中国几乎已经追平：美国不会“赢得”AI战，而是应该这么做

背景与摘要

本文探讨了随着Moonshot.AI（月之暗面）、Z.ai和阿里巴巴等中国企业发布与美国顶尖模型性能相当的开源模型，美国在大型语言模型（LLMs）领域的“护城河”正在消失的现状。作者Gary Marcus指出，寄希望于通过LLM确立永久垄断地位的设想已被打破，美国AI产业正面临利润稀缺和全球激烈竞争的挑战。文章呼吁美国政府重新审视相关政策，并提出了从“无所作为”、“救市”到建立“AI界的CERN”等七个潜在的战略选项。

By Gary Marcus

By Gary Marcus

📌 Executive Summary

追赶已然完成： 近期，包括月之暗面 (Moonshot.AI) 的 Kimi K3、Z.ai 的 GLM 5.2 以及阿里巴巴的 Qwen 等中国公司取得的突破表明，中国已基本抹平了与美国顶尖 AI 模型之间的差距——并且通常免费提供高性能、开源权重 (open-weight) 的模型。
- The Catch-Up is Complete: Recent breakthroughs by Chinese firms like Moonshot.AI (Kimi K3), Z.ai (GLM 5.2), and Alibaba (Qwen) demonstrate that China has essentially closed the gap with the best American AI models—often offering high-performing, open-weight models for free.
硅谷“护城河”的消亡： 人们曾认为美国将通过大型语言模型 (LLM) 在 AI 领域获得永久性的、垄断性的胜利，但这一信念已被证明是毫无根据的。缺乏盈利能力、技术护城河的缺失以及激烈的全球竞争正在动摇美国的生成式 AI (GenAI) 产业。
- The Death of the Silicon Valley "Moat": The belief that the US would secure a permanent, monopolistic victory in AI via Large Language Models (LLMs) has proven unfounded. Unprofitability, a lack of technical moats, and aggressive global competition are destabilizing the US GenAI industry.
战略政策选项： 美国政府在未来面临着艰难的抉择，从无所作为或建立监管护城河，到将实验室国有化或转向类似于“AI 界的欧洲核子研究中心 (CERN)”的全球公共产品模式。
- Strategic Policy Options: The US government faces difficult choices moving forward, ranging from doing nothing or erecting regulatory moats, to nationalizing labs or shifting toward a global public good model akin to "CERN for AI."

The Illusion of US AI Supremacy

三周前，我曾警告过一个建立在毫无技术护城河、过度竞争且利润稀缺基础上的行业，其不可避免的衰退：

Three weeks ago, I warned about the inevitable fallout of an industry built on zero technical moats, hyper-competition, and scarce profits:

“我自 2023 年夏天以来一直在这里定期阐述的观点——‘没有护城河 \(\rightarrow\) 更多竞争者 \(\rightarrow\) 价格战 \(\rightarrow\) 利润稀缺’——其最终顶点已经到来，这可能会摧毁美国的 AI 产业。”

"The ultimate culmination of the 'no moat \(\rightarrow\) more competitors \(\rightarrow\) price wars \(\rightarrow\) profits are scarce' argument that I have been making here regularly since the summer of 2023, has arrived—and may wreck the U.S. AI industry."

上周，中国公司月之暗面 (Moonshot.AI) 宣布推出 Kimi K3，这是一个与美国顶尖替代产品大致相当的模型。独特的是，它是一个开源权重的模型，消费者可以免费下载并在本地运行（前提是他们有必要的硬件）。这是继 Z.ai 的 GLM 5.2 和阿里巴巴的 Qwen 引发浪潮之后的又一力作。

Last week, Chinese company Moonshot.AI announced Kimi K3, a model roughly on par with top American alternatives. Uniquely, it is an open-weight model that consumers can download and run locally for free (provided they have the necessary hardware). This follows waves made by Z.ai's GLM 5.2 and Alibaba's Qwen.

这种迅速变化的格局破坏了 OpenAI 和 Anthropic 的商业模式，让人们对它们即将到来的 IPO 产生怀疑，并打破了美国能够在零和 AI 竞赛中取得决定性“胜利”的观念。

This rapidly shifting landscape undermines the business models of OpenAI and Anthropic, casts doubt on their upcoming IPOs, and shatters the notion that America can definitively "win" a zero-sum AI race.

Why This Shouldn't Be a Surprise

我在 2025 年 1 月撰写的文章——就在 DeepSeek 发布之后——精确地阐明了今天正在发生的事情。当时，我预测：

My January 2025 essay—written right after the DeepSeek release—laid out precisely what is happening today. At the time, I predicted that: * 美国不会仅仅通过过度专注于 LLM 就能对中国取得决定性胜利。 * The US would not achieve a decisive victory over China by hyper-focusing exclusively on LLMs. * 两国将在技术上达成平手。 * The two nations would converge on a technical tie. * 缺乏技术护城河将使 OpenAI 这样的公司无法盈利。 * The lack of a technical moat would keep companies like OpenAI unprofitable. * 英伟达 (Nvidia) 将面临弱点。 * Nvidia would face vulnerabilities. * 《芯片法案》 (CHIPS Act) 将在无法有效遏制中国的情况下产生反噬。 * The CHIPS Act would backfire without effectively holding China back. * 模型将变得更便宜、更高效，但诸如幻觉和不可靠等顽固问题仍将存在。 * Models would grow cheaper and more efficient, but persistent issues like hallucinations and unreliability would remain. * 无止境地重复扩展 LLM 将会抽干原本可用于更具原创性想法的资金。 * Endless repetition of LLM scaling would starve funding from more original ideas.

不幸的是，一个与硅谷群体思维深度纠缠的政府未能听取这些警告。通过在生成式 AI (GenAI) 上“全盘押注”，美国经济被置于一种危险的境地，而国外的竞争对手却在向前迈进。

Unfortunately, a government heavily intertwined with Silicon Valley groupthink failed to heed these warnings. By going "all-in" on GenAI, the US economy has been placed in a precarious position while competitors abroad pull ahead.

Questions Congress Should Be Asking

在决定前进方向之前，立法者必须调查美国是如何挥霍掉其最初的领先优势的：

Before deciding on a path forward, lawmakers must investigate how the US squandered its initial lead: * 是否犯下了根本的战略或战术错误？ * Were fundamental strategic or tactical errors made? * 将一切押注于一项单一且容易复制的技术是否是个错误？ * Was banking everything on a single, easily replicable technology a mistake? * 美国公司在保护知识产权方面做得足够吗？ * Did US companies do enough to protect their intellectual property? * 中国公司在多大程度上获得了国家补贴？间谍活动在多大程度上参与其中？ * To what degree are Chinese companies state-subsidized? To what extent was espionage involved? * 对移民的限制是否造成了致命的人才流失？（注：Kimi 的创始人在返回中国之前曾就读于卡内基梅隆大学。） * Have restrictions on immigration created a fatal talent drain? (Note: Kimi’s founder studied at Carnegie Mellon before returning to China.) * 有哪些经验教训可以应用于即将到来的技术变革，如量子计算或核聚变？ * What lessons can be applied to upcoming technological shifts, such as quantum computing or nuclear fusion?

Seven Strategic Options for the US Government

随着美国政府评估其下一步行动，出现了七条不同的路径——从自由放任的资本主义到激进的全球合作。

As the administration assesses its next moves, seven distinct paths emerge—ranging from laissez-faire capitalism to radical global cooperation.

1. Do Nothing

让 OpenAI 和 Anthropic 在没有政府救助或后门补贴的情况下自力更生。如果它们找到了利基市场（例如前沿部署的企业合同）或者倒闭了，那就随它去吧。亚马逊、谷歌和微软将继续屹立不倒，处理排除了中国模型之后的国内和机密需求。缺点：将很大一部分全球 AI 市场拱手让给中国。

Let OpenAI and Anthropic stand on their own feet without government bailouts or backdoor subsidies. If they find niches (like forward-deployed enterprise contracts) or fail, so be it. Amazon, Google, and Microsoft will remain standing to handle domestic and classified needs where Chinese models are excluded. Drawback: Concedes a massive portion of the global AI market to China.

2. Outlaw Open Source

试图禁止开源 AI。缺点：祝你好运。覆水难收；禁令只会把开发推向暗网和海外司法管辖区。

Try to ban open-source AI. Drawback: Good luck with that. The cat is out of the bag; bans would merely push development to the dark web and overseas jurisdictions.

3. Build a "Regulatory Moat"

打着安全或保障的幌子，利用官僚国家的强制权力，将更便宜的外国竞争对手清除出美国市场。批评者认为，这相当于“企业共产主义”——粉碎独立的初创企业，扼杀创新，抬高消费者价格，并保护少数几家摇摇欲坠的巨头。

Use the coercive power of the bureaucratic state to clear the American market of cheaper foreign competitors under the guise of safety or security. Critics argue this amounts to "corporate communism"—crushing independent startups, stifling innovation, inflating consumer prices, and protecting a select few failing giants.

4. Bail Out the Big AI Labs

将公共资金投入到拯救陷入困境的专有实验室中。坚决反对。

Pour public funds into rescuing struggling proprietary labs. Hard pass.

5. Ban Chinese Models Altogether

尽管有传言称正在考虑之中，但全面禁令很可能会孤立美国生态系统，扼杀国内创新，并推高消费者成本。

While rumored to be under consideration, blanket bans will likely isolate the US ecosystem, stifle domestic innovation, and drive up consumer costs.

6. Buy Out OpenAI and Anthropic for Pennies on the Dollar

鉴于其大量烧钱的模式以及缺乏已被证明的盈利能力，政府可以以其巅峰估值的零头收购这些实验室，并将它们转变为国家研究实验室。缺点：存在政府过度干预和大规模监控滥用的重大风险。

Given their heavy cash-burn models and lack of proven profitability, the government could acquire these labs for a fraction of their peak valuations and convert them into national research laboratories. Drawback: Significant risk of government overreach and mass surveillance abuse.

7. Shift Toward AI as a Global Public Good (A "CERN for AI")

放弃零和心态，转向国际合作。正如我早在 2017 年在《纽约时报》上所建议的那样：

Give up the zero-sum mindset and pivot toward international collaboration. As I proposed in The New York Times back in 2017:

“我羡慕我在高能物理学领域的同行们，尤其是欧洲核子研究中心 (CERN) 的同行们……他们追求雄心勃勃、定义明确的项目……并与全世界分享他们的成果，而不是将成果限制在一个国家或公司……一项国际性的 A.I. 任务能够真正为了变得更好而改变世界。”

"I look with envy at my peers in high-energy physics, and in particular at CERN... They pursue ambitious, tightly defined projects... and share their results with the world, rather than restricting them to a single country or corporation... An international A.I. mission could genuinely change the world for the better."

有趣的是，近期关于在国际 AI 大会上开幕的讲话表明，中国可能愿意就此事进行对话。

Interestingly, recent remarks by Xi Jinping opening an international AI conference suggest China might be open to dialogue on the matter.

如果美国政府想要实现真正的范式转变，一个大胆的举措将包括在全球范围内取缔闭源 AI——要求所有制造商与经认可的研究人员共享其权重、架构和训练数据——并与中国合作，致力于利用 AI 推动全球科学和医疗的进步。

If the administration wants a true paradigm shift, a bold stroke would involve outlawing closed-source AI globally—requiring all manufacturers to share their weights, architectures, and training data with accredited researchers—and partnering with China to dedicate AI to global scientific and medical advancement.

正如投资者兼 CEO 菲尔·利宾 (Phil Libin) 所指出的：

As investor and CEO Phil Libin noted: “当资本、思想和人员自由流动时，美国体系就会受益。我们不需要去‘赢得’一场零和游戏。我们通过将游戏转化为非零和游戏来赢得胜利。”

"The American system benefits when there’s free flow of capital, ideas, and people. We don’t need to 'win' a zero-sum game. We win by transforming games to non-zero-sum."

Gary Marcus 曾在 2016 年提议建立一个 AI 领域的 CERN，并在 2023 年向美国参议院重申了这一概念。

Gary Marcus proposed a CERN for AI in 2016 and echoed the concept before the US Senate in 2023.