翻阅Anthropic PBC为即将公开上市而发布的冗长招股说明书，人们不禁对这家年轻公司在技术和财务两方面取得的成就印象深刻。
Leafing through Anthropic PBC’s wordy prospectus for its imminent public listing, one cannot help but be impressed by this young company’s accomplishments — both technological and financial.
达里奥·阿莫代伊的这家AI公司（目前）还不是一些市场空头可能预期的那种烧钱机器。营收正在快速增长，考虑到在这个行业竞争所需的巨额资金投入，其盈利能力也算不错。
Dario Amodei’s artificial intelligence firm is not (yet) the cash incinerator some market bears may have been expecting. Revenue is growing at a rapid clip and profitability is decent when you consider the huge financial outlays required to compete in this industry.
对于正在考虑Anthropic在未来几周首次公开募股后约1.5万亿美元估值是否合理的潜在投资者来说，难点在于判断其稳健的盈利轨迹是否可持续——该公司正准备在算力上投入数千亿美元，以保持领先于竞争对手。
For prospective investors pondering whether Anthropic merits a roughly $1.5 trillion valuation after its initial public offering in the coming weeks, the difficulty is knowing whether its solid earnings trajectory is sustainable as it prepares to shell out hundreds of billions of dollars on computing power to stay ahead of the competition.
Anthropic在财务上保持保守的记录（用该公司自己的话说就是“大胆但理性”）为信心提供了依据。同样提供依据的还有，那些大手笔支出的企业客户喜爱其编程工具。该公司有一种令人欣喜的本事，能从相对有限的算力资源中榨取更多收入，使其资产负债表看起来相当稳健。截至6月底，该公司持有超过1300亿美元的现金和短期投资。
Anthropic’s track record of financial conservatism (“bold but sane” in its own words) is grounds for confidence. As is the fact that big-spending corporate customers love its coding tools. The company has the happy knack of squeezing more revenue from its relatively limited computing resources, leaving it with a robust-looking balance sheet. It held more than $130 billion of cash and short-term investments at the end of June.
但就连阿莫代伊也承认，预测其服务需求、并不得不为确保满足预期需求而获取极其昂贵的算力资源，这种难度堪称“地狱级”。危险在于：投入太少会给对手可乘之机，投入太多则会使自身财务健康陷入险境。
But even Amodei has acknowledged the “hellish” difficulty of forecasting demand for its services and having to secure the wildly expensive computing resources to meet that anticipated demand. The danger is investing too little and giving your rivals an opening, or spending too much and putting your financial health in peril.
这位Anthropic的掌门人将这一挑战称为“不确定性锥”。随着建设算力容量的成本飙升至每吉瓦约500亿美元，而且用户越来越担心自己的AI账单，这些风险只增不减。
The Anthropic boss refers to this challenge as the “cone of uncertainty.”These risks have only increased as the cost of building computing capacity has soared to about $50 billion per gigawatt and users fret more about their AI bills.
先说好消息。一家成立仅五年左右的公司，能像Anthropic那样在截至6月的三个月里实现116亿美元的季度营收，实属罕见。更令人印象深刻的是其8.4%的营业利润率。可以说，这些数字甚至可能低估了该公司的财务成功，因为它们包含了用于训练更先进AI模型的支出，而这些模型尚未贡献收入。
First, the goods news. It’s extremely rare for a business founded only around five years ago to bring in $11.6 billion of quarterly revenue, as Anthropic did in the three months to June. Even more impressive was its 8.4% operating profit margin. Arguably these figures may even underplay the company’s financial success, as they include its spending on the training of more advanced AI models that aren’t yet contributing to revenue.
诚然，Anthropic若不在前沿技术上大把烧钱，就难免被同行先驱OpenAI和谷歌超越。
It’s true, too, that Anthropic can hardly avoid splashing out on frontier tech without being leapfrogged by fellow pioneers OpenAI and Google.
然而，并非所有人都需要最先进的智能体。更便宜的开源权重模型——之所以叫这个名字，是因为任何人都可以免费拿它们做任何事——正威胁着限制Anthropic的市场份额，并让评估未来需求变得更加困难。许多商业用户可能满足于“够用”的AI工具，而不是阿莫代伊最新最伟大的作品。
And yet, not everybody needs the most advanced agents. Cheaper open-weight models — so called because they’re free for anyone to do what they like with them — are threatening to limit Anthropic’s market share, and making it even harder to assess future demand. Many business users may be happy with a “good-enough” AI tool, rather than Amodei’s latest-and-greatest creation.
因此，Anthropic的亮眼增长中，很可能有一部分——其中许多来自早期AI采用者——是不可持续的。招股说明书承认，其营收增长率预计将随时间推移而下降。
So it’s probable that some of Anthropic’s stellar growth, much of its from early AI adopters, isn’t sustainable. Its revenue growth rate is expected to decline over time, the prospectus acknowledges.
已经有迹象表明这一点正在发生，据报道，7月至9月期间的销售增长环比减速至仅45%。而第二季度则超过140%。
There are signs of that happening already, with sales growth in the July to September period reportedly decelerating to just 45% quarter on quarter. It was more than 140% in the second quarter.
到目前为止，Anthropic在没有大伤元气的情况下取得了很大成就。自成立以来筹集的约1410亿美元中，它似乎只烧掉了相对微不足道的一部分。最近另一宗大型科技IPO SpaceX也展现了同样令人印象深刻的资本效率。
Until now Anthropic has achieved a lot without breaking the bank. Of the $141 billion or so raised since inception, it appears to have burned through a relatively minuscule sum. SpaceX, the other recent mega-cap tech IPO, demonstrated a similarly impressive capital efficiency.
但埃隆·马斯克的公司近来已成为大手大脚的花钱者，因为它正转向人工智能，并在地球上（最终还将在太空）建设数据中心，外加一座为这些中心供应芯片的工厂。Anthropic 现在面临类似挑战，它试图确保算力容量，并摆脱此前仅从亚马逊等云服务商租用这些资源的做法。
But Elon Musk’s company has lately become a massive spender as it pivots to AI and constructs datacenters on earth (and eventually in space), plus a factory to supply them with chips. Anthropic now faces a similar challenge as it tries to secure computing capacity and departs from its previous practice of just renting these resources from cloud providers such as Amazon.com Inc.
未来，它还计划直接租赁由他人代为运营的数据中心。这些中心的装备成本不菲，不断上升的资本支出可能会侵蚀公司的现金流。令人担忧的是，部分算力将由成立不久、运营历史有限的数据中心运营商提供。Anthropic 的财务报表披露了数千亿美元的未来支出，用于租赁 AI 云算力、尚未开始的租赁合同，以及芯片等采购承诺。
In future it also plans to directly lease data centers operated on its behalf. These won’t be cheap to kit out and rising capital spending could dent the company’s cash flows. Worryingly, some of the capacity will be provided by newish data-center operators that have only a limited operating history. Anthropic’s financial statements disclose several hundred billion dollars of future spending for renting AI-cloud capacity, leases that are yet to start and purchase commitments for chips and such.
在这些各项支出中，它已同意每季度向 SpaceX 支付约 93 亿美元，以使用其数据中心。这将让 Anthropic 满足螺旋式上升的当前需求，但这些高得惊人的租金支出可能会挤压其利润率，除非收入能保持同样惊人的增速。作为参考，其第二季度的全部算力支出仅为 56 亿美元。
Among those various outlays, it has agreed to pay SpaceX some $9.3 billion per quarter to access its data centers. This will let Anthropic meet spiraling current demand, but these eye-wateringly high rental expenses could put pressure on its profit margins unless revenue keeps the same breathless pace. For reference, its entire computing spend in the second quarter was just $5.6 billion.
SpaceX 的算力安排至少可以提前 90 天通知取消，不过即使需求不及预期，Anthropic 可能仍需履行其众多其他支出承诺数年。
The SpaceX computing arrangements can at least be cancelled with 90 days notice, though Anthropic may have to keep honoring its many other spending promises for years even if demand doesn’t live up to its expectations.
阿莫代伊并非对这些陷阱视而不见。在其算力需求更加确定之前，公司不会在数据中心的先进芯片上投入资金，而这正是成本最大的组成部分。此外，若高估需求，它将尝试转嫁多余算力。问题在于，在全行业低迷的情况下，这样做有多大可行性。
Amodei isn’t oblivious to these pitfalls. His company won’t spend on advanced chips for its data centers until its computing needs are more certain, and these are by far the largest component cost. Furthermore, it will try to offload excess capacity in the event that it overestimates demand. The question remains on how practical that would be in an industry-wide slump.
“人工智能指数级发展所带来的极端上行可能性，本质上伴随着一定程度的风险，”Anthropic 创始人在其首次公开募股（IPO）招股说明书的致投资人信中如此阐述投资逻辑。换言之，我们或许即将改变人类的命运，但我们正试图负责任地行事：投资者自负。
“Some level of risk is inherent to the extreme upside possibilities offered by the AI exponential,” is how the Anthropic founders describe the investment case in their IPO cover letter. In other words, we may be about to alter humanity’s fate but we’re trying to be responsible here: Buyer beware.