Gavin Baker: The AI Bubble Analogy Doesn't Hold; No Idle Top-Tier GPUs

This is probably the most counterintuitive yet hardcore wake-up call about the AI bubble across the internet this year. Top Silicon Valley investor Gavin Baker just dropped a historical truth in an interview that many people simply don't know. The whole internet is using the 2000 internet bubble to predict today's AI collapse, but he, who personally invested in that era, offered one fatal contrast: at the peak of the 2000 bubble, a full 97% of the dark fiber laid underground across the US went unused; whereas today, in data centers around the world, you simply cannot find a single idle top-tier GPU—every chip is running at high load, even running so hot it melts. The first thing to be shattered is those lazy, by-rote analogies. The telecom crash back then was essentially humanity building pipelines tens of thousands of times greater than what software at the time could consume; capital buried over ten million kilometers of fiber underground, and over 90% of it rotted in the dirt without even being powered on. Today, anyone who has read a large-model paper knows that compute centers are not idling at all; the biggest headache for the world's smartest engineers every day is how to solve cooling and power supply before tens of millions of chips are completely drained by high-intensity training. This is the most fundamental physical divide between the two technology booms. The hallmark of a hollow bubble is assets gathering dust in warehouses and underground, with demand being nothing but a fictional story on a PowerPoint; the hallmark of a real industrial revolution is demand surging like a tsunami and overwhelming existing capacity, with every piece of hardware built depreciating around the clock without rest. Of course we need to maintain reverence for overheating capital markets, but never turn reverence for the future into blindness to the real technological load right in front of us. In the past, the pipes were built but the water never came; from today on, the surging flood is bursting every pipe. As long as the core underlying assets are still running at 100% load, the most dangerous move in this era is never to fail to keep up with it, but to declare its exit too early. In your own business, do you feel a real hunger for compute and intelligence, or do you think many demands are purely hyped up by concepts? For example, just internally, our daily call volume for running automation scripts and knowledge bases is doubling at a visible rate every month. https://t.co/LFi2arFIdw

View original

这大概是今年全网关于 AI 泡沫最反直觉、但也最硬核的一篇当头棒喝。 硅谷顶级投资人 Gavin Baker 刚在访谈里甩出了一个很多人根本不知道的历史真相。 全网都在拿 2000 年互联网泡沫来预言今天的 AI 崩盘，但当年亲手在那个时代做投资的他只说了一句致命的反差：
2000 年泡沫顶峰时，全美铺在地下没人用的暗光纤高达整整 97%；而今天全世界的机房里，你根本找不到一块闲置的顶级 GPU，所有芯片都在高负荷运转甚至跑到发烫熔化。 最先被击穿的，是那些偷懒的刻舟求剑式类比。
当年的电信崩盘，本质上是人类造出了几万倍于当时软件所能消耗的过剩管道，资本在地下埋了上千万公里的光纤，结果九成以上连电都没通就直接烂在了泥土里；而今天任何一个读过大模型论文的人都知道，现在的算力中心根本不是在空转，全球最聪明的工程师每天最头疼的瓶颈，是如何在千万张芯片被高强度训练彻底榨干前解决散热和供电。 这就是两场技术繁荣之间最根本的物理分野。
空心泡沫的特征，是资产在仓库和地下疯狂吃灰，需求只是PPT上的虚构故事；而真实工业革命的特征，是全世界的需求像海啸一样冲垮现有的产能，每一块造出来的硬件都在全天候不眠不休地折旧。 我们当然需要对资本市场的过热保持敬畏，但千万不要把对未来的敬畏，变成对眼前真实技术负载的视而不见。 以前是管道造好了却等不来水，从今天起是奔腾的洪流正在把所有的水管撑到爆裂。
只要底层的核心资产还在以百分之百的负荷运转，这个时代最危险的动作，就永远不是跟上它，而是过早地宣布它退场。 大家在自己的业务里，感受到的是对算力和智能的真实饥渴，还是觉得很多需求纯粹是被概念炒出来的？ 比如说我光是我们内部每天跑自动化脚本和知识库消耗的调用量，每个月都在以肉眼可见的速度往上翻倍。 https://t.co/LFi2arFIdw