METR/Epoch AI 分析：华为和林格尔算力约为 SpaceX Colossus 2 的 7%

若新建楼宇按计划推进，和林格尔到 2028 年初可达约 141,000 个 H100 等效算力；同期 Colossus 2 预计达到 180 万个以上。分析还指出，美国数据中心每瓦特可获得的 AI 算力是中国的 2.5 倍，芯片是中国的主要瓶颈。

文章同时提示，中国拥有电力资源优势，而美国即将面临的是能源瓶颈。在此算力劣势下，中国实验室过去两年仍持续发布新模型，这一差距的成因值得开发者与企业关注。

METR reports are always worth reading. Just finished reading their latest Substack. If you want to understand how much ahead the US is regarding data center, read those numbers:

Huawei's flagship AI data center has roughly 7% of the compute of SpaceX's Colossus 2, according to a new Epoch AI analysis of satellite imagery.

That's around 105,000 Ascend 910C chips, or roughly 81,000 H100-equivalents. Colossus 2 has about 14x as much.

A new building could push Horinger to about 141,000 H100e by early 2028. Colossus 2 should be at 1.8M or more by then.

However (something I point to repeatedly) China has the power. The chips are the bottleneck: US data centers get 2.5x more AI compute per watt. Makes what Chinese labs shipped in the last two years even more impressive.

The biggest bottleneck for the US will be energy in the very near future, for China the bottleneck stays Chips.

本条情报归属 AITOP「行业」核心栏目，重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中，该动态直接关联「中美大模型技术代差与商业化对照」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：Huawei、Ascend 910C、SpaceX、Colossus 2，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明：本文原始线索采编自「kimmonismus」（@kimmonismus），经 AITOP 多维评分管线结构化提炼。截至目前，全网已有 7 家独立信源对该事件进行交叉印证与跟进报道（含 The Information、theverge、rohanpaul_ai 等），属于近期高关注度行业信号。