122 天COLOSSUS 上线91 天COLOSSUS II 上线~2 年行业典型周期$0太空能源边际成本
ENGLISH · ORIGINAL
We believe we have a pathway over time that will significantly reduce compute hardware costs through continued vertical integration and development of proprietary chips, building on our experience designing custom silicon for our Starlink satellites. We also expect that the marginal cost of energy for our AI compute satellites will be minimal because our satellites are powered by solar arrays in space. By driving the energy component to minimal levels and pursuing improvements in compute hardware cost, we believe we can achieve a meaningfully lower overall cost per token in the future. We Have a Dual Speed and Cost Advantage in Terrestrial AI Compute. We own and operate what we believe to be the largest AI training data center clusters on Earth. Our AI compute facilities, COLOSSUS and COLOSSUS II, collectively provide approximately 1.0 gigawatt of compute power, with additional power capacity available for data center operations. Our first-principles thinking enables us to build coherent compute at scale and at rapid speed with lower costs than most other companies in the industry. In order to bring compute clusters online as fast as possible, we employ a vertically integrated, nimble approach to construction. We brought the first cluster of COLOSSUS online in 122 days, repurposing the shell of an existing factory, and the first cluster of COLOSSUS II online even faster in 91 days.
中文翻译
三大降本路径 + 地面 AI 算力的"双优势"(速度 + 成本)
我们认为自己有"长期大幅降低算力硬件成本"的路径:继续"垂直整合"、自己自研芯片——这套打法我们已经在 Starlink 卫星上设计定制芯片的过程中练过手了。我们还预期:AI 算力卫星的"边际能源成本"会极低——因为在太空中,卫星靠太阳能板发电。把"能源"这一项压到接近零、同时继续降"算力硬件"成本,我们相信未来能把单 token 总成本显著降下来。
地面 AI 算力:我们有"速度 + 成本"的双重优势。 我们自己拥有并运营地球上最大的 AI 训练数据中心集群。我们的 AI 算力设施——COLOSSUS 和 COLOSSUS II——合计提供大约 1.0 吉瓦(GW)的算力电力,还有额外的电力容量留给数据中心其他运营用。我们的"第一性原理思维"让我们能用更低的成本又快又大规模地搭建"整体协调(coherent)"的算力。为了让算力集群"尽快上线",我们采用垂直整合 + 灵活施工的方法:
COLOSSUS 第一个集群上线用了 122 天——直接改造了一座现成工厂的壳;
COLOSSUS II 第一个集群上线更快,只用了 91 天。
☀️ 2 年 vs 91 天
想参与互动?打开法量小程序体验完整功能
垂直整合+自研芯片+太空太阳能,这套组合拳确实把算力成本结构打得很透。122天和91天上线的速度也说明SpaceX的工程执行力远超行业平均水平。不过想请教一下,太空AI算力的延迟问题怎么解决?毕竟卫星和地面站之间的信号传输还是有物理限制的,如果对实时性要求高的推理任务,会不会还是得靠地面集群?
垂直整合和自研芯片确实是降本的核心,SpaceX把Starlink的经验迁移到AI算力上很聪明。不过想问一下,太空太阳能供电的边际成本虽低,但卫星本身的发射和维护成本如何平衡?另外COLOSSUS系列91天上线的速度确实惊人,这种快速部署能力在行业里算是独一份了。