你看到807亿净亏损，第一反应大概是这公司要完。
807亿拆开来：724亿是「估值税」，不是现金流
可转换票据与SAFE的公允价值机制
Anthropic 早期多轮融资大量使用可转换票据和 SAFE，这类工具内嵌转股权。US GAAP 要求每个报告期对嵌入衍生工具按公允价值重新计量。估值从 2024 年的数十亿美元跳到 Series H 的 9650 亿美元 post-money，早期那张低价转股权骤升，会计上直接记进利润表，科目叫「可转换票据、衍生负债及 SAFE 的公允价值变动」。钱没动，数字动了。
公允价值变动如何变成账面亏损
估值越高、账面亏损越大
807.48 亿里 723.88 亿是这笔公允价值变动，占 89.6%。剩下约 84 亿才是经营层面的真实缺口。公司越成功、估值越高，这张报表上的「亏损」越大。反直觉，但逻辑自洽：你在为别人手里的期权付账面价格。
九成是估值涨出来的，不是烧出来的
163亿营收的真实增速：从4100万到13倍
2023→2026四年跳跃的时间线
S-1里有一条干净的营收阶梯。2023年Claude商业化元年，全年4100万美元；2024年3.86亿；2025年45.85亿；2026年上半年162.86亿，去年同期只有11.63亿，同比增长约1300%。
四个数字摆出来，曲线形状不用画
更刺眼的是单季拆分。2026年Q2营收115.6亿美元，一个季度超过2024年全年近30倍。S-1把这个数字单独列出来，等于告诉投资者：增长不是匀速爬坡，是指数型加速。
从4100万到163亿，大约两年半。放在任何行业里，这个斜率都找不到几个对标案例。
NRR约900%：老客户越买越多，人均营收超660万美元
营收增速本身不稀奇，稀奇的是NRR（净收入留存率）约900%。这个指标衡量的是同一批客户今年比去年多部署了多少。900%意味着老客户平均把Claude的调用量扩到九倍，增长不是靠拉新撑的。
人均营收超过660万美元。按S-1披露的员工规模反推，每个工程师、产品经理、销售创造的收入远超传统软件公司。
但结构上有两个卡点。第一，两家客户贡献了约20%的总营收，集中度不算低。第二，2025年47%的营收（约21.6亿美元）走的是Amazon和Google的云计算市场渠道，Claude在相当大的比例上是借别人的云卖的。
Anthropic营收结构与渠道依赖
这两个结构问题不致命，但它们决定了Anthropic在算力谈判桌上的议价空间。当近一半收入经由Google（同时拥有TPU和自研Gemini模型）的云平台进来，定价权就不是完全握在自己手里的。
算力才是真账单：90亿已花，6000亿在路上
上半年基础设施支出近90亿 vs 去年全年73亿
S-1里有一组数字比账面亏损更值得盯。2026年上半年，Anthropic在算力与基础设施上的实际支出约89.5亿美元；2025年全年为73亿；2025年上半年仅23.3亿。今年六个月花掉的钱，已经超出去年一整年，同比增速接近四倍。
斜率在2025年下半年开始陡峭化，2026年上半年直接跳了一个档位。这不是线性外推能解释的节奏。
与Google的111亿合作及4500亿云承诺的期限结构
真正压顶的是还没落地的部分。S-1披露，未来七至十年，基础设施支出承诺超过4500亿美元。其中与Google的协议约111亿为已签约确定性支出；若计入正在洽谈的额外算力协议，总盘子可能推至7500亿。
这些承诺有一个结构性特征。云租约是take-or-pay，到期前必须按期付租金，跟营收是否继续13倍增长无关。花90亿买算力的人，不会因为某季度现金流紧张就退租。而收入端，两个大客户合计贡献约20%营收，2025年有47%的收入经Amazon与Google云市场结算。算力成本是固定的、多年的、不可逆的；收入端尚未锁到同等年限的契约。
截至2026年6月底，Anthropic账上现金及短期投资合计约1317亿美元（1124亿现金加193亿短期投资）。这笔钱覆盖未来两三年算力支出没有压力，但撑不到4500亿承诺全部履约的那一天。IPO募资的核心用途，很大程度上就是填这个期限缺口。
算力承诺期限与现金流错配
没花出去一分钱。
807亿拆开来：七成是「估值税」
可转换票据与SAFE的公允价值机制
S-1的亏损结构写得很直白。807.48亿净亏损里，723.88亿（约89.6%）对应一项会计科目：「可转换票据、衍生负债及SAFEs的公允价值变动」。这笔钱的产生逻辑是，Anthropic在2024至2025年间以固定价格发行了可转换债券和SAFE（简单股权融资票据），当时约定持有人可以按某个固定价格转股。后来公司估值从几百亿拉到$965B post-money，那些固定价格的转股权公允价值暴涨，会计准则要求每个季度把差额记入损益表。公司没掏出一个新美元，账面亏损却跟着估值一起膨胀。
估值越高、账面亏损越大
这个机制有个反直觉的推论：公司越值钱，报表越难看。$965B的Series H估值意味着早期SAFE持有人的行权价与公司当前公允价值之间的gap被拉到最大。只要公司继续融资或二级市场估值上行，这笔非现金亏损就会继续滚雪球。它不影响现金流，不影响运维，但会在S-1的损益表上持续制造视觉冲击。
S-1留白的那些格子
22家银行承销团为何不亮价
Goldman Sachs和Morgan Stanley领衔的22家承销团已在S-1上签字，但定价区间、发行股数、募资额、Run-Rate Revenue（第195页）全部空白。这不是疏漏，是IPO标准流程。定价要等路演结束、机构认购簿关闭后才落锤。路演前亮价等于把议价空间交给市场，一旦认购需求不及预期，承销商只能被动降价。22家银行的规模本身也暗示发行体量不小，留给各机构分配份额的协调成本更高，定价窗口自然往后推。
S-1留白与定价时序
留白对估值叙事和发行节奏的影响
留白造成的直接效果是，当下所有「Anthropic值多少钱」的数字都锚在\(965B post-money（Series H）和媒体转述上，而非S-1本身。公司6月底\)131.7B的现金短投总和，意味着它没有2021年那批烧钱SaaS的逼宫式募资压力，可以把路演节奏拉到11月甚至12月。但如果AI板块在路演窗口内整体回调，超额的认购溢价会收窄，发行价可能贴着$965B甚至以下成交。承销团留白，本质上是把定价风险从「锁死」变成「浮动」，给自己留了调价空间。
两个客户占20%营收：集中度藏在哪
大客户依赖对续费和砍预算的敏感性
S-1风险因子栏写明，两家客户合计贡献约20%的半年营收，折算大约\(3.2B。再看2025年全年，Amazon和Google的云市场渠道带来了47%的收入（\)2.16B）。两组数字叠加，头部客户+云渠道对总营收的影响面超过六成。NRR 900%的逻辑链建立在老客户持续扩量上，但如果其中任一家在2027财年预算周期里削减Claude API调用配额，或者把workload迁向自研模型，NRR会率先失速。一个$3.2B量级的客户砍20%预算，对公司全年营收增速的拖累是双位数百分点级别的。
上市前披露这一项的信号意义
把「两家客户占20%」写进S-1风险因子，是证券法对重大客户的强制披露义务。但选择以「两家」而非「五家」为粒度披露，配合900% NRR和13倍增速一起呈现，传递的信号是：集中度确实高，但增长斜率足以在短期内把占比稀释下去。真正需要盯的验证节点是上市后首份10-Q。如果这两家客户的ARPU增速跑输公司整体NRR，集中度就从风险因子变成定价折价，承销团的发行价浮动区间也会下修。
参考文献
- Anthropic S-1 Filing (EDGAR, 2026-10-26): www.sec.gov/cgi-bin/bro…
- NDTV Profit, "Anthropic S-1 IPO Filing": www.ndtvprofit.com/markets/ant…
- DataStudios, "Anthropic files S-1 for Nasdaq IPO": www.datastudios.org/post/anthro…
- HackerNoon, "Anthropic's S-1 Shows \(16.3B in Half-Year Revenue": https://hackernoon.com/anthropics-s-1-shows-\)163b-in-half-year-revenue-a-900percent-nrr-and-an-$807b-loss-that-mostly-isnt
- Yahoo Finance, "Anthropic IPO live updates": finance.yahoo.com/technology/…
- Fortune, "Anthropic's IPO prospectus": fortune.com/2026/10/26/…
- CryptoBriefing, "Anthropic set to publish its S-1": cryptobriefing.com/anthropic-s…
延伸入口
- 原文归档：tobemagic.github.io/ai-magician…
- 公众号：计算机魔术师
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。
You saw a net loss of 80.7 billion, and the first reaction was probably to finish the company.
80.7 billion torn: 72.4 billion is "valuation tax", not cash flow.
Convertible instruments and SAFE fair value mechanisms
Anthropic Early Multi-Court Financing makes extensive use of convertible instruments and SAFE, which are embedded in instruments. US GAAP requires fair value remeasurement of embedded derivatives for each reporting period. The valuation jumped from billions of dollars in 2024 to $965 billion post-money in Series H, the early low-value transfer equity surge, which was recorded directly in the profit statement in accounting, under the heading "A change in fair value of convertible instruments, derivative liabilities and SAFE". The money's not moving. The numbers are moving.
How changes in fair value become book losses Loss
The higher the valuation, the greater the book losses
Of the 80.748 billion, 72.388 billion represent this change in fair value, or 89.6 per cent. The remaining $8.4 billion is the real business gap. The more successful the company is, the higher the valuation, the greater the "losses" on this statement. Counterintuitive, but logical: you pay book prices for options in other hands.
Ninety percent of the valuation went up, not burned.
Real increase in revenues from 16.3 billion battalions: 41 million to 13 times
Timeline 2023 to 2026.
There is a clean gate in S-1. The year of the commercialization of Claude in 2023 was $41 million a year; 386 million in 2024; 4.585 billion in 2025; and 16,286 million in the first half of 2026, compared to 1,163 million in the same period last year, an increase of about 1.3 per cent.
Four numbers, no curves.
More striking is the single season split. The Q2 battalion received $11.56 billion in 2026, nearly 30 times more each quarter than it did throughout 2024. S-1 singles out this figure, which is tantamount to informing investors that growth is not evenly climbing, but exponential acceleration.
From 41 million to 16.3 billion, about two and a half years. In any industry, there are few matching cases for this slope.
NRR about 900%: the older clients buy more and earn more than $6.6 million per person
The rate of increase is not unusual per se, but about 900 per cent for NRR (net income retention rate). This indicator measures how many more clients were deployed this year than last year. 900 per cent mean that old clients have on average increased Claude ' s call to nine times, and the increase is not driven by new growth.
Over $6.6 million was received per capita. Inversely, each engineer, product manager, sales generated revenues far beyond conventional software companies, as disclosed by S-1.
But there are two points on the structure. First, two clients contributed about 20 per cent of the total, with a modest concentration. Secondly, 47 per cent of the revenues received in 2025 (about $2.16 billion) were channelled through the cloud computing market in Amazon and Google, and Claude sold it to others in a significant proportion.
Anthropic battalion revenue structure and channel dependence
These two structural issues are not fatal, but they determine the bargaining space of Anthropic at the bargaining table. When nearly half of the income comes through the cloud platform of Google (with the TPU and the Gemini model), pricing rights are not entirely in their hands.
Calculus is the real bill: 9 billion is spent, 600 billion is on the road.
In the first half of the year, infrastructure spending was nearly $9 billion, vs. $7.3 billion over the last year.
There is a set of figures in S-1 that are more noteworthy than the book losses. In the first half of 2026, Anthropic spent approximately $8.95 billion on computing and infrastructure; $7.3 billion in full in 2025; and only 2.33 billion in the first half of 2025. The amount spent this six months is already more than the entire year of last year, nearly four times the same rate of increase.
The slopes began steepening in the second half of 2025 and jumped directly into a slot in the first half of 2026. This is not the rhythm of linear extrapolation.
Structure of the 11.1 billion-dollar partnership with Google and the 45 billion-dollar cloud commitment
The real crushing part is the one that hasn't landed. S-1 disclosed that over the next seven to ten years, infrastructure spending commitments exceeded $450 billion. Some 11.1 billion of the agreements with Google were signed firm expenditures; if additional calculator agreements were being negotiated, the total plate could be pushed to 750 billion.
These commitments have a structural character. The cloud lease is a take-or-pay that must be paid on schedule before expiry, regardless of whether the collection continues to increase 13 times. People who spend 9 billion dollars on computing do not withdraw their rent because of a season of tight cash flows. At the revenue end, two large clients combined contributed about 20 per cent of the proceeds, and 47 per cent of the revenues in 2025 were settled by Amazon and Google Clouds. Energy costs are fixed, multi-year and irreversible; the revenue side is not locked into contracts of equal age.
At the end of June 2026, cash and short-term investments in Anthropic amounted to approximately $131.7 billion ($112.4 billion in cash plus $19.3 billion in short-term investments). This money covers the next two to three years without pressure, but will not last the day when 450 billion dollars are committed to full compliance. The core use of IPO funding is largely to fill this deadline gap.
Calculating the period of commitment to the cash flow
It didn't cost a penny.
80.7 billion torn apart: 70% of the value tax.
Convertible instruments and SAFE fair value mechanisms
The loss structure for S-1 is very clear. Of the net deficit of $80,748 million, $72,388 million (approximately 89.6 per cent) corresponded to an accounting line: "A change in convertible instruments, derivative liabilities and the fair value of SAFEs". The logic behind this money was that Anthropic issued convertible bonds and SAFE (simple equity financing instruments) at fixed prices between 2024 and 2025, when it was agreed that the holder could transfer the shares at a fixed price. The firm ' s valuation was then pulled from tens of billions to $965B post-money, where the fair value of the transferred equity in fixed prices surged, and the accounting standards required that the differences be recorded in the income statement on a quarterly basis. The company did not pull out a new dollar and the book losses expanded with the valuation.
The higher the valuation, the greater the book losses
There's a counter-intuitive inference to this mechanism: the more valuable the company is, the worse it is. The 965B Series H valuation meant that the initial SAFE holder ' s lead price was maximized between the company ' s current fair value. This non-cash loss will continue to roll snowball as long as the company continues to finance or to value the secondary market. It does not affect cash flows, nor mobility, but it continues to create visual shocks on the S-1 income statement.
The ones S-1 left behind.
Why didn't 22 bank underwriters set prices?
The 22 groups of Goldman Sachs and Morgan Stanley were signed on S-1, but the price-fixing areas, the number of issues, the amount of money raised, and Run-Rate Revue (p. 195) were completely blank. It's not an oversight. It's an IPO standard process. Pricing must wait until road shows are over and institutional subscription books are closed. The opening price before the show is equal to giving the bargaining space to the market, and if the demand for the subscription falls short of expectations, the underwriter can only reduce the price passively. The size of the 22 banks itself implies that the distribution volume is not small, that the coordination costs for the allocation of shares by agencies are higher and that the pricing window naturally slips backwards.
Time series for S-1 retention and pricing
Impact of retention on the narrative and issuance rhythm of the valuation
The immediate effect of this is that all the current numbers of "Anthropic values" are anchored in \(965B post-money(Series H) and in the media, not in S-1 itself. At the end of June, the company's cash pool of 131.7B meant that it did not have the pressure to burn SaaS until November or even December. However, if the AI plate is retweeted in the circuit window as a whole, the oversubscription premium will narrow, and the issue price may even be sold at $965B or less. In essence, the marketing community has changed the price-fixing risk from "locking death" to "floating", leaving room for price-fixing.
Two clients account for 20 per cent of receipts: where is the concentration hidden?
Large clients rely on sensitivity to continuity and budget cutting
The S-1 risk factor indicates that the combined contribution of the two clients is approximately 20 per cent of the semi-annual collection, which is approximately \\(3.2B). In all of 2025, Amazon and Google cloud market channels generated 47 per cent of income (\2.16B). Two sets of numbers are superimposed, with head customer plus cloud channels affecting more than 60 per cent of the total harvest. The NRR 900 per cent logical chain is based on the constant expansion of older clients, but if one of them cuts Claude API ' s quota in the 2027 budget cycle or relocates workloads to self-study models, the NRR will be the first to slow down. A $3.2B volume client cut the budget by 20 per cent, and the drag on the growth of the company ' s business throughout the year is at the double point level of hundreds.
Before you go on the market, disclose the signal.
The inclusion of "two customers account for 20%" in the S-1 risk factor is the mandatory disclosure obligation of the securities law for significant customers. However, the choice of the "two" instead of the "five" as a particle scale, in combination with 900% NRR and 13 times the rate of increase, sends the signal that the concentration is really high, but the growth slope is sufficient to dilute the percentage in the short term. The certification node that really needs to be targeted is the first 10-Q to be listed. If the ARPU of the two clients were to run the NRR as a whole, the concentration would change from a risk factor to a price discount, and the distribution price of the group would be repaired downward.
References
- Anthropic S-1 Filing (EDGAR, 2026-10-26): www.sec.gov/cgi-bin/bro…
- NDTV Profit, "Anthropic S-1 IPO Filing": www.ndtvprofit.com/markets/ant…
- DataStudios, "Anthropic files S-1 for Nasdaq IPO": www.datastudios.org/post/anthro…
- HackerNoon, "Anthropic's S-1 Shows \(16.3B in Half-Year Revenue": https://hackernoon.com/anthropics-s-1-shows-\)163b-in-half-year-revenue-a-900percent-nrr-and-an-$807b-loss-that-mostly-isnt
- Yahoo Finance, "Anthropic IPO live updates": finance.yahoo.com/technology/…
- Fortune, "Anthropic's IPO prospectus": fortune.com/2026/10/26/…
- CryptoBriefing, "Anthropic set to publish its S-1": cryptobriefing.com/anthropic-s…
Extension entrance
- Original archive: tobemagic.github.io/ai-magician...
- WeChat Official Account: Computer magician
Auto-fetched and translated by AI Atlas. All rights belong to the original author.