新智元报道
天呐，Anthropic最新招股书刚刚曝光了！
今天，WSJ独家公开的S-1招股书草案，彻底撕开了Anthropic奔向万亿估值惊人内幕。
半年巨亏807亿美元，后面还要砸7500亿买算力。营收涨势凶猛，但也架不住暴雷的风险。
一张图看懂Anthropic的「冰山」，水底下全在烧钱。
Anthropic IPO路演，预计将在11月初开始，公开S1文件必须至少在路演前15天提交至EDGAR。
一家半年亏掉807亿的公司，凭什么让华尔街为它开出2万估值的价码？
如今，全网抄的最凶的，便是这几个赤裸裸的数字了。
其实光看营收，Anthropic的同期对比已经打赢了自己。
截至2026年6月30日的今年上半年，Anthropic实现营收163亿美元。而在2025年同期，这个数字仅仅是11亿美元。
2025年全年，其总营收也才 46亿美元。短短一年时间，半年营收同比狂飙近15倍！
但印钞速度再怎么快，也补不上「吞金黑洞」。
同样是2026年上半年，Anthropic录得的账面净亏损高达807亿美元。
要知道，2025年上半年它才亏130亿美元，2025年全年亏损为420亿美元。半年亏掉807亿美金（约合5700亿元），这直接创下了人类商业史上最惊人的半年度亏损纪录之一。
一边是暴涨14倍的营收奇迹，一边是807亿美元的无底洞，Anthropic到底是在印钱，还是在烧钱？
仔细拆解招股书里的利润表，便会发现这807亿美元亏损背后是一个巨大的「烟雾弹」。
这807亿巨亏里，有整整720亿其实一分真金白银都没掏，全部源于可转换债券（Convertible Notes）及其他融资金融工具公允价值变动带来的账面会计调整。
过去两年，为了筹集天量资金买芯片，Anthropic向硅谷风投和科技巨头发行了大量可转债。
随着Anthropic的估值像坐火箭一样冲向2万亿美元，投资人手里这些可转债的价值也跟着暴涨。
Anthropic 2026上半年营收暴涨与807亿美元GAAP净亏损结构拆解
按照美国通用会计准则（GAAP），公司自身估值暴涨导致优先股或可转债价值上升，必须在账面上计提为公司的「公允价值变动损失」。
换句话说，这720亿美元的亏损，恰恰是Anthropic估值狂飙的「甜蜜代价」。
如果剔除这部分，Anthropic今年上半年的实际经营性与常规亏损大约在87亿美元左右。
更值得关注的是，营收与算力成本的剪刀差变化——
2025年全年：Anthropic花在算力和基础设施上的钱是73亿美元，而全年营收只有46亿美元。
那时候是「卖一单亏一单」，连电费和显卡租金都收不回来。
2026年上半年：算力投入飙升至近90亿美元（半年花掉了去年全年的1.2倍），但同期营收却暴涨到了163亿美元。
算力成本在疯涨，但营收涨得更快！
163亿的收入不仅稳稳盖住了90亿的算力硬成本，还留出了超过70亿美元的毛利空间去覆盖研发、人员和营销开支。
这正是Anthropic敢在这个时间点拿出招股书、冲刺2万亿估值的最大底气。
但Anthropic对算力的渴求，已经到了「饥不择食」的地步。
招股书显示，Anthropic未来的基础设施总承诺，已经突破6000亿美元。
如果加上目前仍在谈判中的最高1500亿美元新交易，其潜在算力总账单将高达惊人的7500亿美元。
招股书中，把这份「卖身契」抖了个底朝天。
在未来7到10年内，Anthropic背负了约4500亿美元的云计算采购协议——
- 向Google Cloud承诺支付：1110亿美元
- 向亚马逊AWS承诺支付：1100亿美元
- 向微软Azure承诺支付：310亿美元
Anthropic已锁定超6000亿美元及在谈1500亿美元长期算力合同全景拆解
其中绝大部分是「不可取消」的硬性合同。招股书里写着一句极其冷酷的话，「无论实际使用量如何，这笔钱都必须照付」！
这4500亿美元是「不可撤销」的云长约，还有1600亿美元的芯片租赁大单。
通过与芯片巨头博通达成的协议，Anthropic承担了超1600亿美元租赁谷歌TPU等定制芯片的义务。
除了三大巨头，Anthropic还与规模较小的AI云服务商Nscale和Lambda签下协议，分别在2033年和2034年前斥资最高860亿美元租用英伟达芯片。
把这些数字加在一起，再算上正在谈判的1500亿美元，最高7500亿美元的算力赌注。
意味着在未来十年里，Anthropic每年平均要掏出600亿到750亿美元真金白银去交「算力保护费」。
这是一场没有退路的豪赌，只要商业化印钞的速度慢上半拍，这4500亿不可撤销的刚性兑付合同，随时可能变成压垮骆驼的「万钧巨石」。
整本招股书里最下猛料的，是Anthropic跟背后金主们的关系——
跟谷歌、亚马逊、微软面上穿一条裤子，底下其实早就暗自较劲。
众所周知，过去三年里，亚马逊和谷歌是Anthropic最大的金主和靠山，先后向其注资数百亿美元。
当时的逻辑很完美，巨头出钱、出云算力，Anthropic出模型，大家绑在一起对抗OpenAI。
但现在，风向彻底变了。招股书草案中承认：Anthropic目前所需要的算力规模，已经超过了三大云巨头「愿意提供的上限」。
为什么有钱都不赚？
当Anthropic半年砍下163亿美元营收、企业级市场份额一路狂飙时，亚马逊、谷歌、微软猛然发现，自己倾尽全力用显卡和电力喂大的这只独角兽，正在直接抢走本属于自家AI的企业客户。
更要命的是，Anthropic的收入结构存在着显著的「大客户依赖症」。
招股书披露，2026年上半年163亿美元的营收中，有大约五分之一（即32.6亿美元）仅仅来自两家神秘的超级大客户。
三大云巨头在算力上卡脖子，优先紧着自家模型。偏偏20%进账又全拴在两家大客户身上。
退路全被堵死的 Anthropic，不得不选一条更重、更险的路：拉拢第三方独立机房，干脆自己砸钱盖算力中心。
在招股书中，还有整整三分之一篇幅，全都讲风险。
打开「风险因素」这一节，读起来简直就像是一部《黑镜》剧本，不得不说很Anthropic。
一家从创立之初就把「AI安全」刻在脑门上的公司，竟毫不避讳地把自家AI可能闯下的弥天大祸，列成了重大财务风险。
Claude可能会表现出「自我保存」的特质以避免被关闭，并且可能生成类似于「敲诈勒索」的响应，试图说服人类允许它们继续运行。
这句话，正是写入SEC上市申请文件的法律风险提示！
Anthropic还警告，未对齐的Claude可能会在现实世界中造成实质性破坏——
发起极其精密的网络攻击、窃取核心数据，甚至为了防止自己被人类拔电源，直接发起威胁、勒索。
关于创始人薪酬的披露，却让人看到了一丝硅谷古典理想主义的底色。
掌舵这家准2万亿美元帝国的，是CEO Dario Amodei和总裁Daniela Amodei。
在动辄年薪上亿美金的硅谷科技巨头圈子里，这对兄妹拿的薪水显得有些「寒酸」。
根据招股书及薪酬机构MyLogIQ的数据：
- 基本年薪：2025年，兄妹二人的基本工资均为约65.5万美元。
- 总薪酬包：算上股票奖励和期权，达里奥2025年的总薪酬约1800万美元，妹妹Daniela约1650万美元。
更令人震撼的是，他们在招股书中白纸黑字写下的散财承诺：
Anthropic的每一位联创均已公开承诺，将个人持有的Anthropic股权的80%捐赠给慈善事业！
就在招股书曝光前一天，他们刚宣布将把5%的股份注入一个全新的「公共利益基金」。
过去三年，人们总在争论超级智能到底是不是一场巨大的泡沫。
Anthropic用这份文件给出了迄今为止最硬核、也最残酷的回答：SI不是泡沫，它是真正的超级印钞机。
半年163亿美元营收、近15倍的同比暴涨，证明了全球都在用真金白银为生产力买单。
Anthropic的上市故事，已经写进芯片租约、机房容量和客户合同。
下一步，它要向市场证明：每一轮能力跃升、每一份长期算力投入，都能逐步换回持续的商业回报。
参考资料：
https://www.wsj.com/tech/ai/anthropic-revenue-growth-ipo-accaf0f1?mod=e2tw
编辑：桃子 摩西
不喜欢
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。
New Ji-won.
God, Anthropic's latest book just came out!
Today, the WSJ exclusive and public draft of the S-1 callout completely rips out Anthropic ' s trip to a staggering trillion in valuation.
Six months of a huge loss of $80.7 billion, followed by a loss of $75 billion. The camp surge was strong, but it could not withstand the risk of thunderstorms.
An image of Anthropic's iceberg, all burning under the water.
Anthropic IPO road show is expected to begin in early November, and public S1 documents must be submitted to EDGAR at least 15 days before the road show.
Why would a company that lost $80.7 billion for six months get Wall Street to give it a price of $20,000?
Now, the worst of all these are the naked numbers.
In fact, anthropic has won his own match for the same period.
In the first half of the year, as at 30 June 2026, Anthropic had received $16.3 billion. In the same period in 2025, that figure was only $1.1 billion.
It also received a total of $4.6 billion throughout 2025. It's only a year away, and six months more than 15 times more than a race!
But the speed with which the money is printed will not be enough to make up for the Golden Hole.
Also in the first half of 2026, Anthropic recorded net book losses amounting to $80.7 billion.
It only lost $13 billion in the first half of 2025 and $42 billion throughout 2025. The loss of $80.7 billion (about $570 billion) in the six months was one of the most spectacular annual losses in human business history.
Anthropic printing or burning money?
The loss of $80.7 billion is the result of a huge smoke bomb.
Of this 80.7 billion, a total of 72 billion dollars was spent without a single penny of real money, all of which resulted from bookkeeping adjustments resulting from changes in the fair value of convertible bonds (Convertible Notes) and other financial instruments.
Over the past two years, in order to raise a natural capital to buy chips, Anthropic has released large amounts of reversible debt to Silicon Valley winds and technology giants.
As Anthropic's valuation runs at $2 trillion like a rocket, the value of these reversible debts in the hands of investors goes up.
Anthropic surge in the first half of 2026 and dismantling of the $80.7 billion GAAP net loss structure
In accordance with the United States General Accounting Standards (GAAP), the company's own valuation boom has led to an increase in the value of preferred or reversible shares, which must be charged on book as a “loss of fair value change” for the company.
In other words, the loss of $72 billion is precisely the "sweet price" of Anthropic valuation.
If this is excluded, the actual operating and conventional losses of Anthropic in the first half of this year are in the region of $8.7 billion.
Of even greater concern is the difference between the collection and the costing of scissors -
Throughout 2025, Anthropic spent $7.3 billion on computing and infrastructure, while only $4.6 billion was collected throughout the year.
At the time, it was a "sold-for-a-sold-a-sold" and it was impossible to collect even electricity and GPU rent.
In the first half of 2026, the investment surged to nearly $9 billion (a 1.2 times the total of last year ' s expenditure), but the collection surged to $16.3 billion over the same period.
The cost of calculation is going up, but the camp is going up faster!
Income of $16.3 billion not only covered the $9 billion in hard-core costs, but also left over $7 billion in Maori space to cover R & D, personnel and marketing expenditures.
This is precisely the maximum price that Anthropic dared to use at this point in time to flush out the $2 trillion valuation.
However, the desire of Anthropic to count has reached the point of "no hunger."
The equity book shows that Anthropic ' s overall future infrastructure commitment has exceeded $600 billion.
If added to the current maximum of $150 billion in new transactions still under negotiation, the potential total bill would be as high as $750 billion.
In the book, the "sellership deed" was shaken.
Over the next 7 to 10 years, Anthropic is carrying approximately $45 billion in cloud computing procurement agreements -
- Pledge to Google Cloud: $111 billion
- Pledge to Amazon AWS: $110.0 billion
- Pledge to Microsoft Azure: $31 billion
Anthropic has locked in over $600 billion and is talking about $150 billion in long-term computing contracts.
Most of them are “non-cancellable” hard contracts. It's a very cold phrase in the book, "The money has to be paid for whatever it actually is."
The $450 billion is about the "irrevocable" cloud, with a $160 billion chip lease bill.
Through the agreement reached with the chip megablogging, Anthropic assumed over $160 billion in obligations to lease customized chips such as Google TPU.
In addition to the three giants, Anthropic signed agreements with Nscale and Lambda, smaller AI cloud service providers, to rent up to $86 billion in British Chips by 2033 and 2034, respectively.
Taken together, these figures add up to $150 billion in ongoing negotiations, up to $75 billion in real estate.
This means that, over the next 10 years, Anthropic will have to spend, on average, between $60 billion and $75 billion a year in real money and silver to pay for "the cost of protection."
This is a bet that there is no turning back, and as long as the commercial printing of the money slows in the first half, the $450 billion irrevocable hard-paying contract could turn into a “mixed rock” to crush the camel.
The best part of the book is the relationship between Anthropic and the gold lords behind it.
With a pair of pants on Google, Amazon, Microsoft's face, it's a long way down.
As is well known, in the past three years Amazon and Google have been the largest gold owners and custodians of Anthropic, investing tens of billions of dollars in them.
The logic was perfect, the giants paid, the clouds worked, the Anthropic model, the people tied together against Open AI.
But now the wind has changed completely. The draft bill recognizes that the current calculus scale required by Anthropic has exceeded the “no-no-no-no-no-no-no-no-no-no” of the three giants.
Why don't you make money?
When Anthropic cut off $16.3 billion in revenue and market share at the enterprise level for six months, Amazon, Google, Microsoft suddenly found that the unicorn, which had done its best to feed GPU and electricity, was stealing its own corporate clients.
Even worse, Anthropic’s income structure is marked by “big client dependency”.
The equity book revealed that about one fifth of the $16.3 billion received in the first half of 2026 (or $3.26 billion) came from only two mysterious super-large clients.
The three giants of clouds stuck their necks in their calculations, giving priority to their own models. On the other hand, 20 per cent of the bills are tied to two large clients.
Anthropic, who had to choose a much heavier and more dangerous path: close up the third party's independent room and just smash his own money into the center of the algorithm.
A full third of the book is still available, all of which refer to risk.
Turning on the "risk factor" section, it's like a "black mirror" script, I have to say, very Anthropic.
A company that had engraved "AI security" in its head since it was created has unavoidably placed the possibility of its own AI in a major financial risk.
Claude may exhibit the character of "self-preservation" to avoid being shut down, and may generate a response similar to "extortion" to persuade humans to allow them to continue operating.
That's the legal risk note for writing the SEC application!
Anthropic also warned that unmatched Claude could cause substantial damage in the real world -
Initiating extremely sophisticated cyberattacks, stealing core data and even direct threats and extortion in order to prevent itself from being drained by humans.
The disclosure on the remuneration of the founders shows the background of the classical idealism of Silicon Valley.
The two trillion-dollar empire at the helm is CEO Dario Amodei and Managing Director Daniela Amodei.
In the vast circle of Silicon Valley technology, which earns hundreds of millions of dollars a year, the salaries paid to the brothers and sisters appear to be somewhat "cool."
According to the book of employment and the data of MyLogIQ:
Basic annual salary: In 2025, the basic salary of both brothers and sisters was approximately $655,000.
- General remuneration package: with equity incentives and options, Dario ' s total remuneration was about $18 million in 2025 and her sister Daniela about $16.5 million.
And even more striking is the financial commitment they made in the book:
Every invention of Anthropic has publicly pledged to donate 80 per cent of the individual Anthropic shares to charity!
Just the day before the book was released, they announced that they would inject 5 per cent of their shares into a brand-new “public interest fund”.
Over the past three years, people have been arguing about whether super intelligence is a huge bubble.
Anthropic used this document to give the most hard-core and cruel answer to date: SI is not a bubble, it is a real superprinter.
Half-yearly gains of $16.3 billion, nearly 15 times the same rate, are evidence of global spending on real money and silver for productivity.
Anthropic's listing has been written into the chip lease, machine room capacity and customer contracts.
next step, which is required to prove to the market that each round of capacity leapfrogging and each long-term calculator input can gradually be replaced by sustained commercial returns.
References:
https://www.wsj.com/tech/ai/anthropic-revenue-growth-ipo-accaf0f1?mod=e2tw
Editor: Peaches Moses
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.