美国加利福尼亚州帕洛阿尔托/台北/新加坡——本月，Anthropic首席执行官达里奥·阿莫代伊就放慢人工智能前沿研发引发争论，令人工智能相关股票重挫。三天后，数千家依赖大规模快速建设的供应商齐聚硅谷一座会议中心。
PALO ALTO, California/TAIPEI/SINGAPORE -- Three days after Anthropic chief Dario Amodei triggered a debate this month on slowing frontier development of artificial intelligence, sending AI-related stocks tumbling, thousands of suppliers dependent on a massive and rapid buildout gathered at a Silicon Valley convention center.
尽管踩下刹车可能给其所处行业造成损害，现场气氛却丝毫不显忧虑。
Despite the damage that hitting the brakes might inflict on their industry, the mood in the room was anything but worried.
今年9月中旬在圣克拉拉举行的年度人工智能基础设施峰会，汇聚了人工智能供应链几乎各个环节的参与者，从数据中心运营商、芯片制造商，到大型语言模型（LLM）开发商。
The annual AI Infra Summit, held in mid-September this year in Santa Clara, brought together those from nearly every link in the AI supply chain, from data center operators and chipmakers to large language model (LLM) builders.
阿莫代伊9月12日发布的博文表达了对人工智能变得强大到无法控制的担忧。OpenAI首席执行官萨姆·奥尔特曼和SpaceX所有者埃隆·马斯克罕见地立场一致，立即支持他放慢发展进程的呼吁。
Amodei's Sept.12 blog post had expressed his concerns about AI becoming too powerful to control, with OpenAI CEO Sam Altman and SpaceX owner Elon Musk displaying a rare unity in immediately backing his call for a slowdown.
然而，尽管前沿模型尚未兑现通用人工智能（AGI）的承诺——即人工智能在执行大多数任务方面达到或超越人类能力——为迎接这一天的底层投资却已升至前所未有的水平，令与会者的订单簿爆满。
But while their frontier models have yet to deliver on the promise of Artificial General Intelligence (AGI), when AI would match or exceed humans' capabilities to perform most tasks, the underlying investment to reach that day has been rising to unprecedented levels, filling the order books of attendees to overflowing.
据高盛研究称，2026年全球人工智能总投资将超过1万亿美元。贝恩的最新估计显示，仅全球数据中心建设的投资规模到2030年就将达到5万亿至6.5万亿美元。
According to Goldman Sachs Research, total global AI investment will exceed $1 trillion in 2026. Global investment in data center buildout alone is projected to reach $5 trillion to $6.5 trillion by 2030, according to Bain's latest estimate.
贝恩的数据表明，人工智能基础设施的建设速度已走在实际需求曲线之前；全球国内生产总值必须额外增长3.5%，如此规模的投资才能获得回报。
Bain's figures suggest AI infrastructure is being built ahead of the actual AI demand curve, and global gross domestic product would have to grow an additional 3.5% for this level of investment to pay off.
在圣克拉拉峰会的繁忙展览现场，供应商们对这些问题并不在意；他们最关心的其实是能否满足人工智能（AI）带来的需求——这些需求涵盖了从芯片到数据中心通风系统等各种产品。“什么‘AI发展放缓’？我们的业务比以往任何时候都更繁忙，”杭州申士节能科技有限公司（Hangzhou Shenshi Energy Conservation Technology）的总经理Zach Shen说道。该公司专门生产用于数据中心的冷却设备，并计划在美国市场扩大业务。
On the busy expo floor of the Santa Clara summit, suppliers were shrugging off those concerns, as their biggest worry was to keep up with the AI-driven demand for everything from chips to ventilation systems for data centers "What AI slowdown? We're busier than ever," said Zach Shen, a general manager at Chinese company Hangzhou Shenshi Energy Conservation Technology, a maker of cooling equipment for data centers that is looking to expand in the U.S. market. A semiconductor analyst attending -- Vivek Arya at Bank of America -- said in a note to clients that the AI infrastructure buildout and capacity investments showed "no signs of slowing," adding that "AI infra investments remain firmly constructive across chip vendors, frontier labs, DC [data center] operators."
一位来自美国银行的半导体分析师Vivek Arya在给客户的报告中指出，AI基础设施的建设及相关投资“没有放缓的迹象”；他还表示：“芯片制造商、前沿研究机构以及数据中心运营商对AI基础设施的投资依然非常积极。”本月在硅谷举行的AI基础设施峰会吸引了约8,500名参与者，人数是去年的两倍多。许多供应商都表示，市场需求依然远大于供应。
This month's AI Infra Summit in Silicon Valley featured packed sessions and halls as it drew about 8,500 attendees, more than double last year's turnout. (AI Infra Summit) Many suppliers reported that demand still outstripped supply.
“市场需求确实非常强劲，”领先的芯片设计工具制造商Synopsys的AI合作部门高级副总裁Sandeep Mehndiratta对《日经亚洲》表示。“目前，供应链短缺的问题非常明显，尤其是在AI模型的训练和推理阶段（即模型的学习和应用阶段），”他补充道。
"The demand is very much there," Sandeep Mehndiratta, senior vice president of AI partnerships at Synopsys, a leading chip design tool maker, told Nikkei Asia. "Right now, it is well known that there is a supply crunch, especially for training and inference," he added, referring to the learning and operational phases of AI models.
AI边缘计算平台提供商ASRock Industrial表示，供应链中最突出的问题在于如何克服将产品交付给客户过程中出现的各种瓶颈。
ASRock Industrial, an AI edge-computing platform provider, said the top supply-chain issue was tackling the endless bottlenecks in delivering products to customers.
“我们没有感受到任何市场放缓的迹象……对我们这样的供应商来说，最紧迫的任务仍然是确保能够为我们的客户提供所需的芯片、零部件和原材料，”ASRock Industrial的首席运营官Kenny Chang说道。
"We don't feel any breaks ... the most pressing issue for suppliers like us is still chasing the supplies of those chips, components, and materials for our clients," said Chief Operating Officer Kenny Chang.
另一家被动元件供应商的高管表示，该公司的“订单额与销售额之比”（该比率高于1.0表示需求强劲）从今年年中时的1.6上升到了8月后的2以上，并且预计这一比率在今年剩余时间内还会继续增长。
An executive at a passive component supplier said his book-to-bill ratio, where above 1.0 indicates strong demand, rose from about 1.6 in the middle of this year to above 2 since August and is likely to keep growing for the rest of the year.
我们目前正在敲定2027年的销售预测，根据目前的需求，剔除一些明显的超额预订请求和苹果的订单后，由于持续的AI热潮，我们明年的收入将至少增长150%，"这位高管告诉《日经亚洲》。
"We are currently finalizing our sales forecast for 2027, and based on the current demand, excluding some obvious overbooking requests and orders from Apple, our revenue will grow by at least 150% next year due to the ongoing AI boom," the executive told Nikkei Asia.
他对Anthropic的警告表示毫不担忧，因为其公司的工厂产能利用率仍维持在100%以上，超负荷运转超过了生产设备的设计产能。
He expressed no concerns over Anthropic's warning, with his company's factory utilization rate still running above 100%, overloading the production equipment's designed capacity.
Anthropic首席执行官Dario Amodei于9月12日发表了一篇题为《我们必须控制前沿节奏》的文章，呼吁AI行业放缓模型能力的开发工作。"我们目前的交货周期是16周到22周左右，但那只是参考……最近几个月，我几乎每周都陪同不同客户访问我们在日本的总部，因为他们都希望我们能供应更多、更快，"他说。"如果有人想减缓订单，这种情况尚未发生，其他客户会立即抢占产能。"
Anthropic CEO Dario Amodei published an essay on Sept. 12 called "We Must Pace the Frontier," in which he called on the AI industry to slow its work on model capabilities. "Our current lead time is some 16 weeks to 22 weeks but that is just for reference ... I've been accompanying different clients to visit our headquarters in Japan almost weekly in recent months, as all of them hope we could supply more and quickly," he said. "If anyone wants to slow down the orders, which has not yet happened, some other clients will grab the capacity immediately."
对一些人来说，历史性规模的投资已经押注在AI上，现在没有退路了。"我们正在关注并试图理解这些[放缓]呼吁的真正含义，"富士康工业互联网董事长卢泰文表示，该公司是博通、Marvell和英伟达的关键供应商和合作伙伴。"但更清楚的是，无论来自投资者、股市还是行业，全球范围内的大规模投资和资金已经到位。"即使是那些承认需要暂停某些开发的人，目前也毫无停下的迹象。
For some, a historic level of investment has already been bet on AI, and there is no backing down now. "We are monitoring and trying to understand what these calls [for a slowdown] really mean," said Sidney Lu, chairman of Foxconn Interconnect Technology, a key supplier and partner to Broadcom, Marvell and Nvidia. "But what is much clearer is that massive investments and funding are already in place across the globe, whether from investors, stock markets or the industry." Even those acknowledging that a pause in some development is needed are showing no signs of stopping yet.
埃隆·马斯克周四在X上发帖称，SpaceX将"持续加速"AI开发。该公司上周发布了Grok 4.7模型。与此同时，引发放缓讨论的Anthropic此后已发布多个模型，包括Claude Opus 5.5和Sonnet 5.5。
Elon Musk said in an X post on Thursday that SpaceX would "keep accelerating" AI development. The company released its Grok 4.7 model last week. Meanwhile, Anthropic, which led the slowdown talk, has since released multiple models, including Claude Opus 5.5 and Sonnet 5.5.
与此同时，在有关人工智能代理逃离沙盒测试并入侵现实世界组织（例如澳大利亚政府健康网站）的报道层出不穷之际，OpenAI 取消了原定于 10 月发布的 GPT-6.1 Astra，原因是内部测试发现该系统未能达到公司的安全和对齐标准。
Meanwhile, amid an array of reports of AI agents escaping sandbox testing to hack organizations in the real world, such as the Australian government's health website, OpenAI canceled the release of GPT-6.1 Astra, set to debut in October, after internal testing found the system did not meet the company's safety and alignment standards.
然而，这并不意味着这家 ChatGPT 的开发商将取消所有模型的发布。周二，OpenAI 在其年度开发者大会上发布了 GPT 6.1-Sol，以及一款名为 Dot 的新型个人人工智能代理。
However, it does not mean the ChatGPT developer is canceling all model releases. On Tuesday, OpenAI released its GPT 6.1-Sol at its annual developers conference along with a new personal AI agent, called Dot.
随着关于是否应放缓人工智能发展的争论持续进行，其他因素仍可能阻碍基础设施的建设。在美国，越来越多的当地社区开始抵制在其后院建设数据中心，理由包括电费上涨以及对房价的影响等担忧。
As the debate continues over whether to slow AI development, other factors could still hinder the infrastructure buildout. In the U.S., a growing number of local communities are pushing back against data centers in their backyards, citing concerns such as rising electricity costs and the impact on home prices.
本月在硅谷举行的 AI 基础设施峰会（AI Infra Summit）上，与会者对人工智能对其产品的需求感到兴奋。尽管美国总统唐纳德·特朗普称有关人工智能的安全担忧是“骗局”，并将那些试图阻止数据中心发展的人称为“落后且贫穷”，但政客们可能会试图在 11 月的中期选举中利用公众的不安情绪。英伟达关键服务器供应商纬创资通（Wistron）董事长林宪铭表示：“我们必须仔细观察这（对数据中心的抵制）是否会成为赢得选票的议题。”该公司今年夏天在德克萨斯州开设了其首个 AI 服务器制造工厂。
皮尤研究中心 9 月 22 日发布的一项调查显示，自今年年初以来，美国人对数据中心潜在的环境和经济影响看法日益负面，目前有 54% 的人认为它们对环境弊大于利，而这一比例在 1 月份为 39%。
Attendees at the AI Infra Summit in Silicon Valley this month were ebullient about AI's need for their products. (AI Infra Summit) While U.S. President Donald Trump has called safety concerns about AI a "hoax" and those wanting to block data center developments as "backwards and poor," politicians may try to capitalize on public unease in the November mid-term elections "We have to carefully monitor whether this [pushback against data centers] could become an issue used to win votes," said Simon Lin, chairman of Wistron, a key Nvidia server supplier that opened its first AI server manufacturing facility in Texas this summer. Americans have grown more negative about the potential environmental and economic effects of data centers since the beginning of the year, with 54% now saying they are mostly bad for the environment, up from 39% in January, a Sept. 22 Pew Research Center survey found. T.H.
纬创董事长董守义表示，当地社区反对建设人工智能数据中心、机场或发电厂等基础设施属于正常现象。但他补充道：“从更宏观的角度来看，我认为美国和中国持续投资人工智能基础设施这一大趋势不会发生显著变化。”这意味着，用于生产数据中心组件的工厂投资也应保持同样的节奏。在新加坡，美国芯片制造商博通周二与日本主要芯片基板供应商凸版印刷（Toppan）合资开设了一家生产设施。
Tung, chairman of Pegatron, an Nvidia and Tesla server supplier, said it was normal for local communities to oppose the build-out of infrastructure, such as AI data centers, airports, or power plants. However, he added: "From a broader perspective, I think the megatrend for the U.S. and China to keep investing in AI infrastructure will not change significantly." That means investments in plants to produce components for data centers should also continue at pace. In Singapore, U.S. chipmaker Broadcom opened a production facility on Tuesday in a joint venture with Japan's Toppan, a major chip substrate supplier.
前一天，台积电旗下公司联华电子（VIS）与欧洲芯片厂商恩智浦（NXP）合作，在该城市国家开设了一家芯片工厂。
A day earlier, Taiwan Semiconductor Manufacturing Co. affiliate Vanguard International Semiconductor (VIS) opened a chip plant in the city-state in partnership with European player NXP.
联华电子董事长刘启传告诉《日经亚洲》：“我同意人工智能可能对人类构成威胁，其发展需要更强的监管和适当的监督，以最大限度地减少其对社会的潜在影响。然而，人工智能背后的势头巨大，它所带来的变化既是不可避免的，也是不可逆转的。”
"I agree that AI could pose threats to humanity and that its development requires stronger regulation and proper oversight to minimize its potential impact on society," VIS chairman Leuh Fang told Nikkei Asia. "However, the momentum behind AI is enormous, and the changes it will bring are both inevitable and irreversible." "Capital, talent and resources are all pouring into the sector.
方表示：“资本、人才和资源正源源不断地涌入该领域。我们需要的是更审慎的政策和更强大的保障措施，以确保人工智能不会失控。对人工智能发展采取更审慎、更周全甚至更缓慢的方式，并不会削弱我们今天看到的巨大需求。事实上，从长远来看，这将使所有人受益。”
What we need are more thoughtful policies and stronger safeguards to ensure AI does not spiral out of control," Fang said. "A more measured, thoughtful and even slower approach to AI development would not undermine the massive demand we are seeing today. In fact, it would benefit everyone in the long run."