240 points
• 5 days ago
• Article
Link
美国在 AI 竞赛中最关键的商业化环节上占据领先。自 2025 年初 DeepSeek R1 引发轰动以来，美国企业提速发展，OpenAI 向智能体领域和 Codex 扩展，Anthropic 把 Claude Code 打造成为商业产品。中国有竞争力的模型，但美国在营收、采用率、工具链和全球影响力上明显领先。对中国而言，DeepSeek 的真正战略价值不在于商业主导，而在于通过推动推理向华为昇腾等国产硬件迁移，减少对 Nvidia 的依赖，支持供应链自主，而非追求以盈利为核心的 AI 霸主地位。
能源对 AI 很重要，但并非决定性因素。廉价电力能降低模型运行成本，美国的零售电价低于德国、英国、法国和西班牙等西欧主要经济体，但中国和俄罗斯的电价更便宜。电力本身不足以决定输赢：即便电价低廉，如果缺乏云规模、平台覆盖、开发者生态和大量可用数据的获取渠道，也会落败。美国在这些层面都形成了协同效应，这恰恰是其最强的优势。
真正决定胜负的是云基础设施和数据。全球性的超大规模云服务商——AWS 、 Azure 和 Google Cloud——掌握了模型触达世界的主要通道。美国的平台同时在生成和组织驱动 AI 所需的数据，从 YouTube 的视频语料，到嵌入日常办公的 Google Drive 与 Microsoft 365，再到深度参与软件开发的 GitHub 。它们既是分发系统，又是数据平台，使得新模型可以迅速推送到人们日常使用的产品中。中国在其庞大的国内市场内也具备许多这样的能力，但欧洲不具备这一整套体系。
欧洲有高水平的工程人才，但仅有人才还不够。美国的超大规模云服务商已占据主导地位，要想迎头赶上，除了要为本土云提供巨额资金，还必须把银行、制造业和公共机构迁移到这些平台上，这一过程可能耗时接近十年。而到那时，美国平台的领先优势将更难撼动。唯一的例外是 Arkady Volozh 正试图把 Nebius 打造为一家欧洲 AI 基础设施公司，但这恰恰说明了欧洲在这场竞赛中的起步仍然很晚。
SAP 的 Christian Klein 主张欧洲不需要更多数据中心，仅靠大语言模型也远远不够——这两点都是对的。但他忽视了一个核心事实：美国之所以领先，是因为它在同时构建每一个关键层面——芯片、能源、数据中心、云平台、开发者工具、消费级平台和企业软件。 AI 只有与真实数据、真实工作流和真实产品结合时才有价值，而美国具备将这些环节在大规模上整合起来的体系。
另一个正在出现的前沿是武器化的 AI 。下一阶段可能会出现国家之间在机器人网络、网络攻防和自主武器上的 AI 对抗。将系统调整为去人性化对手或针对特定人群令人不安地容易，一旦模型被嵌入媒体、网络和武器中，偏见就会转化为力量。这也把 AI 竞赛推向了一场安全竞赛。像 Anthropic 的 Mythos 这样的模型预示着另一次转变：以往的开源冲动会让位于通过"隐藏式安全"实现的封闭体系，采用封闭的软件、工具、固件和芯片。专有堆栈的价值会延伸到硬件层面，因为无法在目标的代码和架构上训练的模型，将缺乏上下文并且运行速度更慢。
The US is winning the AI race where it counts most, commercialization. Since DeepSeek R1 made waves in early 2025, American companies have accelerated faster, with OpenAI pushing into agents and Codex and Anthropic turning Claude Code into a business. China has competitive models, but the US leads clearly in revenue, adoption, tools, and global reach. DeepSeek's real strategic value for China is less about commercial dominance and more about reducing dependence on Nvidia by pushing inference toward domestic hardware like Huawei Ascend, supporting supply chain autonomy rather than profitable AI leadership.
Energy matters for AI, but it is not the decisive factor. Cheap electricity lowers model costs, and the US enjoys lower retail power prices than major Western European economies like Germany, the UK, France, and Spain. However, China and Russia are even cheaper. Power alone does not determine the winner. A country can have inexpensive electricity and still lose if it lacks cloud scale, platform reach, developer ecosystems, and access to large flows of useful data. The US has all of these layers working together, which is where its strongest advantage lies.
The decisive layers are cloud infrastructure and data. The US owns the global hyperscalers, AWS, Azure, and Google Cloud, which serve as the primary channels through which models reach the world. American platforms also generate and organize the data that fuels AI, from YouTube's video corpus to Google Drive and Microsoft 365 embedded in daily office work to GitHub sitting inside software development. These are both distribution systems and data platforms, allowing new models to be pushed into products people already use every day. China has much of this within its large domestic market, but Europe does not.
Europe has strong engineering talent, but talent alone is not enough. US hyperscalers already dominate, and catching up would require not just financing cloud champions but also migrating banks, manufacturers, and public agencies onto those platforms, a process that would take most of a decade. By then, American platforms would be even further ahead. There is one exception, Arkady Volozh's effort to build Nebius into a European AI infrastructure company, but that only confirms how early Europe still is in this race.
SAP's Christian Klein has argued that Europe does not need more data centers and that large language models alone are not enough. He is right on both points, but his broader view misses the main fact. The US is winning because it is building every major layer simultaneously: chips, power, data centers, cloud platforms, developer tools, consumer platforms, and enterprise software. AI only becomes valuable when tied to real data, real workflows, and real products, and the US has the integrated system to make that happen at scale.
There is another frontier emerging: weaponized AI. The next phase may involve country-versus-country AI in bot networks, cyber campaigns, and autonomous weapons. Tuning systems to dehumanize rivals or target populations is disturbingly easy, and once models are embedded into media, networks, and weapons, bias becomes force. This makes the AI race also a security race. Models like Anthropic's Mythos point to another shift, where the old open-source instinct gives way to security by obscurity, with closed software, tooling, firmware, and chips. Proprietary stacks raise their value all the way down to hardware, because a model that cannot train on a target's code and architecture will have less context and speed.
676 comments • Comments Link
• 有人质疑把 AI 发展描述为"战争"的说法，认为这种竞争思维反而会侵蚀全球的善意与信任，从而损害美国自身利益。如果美国和中国都不再被信任来提供数据或可靠服务，双方都会失去潜在的全球影响力，而欧洲等地区可能会因此受益。
• 一些人认为把 AI 描绘成"战争"是媒体和企业利益推动的叙事，AI 公司为耸人听闻的言论提供资金，以转移公众对真正挑战——对齐问题的注意力。人们更应该关心的不是哪个国家率先开发出更先进的 AI，而是这些 AI 是否与人类利益相一致。
• 在硬件未来轨迹上有一种技术论点，预测太空数据中心由于发射成本下降和太阳能充足，最终在经济性和实用性上会优于地面数据中心。
• 美国倾向于将一切视为竞争赛跑，这种心态受到批评。批评者认为它导致了适得其反的政策和态度，甚至在企业内部沟通中也充满了关于"赢"的体育隐喻。
• 对 GPU 性能趋势的详细分析显示，消费级与高端硬件之间存在长期差距，这引发了对当前大规模 AI 投资经济可持续性的质疑。如果消费级硬件在 10–15 年内追上今天的高端 GPU，那么对尖端基础设施的数万亿美元投资可能难以获得足够回报。
• 许多人怀疑前沿 AI 进展是否真构成一场"竞赛"，指出多个国家和公司都在稳步推进，没有任何单一实体取得决定性且持久的优势。"赢者通吃"的说法更像是为了吸引风险资本，而非反映技术现实。
• 对数据隐私的担忧促使人们预测，对美国和中国治理的不信任将推动用户转向本地开源模型，在个人和企业层面避免依赖任何单一国家的云基础设施。
• 鉴于许多公司以亏损方式出售服务且盈利前景不明，人们质疑美国 AI 公司的商业模式。如果中国的开源权重模型持续改进并在价格上形成优势，美国可能会在商业层面失去优势，尽管目前其模型更为先进。
• 出于国家安全考虑，如果模型带来风险（例如可能助长生物恐怖主义），这些模型可能会被限制使用或被降级用于商业用途。类似于核武器的管控，这可能会限制最强大 AI 系统的可及性。
• 本地化 AI 推理正在快速进步，有人认为几年内强大模型就能在消费级硬件上运行，从而减少对云提供商的依赖，并对当前集中式 AI 基础设施模式的长期可行性提出质疑。
整体讨论反映出对"AI 竞赛"叙事的深切怀疑。许多参与者质疑目前的商业化策略是否可持续或是否值得追求，隐私、信任与公平访问的担忧凸显，同时围绕模型分发和硬件未来也存在技术性争论。反复出现的主题是：将 AI 发展框定为零和竞争可能适得其反，会分散对对齐挑战和社会影响的关注。对话中带有地缘政治的愤世嫉俗，无论是美国还是中国，都不被视为值得信赖的强大 AI 技术监管者。 • The framing of AI development as a "war" is questioned, with the argument that this competitive mindset may actually undermine U.S. interests by eroding global goodwill and trust. If neither the U.S. nor China can be trusted with data or reliable service, both lose potential global influence, potentially benefiting regions like Europe.
• Some argue the "war" narrative is driven by media and corporate interests, with AI companies funding sensationalist rhetoric to distract from the real challenge: alignment. The concern is not which nation develops advanced AI first, but whether that AI is aligned with humanity's interests.
• There is a technical argument about the future trajectory of hardware, with predictions that space-based data centers will become economically and practically superior to ground-based ones due to lower launch costs and abundant solar energy.
• The U.S. tendency to frame everything as a competitive race is highlighted, with criticism that this mindset leads to counterproductive policies and attitudes. One commenter humorously notes that even internal corporate communications are filled with sports metaphors about "winning."
• A detailed breakdown of GPU performance trends shows a consistent lag between consumer and high-end hardware, raising questions about the economic sustainability of current massive AI investments. If consumer hardware catches up to today's high-end GPUs within 10-15 years, the trillions invested in cutting-edge infrastructure may not generate adequate returns.
• Skepticism is expressed about whether frontier AI development constitutes a true "race" at all, pointing out that multiple nations and companies are making incremental progress without any single entity achieving a decisive, lasting advantage. The "winner-take-all" narrative is seen as more useful for raising venture capital than reflecting technological reality.
• Concerns about data privacy are raised, with predictions that distrust of both U.S. and Chinese governance will push users toward local, open-source models for personal and enterprise use, avoiding reliance on any single nation's cloud infrastructure.
• The economic model of U.S. AI companies is questioned, given that many sell services at a loss with uncertain profitability. If Chinese open-weight models continue to improve and undercut on price, the U.S. may lose its commercial edge despite having more advanced models currently.
• National security considerations could lead to models being restricted or degraded for commercial use if they pose risks, such as enabling bioterrorism. This mirrors nuclear technology controls and could limit the accessibility of the most capable AI systems.
• Local AI inference is improving rapidly, with arguments that within a few years, powerful models will run on consumer hardware, reducing dependence on cloud providers and raising questions about the long-term viability of the current centralized AI infrastructure model.
The discussion reveals deep skepticism about the "AI race" narrative, with many participants questioning whether current commercialization strategies are sustainable or even desirable. Concerns about privacy, trust, and equitable access feature prominently, alongside technical debates about the future of model distribution and hardware. There's a recurring theme that the framing of AI development as a zero-sum competition may itself be counterproductive, distracting from alignment challenges and social impacts. The conversation is marked by geopolitical cynicism, with neither the U.S. nor China viewed as trustworthy stewards of powerful AI technology.