过去几年，Anthropic逐渐成为企业级AI市场最受关注的公司之一。它在编程领域建立起优势，将模型能力嵌入开发者工作流，并向企业客户提供更高价值的服务。按照路透社披露的数据，Anthropic 2025年营收达到45.9亿美元，2026年上半年营收进一步增至162.9亿美元。
但招股文件同时揭开了另一面：2026年上半年，Anthropic净亏损达到807亿美元。其中，约724亿美元来自可转换票据公允价值变动，不反映现金流出。真实的经营亏损约为6.93亿美元。
但让投资者难以松口气的是，Anthropic在招股书中给出了未来十年至少5180亿美元的算力采购承诺，其中约八成不可取消或无论使用量均需支付。
本次IPO，Anthropic计划募资1000亿美元，自成立至今，Anthropic已经合计融资了1500亿美元。
美股市场本就对人工智能基础设施的巨额支出表现出了担忧，Anthropic的上市更是催动了市场脆弱的神经。
业内人士评价，此次上市事关重大，如果进展不顺，整个市场都会受到拖累。
一度在追随Anthropic的OpenAI近期也频频转向，一方面是转向低价型模型，10月26日，OpenAI又发布了新一代中端模型GPT-6.1 Terra，将更低的调用价格与计算机操作能力作为重要卖点。
在大洋此岸，始终主打性价比路线的DeepSeek连续多次上调了本轮融资计划，最新消息称其新一轮融资超过了1000亿元，并已经大刀阔斧的走在了大扩张的道路上。多位知情人士向凤凰网科技透露，DeepSeek最近团队扩张势头很猛，“总要把融到钱的花到合理的地方”，而融资越多，也意味着DeepSeek要做的事情越多。
全球AI竞赛正在前所未有的激烈。
商业化样板Anthropic，巨亏807亿美元
在全球AI行业，Anthropic一直是一个颇具代表性的参照系。
与主要依赖消费级流量的产品不同，Anthropic从编程和企业服务切入，逐渐建立起自己的商业路径。Claude在开发者群体中积累了用户，企业客户则通过API调用、付费订阅等方式购买模型能力。模型不只是一个聊天窗口，而是进入软件开发、知识处理和业务流程的生产工具。
这条路的吸引力很直接：企业愿意为能够提高生产效率的工具付费，而且高价值场景通常拥有比普通消费者应用更清晰的付费逻辑。
尤其在编程领域，AI的价值不再仅仅体现在回答问题是否准确，而在于能否完成代码编写、调试、测试和复杂任务执行。对于企业来说，如果模型可以缩短开发周期、减少重复劳动，付费便有了相对具体的依据。
一位AI赛道CEO在与凤凰网科技交流时，会认为Anthropic是AI商业模式的重要样本，“大家一开始发现AI coding有很强的商业化转化能力，这最早也来自于Anthropic，现在Anthropic又开始做AI制药，未来肯定也会有人跟随”。
但收入的质量，与利润的质量，是两回事。
按照路透社对Anthropic招股文件的梳理，公司2025年经营亏损约80.6亿美元，2026年上半年经营亏损约6.93亿美元。其同期计算与基础设施支出接近90亿美元，较上年同期约23亿美元大幅增加。
这意味着，Anthropic的业务确实在快速扩张，但支撑业务扩张的资源成本同样不容忽视。
模型公司卖出的每一个Token，都不是凭空产生的。训练需要算力，推理需要算力，用户规模扩大、任务复杂度提升，也会不断增加计算资源的消耗。尤其当AI从简单问答走向编程代理和复杂任务执行，模型可能需要更多轮推理、调用工具和反复校验，单次任务背后的成本结构也随之改变。
企业客户愿意付费，并不意味着这笔生意天然拥有足够高的利润率。若收入增长需要持续投入更多算力才能维持，规模扩大与盈利改善之间，就未必存在简单的正相关关系。
Anthropic面临的另一重压力，是长期算力采购承诺。路透社报道，公司已与多家合作伙伴签订规模巨大的云服务、计算资源和基础设施协议，未来十年的相关承诺超过5000亿美元。
Anthropic的经营并非毫无进展。相反，收入增长和企业客户扩张说明，市场正在为其模型能力付费。但让外界不解的是商业模式本身，当模型能力持续进步，推理成本能否同步下降？当更多工作交给AI，客户愿意支付的价格，能否覆盖模型、算力和服务交付的全部成本？
各种“链主”加持，始终开源的另一条路
如果说Anthropic代表的是从高价值场景切入、逐步建立商业收入的路线，那么DeepSeek带给行业的另一种启示，是成本本身也可以成为模型竞争力的一部分。
DeepSeek早期引发行业震动，重要原因之一就是它让市场重新审视了训练和使用大模型的成本。外界曾将注意力集中在其模型的性能表现，以及训练成本的相关披露上。但对商业化而言，比某一个训练数字更重要的，是低成本能否延伸到模型部署、推理和实际使用环节。
OpenRouter数据，在调用量方面，DeepSeek旗下模型已连续三周位居该平台周榜第一。此外，minimax、智谱与腾讯等模型也长期居于平台调用量前5。
DeepSeek的路线提供了一个值得研究的方向：模型能力并不只有不断增加参数、堆叠算力这一条实现路径。架构设计、训练方法、推理效率和资源调度，都可能影响最终的单位成本。
更重要的是，低成本会改变模型进入市场的方式。
这将是截然不同的两种发展思路。
美国国土安全部近期下令联邦机构全面移除DeepSeek产品及基于其模型开发的第三方服务，而DeepSeek几乎同时宣布将内部推理框架DeepInfer全面开源，兼容英伟达与华为昇腾两大算力平台。戏剧的对比背后，指向的是同一个事实：DeepSeek已经被摆在了全球AI竞争的前排，而它选择的应对方式不是收缩，是把技术底座铺得更开。
近期，多家媒体报道，DeepSeek即将完成约1000亿元人民币融资。从9月开始，这轮融资的规模就在一路被“追高”：最初计划500亿，中途上调到800亿，最终落在1000亿。加上6月首轮约510亿元，DeepSeek在半年内累计吸金约1500亿元，投前估值维持在5000亿元人民币。
投资阵容空前豪华，据自媒体《金钱的潮汐》统计，本轮融资汇集了超过20家投资机构，上一轮老股东IDG资本、拾象科技、正心谷资本、Monolith、宁德时代、腾讯以及国智投等机构继续参与本轮融资。新进投资方则包括中国移动、中国电信、国寿投资、兆易创新旗下的石溪资本、中芯聚源、博裕资本、CPE源峰、高榕资本、君联资本、中信投资、美团龙珠、横店资本、吉利资本等机构。另据了解，本轮融资吸引了大量上市公司及产业资本间接参与，其中包括牧原股份、知乎、世纪华通、完美世界、巨人网络、大华股份等，参与者之众，难以一一列举。
不难看出，梁文锋在选择投资人时尤为看重产业价值。比如，在能源方面有扎实储备宁德系、与国产存储芯片紧紧捆绑的兆易创新系。
在巨额的投资面前，DeepSeek也正面临着“花钱”的烦恼。一位接近DeepSeek的人士告诉凤凰网科技，DeepSeek正在给投资人规划一个新的增长路线，其中就包括大举扩充团队。
6月首轮约510亿元融资刚落定，DeepSeek就在6月25日发布招聘公告，明确提出“将所有部门规模扩大至少一倍”，一次性开放33个岗位，覆盖技术研发、算力基础设施、产品落地、前沿算法研究、综合职能七大板块。甚至在乌兰察布招揽数据中心基建岗位。
到了9月，DeepSeek Harness团队负责人崔添翼在社交平台一次性放出约150个招聘名额，用他的原话说，需求是“前所未有的”。对待员工福利，DeepSeek也堪称大手笔，有消息称，DeepSeek给清华大学姚班一名实习生开出的税前日薪达到5500元，月薪超过12万元；校招核心系统研发工程师岗位的月薪区间则在6万至9万元。
但梁文锋却始终保持着朴素与低调，一位接近DeepSeek人士告诉凤凰网科技，DeepSeek的办公地主要是北京和杭州，而梁文锋长期在北京租房子住，且房间布置一直非常简单。
性价比开始主导一切
在众多AI巨头里，OpenAI一直是最摇摆的那个。
他们最初追求模型智能的上限，还一度想要拿下C端市场，甚至幻想过要做“AI时代的微信”连接一切。
但在巨额的亏损与被反超的收入面前，OpenAI又一度学起了Anthropic，在AI coding领域攻城略地，以期更好的商业化转化。
如今，OpenAI的做法又越来越DeepSeek。今年9月23日，GPT-6 Sol与Luna同时发布，Luna的API输入价被打到每百万token 0.1美元，堪称地板价。
半个月后的10月7日，OpenAI把Luna直接推给免费用户，覆盖ChatGPT全量人群；付费用户拿Sol，免费用户拿Luna。
10月26日发布的GPT-6.1 Terra，定位于OpenAI的中端模型序列，侧重计算机操作和日常办公任务。其API标准价格较前代GPT-5.6 Terra降低50%；OpenAI还宣称，在部分任务测试中，Terra能够以远低于旗舰模型的成本完成工作。
不过有业内人士认为，“OpenAI仍在摇摆阶段， 从他们在做具身智能模型就能看出来，相比于Anthropic，OpenAI很多时候很摇摆，但很快也会被市场教育，去做能赚钱的生意”。
毕竟，大模型市场的卷有目共睹，如果不能持续建立优势做好主页，很快就会昙花一现。
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。
In the past few years, Anthropic has gradually become one of the most concerned firms in the corporate AI market. It creates advantages in the programming area by embedding model capabilities into developers workflow and providing higher value services to business clients. According to Reuters-disposed data, anthropic received $4.59 billion in 2025, further increasing to $16.29 billion in the first half of 2026.
At the same time, however, the equity instrument revealed another dimension: in the first half of 2026, the net loss of Anthropic amounted to $80.7 billion. Of that amount, approximately $72.4 billion was derived from changes in the fair value of convertible instruments that did not reflect cash outflows. The real operating losses were approximately $693 million.
It is difficult for investors, however, to relax that in its equity book, Anthropic has made a commitment of at least $51.8 billion in computing procurement over the next decade, of which about 80 per cent are irrevocable or have to be paid for regardless of usage.
This time, IPO, Anthropic has raised $100 billion, and since its inception, Anthropic has financed a total of $150 billion.
The U.S. stock market has shown concern about the huge expenditure on artificial intelligence infrastructure, and the listing of Anthropic has given rise to a weak nervous market.
Industry assesses that the listing is of great importance and that if progress is not made, the market as a whole will suffer.
Open AI, once following Anthropic, has also been shifting frequently in the recent past, on the one hand to a low-price model, and on 26 October, Open AI, in addition to publish, launched a new generation of medium-end models, GPT-6.1 Terra, with lower call prices and computer capacity as important selling points.
On the coast of the Ocean, DeepSeek, which has always been the main price-for-money route, has made successive revisions to the current round of financing, which, according to the latest news, exceeds $100 billion and is already on the path of massive expansion. Many informed sources have revealed to Phoenix technology that DeepSeek’s recent team expansion has been strong and that the more money goes, the more money goes, the more it goes.
The global AI competition is becoming more intense than ever before.
Commercial model Anthropic, 80.7 billion dollars.
In the global AI industry, Anthropic has been a representative reference.
Unlike products that are largely dependent on consumption-level flows, Anthropic moves from programming and business services to build its own business path. Claude accumulates users in the group of developers, while business clients purchase model capabilities through API calls, paid subscriptions, etc. The model is not just a chat window but a production tool for entering software development, knowledge processing and business processes.
The attraction of this path is straightforward: businesses are willing to use paid as a tool that can increase productivity, and high-value scenarios usually have a clearer paid logic than ordinary consumers.
In the programming area in particular, the value of AI is no longer limited to the accuracy of answers to questions, but rather to the ability to complete code writing, debugging, testing and complex task execution. For enterprises, paid has a relatively specific basis if models can shorten the development cycle and reduce duplication of effort.
In an exchange with Phoenix technology, an AI track CEO would consider Anthropic to be an important sample in the AI business model, “It was first discovered that AI coding has a strong commercial transformation capability, also from Anthropic, and now that Anthropic is doing an AI pharmaceutical and will be followed in the future”.
But the quality of income is different from the quality of profits.
According to Reuters's combing of the Anthropic book, the company lost approximately $8.06 billion in 2025 and $693 million in the first half of 2026. Its calculations for the same period were close to $9 billion in infrastructure expenditures, a significant increase from the previous year ' s same period of approximately $2.3 billion.
This means that the business of Anthropic is indeed expanding rapidly, but the cost of resources to support it cannot be ignored.
Every Token sold by the model company was not created in a vacuum. Training requires calculus, reasoning requires calculus, expanded user size, increased task complexity and increasing consumption of computing resources. In particular, when AI moves from simple questions and answers to programming agents and complex tasks, the model may require more rotational reasoning, call tools and cross-checking, with the resulting change in the cost structure behind individual assignments.
The willingness of the company ' s clients to paid does not mean that the business has a sufficiently high profit margin. There is not necessarily a simple positive correlation between a larger scale and an improved profitability if income growth requires a sustained investment in more computing power to sustain it.
Anthropic faces another heavy pressure on long-term computing commitments. Reuters reported that the company had signed large-scale agreements with various partners on cloud services, computing resources and infrastructure, with commitments exceeding $500 billion over the next decade.
The business of Anthropic is not without progress. On the contrary, income growth and the expansion of business clients indicate that the market is making its model capacity paid. But what left the outside world incomprehensible is the business model itself, and can the cost of reasoning decline simultaneously when modelling capacity continues to improve? When more work is given to AI, are the prices that clients are willing to pay, and can the full cost of models, algorithms and service delivery be covered?
All kinds of "link masters" plus another path that always opens up.
If Anthropic represents a route to build up business income from high-value scenarios, DeepSeek is another inspiration to industry that costs themselves can become part of model competitiveness.
One of the important reasons for DeepSeek ' s early industrial shock was that it allowed the market to revisit the costs of training and using LLM. Attention has been focused on the performance of their models and on the associated disclosure of training costs. But what is more important for commercialization than a given training figure is whether low cost can extend to model deployment, reasoning and practical use.
OpenRouter data. In terms of call volume, DeepSeek has been the top of the platform for three consecutive weeks. In addition, models such as Minimax, Zhipu and Tencent have long been in the top five of the Platform ' s calls.
DeepSeek ' s route provides a direction worth studying: modeling capacity is not only a path to increasing parameters, stacking capacity. Structure design, training methods, reasoning efficiency and resource mobilization may affect the final unit cost.
More importantly, low-cost changes in the way models enter markets.
That would be a very different approach to development.
The U.S. Department of Homeland Security recently ordered federal agencies to completely remove DeepSeek products and third-party services based on its model development, while DeepSeek announced almost simultaneously that the internal reasoning framework DeepInfer would be fully open to Britain and China as a two-pronged platform. Behind the comparison of the theatre is the fact that DeepSeek has been at the forefront of global AI competition, and it has chosen to respond not by contracting, but by laying the technology floor.
Recently, several media reports have reported that DeepSeek is close to completing approximately RMB 100 billion in financing. Starting in September, the scale of this round of financing has been “higher” all the way: it was initially planned at $50 billion, revised midway to $80 billion and finally fell to $100 billion. Together with the first round in June, about 51 billion yuan, DeepSeek accumulated about 150 billion yuan over a six-month period and maintained its pre-investment valuation at 500 billion yuan yuan.
This round of financing has brought together more than 20 investment institutions, according to the media, The Trend of Money, and the previous round of IDG capital, image-picking technology, positive-eval capital, Monolith, the Ningde era, Tencent and national intellectual investment institutions continue to participate. The new investors include China’s mobile, China’s telecommunications, national life investments, Stone Creek capital under the flag of bogus innovation, China’s core concentration, Bong capital, CPE peaks, high capital capital, King’s Union capital, China’s Sino-Intelligence investment, the United States Dragonballs, Yozooka capital, and Jiley capital. It is also understood that the current round of financing has attracted the indirect participation of a large number of listed companies, as well as industry capital, including rangeland shares, Zhihu, China Century, the perfect world, giant networks, Dahua shares, and so on.
It is not hard to see that Liang Wenfeng values industry in particular when choosing investors. For example, in the energy sector, there are strong reserves of the Ningde system, and a strong bond with the national storage chip.
In the face of huge investments, DeepSeek is also faced with the problem of “payment”. One person close to DeepSeek told Phoenix technology that DeepSeek was planning a new growth course for investors, which included a massive expansion of the team.
In June, the first round of about $51 billion in financing came to an end, and DeepSeek announced, on 25 June, a job announcement for publish, which clearly stated that “to at least double the size of all sectors” and open 33 posts once-in-time, covering seven blocks of technology development, computing infrastructure, product landing, forward algorithm research and integrated functions. Even in Ulanchab, the infrastructure of the data centre was sought.
By September, the head of the DeepSeek Harness team, Choi Tim Wing, had once released some 150 recruitments on social platforms, and in his words, demand was “unprecedented”. In the case of employee freebie, DeepSeek is also considered a big hand. According to information, DeepSeek offered $5,500 per day and more than $120,000 per month to an intern at the University of Qinghua in Yao; the monthly pay zone for the position of research and development engineer for the core system at the school is from $60,000 to $90,000.
But Liang Wenfeng has always kept a low profile, and one close to DeepSeek has told Phoenix technology that DeepSeek’s premises are mainly in Beijing and Hangzhou, while Liang Wenfeng has long rented a house in Beijing, and the room has always been simple.
The value for money starts to dominate everything.
Of all the great alms, Open AI has always been the most swinging one.
They first sought the upper limit of model intelligence, and at one point tried to take the C-end market, even fantasizing about being connected to the "Ai-era WeChat."
But in the face of huge losses and counter-income, Open AI learned another time about Anthropic, striking a little bit in the field of AI coding with a view to a better commercial transformation.
Today, Open AI's approach is becoming more and more DeepSeek. On 23 September this year, GPT-6 Sol and Luna, at the same time publish, were hit at an API price of 0.1 per million token, a floor price.
Half a month later, on 7 October, OpenAi pushed Luna directly to free users, covering the full population of ChatGPT; paid users took Sol and free users took Luna.
GPT-6.1 Terra, publish, 26 October, located in the mid-end model sequence of Open AI, focusing on computer operations and day-to-day office tasks. Its API standard price is 50 per cent lower than that of the previous GPT-5.6 Terra; Open AI also claims that in part of the mission tests, Terra was able to complete its work at a cost far lower than that of the flagship model.
However, according to industry sources, “Open AI is still in the swing phase, as can be seen from the fact that they are making smart models, which is often more volatile than Anthropic, but will soon be educated by the market to make a profitable business”.
After all, the volume of the LLM market is well known, and if the advantages of building a home page on a sustainable basis are not established, it will soon be available.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.