BlackRock 的 Tony Kim：AI 下一轮赢家？
- Kim 将2023年标记为 AI 从BCE跨入AD的分界点——“就像BCE、Anno Domini……砰，’23年发生了，一切都变了”，并称市场价值重心已经转向硬件。 他的粗略估算是：软件、服务和互联网约10万亿美元，Mag Seven为22万亿美元至23万亿美元，芯片与硬件超过30万亿美元；“在AI之前、BCE时代，可能正好相反”。底层算力规模约提升1万倍：“1万美元的服务器变成了100万美元的服务器。”
- 数据中心正围绕底层物理规律重建：今年资本开支1万亿美元、未来5年10万亿美元，只为“把数据移动几厘米、几毫米。这就是 AI”。 数据传输距离从公里级压缩到毫米级，但距离越短，带宽、功耗和发热越高，迫使行业从“铜走向光”、采用800V供电架构，并最终转向固态变压器。
- 存储将成为下一主导力量：RAMpocalypse显示，芯片和模型正越来越像人脑，而 Tony Kim 称人脑“存储能力可能远高于算力”。 他认为，智能体框架和机构记忆正在推高存储强度。Molly 提到目前只有3家主要存储芯片厂商，并称 SK Hynix 即将上市，同时承认节目发布时它可能已经上市。当前短缺下，晶圆厂需要3–4年才能建成，形成“需求、供给、周期的错配”，加剧市场焦虑。“今天所有人都在谈算力、算力、算力。我认为，存储的重要性还会进一步上升。”
- 利润率链条已经反转：传统云计算本质上是“转售 CPU 加硬盘”，靠高 SaaS 利润率赚钱，而 AI 算力工厂出售 token、挤压应用层。 模型“不管你愿不愿意，已经从软件和服务业吸走了市值……就像Borg一样”，这正是 SaaSpocalypse 的由来。如果规模化仍能以每年一个数量级推进——“10乘以10再乘以10，3年就是1,000倍”——防御性就必须重新定义：“护城河总会被攻破，不是吗？所以更重要的是进攻。”
- 在组合构建上，Kim约90%的投资配置都投向AI未来3年的“漩涡”，但他也必须相信这条主线之外仍有未来，而许多前沿主题的时间窗口都指向2030年。 百万量子比特、具备逻辑纠错能力的量子计算机、获监管批准的SMR、轨道数据中心、800V，以及不同版本的AGI预测，都集中在这一时间段。他判断退出倍数的标准是：“你总要押注那些今天不酷的东西。5年后它还会酷吗？”
- 在机器人领域，Kim看到亚洲制造业的重大机会：“中国人来了”——中国有130–140家机器人公司，今年潜在IPO数量为30–40家，而美国只有“0家、1家、2家”。 可能的终局是“把中国的实体机器人与西方的大脑混在一起”，而且他表示自己知道这种组合正在发生。他的判断是，最大的突破口可能是陪伴型机器人：亲和、约半人高、类似 C-3PO 的机器，拥有“莎士比亚和爱因斯坦的智慧”，服务于孤独、教育和老年照护；亚洲生育率已低于1.0，而替代率为2.1–2.2。
- 未来12个月，他认为“每6个月就有一次恐慌”（SaaSpocalypse、RAMpocalypse、融资、资本开支），但建设热潮“会一路碾压过去”，基础模型实验室 IPO 和轨道数据中心的验证进展是值得关注的信号。 他也明确表示，当前判断只是阶段性观点：“如果6个月后我们再聊，结论可能完全不同。”
1. 2023是从BCE到AD的分界点——市值重心转向硬件
- Kim 对时代的划分构成本期讨论的主轴：“AI发生了。就像BCE、Anno Domini……砰，’23年发生了，一切都变了。”底层算力规模约提升1万倍——“1万美元的服务器变成了100万美元的服务器”——硅谷也从“软件主导的世界”转向“算力主导的世界”，模型和算力变得“共生且彼此等同”。
- 他对约1,500家市值超过10亿美元公司的粗略地图是：“10、20、30”——不含 Mag Seven 的软件、服务和互联网公司市值超过10万亿美元，Mag Seven 为22万亿美元至23万亿美元，芯片和硬件超过30万亿美元。“我不认为人们意识到，我们现在已经如此以算力硬件为中心……在AI之前、BCE时代，可能正好相反。”
- 这次价值转移的机制在于：“模型本身，不管你愿不愿意，已经从软件和服务业吸走了市值。就像Borg一样。”而模型要存在，就必须寄居在算力堆栈之上。
2. 云计算经济学反转：从转售CPU到出售token
- Kim 对2000–2020年云计算时代的拆解是：“所有人都说那是软件，但本质上不过是在转售 CPU 加硬盘。”算力当时只是被忽略的底层环节，藏在高额 SaaS 利润率之下，支撑了云计算和 SaaS 长达20年的增长。
- 新数据中心“对现在的数据中心而言就像外星来客”：收入正从轻资产、高利润率业务，转向重资产、低利润率、以大额 token 收入为特征的工厂，顶层堆栈因此被大幅压缩利润空间，也引发了 SaaSpocalypse。“末日场景很多，你知道的。”
- 如果规模化定律仍以每年一个数量级推进——“10乘以10再乘以10，3年就是1,000倍”——能力就会持续复合增长，问题也随之变成利润和防御性究竟存在于哪里。他拒绝传统框架：“护城河总会被攻破，不是吗？所以更重要的是进攻。你能不能跑得更快？”
3. 物理规律：1万亿美元，只为把数据移动几毫米
- 数据中心的传输距离正在从公里缩短至米、厘米、毫米；每缩短一个数量级，“带宽上升，功耗上升，发热上升”。他所描述的反差是：“今年1万亿美元的资本开支，以及未来5年的10万亿美元，竟然只是为了把数据移动几厘米、几毫米。这就是AI。”
- 他关注的后果包括：行业“正从铜的时代走向光的时代”；围绕表后发电和电网展开的电力革命；800V的崛起；以及最终采用固态变压器——因为“每次降压都会损失效率和能量”。
- 另一个结构性主题是协同设计：硅片与模型规格紧密匹配，模型也反过来适配芯片，这是许多头部基础模型实验室正在追寻的新路径，也是他在 Broadcom 讨论新款“Jalapeño”芯片的主题。他参加的4场 RAISE 圆桌分别聚焦 D-Matrix 的加速器、PsiQuantum 的量子计算、Lumentum 的光学，以及 Broadcom 的 XPU 和 AI 协同设计。
4. RAMpocalypse：芯片和模型正向人脑收敛
- Kim 表示，芯片和模型的发展“开始映射人脑”。早期模型主要是并行算力，内存很少；如今个人 AI、智能体框架、机构上下文和存储记忆不断增加，模型对存储的需求也越来越高。他说，人脑“记忆存储能力可能远高于算力”，但也承认这取决于如何看待突触和神经元。如今的芯片架构“本质上都是以某种方式在内存和算力之间进行调度”，涉及 SRAM、DRAM、堆叠式 DRAM、HBM 和高带宽闪存。
- 真正可交易的矛盾在于建设周期：“建一座芯片晶圆厂或存储晶圆厂需要3到4年，但今天就已经短缺。”需求、供给和周期的错配“正在给市场带来很大焦虑”。他的结论是：“我认为，存储的重要性还会进一步上升。”
- Molly 的铺垫是，目前只有3家主要存储芯片厂商，且 SK Hynix 即将上市，同时她也承认到节目发布时它可能已经上市。这引出了投资判断的关键问题：建立在供给受限基础上的需求溢价，还能持续多久？
5. 资本配置：90%押注漩涡，但多个主题指向2030年
- 他的三层框架中，约90%的投资配置进入AI未来3年的“漩涡”——谁在赢、谁在输、谁在上升、谁在停滞。但即便在这部分投资中，“也必须押注未来”，因为对第4年至第6年的信念“会对估值倍数产生巨大影响”。看似便宜但正在萎缩的资产，未必是资本配置的最佳去处。
- 他的前沿投资经验早于生成式 AI：2019、2020和2021年，他就已经在做 AI 投资，“当时你并不知道 LLM 这件事会发生”，依据只是相信某种形式的 AI 算力终究会被需要。6、7年后，“AI加速器战争已经开始”。
- 他反复观察到的模式是：“所有道路都汇聚到2030年”——公用事业级、具备逻辑纠错能力的百万量子比特量子计算，获监管批准的 SMR，开始规模化的轨道数据中心，800V，固态变压器，以及从2028年延伸至2030年的 AGI 预测。他衡量一切的标准是：“5年后它还会酷吗？……市场愿意支付的倍数会下降，增长率也在放缓，而你就陷入了困境。”
- 他称这是自己当前的判断，并表示如果6个月后再次交流，结论可能完全不同。
6. 芯片从来不是商品——“权力之戒失而复得”
- 他重新审视半导体行业：“硅谷之所以叫 Silicon Valley，总有原因……人们忘记了。”芯片拥有“所有行业中最高的盈利能力——利润率高于软件、制药、工业、通信，任何行业都比不上”。直到最近，风险资本一直没有给这些公司融资，最终数百家公司收缩为少数幸存者；在他看来，这形成了近似双寡头的格局和巨大的定价权，“与商品完全相反”。
- AI催生了硬件复兴：“服务器很酷，光纤很酷，电力很酷，机架设计很酷……材料科学很酷。”真正的瓶颈是人才——掌握“失传的模拟计算技艺”的人“就像铁匠”。一家最大存储芯片公司的员工曾对他说：“我们找不到人来设计定制存储……得把一部分软件程序员重新培养成存储协同设计架构师。”
7. 机器人：140家中国公司、西方大脑，以及对孤独的押注
- 他对机器人的拆解是“一具身体里装着两套大脑”：一套负责运动和感知的世界模型，另一套是作为翻译器的 LLM；实验室先开发模型，再将其具身化。手部“可能是最难的部分”，但机器人身体本身“是一门制造业硬件生意”。中国正在进入这一领域，拥有130–140家机器人公司，今年潜在 IPO 数量为30–40家，而美国只有“0家、1家、2家”。部分原因是中国较浅的私募市场迫使企业更早上市，“而且它们正成批涌入”。
- 按 Kim 的框架，西方可能在模型开发上领先，但日本、韩国和中国构成的亚洲制造业体系，可以借助类似电动车的规模化实体制造，降低机器人生产成本。Molly 将能搬起冰箱的液压 Atlas 与 Figure 侧重包裹分拣和商业作业进行对比；Kim 则纠正说，Atlas 是韩国机器人，归 Hyundai 所有，具体通过 Boston Dynamics 持有。他确认了一个可能的组合：“可以把中国的实体机器人和西方的大脑混在一起，我知道这种事情正在发生。”
- 他关注的重点不是工业制造机器人，而是社会化具身：解决孤独、教育和老年照护问题。亚洲生育率远低于1.0，而替代率为2.1–2.2；养老院运营商已经是全球表现最好的公司之一，这些人口趋势支撑了陪伴型机器人的机会。理想产品应当亲和、约半人高，更像 R2-D2 和 C-3PO，而不是 Terminator；它“拥有莎士比亚和爱因斯坦的智慧，会说所有语言”，能够带着共情与他的母亲交谈，并记录她的人生经历。
8. Token流是仓位检验，PE滚雪球是下一个冲击
- 他构想的未来企业结构是：token流入，进入承载专有数据和第三方数据的数据层；再进入记录公司累积知识的上下文层，Molly 提到 Palantir 可能将其称为 ontology；然后“agents go wild”。投资规则很直接：“要么你创造 token——算力；要么你服务 token——基础模型实验室；要么你围绕 token 加上 harness、封装和上下文——应用服务……如果你不在这条流里，就有问题。”应用公司正在寻找自己的位置，一些推理和边缘云公司则已经插入，成为 token 的最后一公里供应商。
- Molly 沿着下游展开：语音 API 公司如 AssemblyAI，Databricks 和 Snowflake 的数据层，以及 MongoDB 等数据库。Kim 表示自己投资了其中很多公司，并同意它们都处在 token 流中。
- 另一种路径是完全抽象化：“我来接管你所有的保险业务。你付我 X……你过去付100，我只收20。”对于 Molly 提到的 PE 滚动收购案例——General Catalyst 的 Creation Fund，以及据报道 Long Lake 收购 American Express Global Business Travel——Kim 的坦率保留是：收购拥有客户的公司、剥离低效环节“也许是一门生意”，但“现在还没有定论，我在观察”。更广义的判断标准是：500人的公司创造出10,000人规模的收入，迫使 incumbent 调整战略或出售。
- 他最后表示，尽管每6个月都会出现一次恐慌，建设热潮“会一路碾压过去，我们的恐惧会逐渐消退”；他期待基础模型实验室上市，因为“市场胃口巨大”，也期待轨道数据中心取得进展，这“可能对地面数据中心建设产生巨大影响”。至于导师，真正塑造他的人是他在图书馆研究过的历史人物——Caesar、Alexander、Napoleon、Beethoven 和 Churchill——以及那些愿意给他机会的人。“逝者和善人。怎么样？”
完整逐字稿
AI happens. It’s like BCE, Anno Domini, and bam, ’23 happens. Everything changed. So the base layer of compute went up 10,000×. A $10,000 server is a $1 million server. There’s roughly $10 trillion-plus in market cap in software, services, and internet. There’s $22 trillion or $23 trillion in the Mag Seven, and then there’s another $30 trillion-plus in chips and hardware.
I don’t think people realize that we are that compute-hardware-centric. Before AI, in the BCE era, it was probably reversed, and so you’ve seen, in the last 4 years, a transformation in value that has systematically been happening for the last 4 years. The realization came to me that when we hit the AD era—the AI era—you needed to rethink everything.
The Chinese are coming, and there are 130 or 140 robotics companies in China. I see 30 or 40 potential IPOs this year in China alone, and there are what? Zero, 1, maybe 2 in the United States this year.
Tony Kim, welcome to Sorcery.
Thank you. Pleasure to be here in Paris.
In Paris, at the RAISE Summit. We’re in a secret off-location that has AC and some croissants, so it’s quite nice.
I know. It’s beautiful here. It’s so classic French. I love it. It’s fantastic.
You’re on stage a bit this year. What are you covering?
I’m doing 4 panels in my involvement with RAISE. One is on accelerators with D-Matrix and next-generation compute architectures. Another is with PsiQuantum around quantum computing. A third is with Lumentum, bringing optics to next-generation data center design. And finally, with Broadcom, on XPU and AI co-design for chips.
Light agenda.
Light agenda.
As the head of global tech for BlackRock, what brings you here? How did you get involved with RAISE?
I got involved last year. One of the companies I was involved with, SambaNova, Lipu was supposed to be one of the speakers, and he couldn’t make it, so I decided to fill in for him. Then I saw RAISE, this thing in Paris at the Louvre, and I thought, “Oh, this is interesting.”
It was the second year of development of the event that Henri had pioneered and built. I saw something there. I saw a lot of my colleagues and friends from San Francisco all congregating here in Paris, and I said, “I’d like to help foster this and get it going.” So last year I was here, and this year it’s probably tripled again in size.
It seems to be Europe’s biggest or most targeted AI conference. I continue to help and do what I can to build an AI presence in Europe.
It’s massive. I don’t know how they get all the names that they get, but I remember last year seeing Eric Schmidt on stage, and I had not heard of the conference before. I was just amazed.
No, and I think they’re outgrowing it. I think next year it might outgrow the Louvre even. It’s really becoming something, and you’re here—a testament to where it’s come.
They’re amazing, so I’m happy to get involved in any way that I can.
Between all of the panels that you’re doing, what are the through lines and the macro themes?
1. The Primacy of Compute
Through lines and macro themes. Okay. I think one of the big ideas, clearly, is what we see in the stock market and the investment market: the primacy of compute. You’re seeing already how much the stock market and the capitalization in Silicon Valley have changed.
We went from a software-centric world to a compute-centric world, and you’re seeing the emergence of companies. To me, the models and the compute are symbiotic and synonymous with each other, so that’s the primacy of compute.
2. The Data Center Rebuild
Secondly, since that is now the dominant theme—where the CapEx, the money, and the capitalization have all gone—it engenders a whole rethink of the data center. I think there is a redesign of the data center, and we’re going through stages of the data center rebuild.
Think of data centers pre-AI, kind of like the scramble to build data centers today, where there’s a massive shortage of compute. On the other hand, we’re hitting the laws of physics, which are driving yet another transformation of data center design going forward. The ramifications of this data center redesign flow through every layer of the AI stack.
Then you go another layer deep in this data center redesign. One of those is power density. It’s incredible what’s happening: as we put in more and more computation, the requirements of these models are requiring more and more compute density. You need to pack more and more bits into a smaller footprint, and when you do that, bandwidth, power, and heat become issues. You have these logarithmic effects of data center design driven by AI necessity.
Effectively, the data center is changing. Before, we would transmit data over kilometers; now it’s building to building, then within the building, then rack to rack, then within the rack, and next it’s within the chip. You’re going from kilometers to meters to centimeters to millimeters.
As you go 1 order of magnitude smaller in distance, the bandwidth goes up, the power goes up, and the heat goes up. That’s the irony of it all: the $1 trillion of CapEx this year and the $10 trillion over the next 5 years are coming to move data centimeters and millimeters. That’s AI.
When you think of it in that context, how does the data center need to change? Everything being done is to optimize around these new physics. A lot of the action is going to be around energy, power density, and the grid. Another part will be around chip architectures, and then another will be around the movement of data.
We’re going from a regime of copper to a regime of light, and these things help address the power density, energy, and laws of physics that are pushing data center design to its very limits.
One more thing I would say around AI and this conference is the symbiotic relationship between compute and LLMs and AI models. The best models, obviously, are trained on and built on the best compute and most optimized inference.
But this co-design—the notion of tightly integrating the design of your silicon to match the parameters and specs of the model, with the model specs informing the design of the compute—is the new path that many of the leading foundation labs are pursuing. This is what I’ll talk with Charlie at Broadcom about. Obviously, they did that with the new Jalapeño chip that recently came out.
Those are some of the big ideas. My panels today are mostly all focused, as you can tell, on the physical layer of AI: the compute stack and the co-integration. I’m not doing panels on the software layer at this conference.
It’s okay. Software is a little sleepy right now.
Yeah. A lot is happening. It’s interesting.
There’s a whole other revolution going on in energy and power around the grid, behind the meter, and all kinds of new generation sources. I’m sure you’ve done stuff around nuclear SMRs and things. A new power architecture—the rise of 800 volts—is going to have a transformative effect, and ultimately, solid-state transformers.
Again, this goes to the point that not only is the chip layer changing, but the energy layer is changing as well. Every time you have these step-downs in voltage, you lose efficiency and energy. So, again, things are about making it more efficient, packing more in, reducing the distance, and getting things up.
This is a sub-narrative that is going around in the future design of data centers. That’s interesting. I think you mentioned one thing: RAMpocalypse. Did you say “RAMpocalypse”?
You said that, by the way. We were on a call beforehand, and you brought up RAMpocalypse.
Yeah.
Yeah. Somebody—it wasn’t my quote.
3. The Memory Supply War
There are a lot of the memory companies here. But this also goes to—maybe I’ll add a fifth or sixth topic—around the data center design of the future. Again, it’s back to this co-design element around the model and the compute.
From my observation—I’m not building the models, but I observe the compute architectures and what the model guys are doing—more and more chip and model development is starting to mirror the human brain. In the initial early days, we had a ton of compute, a ton of parallel compute, and the models didn’t have much memory. Now that we’re adding memory to the models, they’re remembering things about your behavior and what you’re doing.
And then you look in the future: everyone talks more and more about having your personal AI, these agentic harnesses that build in institutional context within an enterprise, along with memory and memory storage. The human brain has a lot more memory storage than maybe compute. It depends on how you look at synapses and neurons and things, but the human brain is very memory-intensive. Today, AI is more compute-intensive. But as you see with RAMpocalypse, the memory intensity has just skyrocketed.
Mm.
Going forward, you’ll see more and more and more memory. When you look at the chip architectures, it’s all about arbitrating memory in some form with your compute. Different kinds of memory—SRAM, DRAM, stacked DRAM, HBM, or high-bandwidth flash—all of these memory and storage methods are tightly packed with your chip and your computer architecture, so that they align with how these AIs may be built to start to emulate more and more of the human brain.
I think that’s what’s fueling RAMpocalypse—the shortage of RAM. The other thing about that is that there’s a mismatch in what I call duration. It takes 3 or 4 years to build a chip fab or a memory fab, but there’s a shortage today. You’re trying to build for the future, and it takes 3 or 4 years to build the capacity, but what do you do about today’s demand?
You’ve got to spend so much money to get chip output, and so there’s this mismatch of demand, supply, and duration. This is causing a lot of angst in the market. But underlying all that, I think we are going to have more and more memory—or let’s just say, memory in concert with compute. Today, we’re all talking about compute, compute, compute. I think the primacy of memory will become even more important.
So we’ve been seeing this trend in a lot of our conversations and on the macro side, in the news and everything. We had a conversation with Jaman from Kotu. He’s the CIO of public markets over there, and they were talking about how, with the proliferation of agents, memory is only increasing more.
They also talked about the chip flip, which we’ll talk about a bit. But on the memory side, there are only 3 main players.
There are only 3.
And SK Hynix is about to go public. I don’t know if, by the time we put this out, it might be public. But the big questions around that are: How do you underwrite that, because the demand premium is massive due to limited supply? How long does that last, and how do you catch up to that?
I’d love to go deeper into all of these topics a bit more. But I guess, to start more on the macro side, we are entering the new era of AI. This has necessitated an entire rebuild of everything that’s going on in tech because we need inference, we need things faster, and agents are now coming to market. It’s no longer just chat.
I’m curious, from your standpoint on the investor side, how do you think—and how has your strategy evolved—to now play offense on this type of field?
4. Investing Across Time Horizons
I like your framing: It is a complete rebuild. So let’s start with that, and then how we play offense as an investor.
You’re absolutely right. It is a complete rebuild. The internet as we know it was built, let’s just say, from 2000 to 2020-ish, in one framework, which is basically around the birth and dawn of cloud computing. Cloud computing necessitated a certain kind of data center. You remember the good old classic data center: megawatts, not gigawatts, right? So we have an order-of-magnitude increase today.
These data centers were small. At the end of the day, cloud computing—everyone says it’s software, but it was really reselling CPUs with hard drives. That was the compute stack: a CPU with a hard drive. It was considered a commodity.
Server prices were tens of thousands of dollars, and now those compute servers are millions and tens of millions of dollars. So compute was an afterthought. The other way I think about it is that these clouds built these classic compute stacks, and then they resold that as platform services and databases, with SaaS built on top of that.
The base unit to create a cloud was relatively small: CPUs and hard drives, right? Basically, that was it, and some databases. Then SaaS was king, and the margins went to that because the cost of compute was so low. Everyone said compute was free, cheap, and a commodity, and all the value went to this layer, right?
That’s what cloud computing was. You were reselling—you were building this massive application on a very thin layer of compute. That fueled a 20-year run in cloud and SaaS. All of that data center infrastructure that was built, including AWS, GCP, and Azure, was built for that. It was built to bring on-prem software to hosted cloud services.
That was a great business. Everyone was very happy, and data centers were built to that spec. AI happens. It’s like B.C., A.D. 2023 is going from BC to AD. Bam—’23 happens, and everything changed.
What was called the base layer of compute went up, I don’t know, 10,000 times. A $10,000 server is a million-dollar server. Small HDD, big HDD. Oh, and by the way, DRAM was a commodity only used in smartphones. Now I need to pack all the DRAM and HBM I can—HBM is expensive DRAM—onto this AI thing.
The data center is now like—it was megawatts, and now it’s gigawatts. It’s small data centers in cities to giant server farms in Texas. That is the data center and cloud of tomorrow, and they’re selling tokens.
The data center of the past is like this old cloud, and it facilitated high margins. That still exists, but it’s not going to grow like this. But this business will facilitate radically more capital—a complete rebuild. This is an alien data center to this data center, and that requires massive capital investment.
It also engenders a very different rethink of the margin stack. Before, you would resell this base layer of low-cost compute, with massive SaaS app margins on top. Now you’ve got this massive compute stack, and they’re reselling that as tokens. That compute factory is creating tokens, and then the model guys are selling their tokens.
That takes a lot of margin out of the top layer of the stack. This is driving what you just said: a complete rebuild. You must build these new data centers because, at some point, all of your revenue will become this, and the revenue from the old data center—which was asset-light and high-margin—is now moving to asset-heavy, lower-margin, big-dollar infrastructure. It’s a completely new data center.
Because of this new rebuild, it has triggered a reexamination of value. If you look at the tech stock market today, I’m going to make some approximations. There are about 1,500 companies globally with a market cap of $1 billion or more—maybe 2,000, depending on whether you add China. Not in China. In the U.S., let’s just talk about the global stock market.
There’s roughly $10+ trillion in market cap across software, services, and internet. They were the classic industries where most of the market cap was in the pre-AI era. That’s $10 trillion, plus or minus. There’s $22–$23 trillion in the Magnificent 7, so I just put the Mag 7 in a new category. Microsoft is a Mag 7, so I just have a non-Mag 7 software, services, and internet category of $10 trillion.
There’s another $30-plus trillion in chips and hardware. So, $10, $20, $30 trillion, something like that. $10 trillion in software, services, and internet; $20 trillion in Mag 7; $30 trillion in non-Mag 7 compute, chips, and hardware. I don’t think people realize that we are that compute-hardware-centric now. Ten, 20, 30. Before AI, in the BCE era, it was probably reversed, okay?
You’ve seen a transformation in value systematically happening over the last 4 years. That follows the data-center transformation because of the primacy of compute, the plurality of the dollars, and the creation of the models themselves. The models themselves, like it or not, have consumed the market cap out of software and services. It’s like the Borg. It has consumed it.
For those foundational models to exist, they need to live on the compute stack. To be offensive in this structure, I needed to rethink everything. The realization came to me that when we hit the AD era—the AI era—you needed to rethink everything. Then you need to align offensively, as you say, an investment philosophy and a capital-allocation philosophy that would mirror what is becoming the new reality, where there is this intelligence and an insatiable demand for intelligence. As a function, intelligence begets compute, and intelligence for compute equals basically revenue.
If you think that the basis of many companies is around these compute factories and your ability to resell that intelligence, then that makes you rethink the margin stacking and where the value sits for companies. The market is trying to adjudicate that right now. That’s why you saw SaaSpocalypse earlier this year and last year. That’s why you’re seeing the RAMpocalypse. There are a lot of apocalypses.
If you believe these scaling laws and intelligence is getting better, let’s say at 1 order of magnitude a year—I mean, 10 × 10 × 10—that’s 1,000 times in 3 years. It’s not like these AIs are getting less capable. They’re getting more capable, and more capable means more compute. If they can do more things, then you have to rethink where the margins are and where your defensibility is.
People use the term “moat” a lot. A moat is very defensive. Moats are always breached, aren’t they? So it’s more about offense, in my opinion. Can you move faster? That’s a broad topic. That’s how I think about the macro. Like you said, this complete redesign facilitates a new rethink, and that rethink also has huge implications for the business models and moats of companies.
This is my thesis now; it might change. If we talk again in 6 months, it might be completely different. That’s my current thinking.
Because to your point earlier, we are investing in new areas, or we’re investing in areas that have 3-, 4-, and 5-year lead times—even 10 years if you want to talk about quantum. I mean, that’s always 10 years, whatever year you’re talking about it, to be clear.
Mm-hmm.
But in terms of those types of outward investments and strategies, how do you think about where you’re going to spend time and which one of those is most effective right now? Because you’re obviously taking a risk on that, and I bring that up in the context of quantum as well.
Yeah.
And with these new chips and building them specifically for models, how do you think about that, and how do you weigh the different kinds of risks that come along with it?
Yes.
Yeah, as an investor and portfolio manager, at the end of the day, you’re allocating capital, right? You have only so many bullets. I’m a public investor and a private investor, but at the end of the day, you’re allocating capital and creating whatever portfolio you’re creating for your mandate and your clients.
At the end of the day, you’re trying to arbitrate between risk, as you said: what is today and what is tomorrow. A lot of the things around AI today are about today. Even the 3-year duration mismatch of, let’s say, DRAM and foundries—I call that the now, right? This is the now. The vortex of AI is the now.
In this 3-year window, that’s probably where 90-plus percent of my investment, or 90-ish percent, something like that—the majority—is going. Within this 3-year window, who’s winning, who’s losing, what is on the ascendancy, what is in decline, what is stagnating, and then you’re arbitrating between these ideas.
The second thing, or even within this 3-year window, is: Is there life after the 3 years? You have to believe that this continues 5-plus years, right? Because the belief in a future has a huge impact on your multiple. If they do not believe in that future, even beyond the 2- or 3-year horizon that most investors and Wall Street can forecast, there is an implicit understanding: Do you have a future or not?
I always feel like you must be betting on the future as well. Things look cheap, but it’s atrophying and maybe in decline, or growth is decelerating, so is that the best allocation of capital? Versus the next 3 years, it’s going to be great for memory or compute or data centers, but will that continue 4, 5, 6 years into the future? Question mark: yes or no.
But then you also have to be betting on the frontier of the frontier. I made some of these AI investments pre-generative AI, in 2019, 2020, and 2021, when you didn’t know that this LLM thing was going to happen. Some of these things gestate longer. My intuition was that we would need AI compute of some form, maybe machine-learning AI.
I didn’t know that this LLM wave would happen. But you’re thinking about future architectures, and so now we sit 6 or 7 years later, and the AI accelerator wars have begun and compute has taken off. I think about that in that longer-term context.
The next set of companies that have this longer context is around, let’s say, quantum. I think we’ll see it by 2030. I started getting involved in 2019, so I’m already 7 years in. You’ve got another 5 years to go. So you have some bets on where the frontier is coming next.
Space, right? These orbital data centers, right? That’s also targeting 2030. It’s very interesting when you look at these long-dated technologies. All roads converge to 2030. It’s like quantum computing: a utility-scale, logically error-corrected, million-qubit quantum computer by 2030. SMRs, fusion, small nuclear reactors with regulatory approval—you talk to these companies, and it’s 2030.
Then you ask, “When will AGI happen for classical computing?” 2030, 2029, 2028, whatever. Then you say, “When will we hit 800-volt power architectures?” Late 2020s, 2030. Will we have solid-state transformers? 2030. Fusion is longer. But many data centers in space: 2030, when it starts to really scale.
You’re sitting here, and obviously you have the now, this AI train that is consuming everything—all my time and energy—but that’s 80–90% of it. You must always be betting on tomorrow, and some of these are long-dated things. I spend X percent of my time on the future. What will really be transformative, not incremental? I want nonlinear, asymmetric potential.
Then I bet on what has primacy today, not only with a 3-year financially forecastable window, but with relevance beyond it. For everything else that doesn’t fit in that window, is it worthy of your time and the opportunity cost of continuing to invest in it? There are other strategies, other portfolios, and other things that can pursue those. It’s just not my focus, really.
I hope that gives you a sense of how I frame capital-allocation and portfolio decision-making.
That’s a super helpful explanation. I’m sure a lot of your investment memos have 2030 on them.
Actually, 2030 is not far away.
No, it’s not.
Yeah, absolutely. Many companies—you’ve got to look at 2031.
I mean, 2031 is 5 years. Ten years—2036. So, I mean, 10 years is almost an impossible forecasting period, but 5 years—you know, a lot of companies will not even have free cash flow by 2031. So you then need to have a belief system that it could flip positive beyond that. But yeah, I’d say at least a 5-year window.
You always want to be betting on not what’s cool today. Will you still be cool in 5 years?
Mm.
Then you become yesterday’s news in 5 years, even though you are cool today. So maybe there’s a little bit of that happening too. That really has a huge impact on your exit multiple.
Right? If you’re just following the trend of today, but you know that there’s a half-life to this, it might be difficult to get a good return on the exit because it will not be what you think it is in 5 years. The multiple that people will pay will go down, and your growth rates are decelerating—
Mm.
And now you’re in a bind.
It’s been really interesting to see how this new era has breathed life into older categories, or categories that have just been around, whether it is chips, whether it’s quantum. But it’s cool to see how entirely new opportunities have formed, and I’m curious to hear your take on those. In that respect, you did mention that orbital data centers really didn’t exist before.
No.
And that is a huge weight and a big weight, especially with SpaceX coming and the whole IPO around that. But there are also fun categories. I recently visited Figure AI, the humanoid robotics company, and I was just at Config, Figma’s conference. Boston Dynamics was there, and one of their heads of design for human-robot interaction was talking about their humanoid robot, Atlas. They’re different.
Mm-hmm.
The Atlas one is hydraulic, so it can pick up a fridge.
Mm-hmm.
Figure AI is more focused on daily use—package sorting, commercial stuff, making cars, and that sort of thing. But of these new categories, which ones are you paying attention to? What are you excited about?
5. Robotics Embody Intelligence
So, robotics, yeah. I mean, the first comment around these older categories: like you said, semiconductors have been around a long time. I don’t understand why people forget. You know, it’s called Silicon Valley for a reason.
People forget.
But people forgot. It’s kind of like the ring of power: it was lost, and then it was found. People always said chips are a commodity, but chip companies have the highest profitability of any sector in the world. They have higher margins than software, pharmaceuticals, industrials, telecom—anything. So this notion that they were a commodity was just a false notion, in my opinion. It really never was.
The other thing is that this industry is very interesting. There were hundreds of chip companies, and then systematically, over 20 years, now there are just a few. So in every category, you have duopolistic power. And, by the way, venture capital—that is, Silicon Valley—until recently never funded these companies. So there’s no money going in, and therefore, if you have no money going in, you have very few companies. In fact, what you have is a shrinking effect.
The number of companies has collapsed, and those that have survived are behemoths with huge pricing power—the complete opposite of a commodity. And, by the way, all those people are engineering nerds. So I always say it’s a revenge of the nerds. It’s not a revenge of the nerds; it’s the lost ring of power that was found. It was always there, and so their time to shine is now.
That said, you’re right. These other industries have spawned a renaissance in hardware, right? When you look at that market cap shift that I was talking about, a lot of those go into—servers are cool, fiber is cool, power is cool, rack design is cool.
People love rack design.
People love it. They’re going crazy about rack design. It’s like bending metal, working with copper and heat. Materials science is cool because you need all kinds of new materials. And then substrates and packaging. It’s the physical world.
These are all physical—what I call the physical world, the physical sciences. That’s cool again. Going to school to study material science is probably cool. It’s very cool. There aren’t enough chip designers in the world.
People in the lost art of analog computing are like blacksmiths. How many friends of yours go into studying new memory design? I was talking to someone at dinner last night, here at Ray’s, from one of the biggest memory companies. They said, “We cannot get people to design custom memory,” because this is a custom co-design with the chip; they want to co-design the memory. Well, where are the people? There are no people. We’ve got to repurpose some of these software programmers into memory co-design architects.
And so all of this has happened in the physical world—what I call the physical world. This is the next unlock that AI will do. Obviously, we’re going to go hard in cognitive work and cognitive labor, and you have to build these models for that. But then those models can be repurposed and implemented in robotics. And so the robotics supply chain is very interesting to me, but it’s kind of a parallel to what was going on in AI.
Because if you really think about it, what is robotics? I mean, you have a brain that will be built, and the brain will have kind of 2 parts to it. It’ll be a baseline LLM that you and I will communicate with as a translator, to translate and talk to the robot. It’s like the human brain, but you also have the brain for the motor functions that control our muscles and our bodies and our reactions, and then the brain for language and memory.
So they’ll have 2 brains: kind of a world model to perceive the world and motion and things, and, obviously, an embodiment of intelligence in an LLM. So you’re going to build these 2 brains into 1. And then you take the brain, and those are like LLMs, right? They’ll be like LLMs, and many of the big labs are working on robotic brains. Then you’ll embody those brains into the body.
But then the body—arms, legs, limbs, hands. Hands are probably the hardest thing, as you, I’m sure, know. But the body, the physical embodiment, that is a manufacturing hardware business.
When you think about that, the Chinese are coming, and there are, I think, 130 or 140 robotics companies in China. I’m looking at the current pipeline. I see 30 or 40 potential IPOs this year in China—yes, yes, this year alone. And there are what, 0, 1, 2 in the United States, maybe, this year.
And the reason for that, though, is also the lack of depth in the private markets in China, so they’re using public markets as a funding mechanism, unlike in the US. So they’re earlier. They’re going to come earlier. And they’re coming in waves. There are 140 of them.
And the thing about China is, in that physical layer—the body, the motion—they may be behind. That’s probably what most people would say: the West is ahead on model development. But make no mistake, China and Asia—and you mentioned the Atlas robot. That’s Korean, actually. That’s Hyundai, which owns Boston Dynamics.
But the Asian manufacturing complex—Japan, Korea, and China—I mean, it’s also kind of an extension of EV platforms, right? If you have physical-scale manufacturing, you then avail yourself to potentially have lower costs to manufacture these robots en masse.
And then what you might have, ultimately, is the ability to mix and match a Chinese physical robot with a Western brain. I know that's happening. People are saying, “Those Chinese robots are amazing, right? So why don't we stick a Western brain inside?” The permutations of this will continue.
I think it's a Wild West, with a lot happening, but it will be a huge market. I have a soft spot for it. One of the things I'm most interested in on the robotics side is not so much the manufacturing robot. Of course, that will happen, and you're seeing this come out of China already.
It's around loneliness, social embodiment, education, the elderly, and young people: bringing consumer and/or commercial social robots to people, more so than for industrial use.
That's kind of a hot take.
Well, I think that's—I mean, if you think about aging populations, when you look at Asia in particular, the birth rates are well below 1.0, and you need 2.1 or 2.2 to stay even. So you're facing demographic population cliffs around the world. Some of the best-performing companies in the world are nursing home companies.
When you look at the elderly, they really want companionship. Even among young people, there are loneliness epidemics and things like that. I think robots, even though they might not have perfect motor function, could embody some intelligence and empathy.
They can have many different form factors, too. It doesn't have to be the Terminator-like robot.
Yeah.
A human or humanoid robot—I think that could unlock a really interesting market, a really big market. That's my view.
That's interesting. I've seen—and I know this doesn't really count—but I saw videos on Instagram of a long-distance relationship, and there was a tiny pet robot on the ground. It was like a ball of some sort, and it was the girlfriend yelling at the boyfriend. She was in an entirely different country, following him around the house.
I mean, I'm not sure if you've seen—have you seen Star Wars?
Yeah. Mm-hmm.
Okay. Who are your two favorites? Did you like C-3PO and R2-D2?
Yeah.
Okay. Now imagine—you have the current embodiment of Optimus and many other humanoid robots, all these sleek, amazing, Westworld-like things. But I hearken to R2-D2 and C-3PO.
What if you had a half-sized robot? Even half-sized—it doesn't have to be full-sized. Something approachable, friendly, not masculine, something that—and then that robot has the intelligence of Shakespeare and Einstein and speaks every language like C-3PO. Then you interact with it.
I often think about the elderly. Imagine them having conversations with my mother and others, with the robot being empathetic toward their stories. Then you can record their stories and their life histories.
Mm.
Do you need to have perfect motor function, with all the hand articulation, or could you get something that can appeal to people in that way? I think that's possible. I think that will be a fascinating market to see.
Yeah.
Besides using robots to build the lunar base, which I also think would be cool.
Really cool.
Yeah.
Yeah. It's been interesting to see, because I cover a lot of high-growth companies in Silicon Valley.
Okay.
The proliferation of coding agents is a big thing, right?
Absolutely. Yeah.
But to your theme of the brain, speech-model companies are crushing it.
Yes.
Crushing it.
They are growing faster than most other companies out there.
Yes.
6. The Enterprise Follows Token Flow
There's a company called AssemblyAI that's growing incredibly fast, and they're great. It's also interesting to see the downstream effects of all this.
We talked about it, but we didn't really cover software at all, because it is what it is. But now it's hitting the data layer, because we have so many agents creating so much data. It's coming downstream, and companies like Databricks and Snowflake are getting some of that extra premium in the market and attracting attention.
Then you go downstream to more app-side companies. As agents create more apps, MongoDB and those databases are benefiting. It's really interesting to see how it's streaming downstream. Are you looking at any of the downstream winners?
Oh, yeah. I'm invested in many of those companies. Streaming downstream or streaming upstream, I don't know what's down or up.
I don't know either.
Yeah.
So it could be the opposite.
Well, no. No, 100%. In supply-chain hardware, is it upstream or downstream? I'm thinking vertically. Compute, models, data, apps—something like that.
Okay.
Right? I'm going up the stack.
You're going up.
I'm going up the stack.
I'm going down.
You're going down the stack. Whatever. Okay, so absolutely. This whole data-center redesign thing—I think the whole enterprise is redesigning.
The enterprise itself, if you really think about the future of what a big enterprise will be like, will be bringing in intelligence—tokens. You'll have a data layer, because that intelligence will need to interact with and be orchestrated around this data layer.
This data layer will be the embodiment of your proprietary data and all your external third-party data. Ultimately, what is a company? It's people, distribution, and a brand. But at the end of the day, can you embody all of the knowledge of your company in what they call a context layer?
Mm-hmm.
A layer of the secrets and the ways of your company, where you embody the cumulative knowledge of your employees into its own context layer.
I think Palantir calls this an ontology.
Ontology. Exactly. So you have this ontology layer, this context layer, sitting on the data foundation with tokens in.
Mm-hmm.
Then everyone builds agents. Agents go wild, right? Agents will interact through the context layer with your data, with tokens. That's it. That is the enterprise.
What I call that is token flow. Follow the flow of tokens. Either you create tokens—compute—or you serve the tokens: foundation models. Then you put a harness and package context around the token: app services, et cetera.
You must be in this token flow to either resell or repackage the tokens with your context and your very specific application. You're serving the token with your intelligence. Is it the proprietary, closed-source token? Is it the open-source token sitting on a compute foundation that's creating and firing up the token? If you're not in that flow, it's a problem.
You mentioned these voice APIs and the data foundation. They are in token flow.
Mm.
I think the app companies are struggling to find their place. But some companies have moved into what I call—I don't know what you call it—inference clouds, edge clouds.
Yeah. Mm-hmm.
Edge AI. They're basically that last-mile token. They're providing tokens, developer kits, and things so that small and medium businesses can take it out of the box. They've inserted themselves into this token flow.
To me, that's why I go back to this base foundation: Where can you earn your margin? Or you do the whole thing. You do the whole thing and say, “I will do all of your claims processing. I will do all of your insurance processing.”
You abstract away all of those layers of the stack and just say, “I will take on all of your insurance. Pay me X.” So you don't know what you are. Are you an app company? Are you a service company? Are you a compute company? A token reseller? No, I'm just selling you the whole solution.
Today, you used to pay 100; I'll charge you 20. And now you're seeing this. You're seeing certain private equity firms, and you're seeing some venture firms saying, “You know what? Let's take an old industry. Let's buy these companies, bring in this whole new stack, reimagine the stack, and just sell a whole new solution.”
Do you think—it's really curious with those PE roll-ups, because I think at the end of the day, they're just creating a new product, but they're buying the customers.
Yeah. They're buying customers, or they're buying companies with customers, and they're basically trying to restructure the whole delivery of services. There's a lot of inefficiencies, fat, and cost in there, and then you can rip it all out.
Mm.
Okay. Maybe that's a business. Let's see. I know people doing that or starting to do that. Yeah, that's interesting. I don't know.
I mean, it's working quite well. We've talked to a couple of them.
Yeah.
Some on camera, some off camera.
Yeah.
We talked to General Catalyst's Creation Fund.
Yep. Yep.
They've been doing a lot of PE roll-ups and creating companies. One of them, Long Lake, just bought American Express Global Business Travel. Their travel business, I think.
Okay.
Which is interesting because you have a small—
We use those—
—player—
We use those guys.
—buying a large player, which was really interesting to see.
Yes.
Yeah.
I mean, there's this other framework, this rubric, that is emerging. When you have these traditional industries and a new company with 500 people that can generate the revenue of 10,000 people, they're approaching it through a radically rethought process.
So maybe this is what you're alluding to. Maybe that's the new framework for these newer companies that are going to go after traditional industries. It'll come down to whether the traditional company, the incumbent, can adjust in the face of these kinds of companies coming in. I think the jury's out.
Mm.
I'm very intrigued by that. I'm watching it. You mentioned some of these companies. It could be quite disruptive. I think that's the next shoe to drop.
Yeah.
All of these traditional industries that have very little adoption of AI, that are still using the business workflow—the way they've organized themselves—they could all be completely rethought and refactored with a new kind of approach.
But that would require, like you say, maybe selling products piece by piece and having your old employees drive that change versus just buying the company and doing the change. It's an interesting idea. It's all new. I'm watching it. It's something to look out for.
We only have a few minutes left, but I'll leave you with 2 questions. First, what are you most excited about in the next 12 months? I know 2030—
Next 12 months.
—is a big date, but let's talk about the next 12 months, maybe.
7. The Next Twelve Months
Oh, boy. Next 12 months, between now and June 27th. I mean, obviously, these big foundation model labs. What am I most excited about, or more concerned about?
Take it either way.
I think it's a continuation of the same. It seems like every 6 months there's a scare: Xpocalypse, Xpocalypse, war, interest rates, too much CapEx, not enough financing, and on and on.
But through it all, I'm optimistic that this compute wall, the memory wall, the compute demand, and this data center redesign will just plow through, and our fears will subside. We'll be sitting here a year from now talking about many of the same things continuing. That's number 1. I hope that's what I'm optimistic about.
The second thing is the progression along what I call this whole data center reimagination theme. I'd like to see more continued proof points along that path.
I'm hopeful and excited to see the big labs go public in the next 12 months. I think that would be interesting and exciting, and I think there's huge market appetite for it.
What I'm excited about in the next 12 months is the next forward step toward orbital data centers, because that also engenders a radical change in data centers. If you keep pushing on that progression, it could unlock a rethink: moving the burden of terrestrial compute into space. That could have huge implications for how current data centers are even being built.
I'm looking at that to see the progress being made there. It's going to be an interesting 12 months.
Amazing. It also sounds like a little bit of a manifestation going on over here.
Manifestation?
Yeah, you're manifesting.
I don't know. I'm just contemplating. Yeah.
Okay, so as we close out, final question. I take this on a personal bent, so no pressure here. I believe personal performance really revolves around who you surround yourself with. People say you're a result of your 5 closest relationships and that kind of thing.
I'm curious, from your standpoint—you've built out a legendary career. Who are some of the people—
Mm. I don't know about that.
It is true.
Yeah.
Who are some of the people who have inspired you or mentored you along the way?
Wow, what a question. First of all, I don't have a legendary career. I'm just trying to survive.
I wouldn't say I had mentors, but you know what I had? I had people who believed in me at certain points in my life and basically gave me the freedom, the latitude, the keys to the kingdom, and said, “You know what? I see something in this guy, and I will give you the latitude.”
There was a guy who brought me into BlackRock who basically gave me carte blanche, freedom, and latitude. He believed in what I could do, so that's 1.
I actually did investment banking long ago, and there were a couple of people there who took a shot on me—some engineering kid out of the Midwest, growing up in the Midwest. I thought I wanted to go into consulting back then, and no consulting firm would hire me. I wasn't good enough for them, so somehow I found a fit in investment banking.
Then someone else said, “Go west, young man.” This was right before the dot-com era. Certain people made a bet and just had faith. It wasn't like they were mentoring me per se; they just had a belief.
I try to do that. I try to always work with lots of young people—not to mentor them, but just to encourage them, give them a break if I can, or give them a shot.
But the mentors I have are all dead. My mentors are, you know, people I like to study. I'm a huge student of history: Caesar, Alexander, Napoleon, Beethoven, Churchill. I like these kinds of historical leaders—people who created their own destiny. Those are my mentors.
Growing up where I did, I was not a social kid. I was somewhat ostracized, so I grew up in libraries, and those historical figures and libraries became my mentors.
When I went into the real world, some people gave me a shot. They just believed in me, or showed me kindness. I'll never forget those people.
So there you go: dead people and kind people. How's that?
That's beautiful.
Yeah.
Wow. Great place to end it. Thank you—
Thank you, Molly.
—so much, Tony.
Yeah, it's a pleasure. Yeah.
Amazing.
Yeah.