Anthropic的IPO招股书于周一早上发布，招股说明书读起来就像一位房地产经纪人拼命想把一座纸牌屋当作投资房产来推销：前景巨大。只是有几点披露你应该了解。当然，地基有裂缝；它位于洪泛区，而且你只有拿到那个晋升才真正买得起，但它最终有非零概率会成功，而且看看那景色。每份S-1的风险因素都读起来有点像末日预言。
Anthropic's IPO filing dropped on Monday morning, and the prospectus reads like a real estate agent desperately trying to sell a house of cards as an investment property: The upsides are enormous. There are just a few disclosures you should be aware of. Sure, the foundation is cracked; it's in a flood plain, and you can only really afford it if you land that promotion, but there's a non-zero chance it'll work out eventually, and just look at the views. Every S1's risk factors read a bit like a doomsday prophecy.
Anthropic，和任何其他将在公开市场交易的公司一样，必须为自己开脱。但这并不改变这样一个事实：该公司的成功建立在一场赌注上，即其收入增长足够快，以跟上其飙升的算力支出，而其义务在未来十年左右现在已超过半万亿美元。烧，宝贝，烧 根据S-1招股书的数据，乍一看，对“克劳德之父”来说情况似乎还不算太糟。
Anthropic, like any other company that's going to trade on public markets, has to cover its butt. It doesn't change the fact that the company's success is built on a gamble that its revenues will grow fast enough to keep up with its rampant compute spend, for which its obligations now exceed half a trillion dollars over the next decade or so. Burn, baby, burn Going off the S1 filing's numbers, things don't look that bad for the Claudefather at first blush.
2026年上半年的收入几乎是算力和基础设施成本的2倍，运营亏损相对较小，为6.93亿美元，而去年上半年为21.4亿美元。公司仍在亏钱，但事情终于朝着正确的方向发展。或者说，如果公司的算力支出预计在未来几年趋于平稳，情况就会如此。但事实并非如此。“我们开发、训练、改进和部署前沿模型的能力取决于我们能否获得巨大且不断增加的专用算力，”招股书写道。
Revenues in the first half of 2026 outpaced compute and infrastructure costs by nearly 2:1 and operating losses were a relatively paltry $693 million, compared with $2.14 billion in the first half of last year. The company is still burning money, but things are finally headed in the right direction. Or, they would be if the company's compute spend were forecast to plateau over the next few years. That's very much not the case. "Our ability to develop, train, improve, and deploy frontier models depends on our access to enormous and increasing quantities of specialized compute," the prospectus reads.
Anthropic模型的持续发展取决于越来越多的算力。Anthropic尚未就其涉及大型云服务商、新云和设备供应商的算力承诺给出确切数字，但该公司未来几年的义务可能高达8000多亿美元，具体取决于如何处理这些数字。
The continued development of Anthropic's models hinges on ever larger quantities of compute. Anthropic hasn't put a firm number on its compute commitments involving large cloud providers, neoclouds, and equipment providers, but the company's obligations over the next few years could be north of $800 billion, depending on how you wrangle the numbers.
根据S-1文件，Anthropic在2025年12月31日之后签订了新的或修订的算力和基础设施协议，未来最低付款义务约为4170亿美元。自2026年6月30日以来，该公司又签订了进一步的新的或修订的协议，未来最低付款义务约为2450亿美元，其中部分受合同条件约束。但如果将列出的所有合同加总，总额更高。微软：“至少449亿美元”——2026年至2033年 谷歌：“至少1145亿美元”——2026年至2033年 谷歌和其他提供商的TPU：“约1250亿至1500亿美元”——2028年至2031年 亚马逊：“至少1100亿美元”——2026年至2036年 SpaceXAI：“最高1205亿美元”——2026年至2029年 AMD：“超过”200亿美元——首期从2027年开始 Nscale：“最高约446亿美元”——至2033年 Lambda：“约413亿美元”——至2034年 博通：“约1612亿美元”——从2026年开始 总义务在7820亿至8070亿美元之间，尽管有些合同不如其他合同严密。
According to the S-1, Anthropic entered into new or amended compute and infrastructure agreements after December 31, 2025, carrying roughly $417 billion in future minimum payment obligations. Since June 30, 2026, the company has entered into further new or amended agreements carrying approximately $245 billion in future minimum payment obligations, some of which are subject to contractual conditions. But if you add up all the contracts listed, the total is even higher. Microsoft: "at least $44.9B" — 2026-2033 Google: "at least $114.5B" — 2026-2033 TPUs from Google and other providers: "approximately $125B to $150B"— 2028-2031 Amazon: "at least $110B" — 2026-2036 SpaceXAI: "up to $120.5B" — 2026-2029 AMD: "to exceed" $20B — initial tranche starting in 2027 Nscale: "up to approximately $44.6B" — through 2033 Lambda: "approximately $41.3B" — through 2034 Broadcom: "approximately $161.2B" — starting in 2026 That's a total obligation between $782 billion and $807 billion, though some aren't as airtight as others.
根据S-1文件，Anthropic与微软、谷歌、亚马逊和博通的协议均不可撤销。事实上，与微软、谷歌和亚马逊的合同规定，无论使用情况如何，均须付款。同样，Anthropic与Nscale和Lambda的合同分别延伸至2033年和2034年，其唯一退出方式是对方未能按时履行承诺或违约。Anthropic在其与埃隆·马斯克旗下SpaceXAI租用其位于孟菲斯的Colossus超级计算机算力的广泛报道的合作方面，确实拥有一定灵活性。
According to the S1, Anthropic's deals with Microsoft, Google, Amazon, and Broadcom are non-cancellable. In fact, the contracts with Microsoft, Google, and Amazon dictate payment regardless of usage. Similarly, Anthropic's only way out of its contracts with Nscale and Lambda, which extend through 2033 and 2034 respectively, is if they fail to deliver on their commitments on time or breach them. Anthropic does have some flexibility with regard to its widely reported collab with Elon Musk's SpaceXAI to lease capacity of its Colossus supercomputer in Memphis.
该协议价值约1200亿美元，有效期至2029年，双方自2027年初起可提前90天通知即可取消。与此同时，Anthropic与AMD的协议预计将花费超过200亿美元，用于首批算力容量，前提是其采购全额，但这可以通过从微软等公司租用算力来满足。
The deal, worth roughly $120 billion through 2029, can be canceled with just 90 days' notice by either party beginning in early 2027. Meanwhile, Anthropic's agreement with AMD is expected to cost more than $20 billion for the initial tranche of compute capacity, provided it procures the full amount, but this can be satisfied by leasing capacity from the likes of Microsoft.
而且，尽管Anthropic何时能够履行这些义务的时间表仍不明朗——目前仍有多个相关领域为空白——可以肯定的是，该公司的增长轨迹必须呈天文数字才能跟上其疯狂的支出，更别说实现盈利了。算力=收入，除非它并非如此。这种支出全部基于这样一种想法：更多算力等于更好、更受欢迎的模型，能够扩大客户群并支撑更高利润率的产品，这也是该公司现在向投资者表示将继续获取并修改算力合同的原因。
And while the timetable for Anthropic to meet those obligations remains unclear — several fields relevant to this remain blank at this time — it's safe to say the company's growth trajectory will need to be astronomical to keep up with its rampant spending, let alone ever turn a profit. Compute = revenues, except if it doesn't This spend is all predicated on the idea that more compute equals better, more desirable models that will grow the customer base and support higher-margin products, which is why the company is telling investors now that it's going to keepg and amending compute contracts.
按此速度，我们并不惊讶模型开发者到2040年的义务在今年年底前就会超过一万亿美元。Anthropic是否真的能够兑现这些承诺则是完全另一回事。
At this rate, we wouldn't be surprised if the model dev's obligations through 2040 exceeded a trillion dollars before year's end. Whether Anthropic will ever actually make good on those commitments is another matter entirely.
文件中写道：“我们能否满足这些对算力的重大且日益增长的财务义务，取决于我们实现显著收入增长的能力。”
"Our ability to meet the substantial and growing financial obligations of these requirements to secure compute depends on our ability to generate significant revenue growth," the filing reads.
如果Anthropic的收入跟不上其算力承诺，可能会导致违约，而公司指出，这可能引发连锁反应，使整个纸牌屋一起倒塌。Anthropic确实有一些后备措施，以防其自行支付账单时遇到困难。例如，公司在S1文件中指出：“如有需要，我们有能力向博通发行总额最高达420亿美元的可转换本票。”然而，这仅适用于Anthropic对博通的承诺。公司还可以通过股权或债务融资来偿还其他债务。但如果无法实现收入目标，公司警告称，这样做可能会稀释股份和/或以不利条件承担额外债务。
If Anthropic's revenues can't keep up with its compute commitments, it could end up defaulting, which it notes could result in a cascade that takes the entire house of cards down with it. Anthropic does have some backstops should it run into trouble paying its bills on its own. For example, "we have the ability to issue, if needed, up to $42.0 billion in aggregate principal amount of convertible promissory notes to Broadcom," the company's S1 notes. However, this only applies to Anthropic's commitments to Broadcom. The company could also turn to equity or debt financing to pay its other debts. But if it can't make revenue targets, the company warns that doing so risks diluting shares and/or taking on additional debt on unfavorable terms.
更糟的是，Anthropic最大的算力供应商不仅是其最大客户之一，还是其最大竞争对手。
My friend, my enemy, my patron Worse, Anthropic's biggest compute providers are not only some of its biggest customers, but also its biggest competitors.
我们说的是谷歌、亚马逊和微软等主要云服务提供商，它们规模远大得多，而且实际上是盈利的。这给Anthropic带来了一个不同的问题。云服务很方便，因为Anthropic不需要预先支付算力费用。但如果公司无法从谷歌、亚马逊或微软获得足够的算力，因为这些公司正在使用算力来训练或运行自己的或他人的模型，那么它们可能需要考虑替代方案，包括自建基础设施。
We're talking major cloud providers like Google, Amazon, and Microsoft, which are far larger and actually profitable. This poses a different problem for Anthropic. Cloud providers are convenient because Anthropic doesn't have to pay for the compute up front. But if the company can't get adequate compute from Google, Amazon, or Microsoft because they're using it for training or running their own or someone else's models, they may need to look at alternative options, including building out their own infrastructure.
“我们未来可能会以预付方式融资更大比例的算力，包括通过融资租赁、预付款、债务融资购买设备或类似安排，”该公司解释道。事实上，在某种程度上，Anthropic已经通过与博通的租赁协议在这样做。这似乎与今年早些时候谷歌和博通与Anthropic合作提供基于TPU的算力有关。据报道，Meta也曾就将谷歌TPU部署到其自身数据中心进行过谈判，所以这种安排并不罕见。
"We may in the future finance a greater portion of our compute on an up-front basis, including through capital leases, prepayments, debt-financed purchases of equipment, or similar arrangements," the company explained. In fact, to some degree Anthropic is already doing this through its lease deal with Broadcom. This appears to be tied to Google and Broadcom's collaboration with Anthropic to provide TPU-based compute earlier this year. Meta was also reportedly in talks to deploy Google TPUs in its own datacenters so the arrangement wouldn't be that unusual.
当然，还有外部威胁，例如南海冲突可能切断绝大多数AI加速器的供应链。只有当你相信收益巨大时，才能证明这种支出和风险是合理的。比如，人类历史上最大的经济转型。正如Anthropic所说：“我们相信，这种AI的丰富将以前所未有的速度改变全球经济的每一个领域，其影响将超过工业化、电力、半导体和互联网的影响。”®
And then of course there are external threats, like conflict in the South China Sea severing supply chains for the vast majority of AI accelerators. The only way to justify this kind of spend and risk is if you believe the upside is enormous. Like, the biggest economic transformation in the history of humankind. As Anthropic puts it: "We believe this abundance of AI will rapidly transform every sector of the global economy in ways greater than the impact of industrialization, electricity, semiconductors, and the internet." ®