The SEC Should Block Anthropic's IPO to Protect the Public

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Matt Scherer is co-author of the recent report Brace for Impact: Protecting people and the economy when the AI bubble bursts.

马特·谢勒（Matt Scherer）是近期报告《做好冲击准备：当人工智能泡沫破裂时保护民众与经济》（Brace for Impact: Protecting people and the economy when the AI bubble bursts）的合著者。

Anthropic CEO Dario Amodei looks on as United States President Donald Trump holds a press briefing with AI executives following a meeting on artificial intelligence, outside the West Wing of the White House on September 29, 2026 in Washington, DC. (Photo by Kevin Dietsch/Getty Images)

2026年9月29日，美国总统唐纳德·特朗普在华盛顿特区白宫西翼外举行人工智能会议后，与人工智能高管召开新闻发布会，Anthropic首席执行官达里奥·阿莫代（Dario Amodei）在旁观看。（图片来源：Kevin Dietsch/Getty Images）

Things are seldom what they seem when it comes to the finances of the AI industry. The AI ecosystem’s circular business dealings, extensive use of financial engineering, and otherwise murky finances make it hard for investors, regulators, or workers saving for retirement to get an accurate financial picture of AI companies.

在人工智能行业的财务状况方面，事情往往表里不一。人工智能生态系统中循环的商业交易、对金融工程的广泛使用以及其他模糊的财务状况，使得投资者、监管机构或为退休储蓄的员工很难获得人工智能公司准确的财务状况。

That is particularly true of Anthropic, which filed the registration statement for its initial public offering (IPO) last week ahead of a planned November stock market debut.

对于Anthropic公司而言尤其如此，该公司上周提交了首次公开募股（IPO）的注册声明，计划于11月在股市首次亮相。

The company’s topline financial figures look quite impressive for a startup, with revenues that appear to have grown quickly over the course of 2026 and a fairly healthy balance sheet — on the surface, that is.

对于一家初创公司来说，该公司的营收数据看起来相当令人印象深刻，2026年期间收入似乎增长迅速，资产负债表也相当健康——当然，这只是表面现象。

Dig a little deeper, however, and a far bleaker picture emerges. Footnotes to the company’s most recent financial disclosures revealed nearly $900 billion in spending commitments, most of them stretching into the 2030s. The company’s financial results for 2026 have not been subjected to formal auditing, which makes them less-than-reliable, but the third-quarter results suggest that the company’s revenue growth slowed to a crawl over the summer.

然而，深入挖掘后，一幅更为惨淡的图景浮现出来。该公司最新财务披露的脚注显示，其支出承诺近9000亿美元，其中大部分延续至2030年代。该公司2026年的财务业绩尚未经过正式审计，这使其可靠性大打折扣，但第三季度业绩显示，该公司夏季的收入增长已放缓至停滞状态。

Anthropic will have to reverse that trend and sustain extraordinary revenue growth to meet its vast obligations. But the financial pressures that are squeezing Anthropic — and, indeed, the AI industry as a whole — will make it exceptionally hard for the leading model developer to do so. If the company goes public, workers, retail investors, and retirees will increasingly be left holding the bag when Anthropic’s bills come due.

Anthropic必须扭转这一趋势并保持非凡的收入增长，才能履行其庞大的债务。但正在挤压Anthropic——乃至整个AI行业——的财务压力，将使这家领先的模型开发商极难做到这一点。如果该公司上市，当Anthropic的账单到期时，员工、散户投资者和退休人员将越来越可能成为最终的买单者。

The financial mountain that Anthropic must climb is unprecedented for a company making its stock market debut. While Anthropic has an impressive $96 billion in cash reserves, its amended registration statement reveals at least $582 billion in non-cancelable spending commitments to just five companies (Amazon, AMD, Broadcom, Google, and Microsoft) to gain access to the vast computational power required to train and run Anthropic’s generative AI models.

对于一家首次亮相股市的公司来说，Anthropic必须攀登的财务高峰是前所未有的。虽然Anthropic拥有令人印象深刻的960亿美元现金储备，但其修订后的注册声明显示，为了获得训练和运行其生成式AI模型所需的巨大计算能力，该公司仅对五家公司（亚马逊、AMD、博通、谷歌和微软）就有至少5820亿美元的不可撤销支出承诺。

Even though they do not appear on Anthropic’s balance sheet, those obligations are effectively debts in a legal and practical sense: they are payments that Anthropic is contractually required to make in the future, regardless of whether the company can afford to do so.

尽管这些义务没有出现在Anthropic的资产负债表上，但在法律和实践意义上，它们实际上就是债务：无论公司是否负担得起，这些都是Anthropic在合同上被要求在未来支付的款项。

Moreover, it is not clear how much of the remaining $300 billion or so of the $895 billion in contractual commitments will ultimately land on Anthropic’s balance sheet. With the company continuing to negotiate further massive infrastructure and compute contracts, Anthropic may well end the year with a trillion dollars or more in contractual commitments.

此外，在8950亿美元的合同承诺中，剩余的约3000亿美元最终会有多少落入Anthropic的资产负债表尚不清楚。随着该公司继续谈判更多大规模的基础设施和计算合同，Anthropic到今年年底的合同承诺额很可能达到一万亿美元或更多。

Generating the revenues necessary to meet those obligations would be, to say the least, quite difficult for any company. But the margins of AI model developers are being squeezed by several forces that will make that task extraordinarily challenging for Anthropic.

对于任何公司来说，产生履行这些义务所需的收入，至少可以说是非常困难的。但人工智能模型开发商的利润率正受到多种力量的挤压，这将使Anthropic完成这项任务变得异常艰巨。

First, it isn’t clear that there is nearly enough demand for generative AI products and services to justify Anthropic’s spending. Multiple studies, including from organizations notable for their bullishness on AI, suggest LLMs are not accelerating productivity or otherwise generating positive returns on investment for most businesses. Worryingly for enterprise-focused Anthropic, one study found that the number of Americans using AI at work fell significantly between the summers of 2025 (46%) and 2026 (33%).

首先，目前尚不清楚生成式AI产品和服务的需求是否足以证明Anthropic的支出是合理的。多项研究（包括来自以看好AI而闻名的机构的研究）表明，大语言模型（LLM）并没有提高生产力，也没有为大多数企业带来正向的投资回报。对于专注于企业的Anthropic来说，令人担忧的是，一项研究发现，在2025年夏季（46%）到2026年夏季（33%）之间，在工作中通过AI辅助的美国人数量显著下降。

Anthropic’s spending commitments suggest that the company believes it will be able to convince businesses to pay for ever-larger and more expensive models. But there are growing signs that most businesses are not interested in paying for more powerful and expensive models. When Anthropic released its much-hyped Fable 5 frontier model this summer, its commercial reception was tepid.

Anthropic的支出承诺表明，该公司认为它能够说服企业为规模更大、价格更昂贵的模型买单。但越来越多的迹象表明，大多数企业对支付更强大、更昂贵的模型并不感兴趣。当Anthropic今年夏天发布备受炒作的Fable 5前沿模型时，其商业反响平平。

You have successfully joined our subscriber list.

您已成功加入我们的订阅列表。

Even if a future Anthropic model captivates businesses, the company would run into a second obstacle: the AI industry’s frenzied competition. Anthropic is part of a crowded Silicon Valley ecosystem where numerous well-heeled corporations (including Google, Meta, Microsoft, OpenAI, and SpaceX, along with Anthropic) offer generative AI models that are largely indistinguishable in their quality and capabilities. As a result, AI companies can quickly offer competing products whenever a rival identifies a use case that shows even a hint of commercial potential.

即使未来Anthropic的模型能吸引企业，该公司也会遇到第二个障碍：人工智能行业疯狂的竞争。Anthropic是硅谷拥挤生态系统的一部分，众多财力雄厚的公司（包括谷歌、Meta、微软、OpenAI和SpaceX，以及Anthropic）都在提供在质量和能力上几乎没有区别的生成式AI模型。因此，每当竞争对手发现一个表现出哪怕一丝商业潜力的用例时，人工智能公司都能迅速提供竞争产品。

In such an environment, companies can rarely afford to raise prices because doing so will prompt customers simply to switch to a rival. The rush to grab customers ahead of their IPOs led Anthropic and OpenAI to launch a price war this summer that remains ongoing. Competition from Chinese companies offering open-weights models puts further downward pressure on prices, as does the increasing availability of models that can run locally on phones and laptops — no expensive data centers required.

在这种环境下，公司几乎无法承受提价，因为这样做只会促使客户转向竞争对手。在IPO前争夺客户的冲动导致Anthropic和OpenAI在今年夏天发动了一场价格战，且目前仍在持续。提供开放权重模型的中国公司带来的竞争进一步压低了价格，可以在手机和笔记本电脑上本地运行（无需昂贵的数据中心）的模型日益普及也产生了同样的影响。

That leads to the third problem: the extraordinarily high costs of the chips and data centers needed to train and run AI models, which put powerful and continuous upward pressure on Anthropic’s costs. Between the chips, servers, networking equipment, electricity and cooling infrastructure, construction costs, and other expenses, a 1 gigawatt (GW) data center costs an estimated $31 billion, and the prices for some components are soaring.

这就引出了第三个问题：训练和运行人工智能模型所需的芯片和数据中心成本极高，这给Anthropic的成本带来了强大且持续的上涨压力。在芯片、服务器、网络设备、电力和冷却基础设施、建设成本和其他费用之间，一个1吉瓦（GW）的数据中心的成本估计为310亿美元，而且一些组件的价格正在飙升。

That is for 1 GW. Anthropic has committed to 15 GW of data center capacity in the coming years and the company estimates it will ultimately need 20-25 GW, which explains the company’s

这还只是1吉瓦的情况。Anthropic已承诺在未来几年内建设15吉瓦的数据中心容量，该公司估计最终将需要20-25吉瓦，这解释了该公司为何……