埃隆·马斯克: “如果到明年年底我们上线的算力少于 13 吉瓦，我会感到意外。……我认为我们将面临超级智能算力的短缺，而且这种短缺会持续很长时间。”

由 AI 翻译，经编辑部审校

原文

“So I've been spending personally a lot of my time in Memphis, often four days a week, night and day, just getting the Memphis Supercompute cluster Colossus II online. We're tracking to be able to bring roughly 200,000 GB300s online per month for every month for the remainder of this year and continuing that pace next year. With next year being primarily Rubin, Vera Rubin compute. At this point, I would say -- I would say I'm fairly confident of being able to bring an incremental 10 gigawatts of compute online in 2027. That would be in addition to the roughly 3 gigawatts, maybe 2.5 to 3 gigawatts that we'll have online by the end of this year. So I put it this way. I would be surprised if we brought on less than 13 gigawatts of compute by the end of next year. [...] Our goal is to kind of split it 50-50, more or less, where half of the compute is used internally and half of it is rented to others. But that number could shift 60-40 one way or the other. It could -- it's going to be dynamic. And obviously, our approach is to rent at the -- effectively the spot market rate as opposed to sign these long-term deals. Our belief is that the value of compute per watt will increase over time, not decrease. And the reason for that, I think it should be obvious to people really is that as you improve the AI software, as the models get better and better, they're able to -- you get more useful intelligence per watt. And the value of that compute is proportionate to the useful intelligence per watt. And that's why you see things like even old computers like H100s renting for more than they rented when they were new. And I think that's likely to continue. I think we're going to face a shortage of super intelligence compute. And that shortage is going to remain for a long time. So I think the smart move is to do short-term contracts and to and essentially to make -- to place the bet on useful intelligence per watt increasing rapidly, which obviously has done. [...] So I don't think we're going to have a demand issue here. And like it's not impossible. I'm not saying this will happen. It's not impossible that we will have a six-month payback period for compute next year.”

— 埃隆·马斯克, SpaceX, 20.10.2026 ↗

译文

“所以我个人花了大量时间待在孟菲斯，常常一周四天、日夜不停，就为了让孟菲斯超级计算集群 Colossus II 上线。按目前的进度，今年余下的每个月我们都能每月上线大约 20 万块 GB300，明年也将保持这一速度。明年将主要是 Rubin，也就是 Vera Rubin 算力。目前我可以说——我可以说，我相当有信心在 2027 年新增上线 10 吉瓦算力。这还要加上到今年年底我们将上线的大约 3 吉瓦，也许是 2.5 到 3 吉瓦。 所以我这么说吧。如果到明年年底我们上线的算力少于 13 吉瓦，我会感到意外。[...] 我们的目标大致是五五开，一半算力内部使用，一半租给别人。但这个比例可能会向某一方偏到六四开。它可能——它会是动态的。显然，我们的做法是按——实际上是按现货市场价格出租，而不是签长期合同。我们相信，每瓦算力的价值会随时间增加，而不是减少。 原因在于，我认为人们其实应该很清楚：随着 AI 软件的改进，随着模型越来越好，它们能够——你每瓦能得到更多有用的智能。而算力的价值与每瓦有用智能成正比。这就是为什么你会看到，即便是像 H100 这样的旧计算机，现在的租金也比它们刚推出时更高。我认为这种情况很可能会持续下去。我认为我们将面临超级智能算力的短缺，而且这种短缺会持续很长时间。 所以我认为明智的做法是签短期合同，并且基本上是——把赌注押在每瓦有用智能的快速增长上，而这显然已经发生了。[...] 所以我认为我们在需求方面不会有问题。而且，这并非不可能。我不是说这一定会发生。明年我们的算力投资回收期只有六个月，并非不可能。”