主要AI实验室的财务有很多谜团，特别是，该公司最终披露更多关于计算经济学的信息，以及Anthropic暴涨的收入损失了多少来计算成本。
在周一的新存档中，Anthropic 发布数据显示计算时这些差值实际上越来越差。 上星期的申报是S-1 发布的更新，但当申报的财务披露仅涵盖到今年6月30日时，新数据将延长一个季度，即9月30日。 大致来说，Anthropic有一个很好的季度：收入达到170亿美元，比上一个季度的115亿美元有所增加，这表明持久的增长。 但运营损失同时增加，这在很大程度上要归功于计算成本的增长。
看着计算边际会使画面更糟糕。 在新季度，计算占了公司收入的58%，比前季度的48%有所增加。 当公司通过显示利润增加来赢得IPO投资者时
显示档案中一张图表更清楚地记录了公司每花费一美元计算的收入。 自2025年年初起
许多迹象表明，计算边际的这种变化，与其说是长期的问题，不如说是一个不成熟的问题，但是它仍然告诉我们有关经营一个大实验室的重要事务。 为了跟上需求，Anthropic的计算账单为本季度投资高达100亿美元，包括作为申报的指出，“短期承付款的计算额为50亿美元，目前的费用远远高于我们多年期承付款的可比计算额 ”。 基本上，需求增长如此之快，以至于Anthropic不得不拼命计算，这意味着支付更高的费率。
其中一些交易已经公开进行。 Anthropic一直支付SpaceX每月12.5亿美元进行计算，同时与Nscale,Akamai和Volta进行数十亿美元的交易。 在行业中，这些是新云而不是超规模的——更小的服装，以更高的费率提供更灵活的计算交易。
Anthropic的差值在Q3中恶化，正如新的申报所指出，“我们期望计算和基础设施费用以绝对美元计算，在可预见的未来，随着根据与社区支助方案的多年安排，我们可获得额外承付的计算，以及我们继续投资于额外计算，我们仍将是最大的业务费用。”
Anthropic仍在试图降低其计算成本。 特别是，该公司在所谓的“Anthropic Direct”项目下，谈到了它为建立自己的计算基础所作的努力。 但随着需求增长如此之快，这些努力将无法跟上步伐，新云将继续成为他们的最佳选择。
从某种意义上说，这些都是香槟问题。 只要收入以接近每季度50%的速度增长，投资人就不会担心高计算成本 — — 而当需求放缓时，投资人可能会因计算画面变得更阳光明媚而感到欣慰。 这提醒了人工智能实验室还有多少依赖于数万亿美元的基础设施建设，如果债券市场变得滑稽或当地反对情绪更加强烈，Anthropic无法应对其中任何一种风险
在此期间，Anthropic和其他实验室的未来将受计算价格的支配。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
There’s a lot of mystery around the finances of major AI labs, and as Anthropic inches towards its IPO, we’re starting to get some answers. In particular, the company is finally disclosing more about the economics of compute, and how much it of Anthropic’s skyrocketing revenue is being lost to compute costs.
In a new filing on Monday, Anthropic released data that showed those margins on compute actually getting worse. The filing was an update to the S-1 released last week — but where that filing’s financial disclosures only covered through June 30th of this year, the new data extended to disclose results through an extra quarter, to September 30th. Broadly, Anthropic had a good quarter: revenue reached $17 billion, up from $11.5 billion in the previous quarter, suggesting durable growth. But the operating loss grew at the same time, thanks in large part to growing compute costs.
Looking at the compute margins makes the picture even worse. In the new quarter, compute took up 58% of the company’s revenue, up from 48% the quarter before. At a moment when the company could win over IPO investors by showing margins improving, they’re getting worse.
A chart in the filing shows this even more clearly, tracking the company’s revenue for every dollar spent on compute. Since the start of 2025, it’s been steadily growing, peaking at just over 2x this past quarter, then falling back down to 1.72x
Many indications imply that this shift in compute margins is more of a blip than a long-term problem, but it still tells us important things about the business of running a major lab. Trying to keep up with demand, Anthropic’s compute bill shot up to $10 billion for the quarter, including, as the filing notes, “$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments.” Basically, demand was growing so fast that Anthropic had to scramble for compute, which meant paying higher rates.
Some of those deals have already played out in public. Anthropic has been paying SpaceX $1.25 billion a month for compute, alongside multi-billion dollar deals with Nscale, Akamai and Volta. In the parlance of the industry, these are neoclouds rather than hyperscalers — smaller outfits that offer more flexible compute deals at higher rates.
Anthropic’s margins got worse in Q3 because of deals like these, and there’s no sign they’ll be able to do without them anytime soon. As the new filing puts it, “We expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future as additional committed compute becomes available to us under multi-year arrangements with CSPs, and as we continue to invest in additional compute.”
Anthropic is still trying to lower its compute costs. In particular, the company talks quite a bit about its efforts to build its own compute base under a project it calls “Anthropic Direct.” But with demand growing so fast, those efforts aren’t going to be able to keep pace, and the neoclouds will continue to be the best option they have.
In a sense, these are champagne problems. As long as revenue is growing at nearly 50% a quarter, few investors will worry about high compute costs — and the moment demand slows down, investors may be comforted by a compute picture that has gotten a lot sunnier. But it’s a reminder of how much AI labs still depend on the trillion-dollar infrastructure buildout — and how vulnerable they are to hiccups in that process. It’s easy to imagine the data center buildout slowing down over the next year if bond markets get skittish or local opposition gets more intense. There’s not much Anthropic can do to address either of those risks, and either one would make a messy compute picture a whole lot messier.
In the meantime, the future of Anthropic and other labs will be at the mercy of compute prices.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.