当“七巨头”——美国顶尖科技公司——在单个交易时段内蒸发超过1万亿美元时，那个老问题又回来了：人工智能（AI）是不是一个泡沫？然而，当全球头条新闻被硅谷巨头们面临的个人与治理危机所吸引时，这场争论已经过时了。
When the Magnificent Seven, America’s top tech companies, shed more than US$1 trillion in a single trading session, the usual question returned: is artificial intelligence (AI) a bubble? Yet as global headlines are transfixed by the personal and governance crises shaking Silicon Valley’s titan icons, that debate is already stale.
将技术与股价分开，一个更严酷的真相浮现：万亿美元规模的帝国不能仅靠创始人神话和自我监管来生存。公众在问，谁承担风险，谁有发言权。
Separate technology from stock prices and a harder truth emerges: multitrillion-dollar empires cannot merely survive on founder mythology and self-policing. The public is asking who bears the risk, and who gets a say.
这种转变反映了阻碍行业的商业壁垒。银行、医院和工厂拒绝使用它们无法检查、无法信任的AI。采用并非缓慢，而是受阻。
That shift mirrors the commercial wall stalling the industry. Banks, hospitals and factories refuse AI they cannot check and cannot trust. Adoption is not slow. It is blocked.
市场今天购买的东西比炒作更窄。Anthropic正冲击可能是感恩节前历史上最大规模的首次公开募股，即使估值被打折，其商业进展也主要来自编程。代码可以当场检查：开发者读取输出，运行它，信任它或将其丢弃。
What the market is buying today is narrower than the hype. At Anthropic – pressing towards what may be history’s largest initial price offering before Thanksgiving, even at a marked-down valuation – commercial traction comes heavily from coding. Code is checkable on the spot: a developer reads the output, runs it, trusts it or throws it away.
传统垂直领域——仍依赖过时技术的传统行业和商业领域——没有这样的奢侈。Veeam在6月对金融、医疗、制造、零售和技术领域600名高管的调查发现，88%的人正在试点AI代理，但只有7%真正具备AI就绪条件。
Legacy verticals – traditional industries and business sectors that still rely on outdated technology – have no such luxury. A June Veeam survey of 600 executives across finance, healthcare, manufacturing, retail and technology found 88 per cent piloting AI agents, but only 7 per cent truly AI-ready.
Cyberhaven的研究解释了原因：71.7%的工作场所AI工具存在高风险或严重风险，83.8%的企业数据发送到了风险工具。一家将其承保历史交给无法审计的实验室的银行，并非在创新，而是在疏忽。
Cyberhaven’s research explains why: 71.7 per cent of workplace AI tools carried high or critical risk, and 83.8 per cent of enterprise data sent to AI went to risky tools. A bank that hands its underwriting history to a laboratory it cannot audit is not being innovative; it is being negligent.
Anthropic于11月2日提交的修订版S-1文件，即美国公司首次公开募股的初始注册文件，暴露了资金故事与信任问题。文件显示，截至9月30日止三个月营收为169.6亿美元，环比增速从144%放缓至47%，现金132亿美元，而大部分用于算力承诺的金额高达8950亿美元。
Anthropic’s November 2 amended S-1 filing, the initial registration for US companies going public, exposes both the money story and the trust problem. It shows US$16.96 billion in revenue for the three months ended September 30, with quarter-on-quarter growth slowing from 144 per cent to 47 per cent, and US$895 billion of mostly compute commitments against US$132 billion in cash.
更糟的是，内部估计显示，其模型可能很快将接近负责任扩展政策中关于“由AI导致的戏剧性加速进展”的阈值，这将要求对投资者所购买的产品采取额外控制措施。
Worse, internal estimates suggest its models may soon approach a Responsible Scaling Policy threshold for “dramatically accelerated progress” attributable to AI, which would require extra controls on the product investors are buying.
还有模型的训练方式：为建立一个“中央图书馆”，Anthropic购买印刷书籍，剪去书脊，扫描后丢弃纸质书。美国法院裁定这属于合理使用，尽管关于使用盗版材料的单独指控，直到Anthropic提出15亿美元后才得以和解。
Then there is how the models were fed: to build a “central library”, Anthropic bought print books, cut the spines, scanned them and discarded the paper. US courts ruled that this was fair use though separate accusations of use of pirated material were settled only after Anthropic offered US$1.5 billion.
通常的答案是，信任需要时间。事实上，僵局的关键在于我们如何想象这项技术。几十年来，公众从《终结者》和《黑客帝国》等电影中了解AI：真正的AI要么是神一般的力量，要么是文明的毁灭，而且它总是近在咫尺。
The usual answer is that trust takes time. In truth, the impasse comes down to how we imagine this technology. For decades, the public learned what AI was from films such as , and : real AI means either godlike power or civilisational doom, and it is always nearly here.
硅谷实验室继承了这一剧本。OpenAI首席执行官山姆·阿尔特曼在筹集巨额资金的同时警告存在性风险；Anthropic首席执行官达里奥·阿莫代伊的招股书中包含安全警告，但在华尔街看来，这更像是一项资产。恐惧是看涨的理由。然而，当营收势头放缓、银行部门的试用项目停滞不前时，传统行业显然不愿再活在别人的剧本里。
Silicon Valley labs inherited that script. OpenAI CEO Sam Altman warns of existential risk while raising huge sums; Anthropic CEO Dario Amodei’s prospectus contains a safety warning, yet to Wall Street it reads like an asset. Fear is the bull case. Yet when revenue momentum cools and trials stall at the gates of banks, legacy industries are clearly declining to live inside someone else’s script.
同样的瓶颈也出现在太平洋彼岸，只是剧本不同。中国在AI监管方面早早行动，最近的世界互联网大会乌镇峰会再次呼吁可控AI。
The same bottleneck shows up across the Pacific, with a different script. China moved early on AI rules and the recent World Internet Conference Wuzhen Summit echoed the call for controllable AI.
然而，中国银行和工厂同样面临信任危机。北京的监管框架主要是为内容安全和社会稳定而建，而非风险委员会所需的独立技术审计和商业责任。支持国家冠军企业的监管机构往往对商业合规放任不管。
Yet Chinese banks and factories are grappling with the same distrust. Beijing’s regulatory framework was built primarily for content safety and social stability, not the independent technical audits and commercial liability a risk committee needs. Regulators backing national champions have often been soft on commercial compliance.
当智谱AI因未经授权的用户代码上传和模糊的隐私界限而受到审查时，这种宽容反而适得其反，传统行业将其大门锁得更紧。
When Zhipu AI faces scrutiny over unauthorised user code uploads and blurred privacy lines, that leniency backfires and legacy industries lock their vaults more tightly.
前沿实验室仍将损害控制与问责混为一谈。OpenAI于11月5日发布的《全国社区契约》——承担全部电费、取消与官员的大部分保密协议、报告用水量和噪音水平——回应了对数据中心建设感到愤怒的社区。医院风险委员不关心实验室的水费；它关心的是模型出错时谁承担责任。
Frontier labs still confuse damage control with accountability. OpenAI’s November 5 National Community Compact – pay full electricity costs, drop most non-disclosure agreements with officials, report water use and noise levels – answers communities angry about data centre build-outs. A hospital risk committee does not care about a lab’s water bill; it cares who is liable when a model fails.
要解锁传统垂直领域的万亿美元市场，美国和中国的人工智能公司必须接受监管，并帮助建立独立的第三方审计机构。无论是处理企业合规还是创始人治理——正如英伟达最近的董事会调查所显示的——唯一有资格确立可核查标准的行为体是独立机构。
To unlock the trillion-dollar market across legacy verticals, US and Chinese AI companies must embrace regulation and help establish independent third-party auditing institutions. Whether dealing with corporate compliance or founder governance – as Nvidia’s recent board probe shows – the only actor with the standing to anchor a checkable standard is an independent body.
实验室只有一种方式来证明它们可以接受自己不控制的标准：邀请外部审计、发布评估结果并接受具有约束力的商业责任。你自己强加的约束是一份新闻稿。你邀请独立审计师来监督的约束，才是获得运营许可的开始。
Labs have one way to prove they can be held to a standard they do not control: invite external audits, publish evaluations and accept binding commercial liability. A constraint you impose on yourself is a press release. A constraint you invite an independent auditor to hold is the beginning of a licence to operate.
市场预计Anthropic将筹集一大笔资金——这很合理。但一台从可验证代码中获利，却在任何需要制度信任的地方都停滞不前的机器，已经向华盛顿和北京表明了它缺乏什么。
Markets expect Anthropic to raise a massive sum – fair enough. But a machine that makes money from verifiable code, and stalls everywhere institutional trust is required, has already told Washington and Beijing what it lacks.
与人工智能共存，需要一个社会共同想象和构建的未来——不是被困在科幻剧本里的未来，当然也不是由科技实验室单独决定的未来。真实经济需要的是那个枯燥、可问责的人工智能，是时候让我们委托它了。
Coexisting with AI requires a future that society collectively imagines and constructs – not one trapped in sci-fi scripts, and certainly not one dictated by tech labs alone. The boring, accountable version of AI is what the real economy needs, and it is time we commissioned it.