AI 是泡沫吗？我们为何押注于安装阶段
1 分•作者: rchachra•20 天前
如果你现在和资产配置者交流，谈话最终会落到同一个问题上：“这是否是一个泡沫？如果是，我们为什么要部署资本？”
这是一个正确的问题。对我来说，这很私人。
我以前经历过这种情景。在互联网泡沫时期，我创立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经营着一只不良债务对冲基金，在音乐停止后，我开始清理残骸。今天，作为一名 YC 创始人（Vantedge AI）和 Eight Capital 的普通合伙人（230 多个 YC 投资），我从内部观察当前的周期。
我们今天看到了同样的模式：巨大的资本支出和在“这改变一切”和“这会以悲剧收场”之间摇摆的叙事。
在最近的一份备忘录中，霍华德·马克斯引入了一个区分，为我的论点奠定了基础：
均值回归泡沫：金融时尚（如次贷）在没有从根本上改善世界的情况下推高资产价格。这些最终导致价值毁灭。
拐点泡沫：变革性技术（铁路、互联网）被大规模过度建设的时期。这些泡沫在短期内摧毁了投资者资本，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，但它会在这个过程中烧毁资本。
以下是我们如何应对。
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）。
我们不为数万亿美元的资本支出提供资金。我们的策略是：
投资于应用层：我们支持使用人工智能基础设施而不是构建人工智能基础设施的早期软件公司。
小额、多元化的投资：我们在演示日之前进入。我们避免了后期融资轮的估值扭曲。
如果人工智能基础设施建设最终被过度建设，痛苦将落在那些为 5 万亿美元的数据中心提供资金的人身上——而不是一家使用廉价计算能力向银行销售工作流程自动化的 YC 初创公司。我们不想成为下一个英伟达；我们支持那些在其基础上进行建设的创始人。
2. 安装 vs. 部署
技术革命始于安装阶段——过度投资的狂热，为技术奠定了基础。这个阶段是混乱的，容易崩溃。随后是部署阶段：将新技术嵌入经济中的盈利时期。
我们倾向于部署。我们的投资组合公司不需要世界对人工智能基础设施估值完美。他们只需要有真正问题的客户和支付意愿。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效益。
3. 远离赌场
在经历过全球金融危机后，我非常厌恶以悲剧收场的行为：
彩票思维：我们避免押注于几乎没有可能性的巨大结果。我不需要一家公司就能为基金带来回报；我们的目标是实现高命中率的稳健业务。
产品发布前的巨额融资：我们不参与对没有已发布产品的公司进行高价“种子”轮融资。
4. YC 作为叙事过滤器
在泡沫中，资本涌入拥有好故事的弱势团队。虽然没有完美的过滤器，但 YC 几乎做到了：
筛选：将数千个项目筛选到前 1.5% 左右。
纪律：为期 3 个月的批次迫使创始人发布产品，而不仅仅是幻灯片。
数据：在 11 个以上的批次中进行投资，使我们能够发现真正的吸引力与噪音。
结论
我同时持有两种相互矛盾的想法：
“人工智能的热情几乎肯定会过头。”
“人工智能是我们一生中最重要的一次技术变革。”
我们的工作不是预测泡沫何时破裂。而是避免成为为过剩提供资金的边际资金，而是确保我们拥有一篮子将定义未来十年的公司。
我们没有押注泡沫。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.