人工智能是泡沫吗？为什么我们看好其安装阶段
如果你现在与投资者交谈，最终会落到同一个问题上：“这是一个泡沫吗？如果是的话，为什么我们还在投入资本？”
这是一个正确的问题。对我来说，这也是一个个人的问题。
我曾经历过这样的情形。我在互联网泡沫时期创办了一家移动钱包初创公司，乘着狂热的浪潮。几年后，我在全球金融危机期间管理了一只困境债务对冲基金，当音乐停止时，我在废墟中寻找机会。如今，作为一名YC创始人（Vantedge AI）和八资本的普通合伙人（230多项YC投资），我从内部观察当前的周期。
今天我们看到相同的模式：巨额的资本支出和在“这改变了一切”和“这将以悲剧收场”之间摇摆的叙述。
在最近的一份备忘录中，霍华德·马克斯提出了一个区分，这为我的论点提供了框架：
均值回归泡沫：金融时尚（如次贷）在没有根本改善世界的情况下抬高资产价格。这些泡沫最终导致价值毁灭。
拐点泡沫：变革性技术（如铁路、互联网）被过度建设的时期。这些泡沫在短期内摧毁投资者资本，但永久性地提高了生产能力。
我的看法是：人工智能几乎肯定是一个拐点泡沫。它将改变世界，但在此过程中会烧毁资本。
以下是我们如何应对这一情况的。
1. 两个泡沫，而不是一个
马克斯区分了“公司行为”泡沫（超大规模企业、GPU建设、债务）和“投资者行为”泡沫（定价、彩票思维）。
我们不资助万亿美元的资本支出。我们的策略是：
投资于应用层：我们支持那些消耗AI基础设施而非建设它的早期软件公司。
小额多样化投资：我们在产品展示日之前进入。我们避免晚期融资轮的估值扭曲。
如果AI基础设施的建设被证明是过度的，痛苦将落在那些融资5万亿美元数据中心的人身上，而不是那些利用便宜计算资源向银行销售工作流自动化的YC初创公司。我们并不想成为下一个英伟达；我们支持那些在其上构建的创始人。
2. 安装与部署
技术革命始于安装阶段——一种过度投资的狂热，为未来铺设轨道。这个阶段是混乱的，容易崩溃。随后是部署阶段：新技术嵌入经济的盈利期。
我们偏向于部署。我们的投资组合公司不需要世界对AI基础设施的估值完美无缺。它们只需要有真实问题并愿意支付的客户。随着“安装泡沫”过度建设产能，我们的公司将受益于更好的单位经济。
3. 避免赌场
经历过全球金融危机后，我对以泪水收场的行为有着深深的厌恶：
彩票思维：我们避免对几乎没有可能性的巨大结果下注。我不需要一家公司回报整个基金；我们的目标是高命中率的稳健企业。
产品前的巨额融资轮：我们不参与没有已交付产品的公司进行的高价“种子”轮融资。
4. YC作为叙述过滤器
在泡沫中，资本涌入那些有好故事但团队薄弱的公司。虽然没有过滤器是完美的，但YC接近这个标准：
选择：从数千个项目中筛选出约1.5%的顶尖项目。
纪律：3个月的批次迫使创始人交付产品，而不仅仅是幻灯片。
数据：在11个以上的批次中投资使我们能够区分真实的市场吸引力与噪音。
结论
我同时持有两个相互竞争的想法：
“人工智能的热情几乎肯定会过度。”
“人工智能是我们这一生中最重要的技术变革之一。”
我们的工作不是预测泡沫何时破裂，而是避免成为资助过剩的边际资本，而是确保我们拥有一篮子将定义下一个十年的公司。
我们并不押注于泡沫。我们押注于建设者。
查看原文
If you talk to allocators right now, the conversation eventually lands on the same question: “Is this a bubble? And if it is, why are we deploying capital?”
It is the right question. And for me, it’s personal.
I’ve lived through this movie before. I founded a mobile wallet startup during the dot-com mania, riding the wave of euphoria. Years later, I ran a distressed debt hedge fund through the Global Financial Crisis, sifting through the wreckage when the music stopped. Today, as a YC founder myself (Vantedge AI) and GP at Eight Capital (230+ YC investments), I see the current cycle from the inside.
We are seeing the same patterns today: enormous capital spend and narratives swinging between “this changes everything” and “this ends badly.”
In a recent memo, Howard Marks introduces a distinction that frames my thesis:
Mean-Reversion Bubbles: Financial fads (like subprime) that inflate asset prices without fundamentally improving the world. These end in value destruction.
Inflection Bubbles: Periods where a transformative technology (railroads, internet) gets massively overbuilt. These bubbles destroy investor capital in the short term, but permanently raise productive capacity.
My take: AI is almost certainly an Inflection Bubble. It will change the world, but it will incinerate capital along the way.
Here is how we are navigating it.
1. Two Bubbles, Not One
Marks distinguishes between a "company behavior" bubble (hyperscalers, GPU build-outs, debt) and an "investor behavior" bubble (pricing, lottery-ticket thinking).
We don't fund trillion-dollar CapEx. Our strategy is to:
Invest in the Application Layer: We back early-stage software companies that consume AI infrastructure not build it.
Small, Diversified Checks: We enter pre-Demo Day. We avoid the valuation distortion of late-stage rounds.
If the AI infra build-out turns out to be overbuilt, the pain sits with those financing $5T of data centers—not with a YC startup using that cheap compute to sell workflow automation to banks. We are not trying to be the next Nvidia; We are backing the founders building on top of it.
2. Installation vs. Deployment
Technological revolutions begin with an Installation Phase—a mania of over-investment that lays the rails. This phase is chaotic and prone to crashes. It is followed by the Deployment Phase: the profitable period where the new technology is embedded into the economy.
We are biased toward Deployment. Our portfolio companies don't need the world to be perfect for AI infra valuations. They just need customers with real problems and willingness to pay. As the "Installation Bubble" overbuilds capacity, our companies benefit from better unit economics.
3. Avoiding the Casino
Having navigated the GFC, I have a deep aversion to behaviors that end in tears:
Lottery-Ticket Thinking: We avoid betting on massive outcomes with near-zero probability. I don't need a single company to return the fund; We aim for a high hit rate of solid businesses.
Pre-Product Mega-Rounds: We do not participate in very high priced "seed" rounds for companies with no shipped product.
4. YC as a Narrative Filter
In a bubble, capital floods into weak teams with good stories. While no filter is perfect, YC comes close:
Selection: Screens thousands down to the top ~1.5%.
Discipline: The 3-month batch forces founders to ship product, not just slides.
Data: Having invested across 11+ batches has enabled us to spot real traction vs. noise.
Conclusion
I hold two competing thoughts at once:
"AI enthusiasm will almost certainly overshoot."
"AI is one of the most important technology shifts of our lifetimes."
Our job isn't to predict when the bubble pops. It is to avoid being the marginal dollar funding the excess, and instead ensure we own a basket of companies that will define the next decade.
We are not betting on the bubble. We are betting on the builders.