Andrew Feldman，Cerebras 联合创始人兼 CEO：AI 芯片战争与打破 Nvidia 主导地位的计划
- Feldman 对 NVIDIA 的核心攻势是：GPU 的片外 HBM 内存是生成式推理的根本性架构限制——从一个 700亿参数模型生成一个词，意味着要把约 140GB 权重从内存搬到计算单元，而且每生成下一个词都要再搬一次。 晶圆级设计让 Cerebras 得以在容量规模上使用高速 SRAM；“过去是他们优势的东西现在成了弱点……它可以被击败，我认为他们知道这一点。”他对市场结构的判断是：NVIDIA 将在 5 年后从今天“几乎全部”的市场份额降至 50%-60%，介于 Uber 的 90/5 和云计算市场更均衡的分割之间。
- 在推理场景，CUDA 护城河“根本不存在”——你可以用 10 个按键，把 OpenAI 从 Nvidia GPU 切换到 Cerebras。 真正、却很少被讨论的护城河，是市场份额本身的领先地位：Intel 做了“接近 10 年的灾难性决策”，仍然掌握约 75%-80% 的 x86 市场。“我可以连续 10 年做出一堆糟糕决策，却只丢掉 20% 的份额——这护城河简直不可思议。”
- Scaling law 带来的收益仍在延续，Feldman 直截了当地否定了“算力、算法和数据都已经走得很远”的共识。“我认为他们错了。我认为我们在这3个方面都还处于早期。”GPU 执行推理时利用率只有 5%-7%——意味着95%或93%都被浪费；OpenAI 的 o1 表明，推理 scaling law“完全有效”；3-5 年后，我们对 Transformer 的依赖会显著下降，“100%”。5 年后，训练数据将“几乎全部是合成数据”。
- 推理市场的方程是：用户数 × 使用频率 × 单次使用算力——三者同时增长，这种条件极其罕见。 AI 在 2024 年Q4 从“新奇玩意”变成了有用工具，5 年后的市场将“大幅超过现在的 100 倍”；速度更快、价格更低只会进一步扩大市场：“过去 50 年的算力行业里，没有任何一个例子是因为把东西做得更便宜、更快，市场反而变小。”
- 以 5 年视角看，芯片供应商的价值将高于模型供应商。 今天的模型估值本质上是在给期权定价——“不确定性是期权价值的朋友”——但最终巴菲特所说的称重机仍会启动。对于模型行业，领先竞争对手 4 个月的发布节奏并没有太大价值；多年保持前10%的位置，才有价值。
- Cerebras 已实现正现金流，而同行仍在大量失血。“毛利率过去是衡量技术差异化的指标……如果你做的是负毛利业务，那你卖的就是大宗商品。”另一面是，G42 占收入的 87%（这笔交易据估算超过 10亿美元），Feldman 将其解释为一种可以习得的能力：“拿下3个大客户的方式，是先拿下1个”，未来 24 个月内将争取建立数个客户关系。
- 在中国市场，Cerebras 拒绝了销售。 这笔交易“不会被用于善事”（包括对少数族裔进行人脸识别、军事用途），也没有通过他的“母亲测试”；但他认为美国对中国的低估程度是“100%”，出口管制可能并不是“一个可处理的问题”，而切断 EDA 工具只是在深圳催生了由美国风险资本支持的 EDA 初创公司。
1. 2015年的下注：AI的难题是搬数据，而不是计算
- Feldman 对芯片设计的框架是：芯片“做计算，也搬数据”——而 AI 颠倒了两者的难度。数学本身很简单（“矩阵乘法和 FMAC，任何一个大二电气工程学生都能做出来”）；真正困难的是，结果和中间结果必须持续搬运——搬到内存、从内存搬回来、在 GPU 之间搬运。Cerebras 的下注是：解决数据搬运问题，就能造出更快、功耗更低的 AI 计算机。
- 作为第5次创业者，他唯一承认的一次失误尤其值得注意：“第一次，我大幅低估了市场规模。”团队判断正确的是，AI 将对内存带宽和通信结构施加压力，而这正是 GPU 从设计之初就没有针对的维度。
- 训练和微调在计算量上“基本相同”，生成式推理则是例外。从一个 700亿参数模型生成一个词、且权重为16-bit 时，需要把约140GB数据从内存搬到计算单元——生成下一个词时还要再搬一次。“这叫内存带宽；如果你的架构是 GPU，这就是根本性限制。”
2. 晶圆级设计与70年来无人解决的良率问题
- 内存的取舍在于：HBM“非常出色……但速度慢”——容量大，适合图形处理，因为图形场景很少反复回读内存。SRAM“快得不可思议，但容量很低”——一块普通尺寸的 SRAM 芯片要服务 400B 参数模型，需要约4000块芯片；DeepSeek 671 则需要“6000块或8000块……这会变成多么可怕的行政管理噩梦”。晶圆级设计让 SRAM 的速度与足够的容量可以集中在1片、2片甚至10片晶圆上，同时功耗更低，因为片外 I/O 是芯片上最耗电的操作之一。
- 过去没人做到这一点，问题在良率。他的比喻是：晶圆像饼干面团，缺陷就像母亲蒙着眼随机撒下的 M&M；“饼干越大，撞上 M&M 的概率越高”，传统做法是把有缺陷的芯片分级处理或直接丢弃。Cerebras 借鉴内存制造的办法是，用数十万块完全相同的晶圆级单元，并设置冗余行或列来制造处理器；关闭有缺陷的单元，再绕开它。“在我们这个行业70年的历史里，从来没人能做到这一点”——很可能是 Gene Amdahl 的公司 Trilogy 曾尝试过，但“撞得头破血流”。
3. 速度不是规格，而是创造产品类别的东西
- Feldman 拒绝对快、便宜、准确做单一排名：癌症诊断中，“93%的准确率就是不如94%”，用户愿意付钱并等待；而对 Llama 405B 为 70B 模型生成调优数据这类批处理任务，价格可能更重要。但在交互式使用中，“毫秒很重要”——Google 多年前就证明，“几毫秒的延迟就能摧毁用户的注意力”。“如果要等8分钟，就不会有搜索。”
- 他更倾向于使用这样的类比：互联网速度很慢时，Netflix 通过邮寄 DVD 运营；宽带到来后，“亚马逊突然变成了电影制片厂。它改变了一切。推理速度也会产生同样的效果”——GPU 延迟无法支持的新应用，会因此出现。根据 Artificial Analysis，自 8 月 26 日发布以来，Cerebras 声称自己在“一整组模型”上实现了最快推理。
4. 推理方程：3个乘数同时增长
- 他的市场规模公式是：推理市场 = 使用 AI 的人数 × 使用频率 × 单次使用算力。“我们正处于一个罕见时期”，3个变量同时增长，因此市场出现离谱的增长。5 年后，市场将“大幅超过现在的 100 倍”。
- 转折点并非技术突破：“ChatGPT 并不是真正的技术创新，而是用户界面创新。”直到 2024 年年中，AI 还是“新奇玩意……哇，这很酷”。从 2024 年Q4 开始，它变得有用——“如果你的营销团队没有使用 LLM，每个人每天用几次，那他们就没有在做自己的工作”——它开始触达“我的父亲、我的兄弟、医生们”；“当你触达这些人时，市场就会爆发。”
- 他承认，行业“消耗了巨量电力……以及一些水”，因此“交付非凡价值的责任在我们身上”——包括治愈疾病、解决社会问题。美国的问题不是没有电，而是地理位置和流程：“尼亚加拉有电……但我们没有电力供应在你想建数据中心的地方”，也没有全国性机制能够绕过地方监管和既有利益。
- 对 Groq 的 Jonathan 所说的“游客数据中心”，Feldman 大体上不同意。最早进入这一领域的是 TeraWolf 和 Crusoe 这样的比特币矿企，“当然不是游客……而是极其成熟的数据中心建设者”，如今正在主导 GW 级项目。“当然会有一些游客”，但很多设施最终会运行良好。
5. Scaling law 带来的收益仍在延续——“我认为他们错了”
- 针对“算力、算法和数据都已经走得很远”的反复论调，他说：“我认为他们错了。我不认为我们已经走得很远……我认为我们在这3个方面都还处于早期。”最直接的证据是：GPU 执行推理时“利用率只有5%或7%——这意味着95%或93%都被浪费”。成本下降来自更便宜的计算机、更低 PUE 的数据中心，以及更好的算法，三者会相互叠加。
- 关于 scaling law 本身，业界确实在争论数据是否已经“耗尽了魔力”；但“OpenAI 在 o1 上的工作让我看到，scaling law 至少在推理上完全有效——投入更多推理算力，就能得到更好的答案。”
- 算法仍有大量提升空间：许多模型依然是全连接的——“有些连接并不会产生任何结果，但我们仍然要对它们做数学运算”。他的类比是：学习某件事，可以读50本书，也可以只读真正重要的3本，或者读那3本书的摘要——“问题在于，一开始我们不知道哪3本才重要”。MoE、dropout、稀疏化都只是早期步骤。至于架构，“3年或5年后，我们不会像现在这样依赖 Transformer——100%。它们不是终极答案”。他不愿猜测继任者会是什么（“我不知道会不会是 state-based models”），但 attention head 的二次复杂度效应是已知弱点，人们“拼命想要克服”它。
- 5 年后的合成数据将是“几乎全部是合成数据”，而且他认为其效用可以与人类数据相当。他完整保留了飞行员训练的类比：真实驾驶数据是“人们在高速公路上直线驾驶——并不困难”；真正需要的是“雪天无保护左转……成千上万种、数百万种不同方式”。就像飞行员模拟器和外科医生面对的罕见病例，合成数据正好补上那些昂贵或痛苦、难以采集的数据。
6. DeepSeek 是“专注的工程”——而蒸馏指控经不起一致性检验
- 让他印象深刻的是：“他们没有沉迷于做模型知识分子……他们只想变得更好。从发明角度看，这有点无聊；但从工程角度看，这是非常漂亮的努力。”DeepSeek 证明了“你不需要5000人和数十亿美元的设备——200个聪明人就能做到，设备比 DeepSeek 说的多，但比其他人少”。在就职典礼前后发布的时点，他将其归为政治因素。
- 关于蒸馏：“摘要难道是错的吗？……我不认为蒸馏是错的；如果蒸馏是错的，那使用人们受版权保护的数据当然也是错的。你必须保持一点一致性。”至于影响力，“很少有开源项目能像那个模型一样产生如此即时的影响……它引发了巨大的轰动。”
7. 价值在哪里累积：芯片胜过模型，以及无人点名的护城河
- CUDA 在推理领域的锁定效应“根本不存在……完全没有——你可以用10个按键，把运行在 Nvidia GPU 上的 OpenAI 切换到 Cerebras，再切换到 Fireworks 或 Together”。大多数 AI 都是用 PyTorch 编写的；编译器“很难，但可处理”。NVIDIA 真正的护城河是成为默认选项，而 Feldman 认为挑战者没有充分研究这一点：在 Lip-Bu 之前，Intel 做了“接近10年的灾难性决策”，但仍然掌握约80%的 x86 市场，AMD 也只 claw back 到25%-30%。“这就是护城河……作为挑战者，我们必须准确理解它，因为我们需要为此搭一座桥。”
- 5 年后的市场结构将介于 Uber 的 90/5/5 和云计算的共享式寡头格局之间——“Nvidia 将占据50%-60%的市场份额……而现在它几乎拥有全部市场。”他强调，NVIDIA 不会在推理领域“躺平装死”——这家公司已经经历了“历史上任何公司都称得上伟大的10年之一”，从 2014 年约 10亿美元起步。只是 NVIDIA 的漫长交付周期构成了“巨大的机会窗口”：“当恶霸倒下时，所有人都想上去踢一脚。”
- 从 5 年后的企业价值看，芯片供应商会大于模型供应商，答案是肯定的。他解释模型今天为何显得昂贵：“给期权定价时，方差和不确定性会增加期权价值……这些极高价格的一部分，正来自这种巨大的波动。”长期来看，巴菲特的逻辑会生效：市场短期是投票机，长期是“称重机制”——“某个时候，称重会启动，而通常是在公开市场上启动。”
- 模型公司的防御力取决于什么？“你是在和别人的发布节奏竞争——你领先4个月，他们领先6个月。如果这就是你的全部位置，那价值不大。但如果你能多年保持前10%，而排在你前面的人不断变化，我认为那就很有价值。”硬件能够长期存在——Cisco、Juniper、Apple、NVIDIA——因为“它们做的事情很难。这就是为什么它值得被挑战。”
8. 这门生意：正现金流、1个大客户，以及为何上市
- 对于同行还在失血、Cerebras 却实现正现金流这一点，他说：“传统上，毛利率是衡量技术差异化的指标……如果你做的是负毛利业务，那你卖的就是大宗商品——你创造的价值没有被市场识别。”
- G42 占87%的收入（宣布交易时据估算超过 10亿美元）既是优势也是弱点：“拿下3个大客户的方式，是先拿下1个……成为战略合作伙伴是一项可以习得的技能。”目前的证明包括：已部署数十 exaflops，“远超 AMD 或 Nvidia 之外的任何公司”；软件已在一些最大的 AI 集群上完成强化；制造规模实现了“2倍、5倍、再2倍”的增长。目标是在未来 24 个月建立“数个”客户关系。
- 对 Harry 提出 IPO 文件似乎过于提前、并向竞争对手提供非对称信息的质疑，他的回应是：“我们没有什么可隐藏的……我们拥有非对称技术。”支持上市的理由包括：Cerebras 是这一类别的先行者，而且“我们最大的几个目标客户中，有些明确偏好与上市公司做生意”。
9. 中国：拒绝销售、尊重对手、质疑管制
- 对出口管制，一位硬件出身的人会区分：一台 500-600 磅的服务器“会放在托盘上送到现场——你可以让使馆的人来查看，每月拍一次照片。它不会被带走”。软件和开源则是“完全不同的层级”。他提到，DeepSeek“可能确实使用了在新加坡的芯片”。更深层的怀疑是：“我不知道延缓另一个国家在一条技术轨迹上的进步，是否是一个可处理的问题。”限制 EDA 工具的结果只是“美国风险投资家支持了深圳大量中国公司去开发 EDA 工具”。他认为，本届政府对 AI 的态度“可能明显好得多”，上届政府则“站到了大科技公司的对立面”。
- 但 Cerebras 自己拒绝了中国的一笔交易——他的测试是：“只要问问自己，我的母亲会为此感到骄傲吗？”他看到或无法忽视的用途，包括“识别少数族裔、用于迫害”的人脸识别，以及军事设备，都没有通过测试。“这比钱更重要。”他同时保留了两种判断：管制可能徒劳，但是否出售这笔交易，是他自己的选择。
- 我们是否低估了中国？“100%……这是最明显、最常见的判断错误之一。”他列举的证据包括：规模惊人的基础设施投资、极强的工程人才培养能力、国家支持的风险资本、国家冠军企业，以及“一套让第三世界大部分地区依赖中国的双保险战略”。深圳经济特区“显然是一个具有远见的举措”——美国在需要时也做过类似的事（特朗普第一任期放宽疫苗规则）。他提出了令人不安的问题：为什么美国造不出火车？为什么“我们的桥梁和高速公路都在失修”？
10. 快问快答：犯错、微型芯片，以及经验的价值
- 他最有记录可查的错误，是反对联合创始人 JP 在 2016 年提出的水冷方案：“我拼命反对，而我完全错了。”Google 在1、2年后宣布采用水冷 TPU；“现在 Nvidia 卖的全是水冷部件。我大错特错，JP 才是对的。”他总结道：“如果你没有准备好经常犯错，就不应该做太多决策。”CEO 必须“多数时候都能做出正确决策”——这与 VC 不同，VC“平均而言一直在犯错，他们关心的是偶尔有一次极其正确”。
- 芯片市场中被低估的角落，是贴近传感器、功耗低于1毫瓦的推理芯片：它们“只把有用的数据发回去”，规模巨大，“对机器人至关重要”。不过这不是他的市场——“我喜欢制造更大的东西，再把它们卖给数据中心。”
- 对 Dario 的预测，他不认同人类能活到150岁，也不认同今年90%的代码会由机器编写；但他预计，AI 的渗透率将在1、2年内接近手机。他的逆向判断是，中东和平比人们想象的更近，关键在于“一个温和、以商业为导向的阿拉伯国家正在崛起”，范围包括 UAE、Qatar，甚至 KSA。至于第5次创业，他认为，只要企业涉及制造、供应链，以及数百或数千名工程师按计划协同工作，经验就会复利——“没人会面不改色地说：我想找的是一个没有经验的工程负责人。”
1. What Will Be the Ratio of Synthetic to Human Data Used in 5 Years?
Our AI algorithms today are not particularly efficient. In a GPU, most of the time it's doing inference, it's 5% or 7% utilized. That means it's 95% or 93% wasted.
We won't be as dependent on transformers in 3 years or 5 years as we are now—100%. The fundamental architecture of the GPU with off-chip memory is not great for inference. Now, they will continue to do well in inference, but they can be beaten, and I think they know it.
Andrew, it is such a pleasure to meet you. I've wanted to do this one for a while, and I've heard so many good things from Eric for a long time, so thank you so much for joining me.
2. Where Was AI Landscape in 2015 When Cerebras Founded
I have my pen ready. I feel like this is going to be a learning experience for me. I want to go back to 2015. What did you and the team see in the AI landscape in 2015 that led to the founding of Cerebras?
We saw the rise of a new workload, and this is every computer architect's dream. We saw a new problem to solve, and what that means is maybe you can build a new machine better suited to that problem.
In 2015—and the credit goes to Gary, Shan, JP, and Michael, my co-founders—they saw on the horizon the rise of AI. What that meant was there'd be a new problem for computers, and what the AI software would ask from the underlying chip or processor would be different. We came to believe that we could build a better machine for that problem.
That's what we saw. Obviously, we didn't see it exactly right. I underestimated it. This is my 5th startup, and the first time I underestimated the size of the market by a lot. What we did get right was that this was going to be big, that it would put a different type of pressure on a processor, that it would put pressure on the memory bandwidth, and that it would put pressure on the communication structure.
3. NVIDIA’s Biggest Strength Has Become Their Biggest Weakness
That's what we saw. We dove in, and it's been an extraordinary 9 years.
Can you help me understand how the movement into an age of AI changes the requirements from a chip perspective of what is needed for a provider, and how that resulted in how you built Cerebras?
The way to think about a chip is that it does 2 things: it does calculations and it moves data. Sometimes, along the way, it stores data. That's what a chip does.
What AI presented was a very unusual combination of challenges. First, the underlying calculation is trivial. It's a matrix multiplication, and an FMAC can be developed by any 2nd-year electrical engineering student. You say to yourself, “Holy cow, this has a huge number of very, very simple calculations.”
The hard part with AI work is that results and intermediate results have to be moved a lot. They have to be moved to memory and from memory, and they have to be broken up and moved among GPUs. What we saw was that this was going to be the hard problem, and that if we could solve for that problem, we would build an AI computer that was faster and used less power.
4. What Happens to the Cost of Inference?
When we think about what we're going to build and what we're building for, there are a couple of core elements. Where are you going to focus? Are you focusing on fine-tuning, training, or inference?
You chose all 3. Why? I'm sorry for my basic questions, but I thought GPUs were specialized toward training and weren't specialized toward inference. Can you have a mono-architecture that does all 3 best?
The first step in computer architecture is deciding what you're not going to do. What are we not going to be good at? That's really the first important question to answer.
To answer your question, is the computational work for training from scratch different from fine-tuning? The answer is that it's not different. It's approximately the same.
Inference and training have some different requirements, and generative inference in particular has some very challenging requirements on exactly the communication dimension that I mentioned. In generative inference, you have to move all the weights from memory to compute to generate a single word. You have to move them again to generate the next word, and again and again.
If you have a 70-billion-parameter model—not a giant model—and each weight is 16 bits, you're moving 140 gigabytes of data to generate 1 word. This is an enormous amount of data movement across memory, and that needs memory bandwidth.
If you have an architecture like we saw in the GPU, that's your fundamental limitation. It's a fundamental architectural limitation. That was what we went to wafer scale to solve.
5. Why Are AI Algorithms So Inefficient?
They use a memory called HBM, a type of DRAM, and it's phenomenal memory, but it's slow. It's slow and high-capacity. When they set the architecture for graphics, that's what you wanted. You didn't have to go back and forth to memory very often.
SRAM, on the other hand, is unbelievably fast but has low capacity. We wanted to use SRAM, but if you build a normal-sized chip, you can't hold a model. By going to wafer scale, we were able to put down a huge amount of SRAM and get the benefits of speed and enough capacity.
If you build a normal-sized chip with SRAM and you want to do a 400-billion-parameter model in inference, you might need 4,000 chips. If you want to do a DeepSeek 671, you might need 6,000 or 8,000 chips. What an administrative nightmare.
You can keep it on 1 wafer, 2 wafers, 4 wafers, or 10 wafers. You get all the benefit of the SRAM, and because you've been able to use the wafer, you get this tremendous capacity as well.
I totally get you on HBM and the slowness of it. Why is it, then, that so much of the market just continues to use it? Forty percent of NVIDIA's revenue is using those chips for inference.
There wasn't really, unless you went to wafer scale, a credible other choice. This is the way GPUs had always been made. It's called a graphics processing unit. That's the way they were built, and it was part of their advantage against a CPU: they were built this way.
But now there are dedicated chips like ours, and what used to be their advantage is now a weakness. That's a fun market to be in, when over a very short period of time what you're good at becomes your weakness.
With a market cap like theirs, and with Jensen as good as he is—which I'm sure we both agree with—they must know this.
They do know this. There aren't a lot of choices. They don't make memory, so they're a consumer of other people's memory. That's SK hynix, Samsung, Micron. There are only 3, 4, or 5 companies that make huge amounts of memory. There aren't many choices.
It's part of a complex architectural tradeoff. The flip side is that it's worked really well for them. Look at where it's taken them.
In comparison to those of us who are wafer scale, it's a small set. It's a set of 1: us. We have a real advantage against them on inference.
How do LPUs fit into this? We've got HBM, we've got SRAM with you, and, bluntly, we have many more of them to make it work and scale. Where do LPUs fit into this mix?
There are a lot of ways to skin a cat. Our way is different from NVIDIA's way. It's different from the TPUs, and it's different from Trainium. They're different.
Right now, and every day since August 26, when we launched inference, our way has been the fastest way across a whole set of models tested by Artificial Analysis and others.
When we think about that speed, you said that you're 1 of 1 with wafer scale and the associated architecture. What does that mean in terms of cost? With such efficiency, is it inherently more expensive, and what does that look like from a cost profile?
This isn't our first dance. We've been building computers for a long time, and when you make a choice like wafer scale, you have to weigh the tradeoffs.
We use less power. We use less power because 1 of the most power-hungry things on a chip is the I/O, moving data off-chip. If you're moving data off-chip frequently, you're using more power than if you can keep it in the silicon domain, on-chip.
We knew we would use less power. We knew that if you went to wafer scale, you had to solve some problems that people said were impossible to solve, like yield. We had to invent techniques that allowed us to yield wafers. In fact, we invented techniques that allow us to yield as well as, or better than, others who are building much smaller chips.
Can I interrupt and ask what yield is, and why is it impossible to solve?
A wafer begins as a 12-inch-diameter circle, a slice of silicon, and your chip is punched out of this. It's the way your mother might take a cookie cutter and cut out cookie dough.
During the process, at some point, just like your mom might have done, she lifts up the edges and all the little bits are removed. What's left are just the cookies. Those are your chips.
What happens is there are a set of naturally occurring flaws. It's like your mother closing her eyes and throwing up a handful of M&M's. The bigger the cookie, the higher the probability you hit an M&M. The bigger the chip, the higher the possibility that you have a flaw.
Traditionally, when you had a flaw, you threw away the chip or sold it as a less valuable part. You shut down part of the chip and sold it as a less valuable part, something called binning.
Every wafer is going to have flaws. The bigger your chip, the higher the probability you hit a flaw, and the more of the silicon is wasted when you throw it away. This is what everybody thought was known truth.
One of the things our team realized was that there are other ways to handle flaws. What if, instead, you built your computer—your processor—out of hundreds of thousands of identical tiles? If there was a flaw, you could shut down that tile and work around it. You could have a row or a column of redundant tiles that, when you needed them, you could pull in.
That had traditionally been the technique used in memory-making, and memory yields are extraordinary. It occurred to us that if we could build a processor out of hundreds of thousands of identical tiles, we could use redundancy. When there was a flaw, we could leave it there, shut it down, work around it, and pull in 1 of the redundant tiles.
That had never been done in a computer before, and that's at the heart of our architecture. It allowed us to yield and deliver whole wafers.
Nobody had ever been able to do that in the 70-year history of our industry. Really, really smart people struggled. Likely Gene Amdahl, one of the fathers of our industry, had a company called Trilogy that crashed and burned trying to do this. We figured it out.
When you speak about being the fastest, and across all benchmarks being the fastest, what matters the most? Is it being the fastest, being the most efficient, or being the least costly? How do you think about the stack of prioritization for your customers?
I think it varies. If you go to get a cancer diagnosis—for God forbid, your mother or your wife—I think 93% accuracy is just plain not as good as 94% accuracy. You pay a lot and wait another week to understand what the accuracy is. You pay a lot.
On the other hand, if you want Llama 405B to generate data to help you tune Llama 70B, maybe you can wait a few days, 3 days, or a week more. There's no urgency there.
If you want an answer from Perplexity, you don't want to wait 45 seconds for a search answer. You don't want to wait in a chat. You don't want to wait 3 minutes for R1 on GPUs to give you an answer.
In interactive mode, milliseconds matter. In interactive mode, what Google showed years ago was that you can destroy your user's attention with milliseconds of delay. Being the fastest matters in that domain.
You have to be thoughtful and say that in some cases being the fastest doesn't matter. We'll call those batch. Maybe cheapest matters there. In other domains, there is no search if you have to wait 8 minutes to get an answer. That's not a product.
When you go fast, a whole set of new opportunities open up. Netflix used to mail DVDs. That's what happened when the internet was slow: they'd mail DVDs. I look young, Andrew, but I'm not that young. I remember Blockbuster.
First, we used to drive to Blockbuster to get a DVD or a video. Then Netflix was mailing them to us. Then we got broadband, and suddenly Amazon is a studio. It changed everything, and speed in inference does the same thing.
When we chatted before, you gave this great equation for inference. What was the equation? It was really helpful for me in understanding it.
It begins with the following: training makes AI. That's how we make AI. Inference is how we use or consume AI.
Understanding how big the inference market is means understanding the number of people who are going to use it, how often they're going to use it, and how much compute each use takes.
Right now, we're in this rare time where the number of people using AI is growing, the frequency with which they use it is growing, and the amount of compute used in each instance of use is growing. That's why you're getting this extraordinary growth, and that's why it's off the charts right now.
When we think about the distribution of resources between training and inference, what will that look like in 5 years? We've seen a lot of focus go to training and not as much go to inference. What does that look like?
What we made in AI until the middle of 2024 was a novelty. What we made in AI late in 2024 began to be useful.
What was the turning point?
If you look at the models, they became useful. ChatGPT wasn't really a technical innovation; it was a user-interface invention. It gave more people access, but we didn't really know what to do with it right away. It was cool. That's what I mean by novelty: “Whoa, this is cool.”
Now, if your marketing team isn't on an LLM each person several times a day, they're not doing their jobs. The difference between “It's cool” and “This is part of everyday workflow” is what changed, starting sometime in Q4 last year and running into this year.
AI became useful not just to a select group in Silicon Valley, but to my dad, my brothers, doctors, and ordinary people who aren't buried in the Silicon Valley discussion. When you get them, the market is ripping.
Do you not still think we're incredibly early? Going back to your point, how many times bigger are we in 5 years? Are we 100 times bigger? Are we 1,000 times bigger?
I think we're way over 100 times bigger.
What does that mean in terms of what we need to equip ourselves to deliver these? They're incredibly energy-intensive, it's incredibly difficult, and our industry consumes a lot of power.
6. Why is it Total BS That We Have Hit Scaling Laws?
We're seeing some water usage come down, but are we equipped from an energy and data-center standpoint to deliver the inference requirements for a population that is as AI-hungry as we are?
The first thing is to admit that this is a power-intensive problem. Our industry consumes an enormous amount of power.
The second thing to say is that the burden is on us to deliver exceptional value as an industry. You take both the good and the bad. In order to make it worthwhile from a societal perspective to expand all this power, we better deliver the goods.
We better use AI to find cures for diseases. We better use AI to solve a bunch of different societal problems. That's the macro view.
Do I think we're equipped? I think we're in a very unusual situation in the US, where we have plenty of power, but it's in all the wrong places. We have power in Niagara. What we don't have is power where you want to build data centers, where we have good fiber.
What we also don't have is a national way to relax the local regulations that make getting power difficult. When you go to Silicon Valley, if you want to build a data center, you're dealing with local government and vested interests. That's not an efficient way to decide if you want to build a power plant or put a new data center in, especially if it's large.
I think those places that have removed some of that burden—for example, through taxes—are getting a huge number of data centers built.
When I spoke to Jonathan at Gro, he said there were a huge number of data centers being built that weren't really equipped properly. We've seen this massive supply of data centers that are done by tourists, so to speak, and that is a massive problem. The provisioning of these data centers isn't there. Do you agree?
A data center is, to begin with, a construction project. It's access to power, a construction project, and a design and engineering component.
I think there's been a huge push for new-construction data centers. We don't know if they're going to be good enough. Many of them will be fine.
The guys who were there early were some of the Bitcoin-mining companies, like TeraWulf, the guys at Crusoe, and others. There were guys in Europe who were early in building buildings near low-cost power in order to run compute that used a lot of power. They are some of the leaders now in some of the largest projects.
Those are certainly not tourists. They're extremely sophisticated data-center builders. There are some tourists, but there are a lot of very knowledgeable data-center builders building huge facilities right now—gigawatt-scale facilities, both domestically and internationally.
How do you think about how the cost of inference goes down with the surge of demand that we mentioned—over 100 times? Does the price reduce 100 times? Does it follow Moore's law continuously? How do we think about the ever-reducing price of inference?
The cost of inference is built up of several pieces. There's the power and space consumed to generate the response. That's a data-center cost and an OPEX item.
Second, there's the cost of the computer. We can drive down the cost of the computers with each generation by driving up their performance.
The other thing we can do is develop more efficient algorithms. Our AI algorithms today aren't particularly efficient. In a GPU, most of the time it's doing inference, it's 5% or 7% utilized. That means it's 95% or 93% wasted.
Over time, I think that as an industry we get better at things. We can drive the cost of compute down, build more efficient data centers with lower PUEs, and make our algorithms more efficient, so that utilization on our now-cheaper computers is higher.
You get a higher percentage of the maximum number of FLOPS. You get more tokens per unit time for the same power.
When you look at the inefficiency of the algorithms, and what that means for the utilization of the chips, why are people suggesting that we're at scaling laws already? That seems to suggest there is so much room for improvement.
How do you think about what you just said in conjunction with the idea that we're hitting this asymptote point?
I don't think there's a lot of debate among senior ML thinkers that we have tremendous room for algorithmic improvement. I don't think there's a lot of debate there.
There's even debate about whether the scaling laws are over, whether we've run out of mojo to keep making or gathering data to fill these ever-bigger models. But OpenAI's work on o1 shows me that the scaling law, certainly for inference, is fully functional. The more compute you put on inference, the better answer you get.
Many of the leading models are now MoEs, so they're not presenting all of the weights to each token. That's 1 way to do it: present the important stuff, not the unimportant stuff.
There are other ways to do it that we will invent and learn over time. We have human models that aren't all-to-all connected. Many of our models today are all-to-all connected. That's a lot of unnecessary connections, connections that don't produce anything but that we still end up doing math over.
What does “all-to-all connected” mean?
In many of the layers in a neural network, every element is connected to every other one. That's not actually the way the learning happens. Some are more valuable, and some are not valuable at all.
Imagine you're going to read 50 books because you want to learn something. You can read all 50 books, or you could read the 3 books that are really important, or you could read summaries of the 3 books that are the most important.
The problem is that we don't know which they are at the beginning. There's a process that you could learn. There are things called dropout and all these other techniques to use sparsity to help solve these problems.
We are early in the evolution of AI, and that plays right into this point that we'll get better at these algorithms. Transformers aren't the end of the world. We'll get better. Better will mean faster, more accurate, and more efficient.
That's what's exciting about an ever-changing industry. That's why I'm not in all these other industries that don't change quickly. They're the same 9 years ago as they are today.
This show is kind of strange to me because I speak to a lot of people, and they think about the 3 pillars—compute, algorithms, and data—and the common refrain is that we're actually very far along in all of them.
When I hear you, it's actually very exciting. I think they're wrong.
I think they're wrong. I don't think we're very far along, and it's very difficult to say that we're early in an industry but far along on all of its underpinnings. I think we are early in all of them.
If we take them 1 by 1, in 5 years, how much synthetic versus human data will be used to train models? If you had to put a percentage on it?
Almost all synthetic.
And is the utility value of synthetic data the same as human data?
I think so. When you teach a pilot to fly in a simulator, there is a lot of potential data that isn't very useful in teaching a pilot to fly. They spend a lot of time going straight and doing nothing.
Takeoffs and landings are where you want to spend your time, and that's why, when we put pilots in simulators, that's what we have them doing. In simulators, we can create data where engines blow, where there are a whole set of problems, and where learning can take place. That's simulated data.
In the same way, when we think about creating data—whether it's for self-driving or other forms of AI—we want the data that's hard to gather. Otherwise, we just have a bunch of data of people driving straight on a freeway. That's not difficult. We've been able to do that for a decade.
What we want is an unprotected left turn in the snow. It's snowing, it's hard to see, and you've got an unprotected left turn. That's a difficult thing, and you want that thousands or millions of different ways. That's where the synthetic data comes along: to fill in the empty parts where it's really expensive or painful to get that type of data.
Think of the pilot. You want them spending a huge amount of time on things that are rare in their training. It's the same with a surgeon: a huge amount of time on things that are rare. Most of the time it's carpentry, but their expertise is only needed when something rare happens.
That's when their mettle is shown, when the unexpected occurs. I think we will get better synthetic data by a great deal.
I get it from a consumer perspective and from an expectations perspective. If we move the needle on compute, algorithms, and data, what does that mean for the experience of AI?
It gets faster and cheaper. Faster and cheaper is the first answer.
The second is that when things become faster and cheaper, new applications emerge. It's used everywhere.
When computers became faster and cheaper, suddenly they were in cars, then they were in your pocket, then they were in your dishwasher and your TV. We were saying 30 years ago, “I need a computer in my TV? Are you kidding me? I need one in my pocket?”
Now you've got powerful computers in your pocket, in your TV, in your kids' toys, and in the car. That's what happens. Diffusion of innovation accelerates when you make things faster and cheaper.
This is Jevons's paradox and Satya's belief, isn't it?
I know that in the VC community you have to cite 19th-century English economists.
I'm English. I'm English. Come on. If I'm not allowed to cite an English philosopher, what am I here for?
I think there are very few examples in our industry—actually none in compute in 50 years—in which, by making things cheaper and faster, the market got smaller. The market always gets bigger. It always does.
From an architectural standpoint, you mentioned transformers. Is there a world where we move past transformers?
Transformers, 100%. We won't be as dependent on transformers in 3 years or 5 years as we are now—100%. They're not the end-all and be-all.
Why? What will replace them, and what does that look like?
I don't know whether they're going to be state-space models or other types of models. What I know for sure is that innovation doesn't stop, and the transformer has some weaknesses that people are desperate to overcome.
There's a quadratic effect in the attention head. There are all sorts of things that could be improved. But it's pretty darn good now. It's the best we have, and that's what you run with. You run with the best you have, and the minute it's not the best you have, you drop it in favor of the best you have.
I think that's what we're seeing. We're seeing a large number of innovative companies designing models.
What DeepSeek showed us is that you don't need 5,000 people and billions of dollars a year. You can do it with 200 smart people and more hardware than DeepSeek said they had, but less hardware than others had.
Were you very impressed with DeepSeek, and what impressed you most?
I think it was the result of focused engineering, and that impressed me. It was designed to be better. They weren't confused about being model intellectuals, and they weren't confused about whether it was important to break new ground. They were interested in being better.
From an invention standpoint, that's a little boring. From an engineering standpoint, that was a sweet effort. They really built a model that was just plain better at many, many things, and that's cool. I like good engineering projects.
They chose to announce it right around Trump's inauguration, and the politics of it are a separate matter. We can talk about that later.
Did distillation rile people up?
I don't think distillation is wrong. Is summarization wrong? I'm a VC. Are you kidding me? That's what we do. If you didn't summarize, you wouldn't know anything.
7. What Specifically Was So Impressive About DeepSeek?
Exactly. That's exactly right. I don't think distillation is wrong. If distillation is wrong, then certainly using people's copyrighted data is wrong. That's the problem. You've got to be a little bit consistent.
I think neither is wrong, actually, but you have to be consistent.
The thing with it, bluntly, is that DeepSeek is open. Everything that they innovated on, OpenAI can learn from and take too.
8. Why is Distillation Not Wrong and OpenAI Need to Look in the Mirror?
I think there are few examples of an open-source anything having the sort of immediate impact that model had. That model had a giant impact in a technical community of really smart people.
There are very few examples of other open-source software projects that had that type of impact in that amount of time. Usually, you're in the business of betting on these guys: they ramp up, and they go from 10,000 users to 100,000 users to 1 million users. Then you better start a company around that and get those graduate students.
This had a loud boom in the industry immediately. It was, “Whoo, the thing!”
9. Where Will Value Accrue in a World of AI?
The thing I have to think about as a venture investor is where enduring and defensible value is, and how I get in early and build that over time. In hardware, that's well understood.
But on the model side, do you think there is value when you look at the sheer number of players with relatively comparable models?
To demonstrate enduring value, you need both immediate value and a trajectory for more.
The problem in some industries is that you're capable of demonstrating a leadership position for a short period, and then someone else—maybe the next generation—generates the next one, and the next generation generates the next one.
I think that, in the software world, you end up competing against other people's release cadences. You're 4 months ahead; they're 6 months ahead. If that's really where you are, there's not a lot of value.
But if you can stay at the top over years—even if you're not the best, even if you're in the top decile over years, while the people above you are changing constantly—I think there's a lot of value.
Very large Silicon Valley companies have been built with technology that was not the most compelling. It might have started as the most compelling technology, and then it got to a point where it was good enough, easy enough to use, and well distributed. That's when you're at the mature market.
10. How Will NVIDIA’s Market Position Change Over the Next Five Years?
We're a long way from there right now. Right now, we're in the early phases. You characterized my position exactly right: data, compute, and algorithms. I think we have a ton of room for improvement on all of them.
You said that computing hardware is where the value is. How does that value distribution shake out? We've obviously got the 800-pound gorilla that is NVIDIA. How do you think about how the distribution of value shakes out in hardware and compute over the next 5 years?
Historically, 1 of the barriers to entry was the capital intensity of a project. In the world of building chips, there are both scarce resources in expertise and high expense.
Historically, it hasn't fit comfortably in a software company. The things that modern software companies value aren't entirely conducive to chip-making.
When I look down the road, I think that people who build systems endure. Cisco and Juniper endure. Chipmakers have endured. There's a reason Apple and NVIDIA are among the most valuable companies on Earth. What they do is hard, and I think that's why it's worth challenging.
If it weren't hard, enormous, and difficult, why spend time being the underdog and challenging it?
A lot of people place defensibility around NVIDIA's CUDA lock-in. To what extent is that real versus hype in inference?
In inference, it's not real at all. There's no CUDA lock-in in inference. You can move from OpenAI on an NVIDIA GPU to Cerebras, to the Fireworks service on something else, to Together, to Perplexity with 10 keystrokes.
Anybody who actually uses AI knows there's no CUDA lock-in in inference.
There was a fundamental effort to disintermediate CUDA, first by Google with TensorFlow and by some graduate students with Caffe, and later by Google with TensorFlow and Facebook, or Meta, with PyTorch.
Today, most AI is written in PyTorch. You ought to be able to compile it and run it on your hardware.
NVIDIA has many moats. When you're a dominant market-share leader, that in itself is a moat. Being the default solution is a moat. Everybody learns to think about AI in your structures. Those are moats.
The software—compilers are hard, but they're tractable.
I completely agree with you that being the leader is a moat in itself. It's never talked about that way.
Look at Intel. Intel has made nearly a decade of catastrophic decisions until hiring Lip-Bu Tan, and they still own 80% of the x86 market. AMD has worked up to perhaps 25% or 30%. After a decade of screwing up, Intel only lost 20% of its share. That's a moat.
That moat is just unbelievable. You can make a bunch of bad decisions for a decade and only lose 20% share. That's extraordinary.
I'm a huge fan of Lip-Bu Tan. He's an investor in our company, and I wish him well. I think if anybody can change that company, he can.
I think we rarely talk about what being the market-share leader means in terms of a moat. As a challenger, we have to think about it exactly, because it's exactly those characteristics of the moat that we need to get over.
In 5 years, is it Uber, or is it like AWS and cloud? Cloud is an interesting market where a couple of players have relative segments—25% or 30%—and it's shared relatively evenly between them. Or is it 1 like Uber, where Uber has 90%, Lyft has 5%, and alternative providers have the other 5%?
I think it's going to be between those 2. In 5 years, NVIDIA is going to have 60%, somewhere between 50% and 60% of the market. Right now, they have approximately all of it.
Of NVIDIA's usage, what percentage will be training versus inference?
They'll continue to have a meaningful business on both sides. They're exceptional at training. They will not roll over and play dead in inference.
They're a world-class company. They've had 1 of the great decades of any company in history. From 2014, when they were worth $10 billion, to where they are right now, it's 1 of the great decades in corporate history.
I don't think they're going to roll over and say, “We're not going to be in the inference market.” That's not going to happen. They're going to have a meaningful share, but the market is growing, and we'll have a piece. Others will have a piece. I think there'll be some very big companies made in this 100× growth.
Do you think chip providers will be far larger than model providers in terms of enterprise value in the 5-year timeframe?
Yes.
How does that prediction change on a different timeline?
In a shorter timeline, when you price an option, variance and uncertainty increase the option's value. If you look at the way Black-Scholes works, or at any option-pricing model, uncertainty and variability are friends of the value of the option.
When people are paying these extraordinarily high prices for model companies right now, I think part of that is this extraordinary uncertainty and wild variance. In the shorter run, it might not be the case.
In the longer run, as markets mature and we begin to understand the value of these models, their businesses, and their long-term net profitability, we'll have a better understanding.
11. Why is the CUDA Locking for NVIDIA BS? What is Their Weakness?
What did Warren Buffett say about markets? In the short term, they're a voting mechanism, and in the long term, they're a weighing mechanism. At some point, the weighing kicks in. Usually, it's in the public markets, and then investors say, “Which is likely to give me better growth in the future?”
You mentioned the word “public.” I do want to hone in on your business. You're cash-flow positive in a world where everyone else literally bleeds cash.
Help me understand: how did you become cash-flow positive when everyone else is bleeding or hemorrhaging cash?
Traditionally, gross margins were a measure of technical differentiation. If you're running a negative-gross-margin business, I think it speaks for itself. You're selling a commodity. Your value creation isn't being recognized in the market.
12. Why is Trump Better for Business than Biden?
I think our technology is creating an opportunity for us to maintain margins where some others can't.
A lot of your revenue is concentrated in the G42 deal. To what extent is that a strength or a weakness?
It's both. The way you catch 3 large customers is to catch 1 first. The way you build 3 large strategic partners is to learn to be a strategic partner. That's a learned skill.
We didn't arrive knowing how to be a strategic partner at G42. Now that we've worked at it, it's a muscle we can replicate. We could be a better partner to any of a dozen different companies in the world.
What have you learned in the G42 relationship-building process that makes Cerebras a good partner in a way that you weren't before?
We've deployed tens of exaflops of compute, vastly more than anybody else that isn't AMD or NVIDIA. That's a huge amount of compute.
Our software has been hardened on some of the largest AI clusters in the world. We've gone through the growing pains of increasing manufacturing 2×, 5×, and 2× again through unbelievable growth in manufacturing.
We've worked with our supply-chain partners to be sure that they're ready for this extraordinary growth.
When you work with a strategic partner of this size, your organization comes out different on the other side. There are things you've learned and mistakes you've made.
I hadn't done a big relationship in the Middle East. There was a huge amount to learn. I think you come out a much better company and much better prepared to do business with a hyperscaler, another massive partner, or another sovereign.
But it takes real work, and your team has to learn.
You said you come out better. Why go public when you did? When it happened, I thought it seemed preemptive, respectfully.
My question now to companies is: why go public at all? There is so much private capital. The decisions have shown very clearly that you can stay private for a lot longer than you planned to. Databricks has certainly shown that.
Those were historically public-market valuations, and the valuations that Anthropic, OpenAI, and some of the others are getting are historically public-market-only valuations. Your S-1 is live; anyone can read it. I wouldn't want people reading mine.
We have nothing to hide.
No, but your competitors have asymmetric information.
Yes, we've got asymmetric technology. I think you have to be pretty transparent to be public.
You have to be ready organizationally. You have to be ready with your processes. You need to be ready to forecast and predict, and to be held accountable in a way that private companies historically haven't been.
We think there's tremendous value. We think we'll be among the first in the category. We think some of our largest targets would have a stated preference for doing business with public companies. Large enterprises in the US have done that historically.
Those were some of the reasons that led us to it.
How many G42 relationships shall you have in the next 24 months? How fast can you ramp them?
That's a good question. Several.
Remind me, how big is the G42 deal?
It was 87% of revenue.
I know it was big.
When we announced it, some estimated it was north of $1 billion.
Well done. That must be a bit of a high five.
There's tremendous excitement, and then there's every entrepreneur's reality: I have to make a lot more gear.
You make a list of your top 10 vendors and fly it to them all, saying, “Big orders are coming. Be ready.” You work with all your partners to get ready because you need to make a great deal more stuff.
That's 1 of the real differences between hardware and software. When we grow fast, the number of people you need to work with in your supply chain, and the amount of collaboration that needs to happen, is truly extraordinary.
Are NVIDIA going to have a cluster of unhappy customers who, bluntly, have waited so long for chips that by the time they get them, the chips are outdated? Are they going to say, “What happened?”
All of that is an opportunity for us and others. Being a market-share leader isn't easy either.
When the bully falls, everybody wants to give him a kick. A lot of that happened at Intel. They'd been the dominant player, and when they fell, everybody was happy to jump in and kick them when they were down.
I think there's a real opportunity in the potential for NVIDIA customer unhappiness. For those of us competing with them, if you can't get your gear, you may as well test somebody else. That's a huge opening.
In hardware, you mentioned the complexity. Are export controls being implemented properly? Do you think they're a good idea?
13. Quick-Fire Round
Everyone was looking at DeepSeek and saying, “How did this happen? They must have stolen chips. How could this be?” What do you think about that?
It turns out that they probably did use chips in Singapore.
I think managing software compliance and managing hardware compliance are extremely different things because their vectors of diffusion are different. There's a different weight.
If you sell a server that weighs 500 or 600 pounds and arrives on a pallet, you can go visit it. If you want to deploy it in Kazakhstan, you can put it in a data center and have somebody from the embassy visit it and take photos once a month. It's not going anywhere.
You can keep track of who uses it and provide logs. That's much harder with software, and open source is a whole other level.
That's the first observation. The second is that I got to know the leadership in Commerce in the previous administration. I didn't always agree with their policies, but it is a world of unintended consequences.
You sought to limit Chinese access to EDA tools to delay the growth of a Chinese chip market, and US venture capitalists backed tons of Chinese companies in Shenzhen to build EDA tools. This is an unbelievably slippery, dynamic, challenging problem.
I don't know if it's a tractable problem. Delaying another nation's progress on a technical trajectory is an enormously challenging thing.
I came to appreciate just how difficult it was for well-meaning people to predict the impact of policy during the last 2 years.
Do you think this administration is better for AI than the prior administration?
I don't think there's any doubt that's the case.
What makes you say that?
The past administration lined itself up against Big Tech, and that was a mistake.
AI is also in a different place, so it's easier to be for it. It's less scary now than it was in 2021. We have a better picture of the trajectory, both the risks and the benefits.
I think this administration had the foresight to put an AI czar or leader in place as a focal point for discussions. It's probably net a fair bit better.
You said it's very challenging to hinder a nation's development, adoption, and progression of a technology. Respectfully, you chose not to sell to China. Why was that, and does that not go against the difficulty of hindering progression?
I have a very simple rule, and I encourage your team to use it. You don't need a big handbook to help you make good decisions in a company. Just ask yourself, “Would my mother be proud?”
Would she be proud if I did this? Would she be proud if I explained the exact situation? Would she look at me and say, “I'm proud you're doing this, son?”
I asked myself that, and I came to believe that the deal on the table wouldn't be used for good. I wasn't comfortable with that. I wouldn't have been able to explain it to my mother.
That's a moral compass. It wouldn't have been used for good.
I'm naive. What would they use it for?
They could use it to power drones. Some use it for facial recognition to identify minorities for persecution, or to build military equipment—to do things that I either couldn't see or that, from what I saw, didn't feel right.
14. Do We Underestimate China in a World of AI?
It's more important than money.
Do you think we fundamentally underestimate Chinese capabilities?
100%. It is 1 of the most obvious and frequent errors in judgment: you underestimate the other side.
You have to look carefully at what they're doing. Their investment in infrastructure has been extraordinary. The rate at which they generate engineering talent is exceptional.
The government's ability to have a policy and implement it is extraordinary. They're not a democracy; they weren't designed to have checks and balances there.
The funding that flowed into the development of AI technology was significant. Their venture capitalists were backed by their government. They have national champion companies. They've developed a belt-and-suspenders strategy to make much of the developing world dependent on them and their technologies.
They absolutely should not be underestimated. They have a lot of people, and we see a tiny fraction of it.
I think they have produced industrial policy that has moved their nation forward.
What was the most significant part of that, do you think?
The creation of economic zones like Shenzhen was clearly a visionary move. They knew that their own system was in the way, so they created zones that relaxed their own system.
Could the US learn from them in that way?
We did some of the same things in the 1st Trump administration. What did we do? We relaxed our own rules in the development of vaccines. We knew that, in that time, it would be very difficult to go through the steps that we always go through, and we tried to implement thoughtful shortcuts, or workarounds.
Why are they committed to trains as a mode of transportation, and we can't build a decent train system in the US or in California? Why can't we build infrastructure when the rest of the world can build extraordinary high-speed trains linking important cities?
Why do we have 3 different standards for train rails? Why are our bridges and freeways in disarray?
I think those are questions we have to ask ourselves when we see other people doing it differently.
If you watch a good football team, you say, “That's interesting offense.” You're not thinking, “How could our team learn? What could we do? Why did that work? What was it about the people they had, the talent, or the structure that made that a successful series of plays?”
What can I take away from that? How can that inspire me to do better?
I'm always looking for inspiration in others, competitors, and partners. Some of our partners at G42 have an unbelievable work ethic. It inspires me. The scope of the challenge they've undertaken inspires me.
What do you believe that most around you disbelieve?
I think we're closer to peace in the Middle East than people believe.
There is a rise of a moderate, business-focused Arab state that wasn't there 25 or 30 years ago. If you visit the UAE, Qatar, or even Saudi Arabia, what you see is amazing transformation.
I think there's a desire to be included in the West in their own way, but also to enjoy the benefits of it. We are closer than people may think.
What's the most underrated threat to NVIDIA's market-share dominance?
The fundamental architecture of the GPU with off-chip memory is not great for inference. They will continue to do well in inference, but they can be beaten, and I think they know it.
What's a crazy AI prediction you have that most people would call science fiction?
Dario at Anthropic says that we'll live to 150. I don't think we're going to live to 150.
I don't think that 90% of our code will be written by machines this year. But I do think that within a year or 2, AI's penetration will be approximately the same as telephones—cell phones.
What have you changed your mind on in the last 12 months?
There are lots of things. Many decisions I made turned out to be wrong.
There are 2 ways you can be wrong. You can actively be wrong, or you can fight against what was right.
In 2016, JP, 1 of our co-founders and chief system architect, laid out a plan that would have us doing water cooling for our systems. Nobody else was doing it, and I fought so hard. I was so wrong.
JP was right. About a year or 2 later, Google announced that the TPUs were going to be water-cooled. We were 1st, and now NVIDIA is only selling water-cooled parts.
I was dead wrong, and JP was right.
When you make a lot of decisions every day, there are many instances where you're wrong. I've been wrong about people. People I thought were pretty good turned out to be extraordinary. People I thought would be extraordinary were really smart but couldn't finish projects or get things done.
If you're not prepared to be wrong a fair bit, you ought not to be making a lot of decisions, because it comes with the territory.
As a venture capitalist, I'm never wrong, so—
As a venture capitalist, you're wrong 9 times out of 10, and everybody forgets as long as you're really right.
I get a picture of you signing the term sheet with me.
Good. Yours is a perfect industry in which nobody cares about the average. On average, you're wrong all the time. What they care about is the occasional time you're really right, and that moves a fund.
That's different from being a CEO. I think we've got to be mostly right most of the time, but if you're making a lot of decisions, you're still making a ton of mistakes.
This is your 5th startup. You are a sucker for punishment, aren't you? Really—5 times? Did you not get beaten alive enough?
My question is about the value of serial entrepreneurship. I've spoken to many people who don't believe in it, respectfully. How do you think about the inherent benefits of having done it 4 times before?
If you're in a business in which running a business is a benefit, then experience matters a great deal.
If you're in a business in which you look like your customer, there was a reason social networks were started by people right out of college or in college. Dating is top of their mind. They look like their customers, and that was more important than knowing anything about running a business.
In that environment, it will select for people who are of the demographic their customers are. They know that backwards and forwards.
But if you want to have a business that has manufacturing, a supply chain, and hundreds or thousands of engineers managed to a timeline and a schedule, I don't think anybody would turn around your statement and say with a straight face, “What I'm looking for is an engineering leader with no experience.”
You wouldn't want somebody who had only led a team of 400 or 500 people. You'd want somebody who had experienced the challenges of growth.
No, I don't want somebody who's led a team of 400 or 500. What I'm looking for is somebody with no experience. Naivety is a bonus here.
The people who sell that are sometimes consultants. “My guys have no experience in your industry, so they aren't biased.” Maybe a little bit of experience in the industry would help.
Where are people investing today in AI across the stack where you're thinking, “Why is so much cash going to that part?” I'm not asking for a company; I mean a category.
Sometimes money needs to find a home. Some people have raised really big funds, and they have to find a home for that money. Some people don't like to be left out. They're willing to make investments for status purposes or other reasons that don't seem to make sense.
I haven't thought about it in detail. I think there are some underappreciated places of investment.
In the chip world, the sub-milliwatt, tiny little chips that live next to sensors and do inference are an extremely interesting market. These are tiny things that will only send back useful data, and they'll sell in enormous volume.
It's not a part of the market I love to play in. I like to build bigger things and sell them to the data center. But I think that part is extremely interesting. I think it will be fundamental for robotics.
That's an area where I think the opportunity is extremely underappreciated.
If we think about Cerebras in 10 years, where do you envision the business? If everything goes well, where are we in business having that conversation?
10 years ago, NVIDIA was worth $10 billion, so that's a long run in our world right now.
In 3 to 5 years, I would like our technology to have been used to solve 2 important societal problems. I would like it to have been used to find a therapeutic for an affliction that impacts more than 1 million people a year.
I would like our inference to be powering a collection of apps that don't exist today. I would like a meaningful portion of the population in the US and Europe to inadvertently use our technology—to use something that we power without even knowing it.
Andrew, I've wanted to make this show happen for a long time. As I said, I've heard so many good things from Marc for many years.
There have been so many requests to have you on the show. My team was saying, “Just get Andrew on the show, Harry.” I was like, “Okay, okay.”
I tweeted it, obviously, which is how we got this conversation. You tweeted it, and around 40 people sent me messages saying, “How come you're avoiding Harry? How come he has to tweet it?” I was just like, “All right, just call me.”
It's good. Send me a note. I'm happy to come on.
Really thoughtful questions, Harry. Really thoughtful and interesting. It was a really good conversation.