人工智能泡沫与人工智能无关。
关于“人工智能泡沫”破裂的无休止讨论，主要集中在ChatGPT的订阅或企业API调用是否能产生足够的收入，以证明数万亿的炒作是合理的。但这完全忽视了关键问题。
即使明天每个前沿实验室，包括OpenAI、Anthropic、xAI、Google DeepMind，突然都没有付费客户，推动所谓“人工智能革命”的金融机制仍会继续运转，基本保持不变。
原因如下：
核心机制并不是依靠当前的产品收入，而是对未来在计算能力、能源、数据和地缘政治定位的巨大平行投资。真正的资金流向如下：
超大规模云服务商（如微软、亚马逊、谷歌、Meta等）每年在数据中心、GPU和电力基础设施上投入数千亿美元，这并不是因为他们的人工智能功能今天能带来现金流（大多数尚未盈利），而是因为他们无法在基础设施军备竞赛中落后。如果某个参与者控制了全球最高质量计算能力的大部分，他们就会赢得分发、人才、数据优势，以及对未来突破的选择权。这是一种强化版的防御性资本支出：要么建设，要么永远被锁在外面。
Nvidia和芯片生态系统处于中间地带，无论最终用户是否为更好的聊天服务支付每月20美元，他们都在收取巨额利润。需求来自于基础设施的建设，超大规模云服务商需要硬件来以惊人的规模进行训练和推理。即使某些设施的当前利用率较低，订单仍在不断涌入，因为没有人想在2028年向董事会解释，为什么在其他人都获得足够的H200时，他们却没有。
地缘政治和国家安全进一步推动了这一切。各国政府（尤其是美国及其盟友）将前沿人工智能计算视为与核技术或能源电网同等重要的战略基础设施。主权人工智能倡议、大规模合作（如Stargate）和出口控制等都在同一循环中发挥作用：现在投资，否则就面临永久的二流地位。尽管受到限制，中国的加速发展只会加剧这一竞争。
估值机制和循环资本使得这一飞轮不断转动。实验室以高估值融资，承诺提供AGI/ASI的选择权（对改变世界技术的看涨期权），超大规模云服务商通过“人工智能顺风”来证明资本支出是合理的，Nvidia在资本支出浪潮中赚取利润，投资者则将收益回流到生态系统中。这一过程在很大程度上是自我强化的，而这一点不会因为“ChatGPT Plus订阅不够”而改变。只有当资本成本飙升、电网达到硬限制，或地缘政治冲击重置优先事项时，情况才会改变。
“人工智能泡沫”关乎谁最终拥有21世纪的数字石油平台、庞大的计算集群、能源合同和人才库。这些投资是基于生存选择价值和战略必要性，而不是季度SaaS年经常性收入。许多人工智能产品可能会永远保持高度补贴或低利润。许多初创公司将会倒闭。估值将会修正（剧烈）。但对计算基础设施的资本涌入仍将持续，直到物理、地缘政治或利率另有说法，而不是因为有人取消了他们的20美元订阅。
查看原文
The endless chatter about an "AI bubble" bursting focuses on whether ChatGPT subscriptions or enterprise API calls will ever generate enough revenue to justify the trillions in hype. But that's missing the point entirely.
Even if every frontier lab, OpenAI, Anthropic, xAI, Google DeepMind, suddenly had zero paying customers tomorrow, the financial machinery driving the so called "AI revolution" would keep spinning, largely unchanged.<p>Here's why:<p>The core scheme isn't powered by current product revenue but rather a massive, parallel bet on future dominance in compute, energy, data, and geopolitical positioning. The real money flow looks like this:<p>Hyperscalers (Microsoft, Amazon, Google, Meta, etc.) are pouring hundreds of billions annually into data centers, GPUs, and power infrastructure, not because their AI features are printing cash today (most aren't profitable yet), but because they cannot afford to fall behind in the infrastructure arms race. If one player controls the lion's share of the world's highest quality compute, they win distribution, talent, data advantages, and optionality on whatever breakthroughs come next. This is defensive capex on steroids: build it or be locked out forever.<p>Nvidia and the chip ecosystem sit in the middle, collecting obscene margins regardless of whether end users are paying $20/month for better chat. Demand comes from the buildout itself, the hyperscalers need the hardware to run training and inference at insane scale. Even with low current utilization in some facilities, the orders keep coming because no one wants to explain to their board in 2028 why they didn't secure enough H200s when everyone else did.<p>Geopolitics and national security turbocharge the whole thing. Governments (especially the US and allies) view frontier AI compute as strategic infrastructure on par with nuclear tech or energy grids. Sovereign AI initiatives, massive partnerships (like Stargate), and export, control theater all feed into the same loop: spend now or risk permanent second rate status. China accelerating despite restrictions only intensifies the race.<p>Valuation mechanics and circular capital keep the flywheel turning. Labs raise at high valuations on the promise of AGI/ASI optionality (a call option on world changing tech), hyperscalers justify capex by pointing to "AI tailwinds," Nvidia prints money on the capex wave, and investors recycle gains back into the ecosystem. It's largely self reinforcing, and that won't change because of "not enough ChatGPT Plus subs." It'll change when the cost of capital spikes, power grids hit hard limits, or a geopolitical shock resets priorities.<p>The "AI bubble" is about who ends up owning the digital oil rigs of the 21st century, vast compute clusters, energy contracts, and talent pools. Those bets are being placed on existential option value and strategic necessity, not quarterly SaaS ARR. Many AI products could remain heavily subsidized or low margin forever. Many startups will die. Valuations will correct (hard). But the underlying capital flood into compute infrastructure? That continues until physics, geopolitics, or interest rates say otherwise, not until someone cancels their $20 subscription.