很久以前就解决的问题。他们会继续改进。所以，作为一个大芯片，需要将数据移出芯片的次数更少是一个巨大优势。我认为这是一个令人兴奋的领域。我认为像许多难题一样，最后的 10% 不是 10%，而是 90%，对吧？自动驾驶就是其中之一，对吧？最后的 10% 我们花了八到十年，现在才刚刚越过那个坎，因为它实际上不是 10%。我把它归为 7 到 10 年的类别。

I think first we have serious advantage in space. In space, some of the most expensive work is the chip-to-chip communication. Right? We've had chips in space for a long time. That's what a satellite is. A satellite looks like a PC motherboard with a big camera stuck on it, a big telescope. Right? If you unpack what's in one of these small satellites, every computer hobbyist will say, "Holy cow, that looks like a server motherboard with a big telescope stuck to it." And then it's hardened. Communicating in building a cluster is actually much more complicated because you have to do a lot of communication in this work. Moving the data from the land to the cluster is a problem we solved a long time ago. They will continue to improve it. So, being a big chip and having to move things off chip less often is a huge advantage. I think this is an exciting domain. I think like many hard problems, the last 10% don't take 10%, they take 90%, right? Self-driving is one of those categories, right? The last 10% we've been sitting at for eight or 10 years and we're just now sort of getting over the hump of the last 10% because it isn't really 10%. I've got it in the 7-to-10-year category.

有趣的是，Andrew，简单提一下，我本以为你会说，对于晶圆级引擎，你必须围绕故障进行设计。在晶圆级你必须非常容错，而在充满电离辐射的太空环境中，你也需要容错，而 Cerebras 凭借其在晶圆级容错方面的经验，就像是高电离环境的计算平台。

It's interesting, Andrew, just to pull on that very briefly, what I would have expected you to say to that would have been something like with the wafer-scale engine, you had to design around faults. You had to be incredibly fault tolerant at wafer scale and in a space environment with lots of ionizing radiation, you also need to be fault tolerant and that Cerebras with its experience with fault tolerance at wafer scale is like the computing platform for highly ionizing environments.

你知道，我认为我们有很多优势，Alex，你指出了一个：在太空中你必须以非常不同的方式屏蔽你的硅片。而且你会遇到更多缺陷。有单比特错误、硬错误，一系列你必须考虑的错误。我们关闭一个核心并绕开它的能力在那个环境中是一个巨大优势。我认为作为一个社区，我们在未来四五年内还有一些工作要做，才能解决将它们送入太空、编排软件、让它们通信的真正困难部分。

There are, you know, I think we have lots of advantage, Alex, and you put your finger on one of them that you have to try and shield your silicon very differently in space. And you will get more flaws. And there are single bit errors, hard errors, a whole collection of errors that you have to contemplate. And our ability to shut down a core and route around it is an enormous advantage in that environment. I think we've got as a community some work to do over the next four or five years before we have the truly hard part of getting them in space, orchestrating the software, getting them to communicate.

所以我认为，距离我们在太空进行生产，大概还需要十年左右的时间。我认为这是一个非常值得追求的项目，但它目前还不在考虑范围内。

So I've got it sort of out the better part of a decade before we have sort of production in space. I think it's a very worthwhile project to pursue, but it's out of the way.

哪些职业实际上会最先被语言模型取代？

Which professions actually fall first to language models?

从每小时一千美元降下来了。是的，这是第一点。第二点，我认为律师和会计师的问题在于他们的业务结构——这是不适合未来的商业模式。

It's down from a thousand an hour. Yes. That's number one. Number two, I think the problem with lawyers and accountants is the structure of their business is the wrong business model for the future.

这完全错了。按小时计费，对吧？他们的业务就是站在普通人和晦涩知识之间。确实如此。你的会计就是干这个的。你不想搞清楚你 2020 年购买或获赠的房产折旧相关的税务规则。对吧，谁想知道那些？所以，他们的业务就是获取晦涩知识，并将这些知识应用于特定问题。这正是语言模型所擅长的。

It's exactly wrong. Something hours, yeah. Right? What their business is to stand between ordinary people and obscure knowledge. That's true. That's what your accountant does. He... You don't want to figure out what the tax rules are with related to depreciation on a property you bought or was gifted to you in 2020. Right, who wants to know that? And so, what their business is is sort of the acquisition of obscure knowledge and the application of that knowledge to particular problems. That's exactly what language models do.

我觉得安德鲁，这可以推广吧？除了作为高深知识的守门人，你还能是什么？

I think Andrew that generalizes though? Like what are you other than gatekeeper of obscure knowledge regarding the high gods?

安德鲁是工程师中的工程师。他们处于最佳状态。实际上，他们之所以被称为顾问，是因为他们提供的是好的建议，而非法律事务。当他们给出好的商业建议，当常识在混乱环境中受到挑战时，那才是他们最出色的时刻。我认为，当你起草所需文件时，大部分东西已经起草好了。对吧？我们不需要另一个律师来审查另一份保密协议。我们不需要……

Andrew is something an engineer's engineer. They're at their best. They are actually, you know, the reason they're called counsel is when they're giving good advice not on legal matters. When they're giving good business counsel, when common sense is challenging in a confusing environment. Those are when they're at their best. I think when you're drafting all the documents you need for most things already been drafted. Right? We don't need another lawyer reviewing another NDA. We don't need...

给你下一个机会。你想选问题二、三还是四？

Let's give you next shot. Which question two, three, or four you want to choose?

我觉得第三个问题很有趣。我认为……

I think number three is interesting. I think...

人们对事物如何运作存在深刻的误解。我们来读一下这个问题。为什么金钱不能为埃隆或扎克买来 AI 领域的领先地位？他们难道不能直接买下最优秀的人才吗？这个问题来自 @novarift。好问题。

There is a profound misunderstanding about how things... Let's read the question. Why can't money buy Elon or Zuck a lead in AI? Can't they just buy the best talent? That's from @novarift. Great question.

不。答案是否定的。我认为你知道为什么英特尔造不出手机处理器？当时他们拥有最好的晶圆厂，最好的计算机架构师。结果他们毁掉了数百亿美元的股东价值，失败了。AMD 也一样。事实证明，在我们这个行业，金钱和获取人才是不够的。那什么才够？还有别的东西。

No. The answer is no. And I think you know why couldn't Intel build a cell phone processor? They had at the time they had the best fabs. At the time they had the best computer architects. And they destroyed tens of billions of shareholder dollars failing. Same with AMD. It turns out in our industry that money and the acquisition of talent isn't enough. What is? There's something else.

MTP？

MTP?

什么？目标。大规模变革性的目标极其重要。不过英特尔本可以对苹果说“是”的。

What? Purpose. Massively transformative purpose is incredibly important. Intel could have said yes to Apple though.

不，不，但他们本可以，但事实是什么让他们相信自己有资格对苹果说“不”？英特尔在追逐利润率。英特尔沉迷于自己的利润。他们有一个 ARM 部门，然后卖掉了。英特尔本可以销售手机芯片……也许吧。也许。或者你的 DNA 里有什么东西让你……

No no but they could have but the truth is what led them to believe that they were in a position to say no to Apple? Intel was chasing margins. Intel was infatuated with its own profits. They had an arm division that they sold off. Intel could have sold cell phone... Maybe. Maybe. Or maybe there is something in your DNA that makes...

运气。大的突变。

Luck. Big mutations.

嗯，运气……你看，我坐在这里整天说，我们宁愿要运气也不要技能。但我也整天说，那些极其勤奋、有巨大毅力的人最终会更幸运。这两点都成立。这就是生活。那些长期勤奋、正直且有道德的人，对吧？他们更常走运。运气在勤奋和不勤奋的人之间分布并不均匀。为什么拥有最多钱的团队一次又一次地失败？但我认为亚历克斯，问题不在于……当然他们可以。为什么他们没有？为什么他们错过了？为什么 AMD 错过了？为什么，比如，为什么英伟达在 GPU 以外的所有领域都失败了几十年？他们没能造出能用的 ARM 处理器。我记得那叫骁龙。我认为他们在北桥南桥部分失败了。而他们在 GPU 上取得了超出所有人预期的成功。我认为同样的问题也适用于洋基队。

Well luck... I look, I sitting here saying all day long we'll take luck over skill. But I'll also say that all day long that extremely hard-working people with tremendous grit end up more lucky. And that both of those are true. That is life. It is really hard-working people over long periods of time who have integrity and ethics, right? They get lucky more often. And luck is not equally distributed to those who work hard and those who don't. And why does the team with the most money lose again and again? But I think Alex that the question isn't... Of course they could. Why didn't they? Why did they miss it? Why did AMD miss it? Why did for example why did Nvidia fail for decades at everything that wasn't a GPU? They failed to build an arm processor that worked. I think it was called Snapdragon. I think they failed at Northbridge Southbridge part. And they succeeded beyond anybody's expectations at a GPU. I think that the same question I think holds true for the Yankees.

对吧？不，不，不。为什么 NFL 中花钱最多的球队不是每年都赢？为什么……我的意思是，我们在思考组织和人才时，有些东西我们描述得不够好。那就是有些东西很难买到，必须被创造出来。而我们似乎无法很好地表达它。买来最多的人才似乎并不足够。你必须拥有大量人才，这是必要的，但显然不充分。

Right? No no no. Why doesn't the team with the biggest money win every year in the NFL? Why... I mean there is something that we have in thinking about organizations and talent that we don't do a very good job of describing. That says there is something that is very hard to buy. And that has to be made. And that we don't seem to be able to articulate it well. And buying the most talent doesn't seem to be sufficient. It is you have to have a lot of talent. It's necessary. But it's clearly not sufficient.

是的，当然。我有些……我不得不说，我对其他一些问题有相当不同的答案，但我认为我必须回答第二个问题，它看起来可能是对我之前播客中一条评论的回应。问题是：“为什么山姆——我想是指山姆·奥特曼——不和贝佐斯及蓝色起源达成协议，成为对抗埃隆的另一股力量？”这个问题来自 Scott Ray Broomfield。所以，我认为答案可能是这已经在考虑之中了。如果我是山姆，我会探索各种潜在的重型发射合作伙伴关系，以成为对抗埃隆和 SpaceX 的 AI 戴森球的制衡力量。我认为重型发射将成为——甚至可以说已经是——未来算力堆栈的关键组成部分。

Yes, of course. I have some... I have to say I have some pretty different answers to some of these other questions, but I think I have to answer question number two, which it looks like might have been a response to a comment that I made in a previous pod episode. So, the question is, "Why wouldn't Sam, I think referring to Sam Altman, cut a deal with Bezos and Blue Origin to become the other counterweight to Elon?" And this is from Scott Ray Broomfield. So, I think the answer is that's probably on the table. If I were Sam, I would be exploring a variety of potential heavy launch partnerships to become a counterweight to Elon and SpaceX AI's Dyson swarm. I think heavy launch is going to become... is already arguably part of a critical element of the stack now for the future of compute.

不，我认为亚历克斯，你的观点都对，但我认为你低估了山姆，这会让你付出巨大代价。我认为山姆在我们行业一次又一次做到的是，看到其他人忽略的拐角。去年和前年，当其他所有基础实验室都没看到时，他就在试图锁定太空中的数据中心容量。

No, I think Alex, all your points are right, and I think that you underestimate Sam at your tremendous cost. I think what Sam has done again and again in our industry is see around corners that other people missed. He was trying to lock down data center capacity in space last year and the year before when all the other foundational labs didn't see it.

真的吗？

Really?

锁定内存。哦，是的，当然。我之前不知道。嗯，他看待指数增长的能力，以及不害怕它两三年后意味着什么，而其他人都害怕说“哦，我们不需要那么多”，这非常了不起。他的影响力也非同寻常。所以，我可以 100% 肯定地告诉你，他正在探索各种可能的方式来获取算力和数据中心容量。我可以这么说，虽然我是从远处观察，没有内部消息，但我被他那种能力所折服。我认为你低估了那个人。我的意思是，你会认为打造资本主义历史上增长最快的公司足以赢得很多尊重，对吧？你会认为那可能是你帽子上一根足够漂亮的羽毛，但我认为他肯定会参与各种对话来获取算力，无论是在太空、海底，还是利用水力发电、地热，他都会参与这些对话。几点想法。有很多选择。是的，我同意 SpaceX 完全主导了入轨质量的论点。

To lock down memory. Oh, yeah, for sure. I didn't know that. Um, his ability to look at an exponential and not be afraid of what it says in two or three years, while everybody else is afraid saying, "Oh, we're not going to need that much." It is extraordinary. And his reach is extraordinary. And so, with 100% certainty, I will tell you that he is exploring deals with every possible way to get access to compute and data center capacity. And I can say that having watched from a distance, I have no inside information, but I've been dazzled by that ability of his. I think you underestimate that guy. I mean you would think it would be enough to build the fastest growing company in the history of capitalism to get a lot of respect. Right? You think that might be a sufficient feather in your cap, but I think he will certainly be in conversations to get compute, whether it's in space or whether it's under the sea or whether it's on, you know, using a falling water, you using geothermal, he will be in those conversations. A few thoughts. There are lots of options. Yes, I agree with the contention that SpaceX is completely dominating mass to orbit.

毫无疑问，包括主导历史性的入轨质量。所以，如果我是 Sam，我可能会探索一个多渠道策略。首先，我会考虑与 Elon 和 SpaceX 达成协议，利用 SpaceX 的发射能力来建设我自己的戴森球。我会探索替代发射方案。然后，如果你真的相信——我认为这是房间里的大象——如果你真的相信我们正处在这个奇点式的指数增长上，那么晶圆厂就不必建在地面上。我们可以在太空中建造晶圆厂，5 到 10 年后我们可能就不再依赖重型发射了。如果你是 Sam，并且你在玩长期游戏，那么你会寻找在月球和近地轨道上建造晶圆厂的方法，而不需要 SpaceX。

No question about it, including dominating historic mass to orbit. So, if I'm Sam, I would be exploring probably a multi-channel strategy. A, I'd be exploring a deal with Elon and SpaceX to leverage SpaceX launch for my own Dyson swarm. I would be exploring alternative launch capabilities. And then, if you really believe, again, I think this is the elephant in this space room, if you really believe that we're on this singularity-esque exponential, then the fabs don't need to be on the ground. We can build fabs in space, and we may not be addicted to heavy launch 5 to 10 years from now. And if you're Sam and you're playing the long game, then you're looking for ways to build fabs on the moon and in LEO that don't need the SpaceX.

我觉得在地面上建晶圆厂对我来说已经够难了。在 5 年内协同投入 400 到 500 亿美元做一件别人——我是说世界上只有一两家公司成功做到过的事。知道那有多难之后，我真的没法认真考虑在太空建晶圆厂。

I think building a fab on land is hard enough for me. Synergy 40 or 50 billion in 5 years doing something nobody else — I mean that one or two other companies in the world have ever successfully done. I can't really think hard about building fabs in space knowing how hard that is.

这正好引向第四个问题。中国在 AI 基础设施方面到底赢在哪里？他们只是缺乏算力。

That's a great segue to question four. What is China actually winning at when it comes to AI infrastructure? They just don't have the compute.

是的，他们迄今为止选择的投资方向是电力基础设施。在这方面，他们现在比我们做得好。他们升级了电网，拥有巨大的电力基础设施，而我们在这方面做了错误的决策。我们的电网还停留在 50 年代的水平，设计上不适合今天的用途，而且在地方、市、州、联邦各级政治上推进基础设施项目都很困难。所以，他们在那里投入了巨资。他们显然缺乏算力，但他们会试图利用已有的东西，也就是极其庞大的电力基础设施。

Yeah, the dimension in which they've chosen to invest so far is in power infrastructure. At that, they're just playing better than us right now. They have upgraded their grid. They have tremendous power infrastructure, and we've made bad decisions there. We are stuck with a grid that's built in the '50s. It's designed not for what we'd like it for today, and we have trouble politically at the local, municipal, state, federal level doing projects like infrastructure. So, what they have done is a tremendous amount of investment there. They are obviously starved of compute, but they're going to try and build on what they have, which is an absurd amount of power infrastructure.

我或许可以补充一点。值得注意的是，在过去一周左右，有不少公开报道称中国正在运营代理服务，以极低的价格向中国用户出售美国 token。比如 10 倍的折扣，目的是为了获取推理轨迹来训练他们自己的模型。这相当具有破坏性，Anthropic 正在追查此事。

I would perhaps just add to this. It's worth noting that in the past week or so, there's been quite a bit of public reporting about how China is operating proxy services that are selling American tokens to Chinese users at incredibly low prices. Like they are 10x discounts in order to siphon the reasoning traces for training their own models. And that is quite disruptive, and Anthropic is pursuing that.

我觉得当 CEO 就够了，老兄。我可不想承担四个小时的责任。通常来说。

I think being a CEO is sufficient, man. I don't want the responsibility of four hours. Usually.

没问题。是的，是的。谢谢你邀请我。非常感谢。保重。这里是 ASI Pill，Alex Wister Gross 与 Cerebras CEO Andrew Feldman 的对话，来自 Moonshots 播客第 256 期的 Alex X Andrew 特别节目。订阅获取更多内容。下期再见。

Easily. Yeah. Yeah. Thank you for having me. Really appreciate it. Be well now. That was the ASI Pill, Alex Wister Gross in conversation with Cerebras CEO Andrew Feldman, an Alex X Andrew special from the Moonshots podcast EP 256. Subscribe for more. Until next episode.