The Mirror Scarcity 镜像稀缺：电与卡

A standard way to write the "AI race" story is to line up the two countries on the same axis — total compute, total models, total megawatts — and declare a winner. Most of those framings quietly assume that what's scarce on one side is also what's scarce on the other. It isn't. The US and China have run into different walls.

The American wall is electrical. The four largest hyperscalers collectively plan to spend close to seven hundred billion dollars on infrastructure in 2026 — roughly three quarters of it on AI compute. The chips exist; the financing exists; the buildings can be poured. What does not exist, in the geographies the chips need to land in, is the connected electricity. PJM Interconnection, which serves Northern Virginia's data-center alley, has watched the time from generator-interconnection application to commercial operation rise from under two years in 2008 to more than eight years in 2025. Its most recent capacity auction cleared at the price cap and still came up six gigawatts short of its own reliability target — the first such failure in PJM's history. Morgan Stanley's standing forecast is that the US will face a forty-seven-gigawatt power deficit by 2028, even after counting every behind-the-meter gas turbine and reactivated nuclear unit on the books.

The Chinese wall is on the other side of the bill of materials. The grid is not the constraint: China sits on roughly thirty-four hundred gigawatts of installed generation — about two and a half times the US total — and added another three hundred and seventy gigawatts in 2024 alone, most of it solar and wind. The 2022 "East Data, West Compute" plan formalised what the topology already implied: route latency-tolerant model training to the western hubs (Inner Mongolia, Gansu, Ningxia, Guizhou) where cheap renewable power is structurally surplus, and keep only latency-sensitive inference in the eastern clusters. Where the US has to wait eight years to plug a transformer in, a Chinese hyperscaler can in principle stand up a hundred-megawatt training campus in Ulanqab inside a year. What it cannot do is fill that campus with accelerators. Logic-die capacity at SMIC is broadly sufficient to ramp Huawei's Ascend 910C into the millions per year. CXMT, the domestic high-bandwidth-memory supplier, cannot produce HBM at anywhere near that rate; its 2025 stack output covers only about two hundred and fifty thousand 910Cs. Without EUV lithography — still subject to export control — yield improvements come from multi-patterning the older tools, which is slower and dirtier. The accelerators that do ship are mostly absorbed by a state allocation system: the "national integrated computing power network" routes capacity preferentially to state-backed labs. Private model companies improvise. DeepSeek trained V3 on two thousand and forty-eight H800s and made the constraint into a research story. Zhipu has, on the record, asked the public for help.

Read horizontally, the asymmetry is almost ironic: the country that can print money and chips can't keep the lights on, while the country that can keep the lights on can't make the chips. Read vertically, on the time axis, the two systems look more similar than either side wants to admit. Both have spent the last three years making policy that flips on the order of months: US export controls on the H20 went on, then off, then on again, then off, in 2025 alone. PJM's reliability auction went from routine to broken inside one delivery year. Beijing's compute-allocation rules have been re-issued three times since the original "East Data, West Compute" plan. The competition is not really between the two grids or the two chip stacks. It is between two improvising bureaucracies, each watching the other for permission to slow down.

What this diagram refuses to do. It does not put a number on the chip-smuggling corridors. The arcs across the middle are real — Operation Gatekeeper, the Supermicro indictment, the Singapore arrests are public record — but the volume that actually arrives is, by definition, not a reportable quantity. Drawing a precise figure there would pretend to a precision the data does not support. Likewise the Chinese hub polygons are deliberately schematic: site-level megawatt and accelerator inventories are simply not in the public record at the resolution the US ones are.

讲「AI 竞赛」常见的写法是把两国放在同一根坐标轴上——总算力、 总模型数、总兆瓦——然后宣布谁赢。这种写法暗设了一个前提： 两边稀缺的是同一样东西。事实并非如此。美国与中国撞到的 是两堵不同的墙。

美国的墙是电。四大 hyperscaler 2026 年的基础设施开支合计 接近七千亿美元，其中约四分之三投向 AI 算力。芯片有、钱有、 楼也能盖起来。在芯片需要落地的那些地理里，唯独缺的是 接入电网的那一段。覆盖弗吉尼亚「数据中心走廊」的 PJM 电网， 从发电机申请并网到商业运行的等待期，2008 年还不到两年， 2025 年已经超过八年。最近一次容量拍卖以价格上限结清， 却仍比 PJM 自己的可靠性目标少了六吉瓦——这是 PJM 历史上 第一次没采购足。Morgan Stanley 长期预测：到 2028 年， 即便算上所有自建燃机和重启核电，美国还将面临约 47 吉瓦 的电力缺口。

中国的墙长在物料清单的另一端。电不缺：总装机约 3,400 GW， 约为美国的两倍半；仅 2024 年就新增约 370 GW，其中大头是 光伏和风电。2022 年的「东数西算」其实是把拓扑学已经决定的 事用文件确认下来——把延迟不敏感的模型训练甩到内蒙、甘肃、 宁夏、贵州这些可再生能源结构性富余的西部枢纽，让东部集群 只承接对延迟敏感的推理。美国侧要等八年才能接进一台变压器， 而中国侧的超大规模公司原则上可以在乌兰察布一年内立起一座 百兆瓦级的训练园区。它做不到的是把这座园区填满加速卡。 中芯国际的逻辑晶圆产能，原则上足以让华为昇腾 910C 一年 量产数百万片；而国内高带宽内存（HBM）供应方长鑫存储， 2025 年的 HBM 产量只够配套约二十五万片 910C。没有 EUV 光刻——仍在出口管制之内——良率提升只能靠老一代设备的 多重曝光，又慢又脏。能下线的那些加速卡，大部分被国家算力 分配机制接走：「全国一体化算力网」把容量优先分配给国家级 实验室。民营模型公司只能自己想办法。DeepSeek 用 2,048 张 H800 训出了 V3，把约束本身变成了一个研究故事；智谱则公开 向外界求救。

横着看，这种不对称几近反讽：能印钞能造芯片的那一方， 点不亮灯；能点亮灯的那一方，造不出芯片。竖着看——把 时间轴展开——两个系统比任何一方愿意承认的都更像。 两边过去三年都在以月度频率翻转政策：仅 2025 年，美国 对 H20 的出口管制就开-关-开-关了四次；PJM 的可靠性拍卖 在一个交付年内从常规变成失灵；北京的算力分配规则自 「东数西算」原稿以来已重发了三次。真正的竞赛并不在 两边的电网之间，也不在两边的芯片栈之间，而在两套 边干边改的官僚体系之间——每一方都在盯着另一方， 看谁先减速。

这张图刻意不做的事。 它没有给走私通道一个具体数字。中间那几道弧线是真实的—— Operation Gatekeeper 起诉、Supermicro 案、新加坡逮捕都 在公开记录里——但真正到岸的体量按定义就不是可报的量。 硬填一个数字反而是对数据的不诚实。同理，中国侧的枢纽 多边形刻意保留示意性：单站点的兆瓦数与加速卡库存， 在公开渠道里没有美国侧那种分辨率。