内容来源：https://www.wired.com/story/europe-race-us-deepseek-sovereign-ai/

内容总结：

随着美欧关系出现裂痕，欧洲人工智能实验室正积极探索创新路径，以缩小与占据主导地位的美国同行之间的差距。当前，从处理器设计制造、数据中心容量到模型与应用开发，美国企业在人工智能产业链的各个环节几乎全面领先欧洲竞争对手。美国亦吸纳了全球大部分人工智能投资，其本土股市表现与经济增速均印证了这一点。

有观点认为，英伟达、谷歌、Meta、OpenAI、Anthropic等美国领军企业地位已难以撼动，欧洲将如同依赖美国云服务一样，无法摆脱对其人工智能技术的依赖。比利时国家网络安全机构负责人年初曾坦言，欧洲已"失去互联网"，需接受一定程度依赖美国基础设施的现实。

然而，英国与欧盟并未放弃自主发展的努力。两国已投入数亿美元资金以减少对外国人工智能供应商的依赖。与此同时，中国实验室DeepSeek的成功案例打破了"算力规模决定胜负"的固有认知，激励欧洲研究人员探索通过创新模型设计打造竞争性产品的新路径。

牛津大学教授罗萨里亚·塔德奥指出："我们过于轻信'创新仅在美国发生'的叙事，这种认为欧洲错过人工智能发展机遇的论调是危险的。"欧洲实验室相较于封闭发展的美国巨头可能具备独特优势——坚持开放开发模式。通过发布可公开使用修改的模型，欧洲研究成果有望在协作中持续进化。"这能产生乘数效应，"德国莱布尼茨汉诺威大学计算机科学教授沃尔夫冈·奈德尔解释道。

面对白宫对欧洲领导力的冷淡态度，以及美国前总统特朗普部分盟友的公开敌意，欧洲推进技术创新与自主发展的紧迫性日益凸显。塔德奥教授强调："地缘政治变化改变了我们对主权的理解……这项技术是我们尚未掌握的关键基础设施，必须立即行动起来。"

跨大西洋摩擦加剧
近期在格陵兰主权、关税政策、移民等问题上的分歧，使欧洲领导人与特朗普政府关系持续紧张，甚至引发对北约联盟未来的担忧。双方在美国科技企业监管问题上冲突尤为明显：欧盟委员会因涉嫌违规对X平台处以1.4亿美元罚款后，美国国务卿马可·卢比奥谴责这是"外国政府对所有美国科技平台及人民的攻击"；英国监管机构就X平台传播AI生成色情图像展开调查后，美国国务院官员萨拉·罗杰斯随即威胁采取报复措施。

在此背景下，欧洲对美制人工智能的依赖日益成为战略弱点。专家指出，美国可能利用这种依赖性作为贸易谈判筹码。塔德奥警告："这种依赖在任何谈判中都是软肋——而未来与美国的谈判只会越来越多。"

为规避风险，欧洲各国正通过资金计划、定向放宽管制和校企合作等方式推动人工智能本土化生产，其中部分项目专注于开发适配欧洲语言的大语言模型。但奈德尔教授指出："这个领域往往赢家通吃。若无法掌握尖端技术，将永远处于追赶状态，最终只能用自己的数据喂养巨头企业，导致差距越拉越大。"

前路挑战
业界观察人士指出，欧洲对"数字主权"的具体定义尚不明确——究竟需要实现全产业链自给自足，还是仅在关键领域提升能力？是否要完全排除美国供应商？计算机与通信产业协会高级政策经理博尼法斯·德尚普里表示："当前这更像是一种理念表述。"

政策层面也存在分歧：部分供应商主张借鉴中国培育本土处理器市场的经验，通过强制或激励措施促使欧洲企业采购本土人工智能产品；而开放市场倡导者则认为排斥美国企业可能导致欧洲企业在全球竞争中处于劣势。德尚普里强调："主权意味着拥有选择权。"

尽管存在诸多分歧，但DeepSeek的成功案例让欧洲业界普遍相信，即使预算和资源有限，缩小与美国领导者的差距仍大有可为。奈德尔教授参与的SOOFI开源模型项目计划在一年内推出约千亿参数的通用语言模型。"这个领域的进步将不再完全依赖最大规模的GPU集群，"奈德尔充满信心地表示，"我们将成为欧洲的DeepSeek。"

中文翻译：

随着美国与其欧洲盟友的关系出现裂痕，欧洲大陆的人工智能实验室正在寻找创新方法，以缩小与目前主导该领域的美国竞争对手之间的差距。

除了极少数例外，美国企业在人工智能产业链的各个环节——从处理器设计与制造、数据中心容量到模型与应用开发——都超越了欧洲竞争对手。同样，美国吸引了流入人工智能领域的大部分资金，这反映在去年其本土股票的表现和经济增长上。

一些观点认为，美国领军企业——英伟达、谷歌、Meta、OpenAI、Anthropic等——地位已如此稳固，以至于欧洲国家不可能打破对美国人工智能的依赖，这类似于云服务领域的模式。今年1月初，比利时国家网络安全机构负责人对英国《金融时报》表示，欧洲已经"失去了互联网"，应该在一定程度上接受对美国基础设施的依赖。

然而，英国和欧盟政府似乎并未准备放弃。他们已经投入数亿美元资金，以尽量减少对外国人工智能供应商的依赖。与此同时，中国人工智能实验室DeepSeek的成功鼓舞了欧洲研究人员。DeepSeek去年的突破打破了"控制最大规模人工智能处理器集群决定企业成败"的教条，欧洲研究人员正在探索通过富有想象力的模型设计来开发有竞争力产品的替代方法。

"我们过于轻信了'创新只在美国发生'的说法——认为我们错过了人工智能列车，甚至不该考虑追赶，"牛津大学数字伦理与国防技术教授罗萨里娅·塔德奥表示，"这是一种危险的说法。"

欧洲人工智能实验室可能拥有的一大优势是愿意开放开发，这与大型美国企业形成对比——后者如同封闭的作坊，极少分享其训练数据或模型设计的复杂性。理论上，通过发布供任何人使用或修改的模型，欧洲实验室取得的突破将在合作者的进一步完善中产生复合效应。"你正在放大这些模型的力量，"德国莱布尼茨汉诺威大学计算机科学教授、L3S研究中心主任沃尔夫冈·内德尔表示。该中心是开发欧洲大型语言模型的联盟成员。

面对白宫对欧洲领导力的冷淡态度——以及美国前总统特朗普部分盟友赤裸裸的敌意——这些创新和实现自给自足的努力显得更加紧迫。

"地缘政治局势改变了我们理解主权的方式……这项技术是一种基础设施——而我们不生产这种基础设施，"塔德奥说，"我们必须开始朝这个方向努力。不能再忽视它了。"

跨大西洋摩擦

近几个月来，欧洲领导人发现自己在格陵兰主权、关税政策、移民等一系列问题上与特朗普政府意见相左，引发了人们对北约联盟可能恶化的猜测——该联盟已维持全球秩序超过75年。双方在监管美国科技公司——尤其是埃隆·马斯克拥有的社交媒体平台X——的方式上公开冲突尤为明显。

去年12月初，欧盟委员会因涉嫌违规对X处以相当于1.4亿美元的罚款后，美国国务卿马可·卢比奥谴责该处罚是"外国政府对所有美国科技平台和美国人民的攻击"。随后，在英国监管机构因X平台上大量传播人工智能生成的女性色情图像（可能导致全国禁令）而展开调查后，美国国务院官员莎拉·罗杰斯威胁要进行报复。

在此背景下，欧洲对美国制造的人工智能的依赖开始显得越来越像一种负担。在最坏的情况下（尽管专家认为可能性不大），美国可能会选择切断人工智能服务和关键数字基础设施的访问。更可能的是，特朗普政府可能利用欧洲的依赖作为筹码，因为双方正在继续敲定一项贸易协议。"这种依赖在任何谈判中都是不利因素——而且我们将越来越多地与美国谈判，"塔德奥说。

欧盟委员会、白宫和英国科学、创新与技术部未回应置评请求。

为对冲这些风险，欧洲国家试图通过资助计划、有针对性的放松管制以及与学术机构的合作，将人工智能生产转移到本土。一些努力集中在为欧洲本土语言构建有竞争力的大型语言模型，如Apertus和GPT-NL。

然而，只要ChatGPT或Claude继续优于欧洲制造的聊天机器人，美国在人工智能领域的领先优势只会扩大。"这些领域通常是赢家通吃。当你有一个非常好的平台时，每个人都会去那里，"内德尔说，"无法在该领域生产最先进的技术意味着你将无法赶上。你只会用你的输入喂养更大的参与者，所以他们会变得更好，而你会更加落后。"

弥合差距

游说者称，目前尚不清楚英国或欧盟打算将"数字主权"推进到何种程度。主权是否需要横跨庞大人工智能供应链的完全自给自足，还是只需要在少数几个学科领域提高能力？它要求排除美国供应商，还是只需要有国内替代方案？"这相当模糊，"技术公司会员组织计算机与通信行业协会高级政策经理博尼法斯·德尚普里斯表示，"现阶段似乎更多是一种说法。"

对于应该拉动哪些政策杠杆来为欧洲实现自给自足创造条件，也没有广泛共识。一些欧洲供应商主张采取一种策略，要求或至少激励欧洲企业从本土人工智能公司购买产品——类似于据报道中国在其国内处理器市场采取的做法。总部位于比利时、开发用于太空的人工智能专用处理器的Magics Technologies公司首席技术官曹颖（音）认为，与赠款和补贴不同，这种方法有助于培育需求。"这比单纯获得资金更重要，"曹说，"最重要的是你能卖出产品。"但那些主张开放市场和放松管制的人声称，试图排除美国人工智能公司可能会使国内企业在全球同行中处于劣势，因为它们只能选择最适合自己的人工智能产品。"从我们的角度来看，主权意味着拥有选择权，"德尚普里斯说。

尽管在政策细节上存在分歧，但人们普遍认为，正如DeepSeek所展示的那样，即使对于预算和资源有限的实验室来说，弥合与美国领军企业的性能差距仍然是完全可能的。"如果我已经认为我们无法赶上，我就不会尝试了，"内德尔说。他参与的开放源代码模型开发项目SOOFI计划在未来一年内推出一个具有约1000亿参数的、有竞争力的通用语言模型。

"该领域的进展将不再主要依赖最大的GPU集群，"内德尔声称，"我们将成为欧洲的DeepSeek。"

英文来源：

As the relationship between the US and its European allies shows signs of strain, AI labs across the continent are searching for inventive ways to close the gap with American rivals that have so far dominated the field.
With rare exceptions, US-based firms outstrip European competitors across the AI production line—from processor design and manufacturing, to datacenter capacity, to model and application development. Likewise, the US has captured a massive proportion of the money pouring into AI, reflected in the performance last year of its homegrown stocks and the growth of its econonmy.
The belief in some quarters is that the US-based leaders—Nvidia, Google, Meta, OpenAI, Anthropic, and the like—are already so entrenched as to make it impossible for European nations to break their dependency on American AI, mirroring the pattern in cloud services. In early January, the head of Belgium’s national cybersecurity organization told the Financial Times that Europe had “lost the internet,” and should make peace with a degree of reliance on US infrastructure.
Yet the governments of the UK and EU do not appear ready to give up. Already, they have committed hundreds of millions of dollars to minimizing their reliance on foreign AI suppliers. Meanwhile, buoyed by the success of China-based AI lab DeepSeek, whose breakout last year shattered the dogma that control over the largest fleet of AI processors determines which firm wins out, Europe-based researchers are pursuing alternative methods for developing competitive products built around imaginative model design.
“We have been too gullible to the narrative that innovation is done in the US—that we lost the AI train and should not even think about it,” claims Rosaria Taddeo, a professor of digital ethics and defence technologies at the University of Oxford. “That’s a dangerous narrative.”
One possible advantage that European AI labs hold over the large American firms—closed shops that share very little about their training data or the intricacies of their model design—is a willingness to develop out in the open. By publishing models for anyone to use or modify, the theory goes, breakthroughs achieved by European labs will compound as they're further refined by collaborators. “You are multiplying the power of these models,” claims Wolfgang Nejdl, professor of computer science at Germany’s Leibniz Universität Hannover and director of the L3S Research Center, part of a consortium developing a large language model for Europe.
In the face of the White House’s lukewarm stance toward European leadership—and a nakedly hostile attitude among some allies of US President Donald Trump—those efforts to innovate and become self-sufficient have taken on a new urgency.
“The geopolitical situation has changed the way we should interpret sovereignty…This technology is an infrastructure—and an infrastructure we do not produce,” claims Taddeo. “We have to start moving in that direction. It’s not possible to ignore it anymore.”
A Transatlantic Spat
In recent months, European leaders have found themselves at loggerheads with the Trump administration over a tangle of issues ranging from the sovereignty of Greenland, to tariff policy, to immigration, leading to speculation about a deterioration in the NATO alliance that has set the global order for more than 75 years. The two sides have clashed particularly openly over the approach to policing American tech firms—especially X, the social media platform owned by Elon Musk.
After the European Commission fined X the equivalent of $140 million over alleged regulatory violations in early December, US Secretary of State Marco Rubio condemned the penalty as “an attack on all American tech platforms and the American people by foreign governments.” Later, after a UK regulator opened an investigation into X over a torrent of AI-generated sexualized images of women distributed on the platform, a precursor to a possible countrywide ban, US State Department official Sarah Rogers threatened retaliation.
Against that backdrop, Europe’s reliance on American-made AI begins to look more and more like a liability. In a worst case scenario, though experts consider the possibility remote, the US could choose to withhold access to AI services and crucial digital infrastructure. More plausibly, the Trump administration could use Europe’s dependence as leverage as the two sides continue to iron out a trade deal. “That dependency is a liability in any negotiation—and we are going to be negotiating increasingly with the US,” says Taddeo.
The European Commission, White House, and UK Department for Science, Innovation and Technology did not respond to requests for comment.
To hedge against those risks, European nations have attempted to bring the production of AI onshore, through funding programs, targeted deregulation, and partnerships with academic institutions. Some efforts have focused on building competitive large language models for native European languages, like Apertus and GPT-NL.
For as long as ChatGPT or Claude continues to outperform Europe-made chatbots, though, America’s lead in AI will only grow. “These domains are very often winner-takes-all. When you have a very good platform, everybody goes there,” says Nejdl. “Not being able to produce state-of-the-art technology in this field means you will not catch up. You will always just feed the bigger players with your input, so they will get even better and you will be more behind.”
Mind the Gap
It is unclear precisely how far the UK or EU intends to take the push for “digital sovereignty,” lobbyists claim. Does sovereignty require total self-sufficiency across the sprawling AI supply chain, or only an improved capability in a narrow set of disciplines? Does it demand the exclusion of US-based providers, or only the availability of domestic alternatives? “It’s quite vague,” says Boniface de Champris, senior policy manager at the Computer & Communications Industry Association, a membership organization for technology companies. “It seems to be more of a narrative at this stage.”
Neither is there broad agreement as to which policy levers to pull to create the conditions for Europe to become self-sufficient. Some European suppliers advocate for a strategy whereby European businesses would be required, or at least incentivized, to buy from homegrown AI firms—similar to China’s reported approach to its domestic processor market. Unlike grants and subsidies, such an approach would help to seed demand, argues Ying Cao, CTO at Magics Technologies, a Belgium-based outfit developing AI-specific processors for use in space. “That’s more important than simply access to capital,” says Cao. “The most important thing is that you can sell your products.” But those who advocate for open markets and deregulation claim that trying to cut out US-based AI companies risks putting domestic businesses at a disadvantage to global peers, left to choose whichever AI products suit them best. “From our perspective, sovereignty means having choice,” says de Champris.
But for all the disagreement over policy minutiae, there is a broad belief that bridging the performance gap to the American leaders remains eminently possible for even budget- and resource-constrained labs, as DeepSeek illustrated. “If I would already think we will not catch up, I would not [try],” says Nejdl. SOOFI, the open source model development project in which Nejdl is involved, intends to put out a competitive general purpose language model with roughly 100 billion parameters within the next year.
“Progress in this field will not to the larger part depend anymore on the biggest GPU clusters,” claims Nejdl. “We will be the European DeepSeek.”