谈中美人工智能，多数人从模型讲起。做跨境对接的人更愿意从钱讲起，因为竞争与相互依存这两个故事，在钱的流向上同时成立，而数字的形状已经悄悄变了。Most accounts of U.S.–China artificial intelligence begin with the models. A cross-border adviser is more likely to begin with the money, because the money is where the two stories, rivalry and interdependence, are both true at once, and where the numbers have quietly changed their shape.

先看竞争本身。按斯坦福的综合评测，2026 年 3 月，美国与中国最好的模型之间，性能差约 2.7%：是差距，不是鸿沟。这点差距背后的花钱，不是一个量级，今年的预算里看得最清楚。美国最大的四家平台公司，今年的资本开支指引合计约 7,300 亿美元，绝大部分投向 AI 基础设施；SemiAnalysis 估算，中国最大的几家科技公司同年约 660 亿美元。SemiAnalysis 还估计，到今年底，美国交付的 AI 数据中心容量约 56 吉瓦，中国 24 吉瓦以上。北京用另一套单位记同一本账：信通院的数字，截至今年 6 月，中国在用机架 1,556 万架，智能算力 2,185 EFlops。哪种算法，不对称都是实的。中国这一侧卡住的不是电，也不是建设，两样都不缺，而是先进加速芯片。今年 9 月的全国算力大会上，官方的题目已经从「建得多」换成「用得好」。据美格视点，建起来的那部分里，有一部分是冲着政策和估值建的，不是冲着需求。用是真用，下面的数字会说话；建的这一层，不全是真的。Start with the race itself. On Stanford's Arena-style evaluations, the performance gap between the best American and the best Chinese model stood at roughly 2.7 percent in March 2026: a margin, not a moat. The spending gap behind that margin is of a different order, and it is clearest in this year's budgets. The four largest American platforms have guided to roughly $730 billion of capital expenditure for this year, most of it directed at AI infrastructure. SemiAnalysis estimates that China's largest technology firms will spend about $66 billion in the same year. By the end of this year, SemiAnalysis further estimates, the United States will have roughly 56 gigawatts of delivered AI data-center capacity against China's 24 or more. Beijing keeps the same ledger in a different unit: the CAICT counted 15.56 million racks in service and 2,185 exaflops of intelligent computing capacity in June this year. The asymmetry is real under either accounting, and it is measured in concrete, power contracts, and above all in chips: China's buildout is constrained not by electricity or construction, both of which are abundant, but by access to advanced accelerators. At China's national computing conference this September, the official theme had shifted from building more to using what has been built. In Megs' observation, part of what was built was raised for policy incentives and valuations rather than for demand. The use is real, as the figures below show. Not all of the building is.

两套体系真正互补的地方，反而较少被平实地讲出来。美国这边，私人资本更深，加速芯片的设计领先，模型前沿略占先手。中国这边，工程师队伍更大，论文与专利在数量上领先（论单篇的分量，美国的研究仍更重），部署的规模没有对手。中国的开源权重模型，如今占全球最大公开路由平台 OpenRouter 调用量的约 65%；但按里昂证券的估算，只占金额的约 30%，因为价格比美国模型低了约八成。国家数据局自己的统计：今年 3 月，中国日均词元调用量超过 140 万亿，两年涨了一千多倍。用量领先，没有变成收入领先。中国最像样的一笔 AI 收入，是字节跳动火山引擎的模型服务：今年上半年 48 亿元，毛利率在 55% 以上；同一时期，Anthropic 一家是 163 亿美元。Where the two systems are genuinely complementary is less often stated plainly. The United States holds the deeper pool of private capital, the leading accelerator designs, and a small lead at the model frontier. China holds the larger engineering workforce, the lead in publications and patents (a lead in volume; paper for paper, American research still carries the greater weight), and a record of deployment without parallel. Chinese open-weight models now account for roughly 65 percent of token volume on OpenRouter, the largest public routing platform, and, on CLSA's estimates, only about 30 percent of the spending, because they are priced some 80 percent below American equivalents. The National Data Administration's own count puts China's daily token calls above 140 trillion in March this year, a rise of more than a thousandfold in two years. Usage leadership has not become revenue leadership. ByteDance's Volcano Engine, the standout Chinese AI business, took in ¥4.8 billion in model-as-a-service revenue in the first half of this year, at a gross margin above 55 percent; Anthropic alone reported $16.3 billion in the same period.

人才是同样的双面格局。卡内基今年更新的 MacroPolo 追踪显示，全球顶尖 AI 研究者中，57% 本科读在中国；中国也已是这批人最大的工作地，占 41%，美国 34%。但流动并没有掉头：测量的那一年，美国净增顶尖研究者 2,145 人，中国净减 1,729 人；从美国流向中国的，每一人对应约三十人反向而行。中国越来越能自己培养、自己留住人，美国仍在收走盈余的那一部分。Talent follows the same double pattern. Carnegie's latest update of the MacroPolo tracker finds that 57 percent of the world's top AI researchers did their undergraduate degrees in China, and that China is now the largest workplace for that cohort, at 41 percent, against 34 percent in the United States. Yet the flow has not reversed: the United States still gained a net 2,145 top researchers in the year measured, while China lost 1,729, and for every researcher who moves from America to China, roughly thirty move the other way. China increasingly produces and keeps its own; America still collects the surplus.

还有一层不对称，只有站在中国里面才看得见。中国 AI 的核心层，外资实际上进不去，这个行业只能靠国内的钱，而国内的钱正没处去。楼市低迷，股市低迷，生育在降，年轻人不敢创业，也不敢消费。有实力的家庭和机构，钱一味涌向少数还被相信的资产，如今就是那几家头部 AI 企业。据美格视点，结果是两头沉：最好的项目，钱进不去，轮次没开就被占满；一般的项目，一分钱也融不到。资本同时过剩、又稀缺，这两个极端，从外面读官方的总量数字，一个都看不见。One further asymmetry is visible only from inside China. Foreign capital is, in practice, shut out of the country's AI core, so the sector runs on domestic money at a moment when domestic money has few places to go. Property is weak and equities are weak; births are falling, and young people hesitate to start businesses or to spend. Savings with means behind them crowd into the few assets still believed in, and those now are the handful of leading AI companies. The result, in Megs' observation, is a barbell. Into the best projects, money cannot get in; the rounds are spoken for before they open. Ordinary projects cannot raise at all. Capital is at once abundant and unavailable, and neither extreme shows in the official aggregates read from outside.

在这个背景下，资本的流向值得细看，因为它没有停，只是换了管道。直接投资很细：荣鼎的跨境监测记下，今年第一季度，中国对北美宣布的投资只有 24 亿美元，是 2023 年以来最强的一个季度，其中大头还是一家食品加工厂。流量转到了新规则留着口的那些通道：到今年 9 月，华尔街银行承销了 172 亿美元的中国科技公司发行；中国大陆与香港的投资人，参与了约 89 亿美元的美国 AI 融资轮次，而 2023 年全年只有 4.36 亿；经内地与香港持有的美股，超过 7,500 亿美元。2025 年 1 月的对外投资规则与同年 12 月的 COINS 法案，划的都是敏感技术的非公开投资；公开证券被留在了外面，钱如今主要就从这个口子走。Against this background, the capital flows repay a closer look, because they have not stopped. They have changed channels. Direct investment runs thin: Rhodium Group's cross-border monitor recorded $2.4 billion of announced Chinese investment in North America in the first quarter of this year, the strongest quarter since 2023, and much of it a single food-processing plant. The traffic has moved to the channels the new rules leave open. Wall Street banks underwrote $17.2 billion of Chinese technology issuance in the year to this September; investors in mainland China and Hong Kong participated in some $8.9 billion of American AI funding rounds over the same months, up from $436 million in all of 2023; and more than $750 billion of U.S. equities sit in mainland and Hong Kong hands. The January 2025 outbound-investment rules, and the COINS Act that followed in December, were drawn around private placements in sensitive technologies. Public securities were carved out, and it is through that carve-out that the money now principally moves.

所以账本上呈现的，不是脱钩，是改道：成本更高，看得更严，每过一道口，合规都是要交的过路费。管着这些通道的规矩，是本系列下一篇的题目。What the ledger shows, then, is not decoupling but rerouting, at higher cost, under closer supervision, and with compliance as the toll paid at every crossing. The rules that govern those crossings are the subject of the next piece in this series.