风险投资与成长资本的苦涩教训：Anthropic 对 OpenAI、Noam Shazeer、World Labs、Thinking Machines、Cursor 与 ASIC 经济学——a16z 的 Martin Casado 与 Sarah Wang
前沿AI融资已经变成风险投资与成长资本的混合体，因为尚未变现的公司几乎从成立之初就需要成长阶段规模的资本和运营支持。 公司成立仅6个月后，融资就可能涉及数亿美元、战略投资者、以股权换算力的谈判，以及市场进入协议。Martin Casado 表示，自己投资10年来“从没见过这种情况”。
看多当下循环资本流动的理由在于，没有“闲置GPU”，不像未被使用的光纤曾经拖长互联网泡沫破裂后的调整。 Sarah Wang 提出的条件同样关键：资金必须持续转化为能力，能力转化为需求，需求再转化为收入。如果 scaling laws 或客户需求失效，支撑整个系统的融资逻辑也会随之失效。
前沿实验室可能不必先达到 AGI，就能吞并自己的应用生态。 其飞轮是：算力融资 → 能力突破 → 第一方应用增长 → “动能最强时”完成更大一轮融资；如果每轮融资扩大3倍，最终超过下游公司可获得的全部资本，模型所有者就能向整个技术栈持续投入并逐一复制。Alessio Fanelli 称之为“苦涩教训在创业行业的应用”。
市场尚未在软件广泛充裕与前沿模型寡头垄断之间作出选择。 swyx 描绘的一种未来是模型扩散、竞争者追赶、软件继续碎片化；另一种未来只需要用多3倍的钱训练，就能产生吞噬所有相邻市场的通用模型。当前收入可能只能覆盖上一代模型，却不足以支付下一代模型的训练成本，公司只能“透支未来”，直到资本重新变得理性，或更便宜的算力修复这笔账。
AI人才市场似乎提高了创业的机会成本，即使2025年最轰动的挖角只是短期异常。 节目提到一笔可能达到50亿美元的挖角、数千万美元的 L5 职级报价，以及手握每年1000万美元报价的投资候选人；Sarah 的结论是，“稳态水平已经被抬高了”。但战略资金和 acqui-hire 也可能让团队收购带来历史上非常出色的风险投资回报。
投资者可能正在忽视稳健的传统软件，同时用仿佛机器人已经迎来“ChatGPT时刻”的方式为其提供资金。 如果一家面向大市场的软件公司增长5倍，Martin 很乐意投资，即使 LP 在一只基金的生命周期内大约只寻求3倍净回报，也不要求公司必须在1年内做到1亿美元收入。机器人需要不同的尽调，因为农业机器人最终要在农业内部竞争，采矿机器人要在采矿内部竞争，而每种机器人都要与人力劳动形成新的均衡。
应用公司的最强防线是聚焦、产品数据，以及向下整合进模型，但第一方模型竞争仍然是结构性威胁。 Cursor 以可能只有前沿模型成本百分之一的投入，做出了接近 SOTA 的模型，一度成为全球最受欢迎的 coding model，同时仍然是一家边界清晰的专业开发者工具公司。Agent 公司可能依据不断上涨的人力成本，而不是不断下降的 token 成本定价；但第一方模型实验室可以补贴自有应用，同时向第三方收取更高费用。
1. 前沿AI融资抹平了风险投资与成长资本的边界
Sarah 的框架是：这些仍然是创始人驱动的下注，但从第一天起，它们的资源需求和增长速度就“算是成长资本的规模”（kind of growth scale）。尚未变现的公司可能已经拥有足够多的用户，需要复杂的量化分析；也可能已经有足够大的算力需求，需要规模非常大的资金。
商务拓展如今从基础设施而非分销渠道开始：算力由谁供应，采购算力是否附带股权安排，哪个战略伙伴参与，以及协议是否包含市场进入支持。价值数亿美元的谈判，在公司注册仅6个月后就可能启动。
Martin 将这套机制与过去2000万至6000万美元的 A 轮或 B 轮融资作了对比：如今的轮次同时纳入财务投资者和战略投资者，而战略部分往往取决于需要数月才能敲定的算力合同。融资工具、基础设施协议和商业合作，已经合并成一笔交易。
对于循环融资焦虑，他给出的答案是有条件的：“只要需求还在。”互联网当年为最终闲置的光纤提供融资，造成持续约4年的供给过剩，即使经历了巨幅崩盘也没有立即消失；而他认为今天“没有闲置GPU”（there’s no dark GPUs），因为融资支持的算力确实被消耗了。
Sarah 补充了这套模式成立的关键条件：资金必须持续带来能力提升，能力提升必须创造需求，scaling laws 也必须继续成立。如果这条链条断裂，融资逻辑也会一同断裂。
2. 资本如今直接复合成能力与市场份额
Sarah 描述了正在成形的循环：为算力融资，把算力转化为突破，再将能力导入 ChatGPT 或 Claude Code 这样的垂直整合产品；如果有必要，就补贴用户采用，然后“在动能最强时融资，再重复一遍，周而复始”（raise money at the peak momentum, and then you repeat, rinse and repeat）。
Martin 还表示，风险投资与成长资本的边界、模型基础设施与应用的边界都在模糊。模型公司既做 API 业务，又在应用层与客户竞争，形成“亦敌亦友”（frenemy）的关系。
过去的瓶颈是工程：增加资金并不会按比例加速软件团队。Martin 认为，如今模型公司可以融资，并在1年内用大约10人或20人做出更好的模型，立即创造需求，再支撑下一轮融资。
Alessio 将这套逻辑进一步推向公司成立阶段：如果从 token 到产品的摩擦接近于零，一家工作室可以说，“我今天要花100万美元做推理，明天就要把产品推出去。”早期风险投资会变得更加迭代化，因为资本可以直接生产并测试产品，而不只是为一个固定组织提供资金。
Martin 尚未解决的情景是：如果 Anthropic 每轮都能多融资3倍，最终超过所有基于其模型构建的公司的总融资额，它就可能“像一颗恒星一样”扩张到每个下游品类。资本转化为能力，再转化为增长，然后转化为更多资本——这将成为创业行业自己的苦涩教训。
3. Character 暴露了 AGI 目标与产品责任之间的冲突
该公司在2023年1月投资 Character；随后 Character 于2024年8月完成与 Google 的 IP 授权交易。Martin 不接受将 Character 的结果简单归结为前沿实验室砸钱超过了整个生态。Sarah 表示，Noam Shazeer 想同时推进产品发布和 AGI，而公司的面向人的产品与数据，正是通往这一目标的载体。
她更深层的复盘是“AGI 与产品之争”。GPU 可以服务当前用户、近期能力工作，或长期研究；但产品使用和收入又要为 AGI 所需的 GPU 提供资金。一家初创公司如果无法展示足够进展来持续融资，就会更早、更尖锐地面对这一冲突。
Martin 认为，如今出现了一批不同寻常的创始人，他们围绕共同的 AGI 北极星组织公司，而不只是单纯地创业。这带来的创始人流动，比他20年创业经历中见过的都更多——最接近的历史类比或许是“Shockley 与叛逆八人”（Shockley and the Traitorous Eight）。
Martin 原本认为，在 Meta 组建团队之后，2025年的极端挖角可能只是短期异常。但 Sarah 表示，她所描述的招聘市场中，每年1000万美元左右的报价仍然存在，甚至 L5 也可能拿到数千万美元；与此同时，战略性 acqui-hire 正在带来异常出色的风险投资回报。
4. “无聊”的软件被错误定价，机器人可能被过度提前定价
Martin 看到的是一根杠铃：一端是 X 上最热门的方向，另一端是深科技，中间的数据库、监控、日志等耐用型软件却缺乏关注。认为低于“0到1亿美元收入增长”就不值得关注的迷因，忽视了这样一个事实：一家面向大市场、增长5倍的公司，完全可以轻松超过 LP 对一只基金约3倍净回报的预期。
swyx 对机器人的担忧不同：市场已经在融资定价中假设硬件会迎来一个他尚未看到的“ChatGPT时刻”。过去无人机和自动驾驶的周期让 Martin 认识到，硬件通常会垂直化——农业机器人最终会变成一项农业投资，遵循农业自身的定价、供应链和竞争格局。
他们偏好的机会是 Applied Intuition、DeepMap 和早期 Scale AI 这类横向赋能者。Elon Musk 的人形机器人项目或许可以“凭意志把一个行业变出来”（will into being an industry），但尽调的关键仍不在于核心技术是否令人印象深刻，而在于机器人与目标市场中的人力劳动之间会形成怎样的竞争均衡。
5. 一次10亿美元训练运行足以为每个模型配置一枚 ASIC
Martin 的算术是：如果训练成本达到10亿美元，推理收入最终就必须超过10亿美元，否则模型在财务上无法成立。哪怕只节省20%，也会创造2亿美元空间，足以完成一枚定制芯片的流片；因此，每个模型配一枚 ASIC 在经济上是合理的。如果效率提升1倍，节省额则大约为5亿美元。
“现在的问题是时间表，而不是钱。” ASIC 必须在模型过时之前交付，但如今通用 NVIDIA 芯片的经济性仍然留下了足够大的价值空间，使定制硅片变得理性。Alessio 提到，截至2025年底，OpenAI 已确认与 Broadcom 及其他公司的定制芯片合作。
Martin 认为，American Dynamism 与其说是一种使命，不如说是针对硬件、监管合规和政府市场的尽调专长；即便客户和供应链遍布全球，该机构历史上的投资仍然以湾区为中心。
6. Anthropic 与 OpenAI 处在一个尚未定型的双重未来市场中
Sarah 亲自验证 Anthropic 产品进展的时刻发生在午夜：她和另外2个人把一份原始客户文件交给 Claude Cowork，一次性得到了准确的 cohort retention 分析。“砰。完全准确。”过去需要一名成长投资人熬夜完成的工作，如今几秒钟就能完成。
swyx 强调 Anthropic 所宣称的企业级重点；Alessio 则反问，Claude Code、Claude Cowork，以及看起来已经出现的消费者广告，是否仍可能代表一条进入 OpenAI 地盘的“创新者窘境”路径。Anthropic 可以从企业和 coding 切入，再将这些能力带向消费者；与此同时，OpenAI 则在多模态方向追求通用智能。
swyx 提出的“西部人，你要往哪边走？”分叉中，一条路径是模型扩散、竞争者追赶，普通软件继续裂变成更多公司；另一条路径是模型凭借连续3倍的训练预算实现足够强的泛化，最终由寡头垄断吞噬所有下游市场。“没人知道答案。”
swyx 认为，公开披露的经济数据掩盖了代际切换：如果只把收入与上一代模型的训练成本比较，毛利率可能为正；但把当前用于训练下一代模型的支出计入后，公司就会转为负值。Sarah 补充说，GPT-4 曾经领先9或10个月，2024年3月开源模型的领先者几乎每天都在变化，而今天的寡头格局仍可能继续移动，只要能力还在进步。
7. 通用智能可能比狭窄任务优化更重要
swyx 修正了自己此前的要求：他原本认为，实验室必须先渐近达到 AGI，之后才能吞噬应用；现在他认为，一个毛利率为60%、70%或80%的 API 业务可以看到客户正在构建什么，如果它能比所有客户加起来融资更多，就能为第一方版本提供资金，无论底层模型是否真正达到 AGI。
Sarah 保留了应用公司的反例：一旦模型的边际改进对某个已经饱和的企业任务不再重要，价值就可能转移到服务、实施和面向客户的定制执行上。因此，法律或类似工作流仍可能支撑高价值的专业公司，尽管她明确承认，模型能力持续进步可能让这个例子失效。
swyx 提出了更难的问题：是否“每项任务都等价于 AGI 完备”（every task is AGI complete）。以他的经验，Codex 在最难的 bug 上胜过 Opus 4.5，而 Opus 4.5 的“床边沟通”（bedside manner）更好；复杂 coding 还需要合规、历史记录、网络研究、头脑风暴和持续协作，而不只是生成代码。
他的结论是，“不存在所谓 coding model”（there’s no such thing as a coding model）——它必须像一个具备普遍能力、同时也会 coding 的人。Alessio 只提出了一个有限的不同意见：他高度确信 OpenAI 会继续同时发布 GPT-5 和 GPT-5 Codex，“一个负责 rizz，一个负责 tizz”（one for rizz and one for tizz），把专业化从数百个维度压缩到可能只有2个维度。
8. 空间智能需要超越语言的表征方式
Martin 参与 SparkJS，这是一个开源 JavaScript 渲染器，用于处理 World Labs 生成的3D场景中的 Gaussian splats。Three.js 和 Unreal 等现有生态都是围绕 mesh 构建的，因此这个库为缺乏传统拓扑结构的辐射场表征提供基础设施。
AI coding 消除了框架学习所需的“激活能”，此前正是这道门槛让他一直没有投入这项工作。如今他可以专注于算法和扩展性，重新拾起自己在1990年代末构建游戏引擎时积累的技能；Andreas Sundquist 仍是 SparkJS 的主要开发者，Martin 则负责配套代码和演示。
swyx 引用 DeepMind 用 Deep Think 取得 IMO Gold 的结果，认为单个 LLM 中更长的推理过程可能取代独立的神经符号系统。Martin 的反例是一间黑屋：关于桌子和障碍物的语言指令并不够用，而开灯之后，距离、曲率和移动路径都会变得清晰。“语言不是描述宇宙的正确原语集合。”
9. 基础模型下注始于成本坍塌与唯一型创始人
Martin 对 World Labs 的判断从经济学开始：通过 Fiverr 把一个房间重建成3D场景，可能需要4000至1万美元；专业制作可能需要3万美元，而生成式系统理论上可能把成本压到不到1美元。这相当于压缩了4到5个数量级，而被压缩的对象本来就已经被游戏、电影、Blender 和 Unreal 购买。
Marble 与重建不同：它接收一张2D图片，然后创造出未被看到的几何结构，例如桌子的背面和底部。投资逻辑不只是更便宜地完成采集，而是以生成式方式生产有用资产，而这些资产的传统边际成本很高。
Sarah 的承保逻辑从“万里挑一的创始人”（N of one founders）开始：他们已经证明了自己的专业能力，比如 Ilya 大约15年的基础研究，或 Thinking Machines 的 Mira 和 John——她称 John 为强化学习领域的教父。该机构并没有把每一个新出现的“Neo Lab”视为可以互换的对象。
第二个前提是，专业化并非零和：即使出现许多音频模型，ElevenLabs 仍然排名第一。基础模型一轮融资中通常约80%会投入算力，但能力突破可能突然创造需求——有一个未具名产品从正式上线到获得数千万美元收入只用了几周，而一些 SaaS 公司则用了7年。
10. Thinking Machines 与 Cursor 更看重执行，而非市场叙事
Sarah 表示，在1月的活动之后，该机构“比以往任何时候都更看好” Thinking Machines，并预计2026年会是重要的一年，理由包括 Tinker、定制模型和未披露的工作。她拒绝提前披露消息，但坚持说团队正在全力推进，“一切都会好起来”（be just fine）。
Martin 更广泛的警告是，行业认知与董事会层面的现实之间的距离“从未如此之远”。X 上的传闻往往从事实碎片出发，在电话传话中不断变形，迫使创始人与虚构的幽灵作战；Sarah 最后短暂地总结道：“Twitter 是精神毒药”（Twitter is mind poison）。Cursor 的 Michael Truell 给出的回应正是他们更认同的方式：“埋头专注业务”（heads down, focus on the business）。
Cursor 展示了与前沿实验室垂直整合相反的路径：先从应用和产品数据出发，再向下整合。它以可能只有前沿模型成本百分之一的投入，做出了接近 SOTA 的模型，一度成为全球最受欢迎的 coding model，同时仍保持为一家专注于专业开发者工具的公司，并收购了 Graphite。
swyx 认为，Agent 实验室可以依据昂贵的人力工时定价，而 token 智能的成本不断下降，因此可能比商品化的模型 API 获得更高毛利。Alessio 接受这一逻辑，但保留了一个条件：第一方实验室可以补贴自有应用，同时向下游竞争者收取更高费用——这正是过去在 EC2 和操作系统中出现过的、微妙的客户与竞争者双重关系。
This is Alessio, founder of Kernel Labs, and I’m joined by swyx, editor of Latent Space.
Hey. We’re so glad to be on with you guys. You’re also a top AI podcast. Martin Casado and Sarah Wang, welcome.
Very happy to be here.
Yeah.
And welcome.
Yes. We love this office. We love what you’ve done with the place. The new logo is everywhere now.
Yeah, yeah.
It’s still getting—takes a while to get used to, but it reminds me of a callback to a more ambitious age—
Right. Yeah, yeah, yeah.
—which I think kind of describes it.
It definitely makes a statement.
Yeah.
Yeah, yeah. Not quite sure what that statement is, but it makes a statement.
Martin, I go back with you to Netlify.
Yep.
And you created software-defined networking and all that stuff. People can read up on your background.
Yep.
Sarah, I’m newer to you. You started working together on AI infrastructure stuff.
That’s right, yeah.
Seven years ago now.
Best growth investor in the entire industry.
Oh, say more.
Hands down. Sarah’s—I mean, when it comes to AI companies, Sarah, I think, has done the most aggressive investment thesis around AI models, right? She worked with Noam Shazeer, Mira, Ilya, Fei-Fei, and these frontier large AI models. I think Sarah’s been the broadest investor.
Mm.
Is that fair?
No, I—well, I was going to say, I think it’s a—
But anyway—
—a really interesting tag team, actually, just because a lot of these big Series C deals, not only are they raising a lot of money, it’s still a tech-founder bet, which obviously is inherently early stage, but—
Cursor, CFL. So many. Oh, my God.
Well, I was going to say, the resources—
It’s all of them.
—the resources, one, they just grow really quickly, but then, two, the resources that they need day one are kind of growth-scale. So I think the hybrid tag team that we have is quite effective.
1. Venture Meets Growth
What is growth these days? You don’t wake up if it’s less than $1 billion or—
No, it’s actually very interesting time in investing. Take the Character round, right? These tend to be pre-monetization, but the dollars are large enough that you need to have a larger fund. The analysis, because you’ve got lots of users and this stuff has such high demand, requires more numerical sophistication. Most of these deals, whether it’s us or other firms on these large model companies, are this hybrid between venture and growth.
Yeah, totally. Stuff like business development, for example—you wouldn’t usually need business development when you were seed-stage, trying to get product-market fit.
Are we talking about BizDev?
BizDev, exactly. But now you sort of—
I’m not familiar. What does BizDev mean for a venture fund? Because I know what BizDev means for a company.
A good example is—we talk about buying compute, but there’s a huge negotiation involved there in terms of: Do you get equity for the compute? What sort of partner are you looking at? Is there a go-to-market arm to that? These are just things that, at this scale—hundreds of millions, maybe 6 months into the inception of a company—you just wouldn’t have to negotiate before.
Yeah. These large rounds are very complex now. In the past, if you did a Series A or a Series B, you were writing a $20 million to $60 million check and you called it a day. Now, you normally have financial investors and strategic investors—
Yeah.
—and the strategic portion always still goes with these large compute contracts, which can take months to do. It’s very different times. I’ve been doing this for 10 years, and I’ve never seen anything like this.
Yeah. Do you have worries about the circular funding from some of these strategics?
I mean, listen, as long as the demand is there, the demand is there. The problem with the internet is that the demand wasn’t there.
Exactly. This is the whole pyramid-scheme bubble thing where, as long as you mark to market on the notional value of these deals, fine. But once it starts to chip away, it really—
Well, no. As long as there’s demand—I mean, listen, a lot of these sound bites have already become clichés, but they’re worth saying. During the internet days, we were raising money to put fiber in the ground that wasn’t used. That’s a problem, because now you actually have a supply overhang.
Mm-hmm.
Even in the time of the internet, the supply and bandwidth overhang, as massive as it was and as massive as the crash was, only lasted about 4 years. But we don’t have a supply overhang. There are no dark GPUs, right? If someone invests in a company, they’ll actually use the GPUs, and on the other side of it is the actual customer. So, circular or not, I think it’s a different time.
I think the other piece, maybe just to add onto this—and I’m going to quote Martin in front of him—is that this is probably also a unique time in that, for the first time, you can actually trace dollars to outcomes—
Yeah.
—provided that scaling laws are holding—
Yes.
—and capabilities are actually moving forward. If you can translate dollars into capability improvement, there’s demand there, to Martin’s point. But if that somehow breaks, obviously that’s an important assumption in this whole thing to make it work. Instead of investing dollars into sales and marketing, you’re investing into R&D to get to the capability increase, and that’s sort of been the demand driver. Once there’s an unlock there, people are willing to pay for it.
Yeah.
Is there any difference in how you build the portfolio now that some of your growth companies are the infrastructure of the early-stage companies? OpenAI is now the same size as some of the cloud providers were early on. What does that look like? How much information can you feed off each other between the two?
2. The Capital Flywheel
There are so many lines that are being crossed or blurred right now, right? We already talked about venture and growth. Another one that’s being blurred is between infrastructure and apps, right? What is a model company?
Mm-hmm.
It’s clearly infrastructure, because it’s doing core R&D and it’s a horizontal platform. But it’s also an app because it touches the users directly. Of course, the growth of these companies is just so high. I actually think you’re just starting to see a new financing strategy emerge, and we’ve had to adapt as a result of that. There have been a lot of changes. You’re right that these companies become platform companies very quickly. You’ve got ecosystem build-out. None of this is necessarily new, but the timescales at which it’s happened are pretty phenomenal. The way we’d normally cut lines before is blurred a little bit.
That said, a lot of it also just feels like things that we’ve seen in the past, like cloud build-out and the internet build-out as well.
Yeah. I think it’s interesting. I don’t know if you guys would agree with this, but it feels like the emerging strategy builds off of your other question. You raise money for compute, you pour the money into compute, you get some sort of breakthrough, and you funnel the breakthrough into your vertically integrated application. That could be ChatGPT, that could be Claude Code, whatever it is. You massively gain share and get users. Maybe you’re even subsidizing at that point, depending on your strategy. You raise money at peak momentum, and then you rinse and repeat.
That wasn’t true even 2 years ago, I think.
Mm-hmm.
That ties into fundraising strategy and hiring strategy, right? All of these are tied. I think the lines are blurring even more today, where everyone is—
But of course, these companies all have API businesses, and so there are these frenemy lines that are getting blurred. They have billions of dollars of API revenue, right? There are customers there, but they’re competing on the app layer.
Yeah, so this is a really important point. I would say for sure venture and growth—that line is blurry.
App and infrastructure, that line is blurry. But I don't think that changes our practice so much. The very open questions are: Does this layer work in the same way that compute traditionally has? During the cloud era, somebody wins one layer, but then another whole set of companies wins another layer. But that may not be the case here. It may be the case that you actually can't verticalize on the token string, like you can't build an app—
Mm.
It necessarily goes down just because there are no abstractions. Those are the bigger existential questions we ask.
Another thing that is very different this time than in the history of computer science is that, in the past, if you raised money, then you basically had to wait for engineering to catch up, which famously doesn't scale. Like The Mythical Man-Month, it took a very long time. But that's not the case here. A model company can raise money and drop a model in a year, and it's better, right? And it does it with a team of 20 people or 10 people.
So this type of money entering a company and then producing something that has demand and growth right away, and using that to raise more money, is a very different capital flywheel than we've ever seen before, and I think everybody's trying to understand what the consequences are. So I think it's less about big companies and growth, and more about these systemic questions that we actually don't have answers to.
Yeah. At Kernel Labs, one of our ideas is: If you had unlimited money to spend productively to turn tokens into products, the whole early-stage market is very different. Because today you're investing X amount of capital to win a deal because of price structure and whatnot, and you're kind of committing to a certain strategy for a certain amount of time.
Yeah.
But if you could iteratively spin out companies and products and just say, “I want to spend $1 million on inference today and get a product out tomorrow”—
Yeah.
We should get to the point where the friction of token to product is so low that you can do this. And then you can change the early-stage venture model to be much more iterative. And then every round is either $100K of inference or $100 million from a Series Z. There's no $8 million C round anymore.
Right.
But there's an industry-structural question that we don't know the answer to, which involves the frontier models. Let's take Anthropic. Let's say Anthropic has a state-of-the-art model that has some large percentage of market share. And let's say that a company is building smaller models that use the bigger model in the background, and you use OpenAI's GPT-4.5, but they add value on top of that.
Now, if Anthropic can raise 3 times more in every subsequent round, they probably can raise more money than the entire app ecosystem that's built on top of it. And if that's the case, they can expand beyond everything built on top of it. It's like, imagine a star that's just expanding.
So there could be a systemic situation where the SOTA models can raise so much money that they can outpay anybody that builds on top of them, which would be something I don't think we've ever seen before, just because we are so bottlenecked on engineering. And this is a very open question.
Yeah. It's almost like the bitter lesson applied to the startup industry, right?
Yeah, 100%. It literally becomes an issue of: Raise capital, turn that directly into growth, and use that to raise 3 times more.
Yeah.
Exactly.
And if you can keep doing that, you literally can outspend the aggregate of companies on top of you, and therefore you'll necessarily take their share, which is crazy.
Would you say that kind of happened to Character.AI? Is that the sort of postmortem on what happened?
No.
No. Yeah, because I think Character—
I mean, the actual postmortem is that he wanted to go back to Google.
Yeah, exactly.
But—
That's a different issue.
You said it, yeah.
We should actually talk about this, yeah.
Yeah. Go for it. Take it however you want.
Well, yeah, I was going to say, I think the Character.AI thing raises a different issue, which the frontier labs will face as well, so we'll see how they handle it. We invested in Character.AI in January 2023, which feels like eons ago. I mean, 3 years ago feels like lifetimes ago.
But then they did the IP licensing deal with Google in August 2024. At the time, Noam Shazeer—he's talked publicly about this, right? He wanted to put products out in the world, and Google wouldn't let him do that. That's obviously changed drastically. But he went to go do that.
He had a product attached. The goal was always—I mean, it's Noam Shazeer. He wanted to get to AGI. That was always his personal goal. But I think that, through collecting data, this very human use case that the Character.AI product originally was and still is was one of the vehicles to do that.
I think the real reason is that, if you think about the stress that any company feels before it ultimately goes one way or the other, it's this AGI-versus-product tension. And I think a lot of the big labs—I think OpenAI is feeling that. Anthropic, if they haven't started to feel it, certainly given the success of their products, they may start to feel that soon.
And there's real trade-offs. When you think about GPUs, that's a limited resource. Where do you allocate the GPUs? Is it toward the product? Is it toward new research or long-term research? Is it toward near- to mid-term research?
And so, in a case where you're resource-constrained, of course there's this fundraising game you can play, right? But the market was very different back in 2023, too. I think the best researchers in the world have this dilemma: “Okay, I want to go all in on AGI,” but it's the product-usage revenue flywheel that keeps the revenue in the house to power all the GPUs to get to AGI.
And so it does make—I think it sets up an interesting dilemma for any startup that has trouble raising up until that level, right? And certainly, if you don't have that progress, you can't continue this fundraising flywheel.
I would say that, because we're keeping track of all of the things that are different, right? Venture versus growth, app versus infrastructure, and one of those is definitely the personalities of the founders. It's just very different this time. I've been doing this for a decade, and I've been doing startups for 20 years.
A lot of people start this to do AGI, and we've never had a unified North Star that I recall in the same way. People built companies to start companies in the past. That was what it was. “I want to create an internet company. I want to create an infrastructure company.” It was more about engineering builders, and this is a different mentality.
And some companies have harnessed that incredibly well because their direction is so obviously on the path to what somebody would consider AGI, but others have not. So there is always this tension with personnel. And so I think we're seeing more founder movement as a fraction of founders than we've ever seen. I mean, maybe since the time of Shockley and the Traitorous Eight or something like that, way back at the beginning of the industry. It's a very unusual time for personnel.
Yeah.
Mm.
Totally, and I think it's exacerbated by the fact that the talent wars—I mean, every industry has talent wars, but not at this magnitude, right? Very rarely can you see someone get poached for $5 billion. That's hard to compete with.
And secondly, if you're a founder in AI, you could fart, and it would be on the front page of The Information these days. There's sort of this fishbowl effect that I think adds to the deep anxiety that these AI founders are feeling.
Mm.
Yes. Just to briefly comment on the founder and talent-wars thing: I feel like 2025 was just a blip. I don't know if we'll see that again, because Meta built the team. I think they're kind of done, and who's going to pay more than Meta? I don't know.
I agree.
Right?
So it feels this way to me too.
Yeah.
Basically, Zuckerberg came out swinging, and now—
Yeah.
He's kind of back to building.
Yeah.
Yeah. You have to pay up to assemble a team to rush the job, whatever.
Yeah.
But now you made your choices, and they have to ship, right?
I mean, the other side of that is that we're actually in the job-hiring market. We've got 600 people here, and I hire all the time. I've got 3 open reqs if anybody listening to this is interested.
For investors?
Yeah, on the team—on the investing side of the team.
Yeah.
A lot of the people we talk to have active offers for $10 million a year or something like that. We pay really, really well, and just to see what's out on the market is remarkable. So I would just say it's actually—
Yeah.
The really flashy one is, “I will get someone for a billion dollars.” But the inflation—
Trickles down.
Yeah.
Yeah.
It is still very active today. I mean—
Yeah. You could be an L5 and get an offer in the tens of millions.
Oh, yeah, easily.
It's—
Yeah.
So I think you're right that it felt like a blip. I hope you're right. But I think the steady state has now elevated.
Everything got pulled up. Yeah, yeah.
Exactly.
We're completely pulled up, for sure. Yeah.
And I think that's breaking the early-stage founder math, too. Before, a lot of people would be like, “Well, maybe I should just go be a founder instead of getting paid $800K or $1 million at Google.” But if I'm getting paid $5 or $6 million, that's different.
But on the other hand, there's more strategic money than we've ever seen historically, right?
Right.
Mm-hmm.
And so the economic calculus is very different in a number of ways.
Yep.
It's causing a ton of change and confusion in the market—some very positive, some negative. For example, the other side of the co-founder acquisition, Mark Zuckerberg poaching someone for a lot of money, is that we're actually seeing a historic amount of M&A for basically acqui-hires, right? Really good outcomes from a venture perspective that are effectively acqui-hires. So I would say it's probably net positive from the investment standpoint, even though it seems from the headlines to be very disruptive in a negative way.
Yep. Let's talk maybe about what's not being invested in—some interesting ideas that you would like to see more people build. It seems, in a way, that as YC has gotten more popular and X has gotten more popular, there's a startup-school path that a lot of founders take, and they know what's hot in VC circles and what gets funded.
Yeah.
There may not be as much risk appetite for things outside of that. I'm curious if you feel like that's true, and what some of the areas are that you think are under-discussed.
I actually think we've taken our eye off the ball in a lot of traditional software companies. Right now, there's almost a barbell: the hot thing on X, or deep tech.
I feel like there's just a long list of good companies that'll be around for a long time in very large markets. Say you're building a database, monitoring or logging, tooling, or whatever. There are some good companies out there right now, but they have a really hard time getting the attention of investors.
It's almost become a meme: if you're not basically growing from 0 to 100 in a year, you're not interesting, which is the silliest thing to say. Think of yourself as an individual person with your personal money. Will you put it in the stock market at 7%, or in a company growing 5x in a very large market? Of course, you're going to put it in the company growing 5x.
Who knows what the margins of those are? Clearly, these are good investments for anybody. Our LPs want 3x net over the life cycle of a fund, right? A company in a big market growing 5x is a great investment. Everybody would be happy with these returns.
But we've got this mania around these strong growth rates. I would say that's probably the most underinvested sector right now.
Boring software. Boring enterprise software.
Just traditional, really good companies.
No AI here.
Well, AI, of course, is pulling them into use cases—
Yeah, yeah.
—but that's not what they are. They're not on the token path, right?
Yeah, yeah.
Let's just say that.
Yeah.
They're software, but they're not on the token path. These are great investments by any definition except for some random VC on X saying, “It's not growing fast enough.” What do you think?
Yeah. Maybe I'll answer a slightly different question, but adjacent to what you asked: an area that we're not investing in right now, but that we're spending a lot of time in, regardless of whether we pull the trigger or not. It would probably be on the hardware side, actually.
Robotics.
Right? In the robotics sector, right?
Robotics, yeah.
I don't want to say that it's not getting funding, because it's clearly almost non-consensus not to invest in robotics at this point. But we spend a lot of time in that space, and I think for us, we just haven't seen the ChatGPT moment happen on the hardware side.
Yeah.
And the funding going into it feels like it's already taking that for granted.
Yeah, yeah. But we also went through the drone era.
There's a Zipline right out there.
What's that? The Zipline?
Yeah, yeah. There's a Zipline.
Oh, yeah. There's a Zipline. Yeah.
We went through the drone era. We went through the AV era. One of the takeaways when it comes to hardware is that most companies will end up verticalizing. If you're investing in a robot company for agriculture, you're investing in an agriculture company, because that's the competition, the pricing, and the supply chain. If you're doing it for mining, that's mining.
The AD team does a lot of that type of work because they actually set up diligence for it. But for horizontal technology investing, there's very little when it comes to robots, just because they're so fit for purpose. We tend to look at software solutions or horizontal solutions, like Applied Intuition, clearly from the AV wave, and DeepMap, clearly from the AV wave.
I would say Scale AI was actually a horizontal one for robotics early on.
That was fair.
That sort of thing we're very, very interested in, but the actual robot interacting with the world is probably better for a different team. Yeah, I agree.
I'm curious who these teams are supposed to be that invest in them. I feel like everybody's like, “Yeah, robotics is important, and people should invest in it.” But when you look at the numbers—the capital requirements early on versus the moment when, “Okay, this is actually going to work. Let's keep investing”—that seems really hard to predict.
Coatue, Khosla, General Catalyst. I mean, these are all invested in hardware companies. You just, you know—
Yeah.
And listen, it could work this time for sure, right?
Right.
The fact that Elon is doing it means that there's going to be a lot of capital and a lot of attempts for a long period of time. So that alone maybe suggests that we should just be investing in robotics, just because you have this north star—Elon with a humanoid—and that's going to basically will an industry into being.
But we're huge believers that this is going to happen. We just don't feel like we're in a good position to diligence these things because, again, robotics companies tend to be vertical. You really have to understand the market they're being sold into. That competitive equilibrium with a human being is what's important. It's not the core tech, and we're more horizontal, core-tech-type investors.
This is Sarah and I.
Yeah.
The AD team is different.
Yeah, yeah.
They can actually do these types of things.
Just to clarify, AD stands for?
American Dynamism.
All right. Okay.
Yeah, yeah.
Yeah, yeah.
So—
3. ASIC Economics Arrive
I actually do have a related question. First of all, I want to acknowledge, just on the chip side—
Yeah.
I recall a podcast where you were on—I think it was the a16z podcast—about 2 or 3 years ago, where you suddenly said something that really stuck in my head: at some point, at some kind of scale, it makes sense to build a custom ASIC—
Yes.
—for each run.
Yes. It’s crazy. We’re here. We’re here.
And I think you estimated $500 billion, or something like—
No, no, $1 billion. A $1 billion training run. A $1 billion training run makes sense for actually doing a custom ASIC if you can do it in time. The question now is timeline, not money.
Yeah.
Because, just rough math, if it’s a $1 billion training run, then the inference for that model has to be over $1 billion; otherwise, it won’t be solvent. So let’s assume that if you could save 20%—which you could save much more than that with an ASIC—20% is $200 million. You can tape out a chip for $200 million, right?
Right.
So now you can literally justify economically—not timeline-wise, which is a different issue—an ASIC per model, which is great.
Because that’s how much we leave on the table every single time we use generic NVIDIA.
Yeah, exactly, exactly.
Yeah.
No, it’s actually much more than that. You could probably get a factor of 2, which would be $500 million.
Yeah. Typical MFU would be around 50%, and—
Yeah, yeah.
—you know, and that’s good.
Exactly. Yeah, 100%.
So, yeah, I just want to acknowledge that here we are at the end of 2025, and OpenAI is confirming Broadcom and all the other custom silicon deals, which is incredible.
Yeah, yeah.
Speaking about AD, there’s a really interesting tie-in that you guys are obviously hitting on, which is this America First movement, or the effort to reindustrialize here and move TSMC here, if that’s possible. How much overlap is there from AD—
Yeah.
—to, I guess—
Nice.
—growth and investing in particularly U.S. AI companies that are strongly bounded by their compute?
Yeah, yeah. I would view AD more as a market segmentation than a mission, right? The market segmentation is that it has regulatory compliance issues or government sales, or it deals with hardware. They’re just set up to diligence those types of companies. So it’s more of a market segmentation thing.
I would say the entire firm, since it was founded, has geographical biases, right? For the longest time, we were like, “The Bay Area is going to be where the majority of the dollars go.”
Great.
And listen, there are actually a lot of compounding effects from having a geographic bias, right? Everybody’s in the same place. You’ve got an ecosystem, you’re there, you’ve got a presence, and you’ve got a network.
I would say the Bay Area is very much back. I remember during pre-COVID, crypto had pulled startups away from the Bay Area.
Miami. Yeah.
Yeah, yeah. New York came up because it’s so close to finance. Los Angeles had a moment because it was so close to consumer, but now it’s kind of come back here.
I would say we’ve historically tended to be very Bay Area-focused, even though, of course, we invest all over the world. If you take the ring out one more, it’s going to be the U.S., of course, because we know it very well. One ring more is going to be the U.S. and its allies, and it goes from there.
Yeah.
Sorry.
No, no, I agree. But I think that’s sort of where the companies are headquartered. Maybe your question is about supply chain and customer base. I would say our companies are fairly international from that perspective. They’re selling globally, right? They have global supply chains in some cases.
I would say the stickiness is also historically very different between venture and growth. There’s so much company-building in venture—so much. Hiring the next product manager, introducing the customer, all of that stuff. Of course, we’re just going to be stronger where we have our network and where we’ve been doing business for 20 years. I’ve been in the Bay Area for 25 years, so clearly I’m just more effective here than I would be somewhere else.
Yeah.
For some of the later-stage rounds, I think the companies don’t need that much help. They’re already pretty mature, historically, so they can kind of be everywhere. There’s less of that stickiness.
This is different in the AI era. Sarah is now the chief of staff of half the AI companies in the Bay Area right now. She’s an ops ninja—biz dev, biz ops.
Are you finding much AI automation in your work? What is your stack?
In my personal stack?
I mean, the reason for this is that it’s triggering—yeah, we are hiring ops people. A lot of founders I know are also hiring ops people, and it’s an opportunity. Since you’re also basically helping out with ops at a lot of companies, what are people doing these days? Because it’s still very manual, as far as I can tell.
Yeah. I think the things that we help with are pretty network-based, in that it’s sort of like, “Hey, how do I shortcut this process?” Well, let’s connect you to the right person. So there isn’t quite an AI workflow for that.
I will say, as a growth investor, Claude Cowork is pretty interesting.
Yeah.
For the first time, you can actually get one-shot data analysis, right? If you take a customer database and analyze cohort retention, that’s just stuff that you had to do by hand before.
The other night, it was like midnight, and the 3 of us were playing with Claude Cowork. We gave it a raw file. Boom. Perfectly accurate. We checked the numbers, and it was amazing.
That was my aha moment. It sounds so boring, but that’s the kind of thing that a growth investor is slaving away on late at night, done in a few seconds.
You’ve got to wonder what Anthropic Labs, their new product studio, would be worth as an independent startup.
A lot.
Yeah. True.
You’ve got to hand it to them. They’ve been executing incredibly well.
4. Anthropic Challenges OpenAI
To me, Anthropic building on Claude Code makes sense. The real pedal to the metal, whatever the phrase is, is when they start coming after consumers against OpenAI. That is red alert at OpenAI.
Oh, I think they’ve been pretty clear that they’re enterprise-focused.
They have been.
It’s been pretty clear publicly—
But here it is: enterprise-focused, it’s coding, right?
Yeah.
And then here’s Claude Cowork.
Hmm.
Apparently, they’re running Instagram ads for Claude AI for people to use their chatbot.
Like they’re the mom and pop.
Right. And so it’s kind of like this disruption thing of—OpenAI has been doing consumer; it’s been pursuing artificial general intelligence in every modality.
Yeah.
And here is Anthropic. They only focus on this thing, but now they’re undercutting and doing the whole The Innovator’s Dilemma thing on everything else.
Hmm. Yeah.
It’s very interesting.
5. Models Face Two Futures
Yeah, but there’s a very open question. Do you know that meme where there’s a guy in the path, and then there’s a path this way and a path this way, and one—
Which way, Western man? Yeah.
Yeah, yeah.
Yeah, yeah. For me, the entire industry hinges on 2 potential futures.
So, in one potential future, the market is infinitely large. There are perverse economies of scale because, as soon as you put a model out there, it kind of sublimates and all the other models catch up. Software is being rewritten and fractured all over the place, and there's tons of upside, so it just grows.
Then there's another path: maybe these models actually generalize really well, and all you have to do is train them with 3 times more money. That's all you have to do, and it'll just consume everything beyond it. If that's the case, you end up with basically an oligopoly for everything.
Hmm. Yeah.
Because they're perfectly general. This would be the AGI path: these are perfectly general, and they can do everything. This other path is actually normal software. The universe is complicated, and nobody knows the answer.
My belief is that, if you look at the numbers of these companies—how much they're making and how much they spent on training the last model—they're gross-margin positive. You're like, "Oh, that's really working." But if you look at the current training they're doing for the next model, they're gross-margin negative.
So part of me thinks that a lot of them are borrowing against the future, and that's going to have to slow down. That's going to catch up to them at some point.
Yeah.
But we don't really know.
Yeah.
Does that make sense?
Yeah, yeah.
It could be the case that the only reason this is working is because they can raise that next round, and then they can train that next model, because these models have such a short shelf life. At some point, they won't be able to raise that next round for the next model, and then things will—
Yeah.
—converge and fragment again. But right now, it's not.
Totally. By the way, a meta point: I think the other lesson from the last 3 years is—we talk about this all the time because we're on this Twitter/X bubble, but—
Very cool.
If you go back to, let's say, March 2024, it felt like an open-source model with benchmark-leading capability was launching on a daily basis at that point. Suddenly, it was like open source took over the world. There was going to be a plethora; it wasn't an oligopoly.
If you rewind time even before that, GPT-4 was number 1 for 9 months? 10 months? It was a long time, right? Of course, now we're in this era where it feels like an oligopoly, maybe with some very steady-state shifts. It could look like this in the future too, but it's so hard to call.
I think the thing that keeps us up at night, in a good way and a bad way, is that capability progress is actually not slowing down. Until that happens, you don't know what it's going to look like.
But I would say for sure it's not converged. The systemic capital flows have not converged. Right now, it's still borrowing against the future to subsidize growth.
You can do that for a period of time, but at some point the market will rationalize it, and nobody knows what that will look like. Or the drop in the price of compute will save them. Who knows?
Yeah. Yeah, I think the models need to asymptote to specific tasks. It's like, okay, now Opus 4.5 might be AGI at a specific task, and now you can depreciate the model over a longer time. Right now, there's no old model.
No, but let me just change that mental model. That used to be my mental model. Let me change it a little bit.
Yeah, yeah.
If you can raise more money than the aggregate of everybody that uses your models, that doesn't even matter. It doesn't even matter. Do you see what I'm saying?
I have an API business. My API business is 60% margin or 70% margin or 80% margin. This is a high-margin business, so I know what everybody's using. If I can raise more money than the aggregate of everybody that's using it, I will consume them whether I'm AGI or not.
And I'll know if they're using it because they're using it. Unlike in the past, where engineering stops me from doing that, this is very straightforward to use as trained.
I also thought it was like, you must asymptote to AGI—general, general, general—but I think there's also just a possibility that the capital markets will give them the ammunition to go after everybody on top of them.
I do wonder, though, to your point, if there's a certain task where getting marginally better isn't actually that much better. We've asymptoted to—you know, we can call it AGI or whatever.
Oli Goldie actually talks about this: we're already at AGI for a lot of functions in the enterprise. For those tasks, you probably could build very specific companies that focus on getting as much value out of that task as possible, where the value isn't coming from the model itself. There's probably a rich enterprise business to be built there.
I could be wrong on that, but there are a lot of interesting examples. If you're looking more at the legal profession or whatnot, maybe that's not a great example because the models are getting better on that front too. But if it's something where it's a bit saturated, then the value comes from services, implementation, and all these things that actually make it useful to the end customer.
Sorry, one more thing I think is under-discussed in all of this is to what extent every task is AGI-complete.
Mm-hmm.
Mm. Yeah.
Right? I code every day. It's so fun.
That's a core question, yeah.
When I'm talking to these models, it's not just code. I mean, it's everything, right?
It's healthcare, it's—
I mean, it's—
Legal.
But it's everything—exactly that.
Yeah, customer support. Yeah.
I mean, it's everything. I'm asking these models to understand compliance. I'm asking these models to go search the web. I'm asking these models to talk about things I know in history. It's having a full conversation with me while I engineer.
So it could be that—
Mm-hmm.
—the most AGI-complete model will always win, independent of the task. I'm not an AGI guy, but the most AGI-complete model will always win independent of the task. And we don't know the answer to that one either.
Yeah.
But it seems to me that Codex, in my experience, is for sure better than Opus 4.5 for coding. It finds the hardest bugs that I work on, and the smartest developers I know work on it. It's great.
But I think Opus 4.5 is actually very—it's got a great bedside manner. It really matters if you're building something very complex because you're a partner and a brainstorming partner for somebody. I think we don't discuss enough how every task kind of has that quality.
Mm-hmm.
And what does that mean for capital investment, frontier models, and submodels?
Yeah.
What happened to all the specialized coding models? None of them worked, right?
Yeah.
They didn't even get released.
There was a whole host. We saw a bunch of them, and there was a whole theory that there could be specialized coding models. I think one of the conclusions is that there's no such thing as a coding model.
Yeah.
That's not a thing. You're talking—
Yeah.
—to another human being, and it's good at coding, but it's got to be good at everything.
Minor disagreement, only because I have pretty high confidence that OpenAI will always release a GPT-5 and a GPT-5 Codex. That's the thing.
Yeah, that's right.
Yeah, yeah.
Yeah, totally.
The way I call it is one for rizz and one for tizz. Then someone throughout OpenAI was like, "Yeah, that's a good way to phrase it."
That's so funny.
Maybe it collapses down to rizz and tizz, and that's it. It's not like 100 dimensions.
It doesn't apply, yeah.
It's 2 dimensions.
Yeah, yeah, yeah, yeah, yeah.
And exactly: bedside manner versus coding.
Yeah, yeah, yeah.
Oh my God. That's rizz and tizz, yeah.
For anybody listening to this—
That's hilarious.
—when you're coding or using these models for something like that, just be aware of how much of the interaction has nothing to do with coding. It turns out to be a large portion of it.
Mm.
I think the best SOTA-ish model is going to remain very important, no matter what the task is.
Speaking of coding, I'm going to be cheeky and ask: What actually are you coding? Because, obviously, you could code anything, and you're obviously a busy investor and a manager of a giant team. What are you coding?
6. World Labs And Thinking Machines
I help Fei-Fei at World Labs—it's one of the investments. They're building a foundation model that creates 3D scenes.
Yeah, we had her on the pod.
These 3D scenes are Gaussian splats, just by the way that kind of AI works. You can reconstruct a scene better with radiance fields than with meshes because they don't really have topology. They produce these beautiful 3D-rendered scenes that are Gaussian splats.
But the actual industry support for Gaussian splats isn't great. It's always been meshes, and things like Unreal use meshes. So I work on an open-source library called SparkJS, which is a JavaScript rendering library for Gaussian splats. You need that support, and right now there's kind of a Three.js moment that's all meshes, so it's become the default in the Three.js ecosystem.
As part of that, to exercise the library, I build a whole bunch of cool demos. So if you see me on X, you see all my demos and all the world-building, but all of that is just to exercise this library that I work on, because it's actually a very tough algorithmics problem to scale a library that much.
And just so you know, this is ancient history now, but 30 years ago, I paid for undergrad working on game engines in college in the late '90s. So I've actually got a background in this. A lot of it's fun, but the whole goal is just for this rendering library to—
Uh-huh.
Are you one of the most active contributors to their GitHub?
SparkJS?
Yeah, yeah.
There's only 2 of us on it.
Okay.
So yes.
No. By the way, the primary developer is a guy named Andreas Sundquist, who's an absolute genius. He and I did our PhDs together, and we set it for constant quality. It was almost like hanging out with an old friend. He's the core guy. I did mostly the side—
But, you know, it's amazing. 5 years ago, you would not have done any of this.
I wrote it for fun.
It's like it brought you back.
No, there's no way.
You're so back.
The activation energy was so high because you had to learn all the framework bullshit, and I fucking used to hate that. Now I don't have to deal with that. I can focus on the algorithmics, and I can focus on the scaling and—
Yeah. And then I'll observe one irony, and then I'll ask a serious investor question. The irony is Fei-Fei actually doesn't believe the LLMs can lead us to spatial intelligence, and here you are using LLMs to help achieve spatial intelligence. I just sort of see some disconnect in there.
Yeah. So I think what she would say is LLMs are great to help with coding—
Yes.
—but that's very different from a model that actually provides that spatial thing.
You'll never have the spatial intelligence—
And listen, our brains clearly have both. Our brains clearly have a language reasoning section, and they clearly have a spatial reasoning section. These are 2 pretty independent problems.
Okay. The one data point I recently had against it is DeepMind's IMO gold. Typically, the answer is that this is where you start going down the neurosymbolic path, right? One sort of abstract reasoning thing and one formal thing. That's what DeepMind had in 2024 with AlphaFold, AlphaGeometry, and now they just use Deep Think and extend the thinking tokens. It's one model, and it's an LLM.
Yeah, yeah, yeah.
That was my indication of, like, maybe you don't need a separate system.
So let me step back. At the end of the day, these things are like nodes in a graph with weights on them, right? If you distill it down. But let me just talk about the 2 different substrates.
Let me put you in a dark room, a totally black room, and then let me just describe how you exit it. To your left, there's a table. Duck below this thing, right? The chances that you're not going to run into something are very low.
Now let me turn on the light and you can actually see, and you can judge distance—how far something is away and where it is—and then you can do it, right? Language is not the right set of primitives to describe the universe because it's not exact enough.
That's all Fei-Fei is talking about when it comes to spatial reasoning: You actually have to know that this is 3 feet away, that far away. It is curved. You have to understand the actual movement through space.
So I do think, at the end of it, these models are definitely converging as far as models, but there are different representations of the problems you're solving. One is language, which would be like describing to somebody what to do, and the other one is actually just showing them. Spatial reasoning is just showing them.
Yeah. Yeah, right. Got it.
The investor question was on World Labs: How do I value something like this? What work does it do? I'm just like, Fei-Fei's awesome, Justin's awesome, and the other 2 cofounders are awesome. But the tech—everyone's building cool tech—but what's the value of the tech? This is the fundamental question of—
Well, let me just maybe give you a rough sketch on the diffusion models. I actually would love to hear Sarah, because I'm a venture person.
Yeah.
You paint a dream, and she has to actually—
She has to make sure—
—make it a reality.
Exactly. So I'm going to say the venture view—
His dream.
—and then she can be like, “Okay, you—”
You little kid.
Yeah. These diffusion models literally create something for almost nothing, and something that the world has found to be very valuable in the past are real markets, right? A 2D image—that's been an entire market. People value them. It takes a human being a long time to create one.
To turn me into a whatever, an image would cost $100 and an hour. The inference cost is a hundredth of a penny, right? We've seen this with speech in very successful companies. We've seen this with 2D images. We've seen this with movies.
Now think about a 3D scene. When's Grand Theft Auto coming out? It's been 10 years. Honestly, how much would it cost to reproduce this room in 3D?
It has been 10 years, yeah.
If you hired somebody on Fiverr, in any sort of quality, probably $4,000 to $10,000. If you had a professional, it'd probably be $30,000.
So if you could generate the exact same thing from a 2D image, and we know that these are used in Unreal and Blender, in movies and video games, if you could do that for less than $1, that's 4 or 5 orders of magnitude cheaper. You're bringing the marginal cost of something useful down by 3 orders of magnitude, which historically has created very large companies.
That would be the venture kind of strategic dreaming map.
And for listeners, you can do this yourself on your own phone with Marble.
Yeah, Marble.
But there are also many NeRF apps where you just go on your iPhone and do this.
Yeah, yeah, yeah. In the case of Marble, though, what you do is you literally give it—
Meaning it has to fill in stuff—
Stuff it can’t see. Yeah.
Like the back of the table, under the table—
Yeah.
The back—images it doesn’t see. So the generative stuff is very different from reconstruction, in that it fills in the things that you can’t see.
Yeah. Okay. So—
Yeah.
All right. Now, the adult perspectives.
No, I mean, I love that.
Well, no, I was going to say, these are very much a tag team. We started this pod with that premise, and I think this is the perfect question to build on that further, because it truly is. We’re tag-teaming all of these together.
Every investment fundamentally starts with the same—maybe the same 2 premises. One is, at this point in time, we actually believe that there are N-of-1 founders for their particular craft, and they have to be demonstrated in their prior careers, right? So we’re not investing in every—you know, now the term is neo-lab—but every foundation model, any company, any founder who’s trying to build a foundation model. Contrary to popular opinion, we’re not invested in all of them, right? We have a very specific thesis around—
I don’t think people say that about you. No, they don’t.
They say that we’re big, we’re in everything. But if you think about Ilya, right, he’s at SSI. He’s been behind almost every foundational breakthrough for the last 15 years.
15 years.
If you think about the Thinking Machines team, right—Mira and John. John is the godfather of reinforcement learning.
I go through this because, if you think about each of the bets that we’ve made, it goes back to a very specific thesis about that person, the team they’ve assembled, and what they’ve done in a prior life. Obviously, we talked about talent wars. We do think that, at this particular moment in time, there are particular people who can move needles. Clearly, other companies believe that too; otherwise, they wouldn’t be willing to pay such crazy prices for single individuals.
And then, 2, we don’t think it’s a zero-sum game, right? If that were true, OpenAI—or actually just DeepMind—would be number 1 in everything, right? There’s clear value to specializations, like ElevenLabs. There have been so—
Oh my God, yeah.
Many audio models have hit the market.
So far, yeah.
They’re still freaking number 1, right? And they’ve created a ton of value for their customers, their investors, and their team. If you think about those 2 things put together, that’s sort of the foundation of our thesis when we back these foundation model companies.
Of course, the valuations sound astronomical when you think about current revenue—the numbers. I would say that’s the market out there, because they are raising larger dollars. They have compute needs, right? That’s 80% of a round that they typically raise.
But I think the thing that gets us excited about backing them is that the revenue growth has typically followed the capability breakthrough. It ties back to that question of the cyclical nature: Are you just funding it and then raising more funding? When there’s a real capability breakthrough, the demand is there, and the revenue growth is much faster than we’ve ever seen once it’s turned on.
There’s a company—I can’t share the name—but its product went GA and, in a few weeks, had tens of millions in revenue, right? We have SaaS companies—
I’ve seen this myself, yes.
Absolutely. We have SaaS companies—
Absolutely.
That have been in business for 7 years, and they get to the same level 7 years later, and the growth is eking along to whatever it is. By the way, they’re great companies—not at all diminishing what they’ve accomplished. But the fact is, to get to that revenue growth that quickly, it’s not just the 2 companies that people talk about. It’s really a lot of these—every domain has a specialist. We think if you can win that, you become very large very quickly, and that’s actually played out in the numbers.
Our viewers are going to roast us if we mention Thinking Machines and don’t discuss what happened. Founder splits happen, obviously. But I guess the question is: Is the thesis unchanged? What’s going on at Thinking Machines?
Yeah. We’re more excited than ever about them. They have some things that we’re not going to break news about on a pod. Obviously, they should share it themselves.
You know, I think when you bring a team of that caliber together, special things happen, and I think 2026 is going to be a big year for them. Obviously, some of the themes that we talked about before, even with the media news story—like the whole “something happens and then it’s everywhere instantly”—that’s a tough situation for any company to be in.
But to come out of that stronger than ever, I think we’re more bullish about Thinking Machines than even before.
And the story is Tinker. It’s around custom models—is that what we’re aiming for?
Yeah, and a bunch of stuff we can’t talk about here.
Okay. All right. Cool.
Yeah, absolutely. But no, that team is cooking, and I think they’ll be just fine. They’ll recover from the events in January.
Yeah.
We have a very privileged position on the boards of these companies, and I’ve never seen the perception of the truth be further from the truth industry-wide, ever.
I guarantee you that, for any of these gossipy things, it’s way off—the general sentiment, way, way off. What happens is that we’ve got this crazy game of telephone right now where there are always seeds of truth, but it gets so warped by the time it reaches us. We hear rumors all the time about stuff we’re directly involved in. We’re literally on the board; we’re the ones who did the thing. By the time it gets to us, it’s gotten so warped and twisted.
I think everybody’s excited. There’s a lot of focus. The schadenfreude is so high that people will just will things into being that didn’t exist. I don’t want to comment specifically on Thinking Machines, but—
It’s an important message to the general audience.
I will tell you, if you hear something on X, the chances that it accurately represents what it’s saying are very, very low.
Yeah.
I have never lost so much faith in the anon accounts on Twitter—
I know.
That just seem very confident in what they’re saying.
I know, yeah.
And could it be further from the truth? I had a couple-day stretch where I was like, “Oh my God, Twitter is mind poison.” And I love X, but—
Yeah, but we talk to each other all the time because we actually know, because we’re there. We’re there seeing these things, and Sarah will text me, whatever. It’s ridiculous.
The problem is that we realize things start taking on a life of their own, and then people assume that they’re real and everything. I think it’s very tough for founders because it’s tough enough fighting the real battle.
Actually building.
Now they’re fighting phantoms too. More and more, we’re just focusing on the business. I got this from the Cursor guys, which I really appreciate. Michael Truell is like, “Listen, heads down, focus on the business.”
Yep.
And I think that’s right.
And he absolutely crushed it. Yeah.
Yeah. And I think that’s right.
Yeah. Absolutely.
I think all founders should do that right now because the noise is so high.
Yeah. No, that team—the Thinking Machines team—has been back to business for weeks, so, yeah.
Yeah. Well, thank you for acknowledging that. It is just the hot topic of the moment.
Oh, for sure.
7. Cursor Owns The Application Layer
We’ve got to address the elephant in the room. Cursor, right? Obviously, you guys are big investors. 2025, I would say it’s Cursor’s year—and maybe decade.
But just going back to the discussion about how AGI would just consume everything, Cursor is the shining example of how you build an application layer that’s a wrapper—
But an extremely damn good one.
Yeah.
And I guess the general analysis of Cursor's development and what it means for everyone is: Is there a Cursor in every industry to be built?
Yeah. What's interesting about Cursor is that they actually developed an almost state-of-the-art model for a small fraction of the cost—1/100th of the cost or less—which, for a period of time, was the most popular coding model in the world. That's really crazy to think about.
I think they're just doing it in reverse. There are 2 approaches: You start with a foundation model and then verticalize up, or you start with the app and all of the product data, and you go down. They're the ones doing that.
I think any company that's building an app has to ask the margin question, which is: How do I extract margin from the tokens that are going through? Everybody has to be on the token path, and everybody has to ask that question.
I've just thought they've been incredibly thoughtful about it. One reason is that if you ask Michael, "What type of company are you?" They are a developer company for professional developers. That's what they are. They're a dev tools company. They're just focused on coding.
I mean, even if you didn't do AI, that's a massive market. They acquired Graphite. Listen, we were investors in GitHub—we know how big this market is. So that's a massive market even without becoming a model company.
But they've also been quite successful at doing their own models. I think it just shows you that if you are focused and you have a large use case, there's a huge opportunity not only to get the application, but to start building your own models. Are these going to be the only models people use? Of course not. But they are in a great position to serve great models, and they've demonstrated that.
Yeah. My thesis, which we're not going to have to go into here, is that what I've been calling agent labs—people who build on top of all the other models—will probably have a better time with the margins because they price against the end-user hours spent, or human labor.
Yeah.
Whereas models get commodity pricing per token.
Yeah. Mm.
And so, margin-wise, we know the inference economics for model labs. But for agent labs, the difference is the delta between token intelligence, which keeps going down, and human costs, which keep going up.
Yeah, yeah.
And so the margin should be higher.
They should be. The caveat to that is if the models go first-party, right?
Yeah, yeah.
What they can do is subsidize themselves.
Which is the Composer dream. Yes.
Yeah. They can subsidize themselves, and then they can charge third parties more. It's a very delicate dance because you're competing with your own customers.
We've seen this historically. We saw this with the cloud, with EC2. So this is not unusual. We saw this with the operating system. It's not unusual, but it's playing out very, very quickly.
Yeah. Thank you for joining us. That's all the time we have today.
It was such a pleasure.
You're welcome back anytime. And thank you for being so open and also leading the industry in so many areas. It's really inspiring to see. So thank you so much.
Thank you so much.
Thank you for having us.
Great. Thank you.