文 | 数读社
Anthropic的招股书中，美股科技巨头可能是最大的赢家。
在风险提示中，Anthropic提到了与巨头合作的“take-or-pay”模式，也就是约定至少支付的金额，无论实际使用情况如何。
招股书中大约提到了6370亿美元的算力承诺。其中，博通、SpaceX、谷歌、亚马逊都达到千亿美元。
到今年7月，Anthropic的ARR（年度经常性收入）只有650亿，差出一个数量级。Anthropic相当于把未来几年钱都给了科技巨头。
AI之王，现阶段还是在给巨头们打工。
01 最佳估值周期
这是Anthropic非常不错的上市窗口期。主要有三个方面。
首先，Anthropic正处在狂奔状态，现阶段的增速是最能打动投资者的。
根据招股书，从2023年到2025年，Anthropic的营收从4100万美元增至45.85亿美元。上半年，Anthropic的营收162.86 亿美元，同比增长了13倍。看似已经快速增长，三季度依然炸裂，结合招股书提到的信息推算，三季度营收约165亿，已经超越上半年的营收，同比依然是10倍增速。
ARR展现的细节也交叉印证了这个结论。2025年年初，Anthropic的ARR只有10亿美元，今年4月，已经突破300亿美元。根据媒体报道，7月底已经达到650亿美元，这种增速，可能在人类史上都是少见的。
其次，Anthropic依然是AI世界最受用户认可的公司。这从用户数据上可以体现。Anthropic用TTM（过去12个月滚动收入）口径统计客户，把客户分为10万以内、10万以上、100万以上、1000万以上。
截至9月30日，10万以上的客户，达到11900家，三个季度涨了10400家。贡献超过1000万美元的已经达到250家，低于10万的更是达到4000万个。
Anthropic对于它的客户蛮横到什么程度？招股书中提到，其收入有一部分为“违反了使用政策或商业条款的客户”。也就是被封号的客户，这些客户哪怕充了钱，也有可能不退款。很多大陆用户，对此可能已经有所感知，甚至怨声载道。可即便如此，很多客户依然离不开Anthropic。
核心原因就是，Anthropic的技术能力依然是领先的。这也是当前Anthropic上市的重要“卖点”。
最后，Anthropic现阶段证明，技术领先能讲通商业故事。上半年，Anthropic的净亏损为807.48 亿美元，但这主要是融资决定的，可转债、衍生负债与公允价值变动的亏损占723.88亿，实际经营亏损只有6.93亿美元。
如果按照经营利润的口径，Anthropic在二季度为9.74 亿美元，首次单季实现经营利润转正。上半年，经营活动现金流净额39.29亿美元，实现了现金流转正，意味着大模型的模式盈利能力较强。
现在这个时机上市，Anthropic能拿到不错的估值。用户认可并且付费意愿强烈，市场议价能力强。处于快速飞奔的阶段，高增速非常亮眼。同时也向资本市场展示了自己商业模式的可持续性——只是暂且处于投入期，商业模式成立。
02“AI之王”的隐忧
如果能赶上SpaceX上市的那个窗口，可能会更好。后者在今年6月IPO时，凭借AI画出的大饼，估值一度冲到1.77万亿美元。
而现如今，据路透社报道，部分参加Anthropic投资者日的投资者，对IPO估值预期有所下降，从此前的2万亿美元降至1.5万亿美元左右。核心原因在于，公司收入环比增速放缓。按照此前的测算，Anthropic在三季度的环比增速已经降到约43%。
当下可能已经是最好时机，否则就要等更长时间，更多不确定性，甚至会更差。
AI世界的变化非常迅猛。在Anthropic之前，OpenAI是这个市场的引领者。后者创造了ChatGPT时刻，并且在较长时间里，模型能力都保持领先。2021年，七位前OpenAI成员出走创业，创立 Anthropic。在短短5年间，就实现了对OpenAI的反超。
而且是全方位反超，5月，Anthropic估值达到9650亿美元，超越OpenAI。7月ARR超越OpenAI。
但Anthropic的领先优势也在缩小。10月的数据显示，中美顶级AI模型之间的性能差距已经缩小至约3%，创下新低。中国的大模型在持续追赶。
在AI模型评测平台LiveBench上，DeepSeek V4.1 Flash排名全球第六。整体性能已经可以与Anthropic和OpenAI的头部模型相提并论。月之暗面推出的Kimi K3.1 Pro。其性能已接近 Opus 5.5。
中国模型的成本更低，性价比更高，在技术能力和商业化两个赛道上，都有可能给Anthropic带来冲击。更为关键的是，中国模型更开放，并不会像Fable 5一样全球出口管制。
或许在接下来某个时段，Anthropic就不再领先，商业变现也会面临巨大的冲击。加之高速增长下，基数会越来越大，增速变缓本身就是不可避免的，越往后的估值，可能会越不如当下。
为了尽可能避免这个状况，Anthropic做出了应对。
Anthropic进行了一定程度的降价，特别是小模型。10月，其发布Claude Haiku 5.5，这是其最小、最经济实惠模型，输入和输出单价较上一代明显下调，标准输入价格降至每百万Token 0.10美元、输出价格降至每百万Token 0.50美元，单价降幅约90%。
旗舰模型也有降价，不过力度很小。比如，9月，Anthropic发布Fable 5.1，定价为10美元/50 美元/0.25 美元（输入/输出/缓存读取，每百万 token，下同）。与6月的Fable 5一样，但是Fable 5.1让典型工作负载的成本比Fable 5降低约 25%，复杂编码和高强度agentic任务的节省最高可达45%，也算是降低了模型的使用成本。
提交招股书前，Anthropic发布了 Claude Opus 5.5，定价为4美元/20 美元/0.20美元，叠加模型效率提升后，单个任务的成本比Opus 5降低约 40%。
但这样的价格，与国产大模型差距巨大。比如新发布的 DeepSeek Flash V4.1 1017，每百万 Token 输入 0.15 美元、输出 0.60 美元、缓存读取 0.003 美元。相比Anthropic，几乎是白菜价。Kimi K3.1 Pro输入 2 美元、输出 10 美元、缓存读取 0.20 美元，输入和输出价格只有Opus 5.5的一半。
Anthropic的降价，只是犹抱琵琶半遮面，其价格相比中国大模型仍然过高。这是成本与商业逻辑决定的，哪怕Anthropic的代码库，9月单月，Anthropic代码库的新增贡献超过90% 由Claude撰写。成本可以压低，可其价格大概率都无法做到中国模型这样极限。
因此，在很长一段时间里，Anthropic都没有价格优势。
那么，想要保证全球领先优势，就只有一条路——技术领先。
03 眼花缭乱的资本手段
技术领先的前提是算力保障。Anthropic已经成为了博通、SpaceX、谷歌、亚马逊、微软等巨头的大金主。
SpaceX三季报显示，其AI业务营收达到75.6亿元，同比增长830%，超越发射、通信，成为第一大业务。SpaceX并没有提起太多客户，Anthropic是唯一被提及、也是最大的外部客户。
招股书中，Anthropic主要提到了7家公司。其中，博通是最大的采购来源，双方签订了五年期主设备融资租赁，租金约1612亿，按季度支付。
SpaceX的采购总额最多为1205亿，合作到2029年，在2027年以前的订单不可撤销。过了期限后，双方任一方90天通知取消。这算得上退出最灵活的合作。
谷歌的算力服务高达1145亿，是典型的不问用量、不足补差，除 Google 重大违约外不可撤销。不仅如此，Anthropic还承诺2028–2031年追加TPU采购1250-1500亿。
亚马逊算力服务为1100亿，与谷歌一样，也是不问用量、不足补差、除重大违约不撤销。微软则是449亿，不问用量，部分可撤销。
仅仅这几家巨头，Anthropic就承诺了最高超过5000亿的采购，哪怕按照千亿ARR，Anthropic也基本上把未来5年的营收交给了巨头们，这基本上可以说是给巨头打工。
而Anthropic并没有足够的资金。截至上半年末，Anthropic现金及等价物1124 亿，短期投资193亿，远不足以覆盖巨额采购订单。
因此，Anthropic只能通过融资手段来解决这个问题。
最典型的就是NPA（票据购买协议），简要来说就是，通过融资，有条件地提取相关款项。比如，与谷歌的NPA就约定了TTM营收目标。与博通的合作，就对营收增速提出了要求。亚马逊的150亿NPA是按交付算力解锁，亚马逊AWS交付多少算力，Anthropic 才能提多少钱。
其次可转债，主要是与亚马逊关联的62亿，转股价取“固定价与 IPO 价孰低者”。
为了避免控制权也被巨头们瓜分，Anthropic采取了5种股权模式。公开发行的是A类，有经济权益。创始人团队是F类，没有经济权，但拥有一般表决事项50.1%的投票权。转股通道是N类，员工激励是E类，拥有董事会选举罢免权的是T类。
可以看出，拥有公司决定权的是七位创始人，而外部科技巨头，没有被列入股东，并且在未来都会有持股上限。
通过这种融资、有条件解锁、同股不同权等方式，Anthropic尽可能维系着算力供应，这其实也让整个AI始终陷入“自娱自乐”的质疑。
拨开这些资本层面的云雾，Anthropic真正的价值，还是客户数、ARR指标。
只看这些指标，Anthropic是处在狂奔阶段的现象级公司。特殊之处在于，其身处AI行业，一个可能决定人类走势的行业，而Anthropic又是这个行业的王者。但这个王者的领先优势，并非永恒。现阶段，反而可能是上市的最佳时机。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Books
In Anthropic's book, American technology giants are probably the biggest winners.
In the risk profile, Anthropic refers to the “take-or-pay” model of cooperation with the giants, i.e. the amount agreed to be paid at least, regardless of actual use.
Some $637 billion in calculative commitments was mentioned in the book. Of these, Motton, SpaceX, Google and Amazon have reached hundreds of billions of dollars.
By July of this year, Anthropic ' s ARR (annual recurrent income) was only $65 billion, one order of magnitude. Anthropic amounts to giving the next few years to the technology giant.
King of AI, still working for the giants at this stage.
01 Best valuation cycle
This is a very good listing window period for Anthropic. There are three main aspects.
First, Anthropic is in a state of raging, and the current rate of growth is the most stimulating for investors.
According to the equity book, from 2023 to 2025, Anthropic ' s collection increased from $41 million to $4.585 billion. In the first half of the year, Anthropic received $16,286 million, a 13-fold increase over the same period. It appears that there has been a rapid increase and that the three quarters have continued to burst and, combined with the information referred to in the book of stock calls, the three quarters have been calculated at approximately $16.5 billion, exceeding the first half of the year, which is still 10 times the same rate.
The details presented by ARR also cross-check this conclusion. At the beginning of 2025, Anthropic ' s ARR was only $1 billion, and in April this year it was over $30 billion. According to media reports, the rate of growth, which had reached $65 billion at the end of July, was probably rare in human history.
Second, Anthropic remains the most recognized user company in the AI world. This is reflected in the user data. Anthropic uses the TTM (rolling income over the past 12 months) calibrated clients, dividing them into less than 100,000, more than 100,000, more than 1 million and more than 10 million.
As at 30 September, over 100,000 clients had reached 11,900, an increase of 10,400 in three quarters. Contributions of more than $10 million have reached 250, or 40 million below 100,000.
How far did Anthropic go with its clients? The equity book mentions that part of its income is “clients who violate the use policy or commercial terms”. That is, the sealed clients, who may not be refunded even if they are charged. Many continental users may already be aware of this, or even complaining. But even so, many clients still have to leave Anthropic.
The central reason is that the technical capacity of Anthropic remains at the forefront. This is also an important “seller point” for the current Anthropic market.
Finally, Anthropic at this stage proves that technology leads business stories. In the first half of the year, the net loss for Anthropic was $80,748 million, but it was mainly a result of financing decisions, with a loss of $72,388 million for reversible debt, derivative liabilities and changes in fair value and only $693 million for actual operating losses.
Anthropic is $974 million in the second quarter, if the business profit is of the calibre, for the first time in a single season. In the first half of the year, net operating cash flows amounted to $3.929 billion, leading to positive cash flows, which meant that LLM was more profitable.
Now it's time to go on the market and Anthropic gets a good valuation. User approval and paid will be strong and the market bargaining power strong. At a fast-flying stage, the growth rate is very bright. It also showed the capital market the sustainability of its business model - - It's only a short-term investment and business models are in place.
"The King of AI"
It might be better to catch up with the window on SpaceX. The latter, in June of this year, with the help of AI's pies, made an impact assessment of $177 trillion.
Today, according to Reuters, some of the investors participating in the Anthropic Investor Day are expected to experience a decline in the IPO valuation, from $2 trillion previously to about $1.5 trillion. The central reason is the slower growth of the corporate income ring. Anthropic ' s ring rate of increase has been reduced to about 43 per cent in the third quarter, according to earlier estimates.
It may be the best time to wait longer, more uncertainty and even worse.
The AI world is changing very rapidly. Before Anthropic, Open AI was the leader of the market. The latter created ChatGPT moments and, over a longer period of time, model capabilities remained in the lead. In 2021, seven former Open AI members ran out of business and created Anthropic. In just five years, the opposite of Open AI was achieved.
In May, Anthropic ' s valuation reached $965 billion, exceeding Open AI. July ARR transcends Open AI.
But Anthropic's lead is shrinking. Data from October show that the performance gap between the top AI models in Central America has narrowed to about 3 per cent, a new low. China's LLM continues to chase.
DeepSeek V4.1 Flash ranks sixth globally on LiveBech, the AI Model Evaluation Platform. The overall performance is already comparable to the head model of Anthropic and Open AI. Kimi K3.1 Pro from the dark side of the moon. It's close to Opus 5.5.
China ' s model is cheaper and more expensive, and has the potential to impact Anthropic on both the technological capability and commercialization tracks. More crucially, the Chinese model is more open and does not provide global export controls like Fable 5.
Perhaps at some point in time, Anthropic will no longer be in the lead, and commercial liquidity will also face a huge shock. Combined with high-speed growth, the base figure will grow, and slow growth will itself be inevitable, and later valuations may be worse than in the present.
Anthropic responded to this situation as much as possible.
Anthropic made some price reductions, especially small models. In October, its publishClaude Haiku 5.5, its smallest and most economical model, reduced significantly the unit price for input and output compared with the previous generation, falling standard input prices to US$ 0.10 per million Token and output prices to US$ 0.50 per million Token, with a reduction of about 90 per cent.
The flagship model has also been priced down, albeit with little strength. For example, in September, AnthropicpublishFable 5.1, priced at $10/50/$0.25 (input/output/cache read per million token, same). As in Fable 5, in June, Fable 5.1 reduced the cost of typical work loads by about 25 per cent compared to Fable 5, with savings of up to 45 per cent on complex coding and high-intensity agglomerations, and reduced the cost of using the model.
Prior to the submission of the equity book, Claude Opus 5.5 was awarded at $4/20/$0.20 for Claude publish and the cost of a single mission was approximately 40 per cent lower than Opus 5 as a result of the increased efficiency of the superimposed model.
But there is a huge gap between such prices and the national production of LLM. For example, DeepSeek Flash V4.1 1017, per million Token input 0.15, output 0.60, cache reads 0.003. Compared to Anthropic, it's almost cabbage. Kimi K3.1 Pro entered $2, output $10, cache read $0.20, and the input and output price was only half of Opus 5.5.
The drop in Anthropic prices is only half-faced and still overpriced compared to China ' s LLM. It's a matter of cost and business logic, even with the Anthropic's code library, in September alone, with over 90 percent of the new contribution from the Anthropic's code library written by Claude. Costs can be kept to a low level, but their price rate is not as high as the Chinese model.
Thus, for a long time, Anthropic had no price advantage.
There is then only one path — technology — to ensure global leadership.
03 Optical capital instruments
Technology leads on the premise of security of calculation. Anthropic has become a major gold owner of giants like Chase, SpaceX, Google, Amazon and Microsoft.
According to the SpaceX Quarterly, its AI operations account for $7.56 billion, an 830 per cent increase over the same period, exceeding launch, communications and becoming the first largest business. SpaceX did not mention too many customers, and Anthropic was the only and largest external customer.
In the equity book, Anthropic referred mainly to seven companies. Of these, Chase is the largest source of procurement, and the parties have signed a five-year financial lease for the main equipment, with a rent of approximately $161.2 billion, to be paid quarterly.
SpaceX procured a total of up to $12.5 billion and worked together until 2029, when the order was irrevocable until 2027. After the expiry of the deadline, either party cancelled it by 90 days' notice. This is the most flexible exit from cooperation.
Google ' s computing services, amounting to 114.5 billion, are typical of no-question, no-compensation, and are irrevocable except for Google ' s major defaults. Moreover, Anthropic committed an additional 12.5-15 billion USD for TPU procurement for 2028-2031.
The Amazon Calculator Service, which is 110 billion, is, like Google, no matter how much it is, not enough to make up for it, and not to cancel any major breach. Microsoft is $44.9 billion, which is partially avoidable, regardless of the amount used.
With these giants alone, Anthropic has committed up to $500 billion in purchases, and even according to hundreds of billions of ARRs, Anthropic has largely handed over the next five years to the giants, which is basically working for them.
And Anthony didn't have enough money. At the end of the first half of the year, the cash and equivalents of Anthropic were 112.4 billion and short-term investments of 19.3 billion, far from covering large purchase orders.
Therefore, Anthropic can only address this problem by means of financing.
The most typical is NPA (instrument purchase agreement), which, in short, makes conditional withdrawal of the relevant funds through financing. For example, we have a TTM target with Google's NPA. Cooperation with Chase has led to demands for the rate of increase. The 15 billion NPA in Amazonia is unlocked by the delivery factor, and the Amazon AWS delivers how much money Anthropic can raise.
Second, debt is reversible, mainly 6.2 billion dollars associated with the Amazon, and the transfer price is “lower than fixed or IPO”.
In order to avoid the separation of control among the giants, Anthropic adopted five equity models. It is publicly available in category A and has an economic interest. The founding team is F and has no economic rights, but has 50.1 per cent of the voting rights in general matters. The transfer route is category N, the employee incentive is category E, and the right to strike on board elections is category T.
It can be seen that the decision-making power of the company is vested in the seven founders, while the outside technology giant, which is not included in the shareholders, will have an equity ceiling in the future.
Through this type of financing, conditional unlocking, and different shares, Anthropic has maintained as much as possible the supply of calculus, which in fact has kept AI as a whole in the challenge of “self-recreation”.
The true value of the cloud at these capital levels is also the number of clients, the ARR indicator.
Looking at these indicators only, Anthropic is a phenomenon-class company in the run-off phase. What's special is that it's in the AI industry, an industry that may determine the movement of humans, and that Anthropic is the master of the industry. But this king's lead is not eternal. At this stage, it might be the best time for listing.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.