【阑夕】美国出口管制意外迫使中国人工智慧实验室转向深度工程优化与开源生态。在不到 18 个月内，中国开源模型在 OpenRouter 上的 Token 消耗市占率从不足 2% 飙升至 45%，大约 80% 的美国人工智慧新创企业已在内部部署至少一款中国开源模型。此外，Qwen 成为 Hugging Face 上最快突破 10 亿次下载的模型家族，显示开源架构正迅速渗透至西方生产环境中。

中美两国在基础设施与商业化表现上呈现显著反差。在电力供给方面，中国于 2024 年新增 429 GW（预计至 2030 年每年新增逾 400 GW），而美国仅约 51 GW，且资料中心面临 241 GW 的电力缺口与 5 年电网排队期。相反地，中国受制于晶片运算瓶颈。营收方面，Anthropic 与 OpenAI 的年化营收分别突破 650 亿与 400 亿美元，而中国领先的字节跳动 Seedance 年化营收仅约 20 至 30 亿美元，纯语言模型公司如 DeepSeek 则接近 5 亿美元。

市场估值倍数与运算设施布局显现出分歧。在估值倍数上，月之暗面达到年化营收的 115 倍，智谱与 MiniMax 约 100 倍，远高于 Anthropic 的 20 倍。在运算基础设施方面，美国顶级实验室各自运作约 2 GW 运算力，锁单逾 5 GW，并计划于 2030 年达到 30 GW。目前跨太平洋双向技术流动持续加深，美国垂直人工智慧深度依赖中国开放权重模型，未来走势将取决于美国电网扩建与中国晶片良率提升的速度竞赛。

U.S. export controls inadvertently forced Chinese AI laboratories to pursue deep engineering optimization and open-source ecosystems. Within less than 18 months, Chinese models expanded their token consumption share on OpenRouter from under 2% to 45%, with approximately 80% of U.S. AI startups deploying at least one Chinese open-source model internally. Additionally, Qwen became the fastest model family to surpass 1 billion downloads on Hugging Face, illustrating how open-source architecture rapidly penetrated Western production environments.

The infrastructure and financial metrics between the two nations reveal stark contrasts. Regarding power supply, China added 429 GW in 2024 (projecting over 400 GW annually through 2030), whereas the U.S. added approximately 51 GW, facing a 241 GW datacenter power deficit and five-year grid interconnection delays. Conversely, China remains constrained by compute chips. Financially, Anthropic and OpenAI reached annualized revenues exceeding $65 billion and $40 billion, whereas China's leading ByteDance Seedance generated $2–3 billion, and pure LLM developers like DeepSeek approached $500 million.

Valuation multiples and computational planning demonstrate divergence across both markets. In valuation metrics, Moonshot reached 115 times annualized revenue, while Zhipu and MiniMax reached approximately 100 times, compared to Anthropic at 20 times. For computing infrastructure, leading U.S. laboratories each operate approximately 2 GW of compute, have booked over 5 GW, and target 30 GW by 2030. Meanwhile, cross-Pacific interdependence deepens as U.S. vertical AI applications leverage Chinese open-weight models, making the midterm landscape dependent on grid expansion versus domestic chip yield breakthroughs.