· 供应链分析

软件利用率对AI算力利润的影响远超廉价电力成本。

涉及标的:

中文翻译

$NBIS 拥有廉价电力,但可能不如 $IREN 或 $CIFR 便宜。但为了让你明白为什么这没那么重要: 软件利用率/编排(Orchestration)是护城河,其重要性比廉价电力高出约 30-70 倍。(请有人核实一下这个数学计算) 一张 H100 功耗约 0.84 千瓦。在 0.05 美元/千瓦时的电价下,成本为 0.04 美元/GPU-小时;即使在 0.10 美元/千瓦时,成本也仅为 0.08 美元/GPU-小时。 鉴于租赁价格约为 4 美元/GPU-小时,电力成本仅占收入的 1-2%。但将利用率从 50% 提升到 85%,会使单 GPU 收入增加 70%,从而大幅改变利润率。 不过,如果你拥有像 $CRWV 那样的所有组件(除了他们的坏账问题),那就是另一回事了,但这巨大的投机性赌注在于所有矿工能否建成这些设施。

英文原文

$NBIS has access to cheap power, but it might not be as cheap as $IREN or $CIFR. But just to give you a sense of why it matters a lot less: Software utilization/orchestration is a moat and matters ~30–70 TIMES more than cheap power. (someone pls fact check the math) A H100 burns ~0.84 kW. At $0.05 / kWh, that’s $0.04 / GPU-hour; even at $0.10 / kWh it’s $0.08 / GPU-hour. With rental prices of ~$4 / GPU-hour, power is just 1–2 % of revenue. But going from 50% to 85% utilization changes revenue per GPU by +70%, swinging margins massively. Then again if you have all the components like $CRWV (aside from their bad debt), then that's a different story, but it's a huge speculative gamble that all the miners can build that out.

在 X 上查看原推 ↗