Anthropic Races Toward IPO, Discloses $80.7 Billion Net Loss; OpenAI Shifts to Low-Cost Models

Anthropic, which lost 80 billion, sprints toward IPO, yet its biggest rival wants to learn from DeepSeek Phoenix Tech Download the app for exclusive early access October 27, 2026 11:04:47 From Beijing Updated October 27, 2026 11:25:40

Summary:

Anthropic was once the commercialization benchmark of the market, but this time OpenAI has turned around and started learning from DeepSeek.

On October 26 local time, according to reports, Anthropic has submitted IPO documents to U.S. securities regulators. This AI company, which rose rapidly on the strength of Claude and coding tools, has finally stepped into the spotlight of the capital markets.

This should have been a story about commercial success.

Over the past few years, Anthropic has gradually become one of the most closely watched companies in the enterprise AI market. It established an advantage in the coding field, embedded model capabilities into developer workflows, and provided higher-value services to enterprise customers. According to data disclosed by Reuters, Anthropic's revenue reached $4.59 billion in 2025 and further increased to $16.29 billion in the first half of 2026.

But the prospectus also reveals another side: in the first half of 2026, Anthropic's net loss reached $80.7 billion. Of that, about $72.4 billion came from changes in the fair value of convertible notes and does not reflect cash outflow. The actual operating loss was about $693 million.

But what makes it hard for investors to breathe a sigh of relief is that Anthropic stated in its prospectus a computing power procurement commitment of at least $518 billion over the next decade, of which about 80% cannot be canceled or must be paid regardless of usage.

In this IPO, Anthropic plans to raise $100 billion. Since its founding, Anthropic has raised a total of $150 billion.

The U.S. stock market was already showing concern about the massive spending on artificial intelligence infrastructure, and Anthropic's listing has further strained the market's fragile nerves.

Industry insiders commented that this listing is of great importance, and if it does not go smoothly, the entire market will be dragged down.

OpenAI, which once followed Anthropic, has also frequently shifted direction recently. On the one hand, it is turning to low-price models. On October 26, OpenAI released a new generation mid-range model, GPT-6.1 Terra, making lower API call prices and computer operation capabilities important selling points.

On the other side of the ocean, DeepSeek, which has always focused on the cost-effectiveness route, has repeatedly raised its financing plan for this round. The latest news says its new round of financing exceeded 100 billion yuan, and it has already embarked on a path of aggressive expansion. Several people familiar with the matter told Phoenix Tech that DeepSeek's team expansion momentum has been very strong recently, "after all, the money raised has to be spent in reasonable places," and the more financing it obtains, the more things DeepSeek means to do.

The commercialization model Anthropic, with a huge loss of $80.7 billion

In the global AI industry, Anthropic has always been a highly representative reference point.

Unlike products that mainly rely on consumer traffic, Anthropic entered through coding and enterprise services and gradually built its own commercial path. Claude accumulated users among developers, while enterprise customers purchased model capabilities through API calls, paid subscriptions, and other means. The model is not just a chat window, but a production tool that enters software development, knowledge processing, and business processes.

The appeal of this path is direct: enterprises are willing to pay for tools that can improve production efficiency, and high-value scenarios usually have a clearer payment logic than ordinary consumer applications.

Especially in the coding field, the value of AI is no longer merely reflected in whether answers are accurate, but in whether it can complete code writing, debugging, testing, and complex task execution. For enterprises, if the model can shorten development cycles and reduce repetitive work, payment has a relatively concrete basis.

A CEO in the AI sector, when communicating with Phoenix Tech, said he considers Anthropic an important sample of the AI business model. "Everyone initially discovered that AI coding has strong commercialization conversion ability, and this earliest also came from Anthropic. Now Anthropic has started doing AI pharmaceuticals, and in the future there will certainly be people following."

But the quality of revenue and the quality of profit are two different things.

According to Reuters' review of Anthropic's prospectus, the company's operating loss in 2025 was about $8.06 billion, and its operating loss in the first half of 2026 was about $693 million. Its computing and infrastructure spending in the same period was close to $9 billion, a sharp increase from about $2.3 billion in the same period last year.

This means that Anthropic's business is indeed expanding rapidly, but the resource costs supporting business expansion cannot be ignored either.

Every token sold by a model company does not come from nowhere. Training requires computing power, inference requires computing power, and the expansion of user scale and the increase in task complexity will also continuously increase the consumption of computing resources. Especially as AI moves from simple question answering to coding agents and complex task execution, models may require more rounds of reasoning, tool calls, and repeated verification, and the cost structure behind a single task changes accordingly.

Enterprise customers being willing to pay does not mean that this business naturally has a sufficiently high profit margin. If revenue growth requires continuous investment in more computing power to sustain, then there may not necessarily be a simple positive correlation between scaling up and improving profitability.

Another pressure Anthropic faces is its long-term computing power procurement commitments. Reuters reported that the company has signed massive cloud service, computing resource, and infrastructure agreements with multiple partners, with related commitments over the next decade exceeding $500 billion.

Anthropic's operations are not entirely without progress. On the contrary, revenue growth and the expansion of enterprise customers show that the market is paying for its model capabilities. But what puzzles outsiders is the business model itself: as model capabilities continue to improve, can inference costs decline in tandem? As more work is handed over to AI, can the prices customers are willing to pay cover the full costs of models, computing power, and service delivery?

Backing from various "chain leaders," and another path that has always been open source

If Anthropic represents the route of entering from high-value scenarios and gradually building commercial revenue, then another inspiration DeepSeek brings to the industry is that cost itself can also become part of a model's competitiveness.

One important reason DeepSeek initially shook the industry was that it made the market re-examine the cost of training and using large models. Outsiders once focused their attention on its models' performance and disclosures related to training costs. But for commercialization, what matters more than any single training figure is whether low cost can extend to model deployment, inference, and actual use.

According to OpenRouter data, in terms of call volume, DeepSeek's models have ranked first on the platform's weekly list for three consecutive weeks. In addition, models from MiniMax, Zhipu, and Tencent have also long remained in the platform's top five by call volume.

DeepSeek's route offers a direction worth studying: model capability does not have only one implementation path of continuously increasing parameters and stacking computing power. Architecture design, training methods, inference efficiency, and resource scheduling can all affect the final unit cost.

More importantly, low cost changes the way models enter the market.

These will be two completely different development approaches.

The U.S. Department of Homeland Security recently ordered federal agencies to fully remove DeepSeek products and third-party services developed based on its models, while DeepSeek almost simultaneously announced that it would fully open source its internal inference framework DeepInfer, compatible with both NVIDIA and Huawei Ascend computing platforms. Behind the dramatic contrast points to the same fact: DeepSeek has already been placed in the front ranks of global AI competition, and the way it has chosen to respond is not to contract, but to spread its technology foundation more widely.

Recently, multiple media outlets reported that DeepSeek is about to complete financing of about 100 billion yuan. Since September, the scale of this financing round has been "chased upward" all the way: initially planned at 50 billion, raised midway to 80 billion, and finally settled at 100 billion. Adding the first round of about 51 billion yuan in June, DeepSeek has raised a cumulative total of about 150 billion yuan within half a year, with its pre-money valuation maintained at 500 billion yuan.

The investor lineup is unprecedentedly lavish. According to statistics from the self-media outlet "Tides of Money," this financing round brought together more than 20 investment institutions. Existing shareholders from the previous round, including IDG Capital, Shixiang Technology, Zhengxin Valley Capital, Monolith, CATL, Tencent, and Guozhitou, continued to participate in this round. New investors include China Mobile, China Telecom, China Life Investment, Shixi Capital under GigaDevice, SMIC Juyuan, Boyu Capital, CPE Yuanfeng, Gaorong Capital, Legend Capital, CITIC Investment, Meituan Dragon Ball, Hengdian Capital, Geely Capital, and other institutions. It is also understood that this financing round attracted a large number of listed companies and industrial capital to participate indirectly, including Muyuan Foods, Zhihu, Century Huatong, Perfect World, Giant Network, Dahua Technology, and others; the number of participants is too many to list one by one.

It is not hard to see that Liang Wenfeng particularly values industrial value when choosing investors. For example, in energy, there is the CATL system with solid reserves, and the GigaDevice system, which is closely tied to domestic memory chips.

In the face of huge investment, DeepSeek is also facing the trouble of "spending money." A person close to DeepSeek told Phoenix Technology that DeepSeek is planning a new growth route for investors, including a major expansion of its team.

Just after the first round of about 51 billion yuan in financing was settled in June, DeepSeek released a recruitment announcement on June 25, explicitly stating that it would "expand all departments by at least double," opening 33 positions at once and covering seven major areas: technical R&D, computing infrastructure, product implementation, frontier algorithm research, and comprehensive functions. It was even recruiting for data center infrastructure positions in Ulanqab.

By September, Cui Tianyi, head of the DeepSeek Harness team, released about 150 recruitment slots at once on social platforms. In his own words, the demand was "unprecedented." DeepSeek is also generous with employee benefits. Reports say that DeepSeek offered an intern from Tsinghua University's Yao Class a pre-tax daily wage of 5,500 yuan, equivalent to a monthly salary of more than 120,000 yuan; the monthly salary range for core system R&D engineer positions in campus recruitment is 60,000 to 90,000 yuan.

But Liang Wenfeng has always remained plain and low-key. A person close to DeepSeek told Phoenix Technology that DeepSeek's offices are mainly in Beijing and Hangzhou, while Liang Wenfeng has long rented a place in Beijing, and the room's furnishings have always been very simple.

Cost-effectiveness begins to dominate everything

Among the many AI giants, OpenAI has always been the most wavering one.

They initially pursued the upper limit of model intelligence, and at one point wanted to capture the consumer market, even fantasizing about building "the WeChat of the AI era" to connect everything.

But in the face of huge losses and revenue being overtaken, OpenAI at one point learned from Anthropic, conquering territory in the AI coding field in hopes of better commercialization conversion.

Now, OpenAI's approach is increasingly like DeepSeek's. On September 23 this year, GPT-6 Sol and Luna were released simultaneously, and Luna's API input price was pushed down to $0.1 per million tokens, a rock-bottom price.

Half a month later, on October 7, OpenAI pushed Luna directly to free users, covering the entire ChatGPT population; paying users get Sol, free users get Luna.

GPT-6.1 Terra, released on October 26, is positioned in OpenAI's mid-range model lineup, focusing on computer operation and daily office tasks. Its standard API price is 50% lower than the previous generation GPT-5.6 Terra; OpenAI also claims that in some task tests, Terra can complete work at a cost far lower than flagship models.

However, some industry insiders believe that "OpenAI is still in a wavering stage, as can be seen from the fact that they are working on embodied intelligence models. Compared with Anthropic, OpenAI is often very wavering, but it will soon be educated by the market to do business that makes money."

After all, the fierce competition in the large model market is obvious to all. If a company cannot continuously build advantages and do its main business well, it will soon be a flash in the pan.