内容来源：https://www.wired.com/story/openai-sovereign-ai-us-china-tech-war/

内容总结：

今年以来，美国人工智能公司OpenAI以“主权人工智能”为名，与多国政府达成合作协议。该公司声称此举旨在让各国政府对这项可能重塑经济格局的技术掌握更大控制权，部分合作项目甚至与美国政府协调推进。

“主权人工智能”已成为华盛顿和硅谷的热门概念。支持者宣称，在美国及其盟友国家开发的人工智能系统必须在全球普及，以“阻止战略竞争对手通过技术输出使盟友依赖外国技术”。值得注意的是，OpenAI选择的合作伙伴包括阿联酋等君主制国家，其高管声称与“非民主政府”合作能推动其“向更自由方向演变”，这种论调与二十多年前某些西方政客对中国的判断如出一辙。

在技术实现路径上存在理念分歧。开源阵营代表企业Hugging Face首席执行官指出“没有开源就没有真正的主权”，而OpenAI则主张开源与闭源模式可并行发展。目前OpenAI与阿联酋的合作包括建设大型数据中心集群，但该国政府仍无法查看或修改ChatGPT的内部运行机制。

与美企着力推动政府合作形成对比的是，中国正通过开源模式在全球人工智能领域快速拓展。阿里巴巴的千问系列模型全球下载量已突破3亿次，基于该模型衍生的子模型超过10万个，连阿联酋研究人员也依托其开发出顶尖新模型。日本等国的初创企业发现千问模型在本地化任务处理上表现优异后已广泛采用。业内观察人士指出，中国开源模型正使全球算力资源产生更高效益——不同实验室无需重复训练基础模型，这种协同创新模式加速了技术迭代。

尽管OpenAI近期也发布了开源模型，但业内人士透露其举措某种程度上是受到中国深度求索等公司开源模型迅猛发展的推动。开源战略正助力中国人工智能实现跨越式发展，有专家预测中国在人工智能整体领域明年实现领先“也不足为奇”。

（根据Model Behavior时事通讯内容整理）

中文翻译：

今年，OpenAI已与多国政府宣布合作建设所谓"主权AI"系统。该公司表示，这些部分获得美国政府协调的合作项目，旨在让各国领导人能对这项重塑经济格局的技术掌握更大主导权。

过去数月间，"主权AI"已成为华盛顿与硅谷热议的话题。支持者主张，民主国家开发的AI系统必须实现全球推广，尤其是在中国加速向海外输出AI技术的背景下。特朗普政府在7月发布的《AI行动计划》中明确指出："美国技术的传播扩散将阻止战略竞争对手使我们的盟友依赖外国敌对技术。"

对OpenAI而言，这项行动也意味着与阿联酋等君主制国家合作。该公司首席战略官贾森·权认为，与非民主政府合作能推动其向更开放的方向演进。他在伯克利Curve大会上接受采访时表示："我们秉持接触优于遏制的理念，这种策略有时奏效，有时未必。"

权的论述与二十多年前某些政要对中国的表态如出一辙。2000年中国筹备加入世贸组织时，美国总统比尔·克林顿曾宣称："我们要么引导中国走向正轨，要么袖手旁观将其推入歧途。"此后众多美国企业通过对华贸易获利丰厚，但中国政治体制却愈发集权。

有观点认为，真正的AI主权需要政府能审查并在一定程度上控制相关模型。开源AI平台Hugging Face首席执行官克莱门特·德朗格指出："在我看来，没有开源就谈不上主权。"这方面中国已占据先机，其开源模型正快速风靡全球。

何为"主权AI"实质？

当前的主权AI项目赋予各国对技术栈从部分到完整的控制权，即政府全面管理从硬件到软件的AI基础设施。大西洋理事会GeoTech中心副主任特丽莎·雷分析道："所有项目的共同基础是法律合规性——通过将至少部分基础设施与地理边界绑定，设计、开发与部署过程就能遵循该国法律。"

OpenAI在阿联酋与美国政府联合宣布的项目，包含在阿布扎比建设5吉瓦数据中心集群（其中200兆瓦计划于2026年投运）。虽然阿联酋正全国部署ChatGPT，但政府似乎无权探查或修改该聊天机器人的内部运行机制。

若在数年前，于威权国家建设AI基础设施的设想可能引发硅谷员工抗议。2019年谷歌员工曾成功迫使公司取消在中国推出审查版搜索引擎的计划。雷指出："当前这些大语言项目情况类似，但并未引发同等程度的抵制。『在境内运营需遵守当地法律』的认知，已逐渐成为常态。"

权坚称即使外国政府提出要求，OpenAI也不会审查信息："我们不会限制信息获取，可能进行内容补充，但绝不删除。"

中国的先发优势

当美国AI企业争相与外国政府结盟时，中国科技公司正将开源模型推向全球。阿里巴巴、腾讯等巨头与深度求索等初创企业发布的基础模型，性能已可比肩美国同类产品。

阿里巴巴透露其Qwen系列AI模型全球下载量突破3亿次，基于该系列构建的衍生模型超10万个。日本等国初创公司发现Qwen能精准完成本地语言任务后已广泛采用。上月阿联酋研究人员发布的尖端模型正是基于Qwen2.5构建。

据此前报道，OpenAI今年发布自GPT-2以来首个开放权重模型，既是长期规划成果，也受到中国深度求索模型1月横空出世引发的热潮推动。

德朗格表示，聚焦开源模式使中国AI企业能快速迭代，同业公司可迅速借鉴彼此的训练技巧。"他们从五年前的严重落后发展到如今与美国并驾齐驱，并主导开源领域。若中国明年在AI整体领域领先也不足为奇。"

他进一步指出，中国的开源战略还放大了其AI基础设施投资效益。"在美国，由于领域内多数技术闭源，每个实验室基本都在重复相同训练。而在欧洲或中国，更开放的开源生态使同等规模的算力能分散于不同实验室——当某个实验室完成训练并发布模型后，邻近实验室就无需重复劳动。"

OpenAI则认为开放与闭源模型均可融入主权AI体系。权表示："我们并非走向单一模型的发展路径。各国既希望采用最优秀的闭源模型，也需要依赖开源模型满足多样化应用场景。"

（本文系Model Behavior时事通讯专题报道，过往内容可通过此处查阅）

英文来源：

OpenAI has announced a number of projects this year with foreign governments to help build out what it has called their “sovereign AI” systems. The company says the deals, some of which are being coordinated with the US government, are part of a broader push to give national leaders more control over a technology that could reshape their economies.
Over the past few months, sovereign AI has become something of a buzzword in both Washington and Silicon Valley. Proponents of the concept argue it's crucial that AI systems developed in democratic nations are able to proliferate globally, particularly as China races to deploy its own AI technology abroad. “The distribution and diffusion of American technology will stop our strategic rivals from making our allies dependent on foreign adversary technology,” the Trump administration said in its AI Action Plan released in July.
At OpenAI, this movement has also meant partnering with countries like the United Arab Emirates, which is ruled by a federation of monarchies. OpenAI’s chief strategy officer, Jason Kwon, argues that partnering with non-Democratic governments can help them evolve to become more liberal. “There’s a bet that you make that engagement is better than containment,” Kwon said in an interview with WIRED last week at the Curve conference in Berkeley, California. “Sometimes that works, and sometimes it hasn't.”
Kwon’s reasoning echoes what some politicians said about China more than two decades ago. “We can work to pull China in the right direction, or we can turn our backs and almost certainly push it in the wrong direction,” US president Bill Clinton said in 2000 when China was gearing up to join the World Trade Organization. Since then, many American companies have gotten rich by trading with China, but the country’s government has only become more authoritarian.
Some people argue that true sovereignty can only be achieved if a government is able to inspect—and to some extent control—the AI model in question. “In my opinion, there is no sovereignty without open source,” says Clément Delangue, the CEO of Hugging Face, a company that hosts open source AI models. In this respect, China is already ahead, as its open source models are quickly becoming popular globally.
What Is “Sovereign AI” Actually?
Today’s sovereign AI projects range from giving countries partial to full control over the entire tech stack, meaning the government manages all of the AI infrastructure, from hardware to software. “The one common underlying thing for all of them is the legality portion—by having at least some part of the infrastructure tied to geographical boundaries, the design, development, and deployment would then adhere to some national laws,” says Trisha Ray, an associate director at the Atlantic Council’s GeoTech Center.
The deal OpenAI announced in partnership with the US government in the UAE includes a 5 gigawatt data center cluster in Abu Dhabi (200 megawatts of the total planned capacity is supposed to come online in 2026). The UAE is also deploying ChatGPT nationwide, but it doesn’t appear that the government will have any ability to look under the hood or alter the chatbot’s inner workings.
Only a few years ago, the idea of building AI infrastructure in authoritarian countries might have sparked worker protests in Silicon Valley. In 2019, Google employees pushed back against the tech giant’s plan to deploy a censored search engine in China, eventually succeeding in getting the project canceled. “What's happening with some of these LLM projects, it's quite similar, but there isn't as much of a backlash,” Ray says. “That notion of, ‘well, yes, if you're operating within a country's borders, you have to adhere to all laws of the land,’ that's become a lot more normalized over time.”
Kwon is adamant that OpenAI will not censor information even if asked by a foreign government. “We're not going to suppress informational resources,” he says. “We might add, but we're not going to eliminate.”
Beijing’s Head Start
While US AI firms race to partner with foreign leaders, Chinese tech companies are spreading their open source models across the globe. Giants like Alibaba and Tencent, as well as startups like DeepSeek, have released open source foundation models that essentially match the capabilities of US rivals.
Alibaba says that its Qwen family of AI models have been downloaded more than 300 million times worldwide, and over 100,000 derivative models have been built using them. Startups in countries like Japan have widely adopted Qwen after discovering that it excels at completing tasks in the local language. Last month, researchers in the UAE released a new state-of-the-art model built on top of the Qwen2.5.
Earlier this year, OpenAI released its first open weight models since GPT-2. While the project had been in the works for a while, OpenAI was also galvanized by the popularity of models released by China’s DeepSeek, which exploded on the scene in January, according to WIRED’s previous reporting.
Focusing on open source models has allowed Chinese AI companies to iterate rapidly, says Delangue, since rival firms can quickly adopt each other’s training tricks.
“They went from being very behind five years ago to now being on par with the US and dominating open source,” he says. “It wouldn't be surprising if China was ahead in AI in general next year.”
China’s open source strategy, Delangue argues, also helps its AI infrastructure investments go further. “One gigawatt in the US, where most of the field is closed source, means that every single lab is basically doing the same training run,” he says. “In Europe or in China, because it's much more open source and open science, the same gigawatt is actually distributed between the labs, because there's one lab doing the training and then they're releasing their models, which means that the neighboring lab doesn't need to do the same training.”
OpenAI, for its part, argues that both open and closed models can be part of the sovereign AI approach. “It doesn't seem like we are on the path where there's just one model,” Kwon says. “You see that when you go to various countries and they want to use both the best models that are closed, and then they also want to have open models that they also rely on for various types of use cases.”
This is an edition of the Model Behavior newsletter. Read previous newsletters here.