图1：数据中心与变电站夜景。马斯克称，大型AI集群上线必须同步建电厂、变压器和冷却系统，电力已成最大瓶颈。
Photo 1: A data-center and substation complex at dusk. Musk said large AI clusters cannot come online without power plants, transformers and cooling; electricity is now the binding constraint.
【天网美西时间2026年9月1日讯】（天网记者 王晶）
【Tianwang West Coast Time, September 1, 2026】(Tianwang reporter Wang Jing)
美西时间9月1日（周二）上午8时19分，SpaceX首席执行官埃隆·马斯克在X平台发文：要把大型人工智能数据中心真正上线，不只是买芯片，还必须同步建设大规模发电厂、变压器、液冷回路、冷水机组，以及训练集群所需的极复杂网络。
At 8:19 a.m. Pacific Time on Tuesday, September 1, SpaceX chief executive Elon Musk posted on X that bringing a large AI data center online is not just buying chips. It also requires power plants, transformers, liquid-cooling loops, chillers, and the extremely complex networks a training cluster needs.
图2：马斯克美西时间9月1日上午8时19分原文：大型AI数据中心上线需要电厂、变压器、液冷回路、冷水机组和训练集群所需的复杂网络。
Photo 2: Musk’s post at 8:19 a.m. Pacific Time on Sept. 1: large AI data centers need power plants, transformers, liquid-cooling loops, chillers and complex training-cluster networks.
马斯克此言，是在回应一条关于他最新表态的视频帖。视频中他说：还没到明年，电力问题就已经出现；正因为如此，谷歌、Anthropic以及许多其他公司正在向SpaceX租用算力。他强调，SpaceX迄今把AI算力送上线的速度比任何人都快——“自己建电厂，是我们能做到这一点的唯一办法。”
The post answered a video of his latest remarks. In the clip Musk said the power problem had already arrived before next year, which is why Google, Anthropic and others are renting compute from SpaceX. SpaceX, he said, has brought AI compute online faster than anyone — “building our own power plants is the only way we can do that.”
图3：X平台视频帖。马斯克说谷歌、Anthropic等正在向SpaceX租用AI算力，SpaceX上线速度最快是因为自己建电厂。
Photo 3: Video post on X. Musk said Google, Anthropic and others are leasing AI compute from SpaceX, and that SpaceX outpaced others by building its own plants.
这一表态把当前AI竞赛里最硬的约束说直白：限制模型训练与部署的，已经不只是GPU，更是电、冷却和电网接入。美国多地数据中心排队等电、变压器交货周期拉长，已是公开事实。谁能先把“电＋冷却＋网络”一次性配齐，谁就能先把算力租出去。
The point is the hard constraint in the AI race: training and deployment are no longer limited only by GPUs, but by power, cooling and grid hookups. Data centers in several U.S. states wait in line for electricity; transformer lead times have stretched. Whoever assembles power, cooling and networking first can rent the compute first.
公开披露显示，SpaceX旗下Colossus、Colossus II等设施已向第三方出租算力。Anthropic于2026年5月与SpaceX达成云服务协议，披露口径约为每月12.5亿美元，覆盖田纳西州孟菲斯一带Colossus集群的大量英伟达GPU；文本可写到2029年，但双方均可在约90天通知后终止。马斯克曾说，短期安排是SpaceX的要求，若自家算力吃紧，不排除收回。谷歌另有一份约每月9.2亿美元的租赁安排，完整费率原定自2026年10月起执行，同样设有退出条款。上述数字是合同上限，实际履约取决于交付与是否提前解约，不能当成已锁定的多年收入。
Public disclosures show SpaceX’s Colossus and Colossus II sites already rent capacity to third parties. In May 2026 Anthropic reached a cloud deal with SpaceX on the order of $1.25 billion a month for large Nvidia GPU capacity at the Memphis-area Colossus cluster. The paper can run to 2029, but either side may exit on about 90 days’ notice. Musk has said the short term was SpaceX’s demand, and that capacity could be pulled back if SpaceX itself runs short. Google has a separate arrangement of about $920 million a month, with the full rate slated from October 2026, also with an exit clause. These figures are contract ceilings, not guaranteed multi-year revenue.
对读者而言，意义不在“谁租了谁的机房”，而在于：人工智能已从软件竞赛，变成发电、变电站和冷却系统的竞赛。马斯克把SpaceX从火箭公司进一步做成“带电厂的算力房东”，说明美国头部科技公司宁可向对手租电租卡，也不愿干等公共电网扩容。谁控制可靠电力，谁就在下一轮模型训练里握有定价权。
The story is not which company rented whose hall. AI has moved from a software race to a race over generation plants, substations and cooling. By turning SpaceX into a compute landlord that builds its own power, Musk is showing that leading U.S. firms would rather rent electricity and chips from a rival than wait for the public grid. Whoever controls reliable power will set the price of the next training run.
六四天网将继续关注美国AI基础设施、电力扩容与算力租赁的后续进展。
Tianwang will continue to follow U.S. AI infrastructure, power expansion and compute leasing.
（完）
(End)
（六四天网 综合马斯克X平台发言及公开合同披露）
(Tianwang, compiled from Musk’s posts on X and public contract disclosures)
天网相关链接：https://64tianwang.net/archives/3944
