跳到正文
Rohan Paul· @rohanpaul_ai · X·· 2 小时前AI 评分60
AI 导读

彭博发布了一篇关于在得州建设 AI 数据中心估算开支的文章。文中提到服务器占账单至少一半,两年内平均价格上涨逾一倍,DRAM 成本是去年的 5 倍以上;电网很少能在一处提供多余的一吉瓦电力,开发商要么以每兆瓦 58.4 万美元购买并网土地,要么自建燃气发电,后者成本近年已涨至三倍。

正文

Bloomberg published a piece on the estimated expenses to construct an AI data center in Texas

Servers make up at least half the bill, and demand has more than doubled their average price in 2 years while DRAM costs over 5 times last year’s price.

Grids rarely have a spare gigawatt in one place, so developers either pay $584K per megawatt for grid-connected land or build their own gas generation, whose cost has tripled in recent years.

About $750B of hyperscaler capex this year chases the same turbines, steel, copper and fiber, so power-gear backlogs run past 2030 and a 1GW site may need nearly $3B for overruns.

GPU racks run hot enough to need liquid cooling, which makes merely installing the servers and cooling units in a 1GW site cost about $1.2B.

来源:Rohan Paul · x.com