AI Interview Wiki
主题综述

AI 资本周期：泡沫还是 S 曲线 · AI Capital Cycle

主题综述

更新日志

主流共识

第一点：用股价定义"泡沫"是错的——要看 API calls / 资产负债表

几乎所有投资人都拒绝"看股价判断泡沫"。State of AI 那期讲得最彻底：

"People equate bubble to stock prices, which has nothing to do with anything other than fear and greed among investors. So I always look at bubbles should be measured by the number of API calls — or in the dot-com bubble, which people refer to, it should be amount of internet traffic. … There's no bubble detected in internet traffic. I would almost guarantee you, you won't see the bubble in number of API calls."
「人们把泡沫等同于股价——而股价跟别的没关系，只跟投资者的恐惧和贪婪有关。所以我一直认为，泡沫应该用 API 调用次数来衡量；在人们常提的 dot-com 泡沫里，则应该看互联网流量。……互联网流量里根本没检测到泡沫。我几乎可以向你保证，你在 API 调用次数里也不会看到泡沫。」
State of the AI Industry (Ep 12) · Episode 12 - State of the AI Industry

Casado 把判断锚在出资方的资产负债表上，并直接对照 dot-com 的债务驱动：

"[WorldCom] had $40 billion in debt. We also had 9-11 … The companies that are investing in these data centers have hundreds of billions of dollars on the balance sheet. … These companies have, you know, great balance sheets, great cash flow. And so like the fundamentals of like who's funding this is quite different."
「WorldCom 背着 400 亿美元债务，还赶上 9-11……而今天投资这些数据中心的公司，资产负债表上趴着几千亿美元现金。……这些公司资产负债表好、现金流好。所以'谁在出资'这件事的基本面，完全不同。」
Martin Casado (a16z) · Why This Isn't the Dot-Com Bubble

第二点：Jevons paradox——降本不减需求，反而触发指数增长

这一条跨多篇，几乎是共识。Nebius 讲得最直接：

"Every time we got intelligence cheaper. Same unit of intelligence cheaper, we are not reducing the consumption but we're increasing the consumption because we can just solve more complex tasks with the same budget."
「每次我们把智能——同样单位的智能——做得更便宜，我们不是在减少消费，而是在增加消费，因为同样的预算就能解更复杂的任务。」
Roman Chernin (Nebius) · 20VC: Nebius Co-Founder on AI Infrastructure Bubbles

State of AI 给了一个更强的版本——需求目前只被算力供给卡住，价格弹性那根杠杆还没开始拉：

"Demand is limited not by anything other than availability of compute today. … There's price elasticity where demand is infinite for compute. … We haven't even started to exercise the price elasticity lever. It's just we can't fulfill demand and it's limited by compute."
「需求目前只被一件事限制——今天的算力供给。……价格弹性那一侧，算力的需求是无限的。……我们甚至还没开始拉价格弹性这根杠杆。我们满足不了需求，它被算力卡死。」
State of the AI Industry (Ep 12) · Episode 12 - State of the AI Industry

2026 年这条得到两个新独立印证。 Accel 增长团队把"被算力卡死"直接说成当下共识：

"Probably at this point a pretty consensus view is that we're rate limited on infrastructure and we need to figure out ways to scale that."
「目前，关于基础设施的共识观点是我们受限于基础设施，我们需要找出扩展的方法。」
同场发言（逐字稿标注 Speaker 3 · Accel 增长团队） · Accel: The Quiet Firm Behind Facebook, Cursor, Nebius

需求那一侧的具体形态，Anthropic 平台团队给了一条纹理——agentic 负载天生"极度吃 token"，而且是靠 shadow IT 自下而上蔓延，企业想按都按不住：

"Oftentimes, the way that AI spend has erupted inside their company has been through some kind of shadow IT. … before you know it, half your org has found some way to install Cloud Code. And in that world, it is kind of hard to manage because these things are again, like they're very token hungry ultimately."
「通常，AI 在他们公司内部的支出方式是通过某种影子 IT 方式出现的。……你一转眼，你的半个组织就找到了安装 Cloud Code 的方法。在那个情况下，管理起来比较困难，因为这些东西最终确实是非常渴求代币的。」
Angela Jiang (Anthropic) · Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden

2026-08，这条共识第一次等来了需求侧的 principal。 此前说"需求无限"的全是投资人和平台团队；这次是把 capex 承诺签在自己名下的人。Altman 把 OpenAI 当年的原始判断讲得很直白：

"We could just tell that we were on this exponential of model improvement. That part we were very confident about. We knew it was going to keep going. We were pretty sure, although as you mentioned, we underestimated, that as the models got better and better, if we could continue to drive costs down, the demand for AI at a sufficiently high level and a sufficiently low price was basically uncapped."
「我们可以清楚地感受到我们正在经历模型改进的指数增长。这部分我们非常有信心。我们知道这会继续下去。我们很肯定，尽管正如你提到的，我们低估了，随着模型越来越好，如果我们能够继续降低成本，AI 的需求在足够高的水平和足够低的价格下基本上是没有上限的。」
Sam Altman (OpenAI)（逐字稿标注 Speaker 3） · Sam Altman - How to Make an Abundant Future (ILTB EP.484)

更值得记录的是，被主持人逼问"如果两年后算力过剩，故事会是什么"时，他没有回避，亲手给出了两个触发条件——这可能是全语料里需求侧最诚实的撤退条款：

"It does feel possible. If the models get so smart and so efficient that they can do everything we need and build every piece of software we want. And if the bounds of our attention are such that they just cannot absorb more than what it turns out a fairly limited amount of compute can do, then we can get into oversupply. Also, if we don't drive the cost curve down because we hit some sort of scaling wall, we could also get an oversupply. The observation about uncapped demand implies a certain price."
「感觉是有可能的。如果模型变得如此聪明和高效，以至于它们可以完成我们需要的所有事情，构建我们想要的每一款软件。并且如果我们的注意力范围如此有限，以致于它们无法吸收超出相对有限的计算能力所能做到的，那么我们可能会出现过剩。而且，如果我们无法降低成本曲线，因为我们碰到某种规模瓶颈，我们也可能会面临过剩。关于无上限需求的观察暗示了一定的价格。」
Sam Altman (OpenAI)（逐字稿标注 Speaker 3） · Sam Altman - How to Make an Abundant Future (ILTB EP.484)

（同场他还把"不需要高毛利"说破——"We have so much usage of our models that we do not need to be a gigantically high margin business to be able to afford model training. So much of our future compute plans will be used to sell inference to customers that even if we can enjoy a modest margin on trillions of dollars of revenue, we can go afford to train some giant models"。这句与阵营 G Giffon 的"沃尔玛化"意外同向：连头部实验室自己都按"低毛利 × 超大规模"在做规划。）

第三点：这是史上最大的预算重配——不是凭空创造市场

Casado 用一个具体的反问把"预算重配"讲活了——大公司不是在新增开支，是在把存量预算从旧业务搬向 AI：

"Do you think Meta is spending more money on VR or AI? … These companies spend lots of money on infrastructure historically."
「你觉得 Meta 现在花在 VR 上的钱多，还是 AI 上的多？……这些公司历来就在基础设施上砸很多钱。」
Martin Casado (a16z) · Why This Isn't the Dot-Com Bubble

（这一点在本轮被 Jeremy Giffon 从根上挑战——他认为因果反了：不是需求把预算搬过来，而是资本找不到出口自己造出了这些高 capex 公司。见下方"分歧"阵营 G。）

2026-08 的补充把"重配"的投资逻辑说得更冷峻。 复盘 Google AI 人才出走时，Friedberg 给出的解释不是技术叙事，而是资本配置——基础设施 capex 已经成了"低 beta"的那个选项，模型研发反而是高风险资产：

"Google has made a commitment to deploy $200 billion in CapEx this year in AI infrastructure, data center build-out. Because of the CapEx and accelerated depreciation, making an investment in AI compute in the US right now is hugely tax advantaged. And because of the extreme demand for compute, It's a pretty obvious kind of ROIC model, return on invested capital. … So the way I would frame it is CapEx is high alpha, low beta in data center infrastructure, that capital. And model development theoretically could be high alpha, but it's very high beta. It's a very risky way to deploy capital."
「Google 承诺今年在 AI 基础设施、数据中心建设上投 2000 亿美元 CapEx。因为 CapEx 和加速折旧的关系，现在在美国投 AI 算力在税务上极其划算。而且算力需求极端旺盛，这是一个相当显而易见的 ROIC（投入资本回报率）模型。……所以我的框架是：数据中心基础设施的 CapEx 是高 alpha、低 beta 的资本；而模型研发理论上可能高 alpha，但 beta 极高——那是一种非常冒险的资本配置方式。」
David Friedberg（评 Google $200B capex 与 AI 人才流向） · Google's AI Brain Drain, SpaceX's Huge Quarter, Airtable's 90% Collapse

（同场 Brad Gerstner 补充 Satya 引 Morgan Stanley 的口径：token-as-a-service 基础设施业务的投入资本回报率已超过 30%——"预算重配"不再只是战略姿态，而是有 ROIC 报表背书的常规资本决策。这也让本点与阵营 G 的对撞更尖锐：若重配自有报表理性，Giffon 的"资本自造供给"至少在超大规模云厂商这一层暂时讲不通；但两说在 neocloud / 中间层身上仍未分胜负。）

分歧在哪

共识止于"不是简单的股价泡沫"。再往下，先前吵三件事——有没有过度投资的系统性风险、价值落在哪一层、用什么单位测增长；上一轮劈出两条新裂缝：同一批赢家能给出相反的风险读数（Accel vs Founders Fund），以及到底是需求拉动建设、还是资本推动供给（Giffon 的逆因果）。2026-08 这轮再添一条：度量问题（Randle：估值软件公司的旧尺子全断了）。

阵营 A · "不是泡沫，是扩张早期"——Nebius / Casado 乐观派

Nebius 明确把当下定义为扩张期而非泡沫，且认为真正的长期威胁不是泡沫破裂、而是过度集中：

"I think that the main threat for Nebius as a business is the world will be too much consolidated. … If you end up in the world where, I don't know, three, five supermodels, super companies, super empires control the world, then Nebius or companies like Nebius will be needed only to help them maybe serve their needs on physical layer."
「我认为对 Nebius 来说，作为一项业务，主要威胁是世界会过于整合。……如果你最终身处于一个世界中，我不知道，三五个超级巨头、超级公司、超级帝国掌控世界，那么 Nebius 或像 Nebius 的公司只会被需要来帮助他们在物理层面上服务。」
Roman Chernin (Nebius) · 20VC: Nebius Co-Founder on AI Infrastructure Bubbles

（"鲨鱼——你动着才活着"确是 Chernin 的原话，但出自访谈结尾谈公司永不停步、庆祝太少的另一段回答——"never stop. We cannot stop. It's like a shark. You're alive when you move, right?"——与集中化威胁无关，此前版本误把两段拼进了同一条引用。）

但 Casado 自己留了一个撤退条款——他把问题问得很诚实：

"The question is, are we over-investing relative to long-term demand? And if so, do we have the economic reserves to stop some sort of, you know, systemic unraveling?"
「问题是：相对于长期需求，我们是不是投资过度了？如果是，我们有没有经济储备去阻止某种系统性瓦解？」
Martin Casado (a16z) · Why This Isn't the Dot-Com Bubble

这期还留下一个所有怀疑论者都该记住的历史模式——琐碎的早期用例不等于不重要。（那句精炼的"Every major technology wave starts with use cases that look trivial and every time skeptics confuse silliness with insignificance"是主持人 Tim Higgins 开场旁白里的台词，不是 Casado 说的；Casado 自己的版本是剑桥咖啡壶的故事：）

"You know, in the mid 90s, you had these things that just looked like toys and they were so silly and you made fun of them. But the reality is, is that video of a coffee pot in no small way became Netflix and people could see that. … But it all turned out to be true in the long run. And I think that that will be the story of this realm, which is that we see a lot of like anime and a lot of silly use cases. And then we tend to poo-poo this as like, oh, this isn't serious stuff. … But like, this is what the future always looks like."
「在 90 年代中期，你看到的这些东西看起来就像玩具，它们非常愚蠢，你会嘲笑它们。但事实是，咖啡壶的视频在很大程度上变成了 Netflix，人们可以看到这一点。……但从长远来看，一切都变成了现实。我认为这将是这个领域的故事，那就是我们看到了很多动漫和很多愚蠢的用例。然后我们倾向于对此不屑一顾，认为这不是严肃的事情。……但这就是未来一贯的样子。」
Martin Casado (a16z) · Why This Isn't the Dot-Com Bubble

阵营 B · "有泡沫，但泡沫对算力买家是好事"——David Cahn 的辩证派

Cahn 不否认泡沫，但把它翻成一个反直觉的判断——过度生产算力会压价、利好算力消费者：

"Consumers of compute benefit from a bubble because if we overproduce compute, prices go down."
「算力的消费者受益于泡沫——因为如果我们过度生产算力，价格就会下降。」
David Cahn (Sequoia) · 20VC: Sequoia's David Cahn

但逐字稿里他的核心贡献，是把"风险在沿着生态链往下传"讲得极其具体——这是摘要层完全压扁的一段：

"Microsoft walked away from two data centers. And it sent a message to the market, like, hey … we're not going to be this risk absorber in the ecosystem anymore. … Oracle obviously stepped up … CoreWeave has really stepped up … Oracle and CoreWeave are a lot smaller than Microsoft and Amazon. They simply can't absorb as much risk as Microsoft and Amazon could. And so the chip companies are now stepping up … to finance this build out where the demand on the other side is not so clear."
「微软放弃了两个数据中心，向市场传了个信号：嘿……我们不再当这个生态里的'风险吸收器'了。……Oracle 明显顶上来了……CoreWeave 也实实在在顶上来了……可 Oracle 和 CoreWeave 比微软、亚马逊小得多，它们根本吸收不了微软、亚马逊能吸收的那么多风险。于是芯片公司现在也顶上来了……去给这场建设出资，而需求那一侧并不那么清晰。」
David Cahn (Sequoia) · 20VC: Sequoia's David Cahn

而且他点出这条链里的"循环融资"——芯片公司既出资又把它记成营收，资金成本被压到近乎为负：

"The chip companies also get to book this as revenue. So their cost of capital is very low. One might even say their cost of capital is negative in some of these deals. … Moving from expensive capital from these big tech companies to cheaper capital from the chip companies themselves who get to benefit from circularity — I think that's probably been the biggest change in the last 12 months in AI."
「芯片公司还能把这笔记成营收，所以它们的资金成本非常低——某些交易里你甚至可以说是负的。……从大科技公司那种昂贵资本，转向芯片公司自己这种因'循环'而便宜的资本——这大概是过去 12 个月 AI 里最大的变化。」
David Cahn (Sequoia) · 20VC: Sequoia's David Cahn

但他对"债务会引爆泡沫"这个流行叙事泼了冷水——他认为真要瓦解，是股权式瓦解，痛的是普通人的股票组合：

"The AI buildout today has been equity funded and cash funded. … To the extent that the bubble unwinds at some point, it's going to be an equity unwind. And what that looks like is 40% of the S&P 500 is basically a bet on AI. … A greater percentage of Americans' net worth is equities than I think ever before in history. And so people are going to feel this in the form of their equity portfolio going down more likely than some credit unwind where the banks get affected and all of that stuff."
「今天的 AI 建设主要是股权和现金出资的。……所以真要在某个点瓦解，那会是一次股权式瓦解。标普 500 有 40% 本质上是在押 AI。……美国人净资产里股票的占比，比历史上任何时候都高。所以大家会以'股票组合缩水'的形式感受到它——而不是 2008 那种银行被波及的债务瓦解。」
David Cahn (Sequoia) · 20VC: Sequoia's David Cahn

阵营 C · "价值落在卖短缺的人手里"——Coatue 的 shortage 框架

Coatue 给了这批里最锋利的分类——sellers vs buyers of shortage，并点明资本如何在两者间转移：

"The buyers are being punished because their spending, their capex is so high that their near-term cash flow is gone. And that transfer is happening directly to the pockets of the sellers of the shortage."
「买方正在被惩罚——因为他们的开支、capex 太高，近期现金流被吃光了。这笔（被惩罚的代价）直接转移进了'短缺卖方'的口袋。」
Coatue CIO (Jaimin Rangwalla) · Exclusive Interview: Coatue CIO on AI's Biggest Winners

稍后展示"卖方 vs 买方"图表的那段发言补上了翻转警告——逐字稿此处只标"Speaker 2"、未具名（按上下文很可能仍是 Rangwalla，但无法从逐字稿确认）：

"But I would say that at one point, there will be a power that shifts the other way. And we need to be extremely, extremely aware and well prepared to do that."
「但是我想说，总会有一个力量将局面转向反方向。我们需要对此保持极度警觉，并做好准备。」
同场发言（逐字稿标注 Speaker 2） · Exclusive Interview: Coatue CIO on AI's Biggest Winners

而且 Coatue 观察到一个被忽略的架构转变——agent 时代算力配比正在从 GPU-heavy 翻向 CPU-heavy：

"For the greater part of the last three or four years, it was very much a one CPU and even eight GPUs … But now the ratio is actually moving from to one CPU to four GPUs … We think it actually has a chance to flip the opposite direction, which is one GPU to four CPUs. Some people, very aggressive, say they could go to one GPU to eight CPUs. … those new tasks … don't require the parallel compute, but require more just serial task completion."
「过去三四年的大部分时间，基本是 1 个 CPU 配 8 个 GPU……现在比例实际正移到 1 CPU 配 4 GPU……我们认为它甚至有机会翻到反方向，也就是 1 个 GPU 配 4 个 CPU，激进的人说可能到 1 配 8。……这些新任务……不需要并行计算，而更多是串行任务完成。」
Coatue CIO (Jaimin Rangwalla) · Exclusive Interview: Coatue CIO on AI's Biggest Winners

（Coatue 同时把度量单位从 GPU 换成 gigawatt——"the gigawatt is almost the atomic unit of where the growth in AI is coming from"；但若上面这个 CPU 配比转变是真的，"卖短缺的人"名单本身也会变。）

2026-08，Sacks 从模型层给了"短缺卖方"一个更结构化的版本——双层市场：

"I think it's a very powerful duopoly. I don't think it's being commoditized. I think that what we're evolving to is a two-tier market structure where there's a market for frontier intelligence and there's a market for let's call it kind of commodity or lagging intelligence, whatever you want to call it, that's six to 12 months behind. There is a market for those tokens, those models, but the reality is you can't charge anything for the weights. You can charge for the compute. You can charge for the inference that you're providing. You can charge for Essentially consulting services to help put the whole thing together. But if you're not at the frontier, you can't charge for the model layer itself."
「我认为这是一个非常强大的双头垄断，我不认为它在被商品化。我认为我们正在演化成一个双层市场结构：一层是前沿智能的市场，另一层是那种落后六到十二个月的、姑且叫商品化或滞后智能的市场。那些 token、那些模型也有市场，但现实是你没法为权重收一分钱。你可以为算力收费，可以为你提供的推理收费，可以为把整套东西拼起来的咨询服务收费。但如果你不在前沿，你就无法为模型层本身收费。」
David Sacks · Google's AI Brain Drain, SpaceX's Huge Quarter, Airtable's 90% Collapse

（他的证据是增长率本身："The latest we heard is Anthropic is now over 80 billion of ARR. Started the year at 10."——前沿层像 Apple，用户不是最多、钱全进它口袋；非前沿层像 Android，量大但只能收算力和实施的钱。若此说成立，Coatue 的"短缺卖方"名单在模型层只剩两个名字——这与阵营 E"剥掉四五家赢家全是赌场钱"竟是同一张图的多空两读。）

阵营 D · "S 曲线 + 硬件去商品化"——Whale Rock Sacerdote 的周期派

Sacerdote 用 S-curve 框架，给了一个反 FOMO 的实操建议——晚入场没关系：

"It's okay to be late. It's okay to miss the first 1, 2, 3 years in a lot of cases because if the top of the S-curve is half a trillion, the growth can go on for a long time."
「晚一点没关系。很多情况下，错过头 1、2、3 年也没关系——因为如果 S 曲线的顶端是五千亿，增长可以持续很久。」
Alex Sacerdote (Whale Rock) · How to Invest Through Technology Cycles

他对价值落点的判断是数据中心硬件的"去商品化"——而且讲得极具体、极生动：

"We call it the decommoditization of the hardware industry. … To do an AI server, it's liquid cooled … It's a two or three hundred thousand dollar piece of machinery, whereas an old server was five thousand dollars. If it breaks, you just throw it away. If this thing breaks, the whole thing goes down. So you become a commodity like supplier to like selling a critical part on a plane. You'll never get swapped out."
「我们管它叫硬件行业的去商品化。……做一台 AI 服务器是液冷的……它是一台二三十万美元的机器，而旧服务器才五千美元。旧的坏了你扔掉就行；这玩意儿一坏，整套系统就宕。所以你就像一个商品供应商，像是在出售飞机上的关键部件——永远不会被换掉。」
Alex Sacerdote (Whale Rock) · How to Invest Through Technology Cycles

（"So you become a commodity like supplier to like selling a critical part on a plane"为逐字稿原文，疑是转写含混——按上下文语义应为"从商品化供应商，变成卖飞机关键零件式的供应商"。）

"A regular server, you need 10 layers. These AI servers, you need a 40 layer [PCB]. And there's very few PCB suppliers that can make this."
「普通服务器需要 10 层 PCB，这些 AI 服务器需要 40 层。而能做 40 层的 PCB 供应商极少。」
Alex Sacerdote (Whale Rock) · How to Invest Through Technology Cycles

但 S-curve 乐观派头上还有一盆冷水——光在好赛道里不够。这个点是主持人 Patrick O'Shaughnessy 先抛出的，Sacerdote 表示同意（"you're right"），并用一串手机时代的阵亡名单把它钉死；"出局"那半句其实是主持人接的话（"And so is out."）：

"But you're right that if you don't have a competitive advantage, you can be in the best S-curve of all time. … But if your name was RIM, Palm, Nokia, HTC, LG, Motorola, I can go on forever. Zero, zero, zero, negative, negative, negative, negative."
「但你说得对——如果没有竞争优势，你可能身处史上最佳的 S 型曲线中。……但如果你的名字是 RIM、Palm、Nokia、HTC、LG、摩托罗拉，我可以不停地数下去：零，零，零，负，负，负，负。」
Alex Sacerdote (Whale Rock) · How to Invest Through Technology Cycles

阵营 E · "价格与现实脱节、像极了 2021"——Founders Fund 的不安派（2026-07 新增）

新增的这期把一位重量级的不安派摆上台面。Trae Stephens 直言当下让他很不舒服：

"I am Very uncomfortable. I am not enjoying this moment at all. Maybe some other people are."
「我非常不舒服。我一点都不享受这个时刻。也许其他人正在享受。」
Trae Stephens (Founders Fund) · Uncapped #53: Trae and Delian from Founders Fund

同场一位在 Benchmark 与 Founders Fund 都待过的发言人（逐字稿标注 Speaker 1）把这种不安提炼成"泡沫的关键是一种二元性"——巨额账面回报被四五家赢家绑架，剥掉它们剩下的全是赌场钱和"一定会涨"的必然感：

"I think this duality is the key to a bubble. These last three years, unquestionably, is going to be the period where venture capitalists have made the most amount of money in history. … But then when you Strip out those four companies or five or six companies, then like you start one playing with house money and two, you start assuming that everything is going to like be daisies and roses because of the success that has been baked in other parts of the market."
「我认为这种二元性是泡沫的关键。过去三年，无疑是风险投资家历史上赚取最多钱的时期。……但是当你剔除那四家公司或五家公司或六家公司后，你就像是开始用房钱在玩，而且，你开始假设一切都会像鲜花和玫瑰一样，因为其他市场部分的成功已经确定。」
同场发言（逐字稿标注 Speaker 1，语境为在 Benchmark 与 Founders Fund 均任职的联席主持） · Uncapped #53: Trae and Delian from Founders Fund

估值端的硬数字最能说明"脱节"。Delian 用 Ramp 的过山车讲清了 2021 式的估值弹跳：

"And so one of those investments was the $9 billion post-ramp round. Nine months later, the company did a down round at $6 billion base gain valuation. … It's now obviously at a $40 billion valuation. And so even the IRR with the down and up, Still make sense."
「其中一笔投资是 90 亿美元的后期融资轮。九个月后，该公司以 60 亿美元的基础估值进行了贬值融资。……但现在显然达到了 400 亿美元的估值。因此，即便是贬值和升值，内部收益率依然有意义。」
Delian Asparouhov (Founders Fund) · Uncapped #53: Trae and Delian from Founders Fund

而入场价整体已被抬到另一个量级——Delian 说最新基金的平均入场估值约 $700M，是前五只基金 $80–100M 的七八倍：

"in our latest venture fund that we kicked off maybe like six or seven months ago, the like average entry price right now, I think is something like 700 million in valuation. And if you compare it to the prior five … it would probably be like 80 to 100 million … entry point average price."
「像在我们大约六七个月前启动的最新风险基金中，现在的平均进入价格，我认为在 7 亿美元估值左右。如果你将它与之前的五项进行比较……它可能大约是在 8,000 万到 1 亿之间……入场点的平均价格。」
Delian Asparouhov (Founders Fund) · Uncapped #53: Trae and Delian from Founders Fund

不过这一派也承认，为什么估值敢这么给——Delian 说 Anthropic 是自 Google 以来第一家收入曲线更陡的公司，这才把"必然感"喂到了整个链条：

"Anthropic in this cycle is literally the first company that since Google has actually shown an even steeper curve. And then I think that feeds into both the like, hey, you're playing with house money."
「在这个周期中，Anthropic 真的成为了自谷歌以来第一家展现出更陡增曲线的公司。我认为这也造成了，嘿，你是在用赌场的钱来做事的感觉。」
Delian Asparouhov (Founders Fund) · Uncapped #53: Trae and Delian from Founders Fund

Trae 还提醒：这轮的热甚至不止 AI，硬科技"如果不是更热"，而且完全没经历过 vertical AI SaaS 那种回调——这与 Coatue、Sacerdote 只盯 AI 层的框架是错位的。

阵营 F · "市场完全理性，我们反而低估了 AI"——Accel 的乐观派（2026-07 新增）

正对面站着 Accel 的增长团队。同一批 AI 赢家里他们也重仓（Cursor、Anthropic、Nebius），但读数完全相反——他们说市场是理性的，错就错在大家还低估了 AI：

"So I actually think the market is totally rational on this. I think if anything, a lot of us have underestimated the potential of AI and the value creation that it offers."
「所以我实际上认为市场在这方面是完全理性的。我认为如果有的话，我们很多人低估了人工智能的潜力以及它所带来的价值创造。」
同场发言（逐字稿标注 Speaker 1 · Accel 增长团队） · Accel: The Quiet Firm Behind Facebook, Cursor, Nebius

支票尺寸的跳变最能量化这份信念——首只增长基金才 $480M，如今单笔投一家公司 $5 亿、乃至超过 $10 亿是常态，逻辑是"如果我们相信这些公司能在很短的持有期内变成万亿/多万亿公司，就该在支票尺寸上体现出来"：

"Our first growth fund was $480 million when we first joined. It is not unusual for us to invest $500 million into a company, actually to invest over a billion dollars into a company. … if we believe that these companies can be trillion, multi-trillion dollar companies within a very short hold period, we should reflect that in check size."
「我们的第一支成长基金在我们刚加入时为 4.8 亿美元。对我们来说，投资 5 亿美元到一家公司并不罕见，实际上投资超过 10 亿美元到一家公司的情况也是存在的。……所以如果我们相信这些公司可以在非常短的持有期内成为万亿和多万亿的公司，我们应该在支票金额上有所反映。」
同场发言（逐字稿标注 Speaker 1 · Accel 增长团队） · Accel: The Quiet Firm Behind Facebook, Cursor, Nebius

万亿公司数量的跃迁，是他们眼里"理性"的锚——从 0 到 5 到 14，下一周期看 $10T：

"10 years ago, there were zero publicly traded companies worth a trillion dollars. Five years ago, there were five companies worth a trillion dollars. As of today, there's 14 with probably three or four private companies that are pre-IPO that we could all point to. And I think it's fair to expect there will be $10 trillion companies and beyond over the next cycle."
「想想看，十年前，市值在一万亿美元的上市公司为零。五年前，有五家公司市值达到一万亿美元。到今天为止，已经有 14 家，还有三到四家在 IPO 前的私有公司，我们都能指出来。我认为合理的预期是下一个周期会有十兆美元的公司甚至更多。」
同场发言（逐字稿标注 Speaker 2 · Accel 增长团队） · Accel: The Quiet Firm Behind Facebook, Cursor, Nebius

capex 的绝对量级他们也直接摆出来——Google 十多年来首次股权融资 $80B、"花得比全部现金流还多"，而他们的解读是"我们才吃第一口、才刚开始"：

"Google announcing that they're going to do their first equity raise in Over a decade for 80 billion dollars to go spend even more than all of their cash flow on infrastructure. So I still view it as we're eating one and we're just getting going."
「谷歌宣布他们将在十多年内首次进行股权融资，筹集 800 亿美元用于基础设施支出，甚至超过他们的现金流。所以我仍然认为我们正在吃一顿大餐，只是在开始。」
同场发言（逐字稿标注 Speaker 3 · Accel 增长团队） · Accel: The Quiet Firm Behind Facebook, Cursor, Nebius

（同场还有几个印证资本集中度的硬数：Nebius 的 $150M PIPE 16 个月里翻 13x、Nebius–Meta $26B 交易；Q1 有近 $2000 亿涌向两三家公司；他们判断这轮后期投资的回报会追平早期——"park a bunch of money in it and still outperform"。这和阵营 D 的"晚入场没关系"从操作上竟然同向，却与阵营 E 的"像极了 2021"正面撞车。）

2026-08，"下一周期看 $10T"遭到正面顶撞。 Elad Gil 承认拐点、否认外推：

"basically what we had is over the last five years or so, we had three companies roughly go from close to zero to a trillion dollars in market cap, right? Anthropic basically didn't exist five years ago. OpenAI was still quite early. … I think a lot of people now are assuming that there's a bunch of other trillion-dollar companies that will be formed in three to five years. That's unprecedented in human history. Usually, it takes 20 years. SpaceX actually took Since the early 2000s and Google took since the 90s, you know, these are usually 15, 20 year arcs. And then we had this weird five-year inflection. … And maybe some of these will over the next decade, but it's unlikely that we'll see that many more in the next three to five years. I mean, there's one I can think of that could maybe get there, but not multiple."
「基本上，过去五年左右，我们有三家公司差不多从接近零涨到一万亿美元市值，对吧？Anthropic 五年前基本还不存在，OpenAI 也还很早期。……我觉得现在很多人在假设，未来三到五年还会再长出一批万亿美元公司。这在人类历史上没有先例。通常这要花 20 年。SpaceX 实际上从 2000 年代初走到现在，Google 从 90 年代走到现在——通常都是 15 年、20 年的弧线。然后我们撞上了这个诡异的五年拐点。……也许其中一些会在未来十年做到，但未来三到五年不太可能再看到那么多。我是说，我能想到一家也许能到，但不会是多家。」
Elad Gil（No Priors，与 Sarah Guo 对谈） · Chasing Trillion-Dollar Companies, Founder Ambition, Token Budgets

（同场还给了一个结构性理由：算力被差异化地配给极少数研究员、compute 短缺本身在强制执行寡头结构——增长斜率是被算力上限锁住的，不是被想象力。注意：Accel 的 0→5→14 外推与 Elad 的"间断平衡的反常"用的是同一组数据——又一个"同一个数字、两种含义"。）

阵营 G · "是资本推动了供给，不是需求"——Jeremy Giffon 的逆因果（2026-07 新增）

Giffon 把整套乐观叙事的因果关系直接掉了个头。共识第三点说"需求真实、大公司只是搬预算"；他说反了——是负利率时代过剩的资本找不到出口，像海绵一样"没地方去就自己造一个地方去"，这些高 capex 的 AI 公司是资本的下游产物：

"businesses and assets are sponges for capital. And that the excess in an era of back then we had negative interest rates and stuff, the capital has to go somewhere. And if it has nowhere to go, it will create somewhere for it to go. … So these companies almost got created downstream of capital, which I think is a little bit different of the narrative than most people would look at."
「商业和资产是资本的海绵。在那时的负利率时代，资本必须发挥作用。如果没有地方投放，它会创造出一个地方来。……因此，这些公司几乎是在资本流出的情况下被创造出来的，我认为这与大多数人所看到的叙事有些不同。」
Jeremy Giffon · The Billion Dollar PDF (Invest Like the Best EP.481)

顺着这条逻辑，他给了一个直击"价值落在哪一层"的经济学判断——软件正从"卖字符串"（零边际成本、高毛利）切换到"卖算力"（每次都要真算、边际成本非零），高毛利常态会消失，取而代之的是低毛利、薄净利、超大规模——软件的"沃尔玛化"：

"now I think we're in an era where we're selling compute. And selling compute, you can't write the prompt once and then sell copies of the output. You have to do the compute every single time. And so the marginal cost obviously is not zero. … this era of high gross margins being the norm is just going to go away. And what's going to make up for it, I think, is lower gross margins, much thinner net margins, and much more scale. … a bit of a Walmart effect in software, which is that the future looks like low gross margins, razor thin net margins, huge scale."
「现在我认为我们正处于一个售卖计算的时代。在售卖计算时，你不能只写一次提示，然后出售输出的副本。你每次都必须进行计算。因此，边际成本显然不是零。……高毛利时代将不再是常态。而且我认为能弥补这一点的，是较低的毛利，更薄的净利，和更大的规模。……这是软件中的沃尔玛效应，未来看起来将是低毛利，超薄净利和巨大的规模。」
Jeremy Giffon · The Billion Dollar PDF (Invest Like the Best EP.481)

他还给资本流向补了一个"叙事驱动"的机制——所谓 billion-dollar PDF：一个恰好在对的时机被说出口的故事就能聚起数十亿资本，资本像追球的小孩一样跟着叙事满场跑：

"you can form billions of dollars of capital one way or another around simply setting a new idea. And then maybe you can think of capital as 10-year-olds playing soccer. They all sort of just follow the ball around. The capital just follows the billion-dollar PDF around the field."
「你可以围绕新想法形成数十亿美金的资本。然后你可能可以把资本看作是 10 岁的孩子在踢足球。他们都围着球转。资本就像跟着十亿美金的 PDF 在场地上移动。」
Jeremy Giffon · The Billion Dollar PDF (Invest Like the Best EP.481)

而这套"叙事定价"的另一面是：他认为市场根本没那么有效，连全球最大的公司都谈不上被合理定价——

"the 52-week variance on these things is like nearly 100% for the biggest companies in the world. And so they're not priced well at all."
「这些企业的 52 周波动率几乎达到 100%。所以它们根本没有被合理定价。」
Jeremy Giffon · The Billion Dollar PDF (Invest Like the Best EP.481)

（Giffon 对当下最诚实的一句是"No one knows if it's the death of software … the most unprecedented and uncertain time since the transition of the internet"——他给创业者的对策是保住 optionality、别被 cap table 锁死，因为 per-seat 定价可能整个消失、要转 usage。这与"市场完全理性"是两个世界。）

阵营 I · "旧尺子全断了"——Benchmark Ev Randle 的范式失效派（2026-08 新增）

Randle 不站多也不站空——他说的是：你们吵架用的仪表盘本身失灵了。规模不再自动降险：

"I think that the thing that's changed massively in the new AI paradigm is that you can have businesses that are well over a billion dollars in revenue that haven't proven out their unit economics. They haven't proven out durable product differentiation. In many ways, it feels like The risk of impairment or a company either going to zero is the worst case scenario, but even just like being devalued from its last round in a significant way, it feels like the risk of that over time is actually sort of flat or even like maybe there's like a weird positive correlated relationship with scale and risk. And it's a very disoriented thing because you're just very used to The fact that the bigger you are, kind of the safer the company is. And it doesn't feel like that anymore."
「我认为在新的人工智能范式中，发生了巨大的变化，企业的年收入超过十亿美元，但尚未证明他们的单位经济学。他们尚未证明可持续的产品差异化。在很多方面，它感觉像是公司出现减值或完全归零的风险是最糟糕的情况，但即使只是从上一轮融资中显著贬值，它似乎随着时间的推移，这种风险实际上是平稳的，甚至可能与规模和风险之间存在一种奇怪的正相关关系。这是一件非常令人困惑的事情，因为你已经习惯于公司越大，相对来说越安全。而现在感觉并不是这样。」
Ev Randle (Benchmark) · Benchmark's AI Bets: Cerebras, Sierra, Legora, Fireworks

连"高毛利=好公司"都翻转成了红旗：

"And then now gross margins, if your gross margins are high, that's actually a bad thing. Because AI inference costs a lot of money. And if you have an AI product with high gross margins, that means that no one's using your AI features. … And so it's like all these bizarro things where it's like all of the golden rules of the past that defined like the spreadsheet investing era are all gone. And actually, the most popular companies and categories are almost the inverse of all these golden rules."
「现在如果你的毛利率很高，那实际上是件坏事。因为 AI 推理成本很高。如果你有一个毛利率高的 AI 产品，这意味着没有人在使用你的 AI 功能。……所以这就像所有这些怪异的事情，过去定义电子表格投资时代的黄金规则都消失了。事实上，最受欢迎的公司和类别几乎与这些黄金规则相反。」
Ev Randle (Benchmark) · Benchmark's AI Bets: Cerebras, Sierra, Legora, Fireworks

而后期轮的绝对量级已经跳出历史坐标系——他自己算过一笔账：

"When you think about the Anthropic $380 billion round, if Anthropic was to go public and get liquid at a $1.5 trillion valuation, … the gross return of their $380 billion round, the $30 billion that they raised at $380, it would return 35 times that of the Snowflake pre-IPO round. So it's 35 Snowflake pre-IPO rounds in a single round. And like I have friends, like I know several people that have like $3 to $4 billion invested into Anthropic. And like, it's like so hard to even talk about, like these numbers are so big that it's hard to even comprehend because five years ago, a growth fund was like a billion dollars."
「当你想到 Anthropic 3800 亿美元的融资轮时——如果 Anthropic 公开上市并以 1.5 万亿的估值获得流动性，……他们 3800 亿那一轮的毛回报——在 3800 亿估值上融的那 300 亿美元——将是 Snowflake 上市前轮次的 35 倍。等于一轮里装下 35 个 Snowflake pre-IPO 轮。我有朋友，我认识好几个人，在 Anthropic 里投了 30 亿到 40 亿美元。这真的很难谈论，这些数字大到难以理解，因为五年前，一支增长基金也就十亿美元左右。」
Ev Randle (Benchmark) · Benchmark's AI Bets: Cerebras, Sierra, Legora, Fireworks

（这一派对阵营 E/F 之争是釜底抽薪：如果 Rule of 40、毛利率、"规模=安全"全部失效，那么 Accel 的"完全理性"与 Founders Fund 的"像极了 2021"其实都在用旧仪表盘读新机器——谁都可能对，谁都无法证明。Randle 自己给出的条件分岔：若递归自我改进兑现，前沿实验室有巨大定价权；若能力触顶、蒸馏把开源追到前沿的 95%，token 溢价被摧毁——同一套资产在这两个世界里差一个数量级。）

历史对照 · "前 17 年每年 1%，最后三年 83%"——Kopelman 的警钟

Josh Kopelman 提供了一个让所有"这次不一样"叙事冷静的历史数据：

"1% a year for the first 17 years, 83% in the last three. And why is that important? Because the last three, it was universally acknowledged, we're in a bubble. … If you had said, oh no, we're in a bubble, and you said, I want to sell and exit now, you would have given up 83% of your profit."
「前 17 年每年 1%，最后三年占 83%。为什么这重要？因为最后那三年，所有人都公认我们身处泡沫。……如果你当时说，不，我们在泡沫里，然后你说，我要现在卖出退场，你就放弃了 83% 的利润。」
Josh Kopelman (First Round) · How To Make Money in Venture

这句话的双刃：它既说明"泡沫期也能赚最多钱"，也说明"被普遍认为是泡沫"和"利润最高"可以同时成立——这恰恰是当下最难拆的结。本轮 Founders Fund 那位不安派的"二元性"其实是同一枚硬币的另一面：所有人都在赚最多的钱，和所有人都隐隐觉得像 2021，是可以同时为真的。

2026-08，David Frankel 站上了同一枚硬币，并把它变成操作手册。 他对这一波的定位比谁都坦率——好莱坞式的幂律，垫底的是尸体：

"The historical or anachronistic view on this would be the bubbles get bigger. This is the wave of our lives. I feel that way, by the way. If I look at internet, SaaS, mobile, AI, nothing looks the same. And will there be roadkill from this wave? Oh my God, there's going to be a lot. You know, again, you look at those stats of 500 companies, less than 100 over 10 billion in the last 25 years. … It doesn't mean that there aren't survivors and companies that are going to change the trajectory of technology forever. And I think in open AI and Anthropic and SpaceX, we're seeing that already. Like these are the Metas and the Googles of our era, highly likely. But wow, like it's Hollywood, man. Like 95% are not going to be there."
「按历史的、或者说老派的看法：泡沫会一个比一个大。这是我们此生最大的一波。顺便说，我自己就是这么觉得的。看互联网、SaaS、移动、AI，没有哪次长得一样。这一波会有'路杀'吗？我的天，会有一大堆。再看看那些数据：过去 25 年 500 家公司里，超过 100 亿美元的不到 100 家。……这不意味着没有幸存者、没有将永远改变技术轨迹的公司。我认为在 OpenAI、Anthropic 和 SpaceX 身上我们已经看到了——它们大概率就是我们这个时代的 Meta 和 Google。但是，哇，这就是好莱坞，兄弟。95% 的公司到不了那里。」
David Frankel (Founder Collective) · 20VC: The AI Boom Will Create Enormous Roadkill

对"何时破"他给出的是 Kopelman 式的双重诚实：

"What's going to happen is there's always boom and bust. So a lot's going to come out of the system at some point. Are we headed for another dot-com crash? Definitely. If is not a question. When, nobody knows."
「接下来会发生的是：繁荣与破灭永远都在。所以到某个时点，会有大量东西从系统里退出来。我们是不是正走向又一次 dot-com 式崩盘？必然。'是否'不是问题。'何时'，没人知道。」
David Frankel (Founder Collective) · 20VC: The AI Boom Will Create Enormous Roadkill

而他的对策恰好补上了 Kopelman 之结缺的那半边——不用猜顶。他说他从未见过这么流动的 secondary 市场，头部 100 个名字都能有效定价，于是："And the ability to give DPI, even in your top names, sometimes taking 20% off the table, if you can return 25% of the fund, particularly if it's a newish fund. So if it's a 2024 fund, and you can give back 25%, why wouldn't you do that? And you're still long, you still own 80% of that company."（在头部持仓上卖掉 20%、给 LP 返还相当于 25% 基金规模的 DPI——你依然做多、依然持有 80%；最好的结局是你卖了之后"错了"、公司继续涨。）"泡沫期利润最高"与"必有崩盘"从此不再互斥：留在场上，但边打边收。 这与 Sacerdote 的"晚点没关系"、Accel 的"后期照样跑赢"构成同一个问题的第三种答案——问题从"何时进场"换成了"如何一直在场"。

都没说透的

我的看法

判断（不是事实）：退库清理后，上一版"时钟"框架的量化支撑（压缩版电信周期、capex-收入三桶账、hyperscaler RPO 的单一支点）已随访谈退库整体移除，我把框架退回"缺口"一代，只保留在库访谈仍能支撑的部分。核心判断不变："泡沫 vs S 曲线"的答案不取决于需求是否真实，而取决于收入的到达速度能不能一直跑赢 capex 承诺的兑付节奏。需求那一半在 2026 年被反复印证——Altman 的 uncapped 论（含他亲手给出的两个撤退条款）、Anthropic 平台层"agentic 负载极度吃 token"、Accel 的"被算力卡死"——这一半我基本接受。速度那一半，语料里现在只剩两组硬数可盯：(1) 两家基础模型公司 ARR 增速的滑落斜率（Sacks 口径 Anthropic 年初 $10B、现已 $80B+——若真，阵营 E 的"脱节"在模型层暂不成立）；(2) hyperscaler 是否需要外部股权融资来续 capex（Google $80B"超过全部现金流"的股权融资——它同时是阵营 A/F"强资产负债表理性投资"的样本与 Cahn"股权式瓦解"的候选引信，两个剧本并行计分）。上一版对"循环融资 + 叙事定价"中间层最先爆的判断保留不动，Randle 的"旧尺子断了"反而加重了它：当估值失去公认量尺，叙事定价的占比只会更高。

把握程度：仍是中等偏低。最强支撑：Jevons/供给受限共识有平台（Anthropic）、云（Accel）、实验室 principal（Altman）三层独立印证；最弱处有二：一是"收入到达速度"的中间读数（token 增速的月度数据、大额承诺的兑现口径）在语料里完全缺失；二是 Altman 亲口给出的两个过剩条件（注意力上限/scaling wall）一旦触发，上面三层印证会同时失效——这是一个相关性极高的尾部，不能因为它由多信源印证就误以为它分散了风险。

还想知道什么

取材

核心 16 篇（2026-08-10 新增 5 篇，均按逐字稿 ## Transcript 层核对；原 11 篇沿用前两轮逐字复核；2026-08-23 退库清理移除 6 篇）：

其余 alias 命中、关联较弱或未引用：Alfred Lin、Sam Altman×Collison、Anish Acharya、Hidden Economics、两期 Coatue $70B、Epoch AI、Atlassian、Redpoint 等。后续 headless 全量重综会自动纳入全文。

继续读
线索
阅读线索：agent 不到 9 nines 可靠，那应该卖给企业还是消费者
线索
阅读线索：如果 coding 是 commodity，AI startup 的护城河长在哪儿
线索
阅读线索：推理成本一年跌 10×，AI app 该按什么定价
线索
摘报
线索
摘报
线索
摘报