文 | 市值榜，作者 | 相青；编辑 | 赵元
过去几年，OpenAI、Anthropic等大模型公司不断刷新融资规模和估值，但外界始终很难看清，这些动辄数百亿美元的投入，究竟能换来多少收入和利润。
如今，Anthropic率先摊开了自己的账本。
近日，Anthropic向美国证券交易委员会（SEC）公开提交S-1招股书，计划登陆纳斯达克。
招股书显示，2026年上半年，Anthropic营收达到162.86亿美元，是去年全年收入的三倍多。第二季度，公司实现9.74亿美元营业利润，首次在披露的季度数据中实现营业盈利。
更重要的是，Anthropic的收入增速已经远远超过算力成本增速。今年上半年与2025年上半年相比，Anthropic算力及基础设施费用占营收的比例，从201%降至55%。此前被认为需要长期烧钱的大模型生意，开始显现盈利的可能。
但这份账本也暴露了另一面。
Anthropic目前的收入高度依赖企业的模型调用量，而模型性能和价格仍在快速变化。为了维持竞争力，公司需要不断推出新模型，投入更多研发和算力。据外媒报道，Anthropic已经签下超过6000亿美元的多年算力及基础设施相关协议。
与此同时，Anthropic近一半收入来自云平台销售渠道。Amazon、Google等科技巨头既向它提供算力，也帮助它销售模型，同时还在与Claude争夺企业客户。
从收入快速增长到季度营业利润转正，Anthropic已经走过了大模型商业化最初的验证阶段。但收入规模越大，对算力和研发的投入也越高。模型公司究竟需要付出多大成本才能维持竞争力，又能从不断增长的收入中留下多少利润？
一、半年营收超160亿美元，Anthropic靠什么赚钱？
在用户规模远不及OpenAI的情况下，Anthropic却实现了营收反超。
2026年第二季度，Anthropic营收115.55亿美元，环比增长144%。据《华尔街日报》报道，OpenAI同期营收67亿美元，环比增长18%。尽管两家公司对部分云平台销售采用不同的收入确认方式，但增速差距依然明显。
Anthropic的增长究竟来自哪里？
招股书显示，2025年上半年，Anthropic年贡献收入超过10万美元的客户只有600多家，到了2026年同期，已经超过6000家。这些客户带来的半年收入从8.7亿美元增至120亿美元，贡献了公司上半年约74%的营收。
同时，单个客户花的钱也越来越多。
2026年上半年，Anthropic约80%的收入来自按使用量计费，订阅收入仅占20%。企业交给Claude处理的任务越多，Token消耗越大，带来的收入通常也越高。
Claude Code进一步放大了这一特点。过去，程序员主要用AI提问、生成代码，如今可以让Claude Code连续执行编写、测试、调试等任务，模型调用开始贯穿整个开发流程。
企业不断增加的AI使用量，也反映在客户账单上。截至9月底，Anthropic年贡献超过100万美元的客户已有1900多家，其中约250家超过1000万美元。同期客户净收入留存率约为900%。
Anthropic还试图让客户使用更多自有产品。
在年贡献超过10万美元的客户中，同时使用Claude模型平台与Claude Code、Cowork等应用的比例，从2025年底约30%升至2026年第三季度的超过50%。
企业从调用模型，到使用编程工具，再到把AI接入更多业务，每增加一类使用场景，都可能带来新的收入。
不过，按量计费虽然能让收入快速增长，但也让模型厂商更容易受到客户使用量变化的影响。
招股书披露，2026年第三季度，Anthropic超过80%的收入来自此前六个月内发布的新模型版本。
这说明最新模型已经承担了绝大部分收入。对Anthropic来说，新模型既要吸引新客户，也要让老客户愿意继续付费。一旦竞争对手推出性能更好、价格更低的产品，企业就可能重新分配调用量。
更何况，Anthropic的大客户收入并没有完全锁定。2026年前九个月，其前五大客户贡献了约25%的营收，而不少大客户没有长期合同约束，可以减少甚至停止支出。
OpenAI也在加快争夺企业市场。今年4月，OpenAI披露企业业务收入占比已超过40%，Codex等编程Agent正在与Claude Code争夺开发者和企业客户。双方从不同市场起步，如今越来越多地争夺同一批企业预算。
企业愿意为AI投入更多预算，已经成为Anthropic收入快速增长的重要基础。但要持续获得这些收入，Anthropic就得不断推出更有竞争力的模型。
二、首次实现营业盈利，Anthropic能持续赚下去吗？
收入快速增长的同时，Anthropic距离盈利也越来越近了。
招股书显示，2026年上半年，Anthropic净亏损高达807.48亿美元，但同期营业亏损只有6.93亿美元。到了第二季度，公司甚至实现了9.74亿美元的营业利润，营业利润率约为8%。
净亏损和营业利润之间的巨大差距，主要来自会计处理。上半年，Anthropic因可转债等融资工具公允价值变动，产生约724亿美元的账面损失。这部分损失主要与公司估值上升有关，并不涉及同等规模的现金流出。
相比净亏损，营业利润更能反映Anthropic的实际经营状况。
过去几年，大模型公司给外界留下的印象是收入增长快，但烧钱速度更快。如今，至少从单季度表现来看，Anthropic已经开始赚钱了。
利润改善的关键，是算力成本开始被收入摊薄。
2025年上半年，Anthropic的算力及基础设施费用达到23.34亿美元，相当于同期收入的201%。换句话说，每赚1美元，就需要支出约2美元的算力费用。
一年后，这个比例降至55%。2026年上半年，算力费用增长283%，达到89.45亿美元，但同期收入增长约1300%。
算力支出仍在增加，只是收入增长得更快。这也让大模型生意开始显现规模效应。算力资源通常需要提前采购，当企业调用量快速增加，已有算力得到更充分利用，收入便可能跑赢成本。
不过，第二季度的盈利能否持续，现在还很难判断。
首先，Anthropic必须不断增加算力投入。
前面提到，Anthropic第三季度超过80%的收入来自过去六个月发布的新模型。要维持这样的产品更新速度，公司就需要持续训练更强的模型。而随着企业把更多工作交给AI，推理需求也在增长。
今年上半年算力费用占收入的比例大幅下降，并不代表未来还能保持同样的趋势。新一批算力投入使用后，相关费用可能继续上升，利润率也会受到影响。
其次，模型价格也在持续承压。
OpenAI、Google以及开源模型厂商都在争夺企业客户。随着模型性能差距缩小、推理成本下降，企业有更多选择，也有更大的议价空间。
降价能够刺激调用量，但如果调用量增长赶不上价格降幅，收入就会受到影响。对于高度依赖按量计费的Anthropic来说，这一点尤其重要。
更大的压力来自已经签下的长期算力合同。
据外媒报道，Anthropic已签订超过6000亿美元的多年算力及基础设施相关协议。其中，与Amazon、Google签订的合同，分别涉及至少1100亿美元和1145亿美元的消费承诺。
这些费用分布在未来多年，不能直接计入当期成本。但部分合同设置了最低消费要求，即使算力没有得到充分利用，Anthropic也可能需要按约定付款。
对Anthropic来说，提前锁定算力可以保障模型研发和商业化需求，却也增加了未来经营的不确定性。一旦收入增速放缓，大量尚未充分利用的算力就可能拖累利润。
因此，Anthropic第二季度实现营业盈利，是一个积极信号，但距离证明大模型能够持续盈利，还有很长的路要走。
三、模型公司的生意，仍然绕不开科技巨头
Anthropic的营收已经达到百亿美元规模，但有近一半收入，仍需要借助其他科技公司的渠道完成销售。
招股书显示，2023年，Anthropic通过云平台销售的收入仅占11%，2025年升至47%，2026年上半年进一步达到48%。
这些收入主要来自AWS、Google Cloud、Microsoft Azure等云平台。
对Anthropic来说，借助云厂商销售Claude，可以省去大量企业获客成本。大型企业通常已经采购AWS、Google Cloud等云服务，能够直接通过现有账户和采购体系接入Claude。
但Anthropic与云巨头的关系，远不止销售合作这么简单。
Amazon、Google既是Anthropic的算力供应商，也是重要的销售渠道。Anthropic向它们采购算力，企业又通过这些云平台购买Claude，两家公司可以同时获得算力服务收入和模型销售分成。
与此同时，Amazon和Google也在发展自己的AI模型，与Claude争夺企业客户。
虽然双方采取联合销售模式，Anthropic直接与客户签约，并保留模型定价权，云厂商则收取渠道费用。但真正的问题是，Anthropic在多大程度上能够独立于这些合作伙伴开展业务？
Amazon、Google、Microsoft都有自己的模型和AI产品，也掌握着大量企业客户。当Claude与其他模型在性能和价格上接近时，云平台选择重点推广哪款产品，就可能影响Anthropic的销售。
算力供应同样如此。Anthropic在招股书中提醒，这些合作伙伴可能优先满足自身业务或其他客户的需求，影响Claude获得算力的成本和供应。
更何况，企业AI市场的竞争，已经逐渐从模型本身延伸到应用和服务。
Google有Workspace，微软有Microsoft 365、Copilot。这些产品已经进入企业日常办公流程，拥有成熟的客户基础。
对企业来说，选择AI服务不能只看模型性能，还要考虑与现有系统的兼容性、数据安全以及使用成本。
Anthropic也在向应用层寻找机会。Claude Code、Cowork等产品，让Anthropic能够直接服务程序员、办公人员，而不只是向其他软件公司提供底层模型。
单纯销售API，客户可以根据性能和价格，在不同模型之间切换。但如果Claude Code已经进入企业的开发流程，涉及代码管理、工具配置和团队协作，更换产品就需要付出额外成本。
Anthropic还在补齐企业服务能力。除了与埃森哲、德勤等咨询公司合作，它还联合私募股权机构成立了Ode with Anthropic，帮助中型企业开发、部署Claude应用。
不过，向应用层扩张，也让Anthropic开始与自己的客户争夺生意。
一些软件公司原本采购Claude API，开发编程助手、办公工具等AI产品。随着Anthropic推出Claude Code、Cowork，双方的产品可能出现重叠。
Anthropic在招股书中也明确提醒，随着公司扩大产品范围，部分SaaS客户可能认为自身业务受到威胁，减少采购甚至转向其他模型供应商。
这让Anthropic处于一种复杂的关系中。Anthropic需要云厂商提供算力和销售渠道，也需要软件开发商持续调用Claude。但随着自有产品不断扩张，它与部分合作伙伴和客户之间的竞争也会越来越直接。
OpenAI、Google、微软也在向模型、应用和企业服务的多个环节扩张。
Anthropic已经证明，独立模型公司可以创造巨大的商业收入。但在科技巨头同时掌握算力、渠道和应用入口的市场里，仅靠模型能力领先，恐怕还不足以建立长期优势。
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。
Zhao Yuan, Editor
Over the past few years, Open AI, Anthropic and others LLM have continuously updated the scale and valuation of the financing, but it has always been difficult to see how much revenue and profits could be exchanged for these billions of dollars in investment.
Now Anthropic is the first to spread his own books.
Recently, Anthropic publicly submitted a S-1 book to the United States Securities and Exchange Commission (SEC), planning to land in NASDAQ.
According to the equity book, in the first half of 2026, the Anthropic battalion received $16,286 million, more than three times the annual income of the previous year. In the second quarter, the company achieved operating profits of $974 million and, for the first time, in the quarterly data disclosed.
More importantly, Anthropic's income growth has gone far beyond the cost of calculation. In the first half of the year, as compared to the first half of 2025, the percentage of revenue earned by Anthropic and infrastructure costs fell from 201 to 55 per cent. The LLM business, which was previously considered to require long-term burning of money, began to show the potential for profit.
But this book also exposes the other side.
The current income of Anthropic is highly dependent on the use of models by enterprises, whose performance and prices are still changing rapidly. To maintain competitiveness, companies need to continuously introduce new models and invest more in research and development and computing. According to external sources, Anthropic has signed over $600 billion in multi-year computing and infrastructure-related agreements.
At the same time, nearly half of Anthropic's revenue comes from cloud platform sales channels. Science giants such as Amazon and Google both provide it with credit and help it sell models, while competing with Claude for corporate customers.
From a rapid increase in revenue to a positive shift in quarterly operating profits, Anthropic has passed the initial certification phase of LLM commercialization. But the larger the income, the higher the investment in computing and research and development. How much is it going to cost a model company to remain competitive and how much profit can it keep from growing revenues?
One or six months, over $16 billion. How does Anthropic make money?
Anthropic achieved a counter-receipt when users were far below Open AI.
In the second quarter of 2026, the Anthropic battalion received $11,555 million, an increase of 144 per cent in the ring. According to the Wall Street Journal, Open AI collected $6.7 billion over the same period, an 18 per cent increase in the ring. Despite the fact that two companies use different methods of revenue recognition for some cloud platform sales, the growth gap remains significant.
Where does Anthropic grow?
According to the book, in the first half of 2025, only 600 clients had contributed more than $100,000 a year to Anthropic, and by the same time in 2026 over 6,000. The annual revenue generated by these clients increased from $870 million to $12 billion, contributing about 74 per cent of the company ' s revenue during the first half of the year.
At the same time, individual clients spend increasing amounts of money.
In the first half of 2026, approximately 80 per cent of Anthropic's income was derived from usage charges and only 20 per cent from subscriptions. The more the enterprise gives Claude the task, the more Token consumes, the higher the income it usually generates.
Claude Code further magnifies this feature. In the past, programmers used AI mainly to ask questions and generate codes, and now Claude Code was able to perform successive tasks of writing, testing, debugging, etc., and model calls began throughout the development process.
The increasing use of AI by enterprises is also reflected in client bills. By the end of September, there were more than 1,900 clients in Anthropic who had contributed more than $1 million annually, of which some 250 had exceeded $10 million. The customer net income retention rate for the same period was approximately 900 per cent.
Anthropic also tried to make more of its own products available to clients.
Of the clients who contributed more than $100,000 annually, the proportion of the Claude Model platform with applications such as Claude Code and Cowork increased from about 30 per cent at the end of 2025 to over 50 per cent in the third quarter of 2026.
Each of the additional types of use that enterprises can generate could generate new income, from the use of models to the use of programming tools to access more businesses.
However, while volume charges allow for rapid income growth, they also make model producers more vulnerable to changes in customer usage.
The equity book disclosed that in the third quarter of 2026, more than 80 per cent of Anthropic ' s income came from the new model version of publish in the previous six months.
This indicates that the most recent model already covers the bulk of income. For Anthropic, the new model must attract both new clients and older clients willing to continue paid. Once competitors have introduced better performance and lower prices, firms may reallocate transfers.
Moreover, Anthropic's large client income is not fully locked. In the first nine months of 2026, its top five clients contributed about 25 per cent of the revenue, while many large clients were not bound by long-term contracts and could reduce or even stop spending.
Open AI is also accelerating competition for the business market. In April of this year, Open AI disclosed that business income had exceeded 40 per cent, and programming Agent, such as Codex, was competing with Claude Code for developers and business clients. Starting in different markets, the parties are now increasingly competing for the same corporate budget.
The willingness of enterprises to invest more in AI has become an important basis for rapid income growth in Anthropic. However, in order to sustain these revenues, Anthropic will have to continuously introduce more competitive models.
Is Anthropic sustainable for the first time in business?
The rapid growth of income has been accompanied by the proximity of Anthropic to profitability.
According to the equity book, in the first half of 2026, the net loss for Anthropic amounted to $80,748 million, but only $693 million for the same period. By the second quarter, companies had even achieved $974 million in operating profits, with a profit margin of about 8 per cent.
The large gap between net losses and operating profits arises mainly from accounting treatment. During the first half of the year, Anthropic incurred a book loss of approximately $72.4 billion due to changes in the fair value of financing instruments such as reversible debt. This portion of the loss was mainly related to the increased valuation of the company and did not involve cash outflows of the same size.
Business profits are more reflective of Anthropic ' s actual performance than of the net loss.
Over the past few years, LLM has left the impression to the outside world that revenues are growing faster, but that the money is burning faster. Now, at least in a single quarter, Anthropic is already making money.
The key to improving profits is that the cost of computing begins to be spread over the income.
In the first half of 2025, the computing power of Anthropic and infrastructure costs amounted to $2,334 million, equivalent to 201 per cent of income for the same period. In other words, for every dollar earned, there would be a cost of about $2.
One year later, the proportion dropped to 55 per cent. In the first half of 2026, real cost increased by 283 per cent to $8,945 million, but income increased by about 1.3 per cent over the same period.
Calculus spending continues to increase, except for faster income growth. It also began to show the scale effect of the LLM business. Calculator resources usually require early procurement, and income can be costed when the number of business calls increases rapidly and the computing power becomes more fully available.
However, it is difficult to judge whether the profits of the second quarter will be sustainable.
First of all, Anthropic must continue to increase its input.
As mentioned earlier, more than 80 per cent of the income of Anthropic in the third quarter came from the new model of publish over the past six months. Maintaining the rate of product renewal requires companies to continuously train more models. And as businesses give more work to AI, the need for reasoning is growing.
The sharp decline in the ratio of revenue to capital costs in the first half of the year does not mean that the same trend can be maintained in the future. The associated costs may continue to rise and profit margins may be affected when the new calculator is put into operation.
Second, model prices also continue to be under pressure.
Open AI, Google and open source model manufacturers are competing for business customers. As model performance gaps narrow and reasoning costs fall, enterprises have more choice and more bargaining space.
Lower prices can stimulate transfers, but if they do not keep pace with price reductions, incomes will be affected. This is particularly important in the case of Anthropic, which is highly dependent on a volume fee.
Greater pressure came from long-term computing contracts that had been signed.
According to external sources, Anthropic has signed over $600 billion in multi-year computing and infrastructure-related agreements. Of these, the contracts with Amazon and Google involved consumption commitments of at least $110 billion and $114.5 billion, respectively.
These costs are distributed over the years and cannot be charged directly to current costs. However, some of the contracts set minimum consumption requirements, and even if the calculations were not fully utilized, Anthropic might have to pay as agreed.
For Anthropic, early calculus ensures demand for model development and commercialization, but also increases uncertainty about future operations. Once the pace of income growth slows, a large amount of under-utilized computing power may delay profits.
Thus, the realization of business profits in the second quarter of Anthropic is a positive sign, but there is still a long way to go before LLM can prove that it will be profitable.
III. THE BUSINESS OF THE MODEL COMPANY, STILL IN FOCUS
Anthropic's revenue has reached a billion dollars, but nearly half of its revenue still needs to be sold through other technology companies.
According to the book, in 2023, only 11 per cent of Anthropic revenues were sold through cloud platforms, rising to 47 per cent in 2025 and reaching 48 per cent in the first half of 2026.
These revenues come mainly from cloud platforms such as AWS, Google Cloud and Microsoft Azure.
For Anthropic, the sale of Claude by cloud manufacturers saves a large number of businesses from the cost of taking clients. Large enterprises have typically procured cloud services such as AWS, Google Cloud and have direct access to Claude through existing accounts and procurement systems.
But the relationship between Anthropic and the cloud giant is much more than just marketing cooperation.
Amazon, Google, is both a calculator and an important sales channel for Anthropic. Anthropic procures computing power from them, and businesses buy Claude through these cloud platforms, allowing both companies to obtain both revenue from computing services and a share of model sales.
Meanwhile, Amazon and Google are developing their own AI models to compete with Claude for corporate clients.
Although the parties have adopted a joint sales model, Anthropic has contracted directly with its customers and retains model pricing rights, cloud manufacturers charge channel fees. But the real question is to what extent can Anthropic operate independently of these partners?
Amazon, Google, Microsoft have their own models and AI products, as well as a large number of corporate clients. When Claude and other models are close to performance and price, the cloud platform selects which product to focus on, and may affect the sale of Anthropic.
The same is true for power supply. In its equity book, Anthropic warns that these partners may give priority to meeting the needs of their own business or other clients, affecting Claude ' s cost and supply of computing power.
Moreover, competition in the corporate AI market has gradually extended from models themselves to applications and services.
Google has a Worldspace, Microsoft has Microsoft 365, Copilot. These products have entered the enterprise ' s day-to-day office processes with a mature client base.
For enterprises, the selection of AI services cannot be based solely on model performance, but also on compatibility with existing systems, data security and the cost of use.
Anthropic is also looking for opportunities at the application level. Claude Code, Cowork, etc., allows Anthropic to directly serve programmers, office staff, not just other software companies.
The simple sale of API allows customers to switch between models based on performance and price. However, if Claude Code is already in the enterprise development process, involving code management, tool configuration and teamwork, the replacement product will entail additional costs.
Anthropic is also completing the capacity of the enterprise. In addition to collaborating with consulting firms like Esenzer and Deloitte and Touche, it has established Ode with Anthropic in private equity institutions to assist medium-sized enterprises in developing and deploying Claude applications.
However, expansion to the application level also allowed Anthropic to start competing with his own clients for business.
Some software companies had originally procured Claude API and developed AI products such as programming assistants and office tools. With the introduction of Claude Code and Cowork by Anthropic, there may be overlaps between the two products.
In its equity book, Anthropic also explicitly warns that, as the company expands its product range, some SaaS clients may feel that their business is under threat, reduce procurement and even shift to other model suppliers.
This puts Anthropic in a complex relationship. Anthropic needs cloud manufacturers to provide computing and marketing channels, as well as software developers to use Claude on a continuous basis. But as home-grown products expand, competition with some partners and clients becomes more direct.
Open AI, Google and Microsoft are also expanding to multiple stages of models, applications and business services.
Anthropic has demonstrated that independent model companies can generate huge business revenues. However, in markets where technology giants simultaneously master the calculus, channels and access points, model capabilities alone may not be sufficient to build long-term advantages.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.