GPU 计算需求持续高涨，Neo Cloud 远非商品Insatiable GPU Demand: Neo Clouds Are Not Commodities
Lambda 联合创始人兼 CTO Stephen Balaban 在 MAD Podcast 中表示，AI 计算需求远未见顶。公司继续低估扩展规律带来的市场扩张。GPU 的经济使用寿命远超会计折旧周期，H100 等芯片租赁价格甚至高于最初水平。Balaban 强调，云计算尤其是 AI 云是高度垂直整合的复杂业务，涉及土地、建设和融资等多层创新，而非简单商品。
他指出，主要瓶颈在于带电力的土地和数据中心建设，而非 GPU 本身。数据中心采用闭环液体冷却，几乎不消耗水，还能强化电网。Balaban 乐观认为，多个大型玩家将共存，就像传统云市场一样。
“The people who are the naysayers, you're gonna throw these GPUs out in five years, are completely wrong.”
他指出，主要瓶颈在于带电力的土地和数据中心建设，而非 GPU 本身。数据中心采用闭环液体冷却，几乎不消耗水，还能强化电网。Balaban 乐观认为，多个大型玩家将共存，就像传统云市场一样。
“The people who are the naysayers, you're gonna throw these GPUs out in five years, are completely wrong.”
Lambda cofounder and CTO Stephen Balaban on the MAD Podcast argues that AI compute demand shows no signs of peaking and that people continue to underestimate scaling laws expanding the addressable market. GPU economic usable life far exceeds accounting depreciation, with H100 rental rates now higher than initial levels. He stresses that AI cloud is a highly vertically integrated complex business involving land, construction, and financing innovations, not a commodity.
Main bottlenecks are powered land and data center buildout. Modern deployments use closed-loop liquid cooling with near-zero water use and grid benefits. Balaban sees room for multiple large players, similar to traditional cloud.
“The people who are the naysayers, you're gonna throw these GPUs out in five years, are completely wrong.”
查看原文 →
Main bottlenecks are powered land and data center buildout. Modern deployments use closed-loop liquid cooling with near-zero water use and grid benefits. Balaban sees room for multiple large players, similar to traditional cloud.
“The people who are the naysayers, you're gonna throw these GPUs out in five years, are completely wrong.”