With a total of 150 billion yuan raised in financing within half a year, Liang Wenfeng, the "technological idealist", is being hotly pursued by capital.
Liang Wenfeng, who once emphasized that "the problem we have never faced is a shortage of capital", is accelerating the capitalization process of DeepSeek.
On October 22, several media outlets reported that DeepSeek's new round of financing has been oversubscribed to 1 trillion RMB, with investors including Seastone Capital, SMIC PE, CPE, Legend Capital, Geely, etc. Existing shareholders such as CATL, Tencent, IDG Capital, Lisi Capital and Shiang Capital continued to participate in the follow-on investment.
In April 2026, DeepSeek launched its first round of external financing since its establishment, which was closed at the end of June with a size of about 510 billion RMB, bringing a post-money valuation close to 4 trillion RMB. Founder Liang Wenfeng personally contributed about 200 billion RMB, Tencent contributed about 100 billion RMB, the CATL ecosystem contributed about 50 billion RMB, NetEase, JD, Lisi Capital and IDG Capital each contributed about 30 billion RMB, and the National Artificial Intelligence Industry Investment Fund also participated in this first round of financing.
Once the new round of financing is closed, DeepSeek will have absorbed a total of more than 1.5 trillion RMB in capital within half a year. However, according to reports, although the size of this round of financing has doubled, the pre-money valuation remains at about 5 trillion RMB, which means DeepSeek has not taken advantage of the market hype to push up its pricing excessively.
Source: AI-generated
It can also be seen from the shareholder list that the investors of this round of financing are mainly industrial capitals. What Liang Wenfeng pays more attention to should be the optimization of the shareholder structure, rather than the maximization of valuation. The shareholder list covers upstream chips, power supply, computing power and downstream application scenarios in the AI industry chain, Liang Wenfeng is building an industrial collaboration network that can support DeepSeek's long-term development.
At the same time, the DeepSeek team is also preparing for listing. Media reports say that since late August 2026, DeepSeek has begun to prepare for listing on the Science and Technology Innovation Board, plans to submit its prospectus by the end of 2026, and achieve listing in 2027. In addition, Yan Wentao, former partner of GL Ventures, has joined the company in September as CFO.
DeepSeek is not the only one preparing for listing. Entering 2026, leading AI companies are accelerating their capitalization process to reserve ammunition for the next stage of competition —
Moonshot AI was recently exposed to have completed its final round of financing with a valuation of about 500 billion US dollars, and plans to conduct an IPO in Hong Kong in the first quarter of 2027; Zhipu AI and MiniMax have completed their listings on the Hong Kong Stock Exchange, and successively launched procedures for A-share listing; Anthropic is also expected to be listed in November with an expected valuation between 1.5 trillion and 2 trillion US dollars, which is expected to become one of the largest technology IPOs in history; OpenAI is accelerating the promotion of financing of about 300 billion US dollars, planning to lock in capital and computing power commitments before Anthropic goes public.
This means that the resource competition of global large AI models has entered a white-hot stage.
But beyond capital, Liang Wenfeng is still building an open and win-win ecosystem relying on DeepSeek. Since the financing news came out in April 2026, the content updated on DeepSeek's official WeChat account has remained nothing but models, open source, and technical reports.
Liang Wenfeng, who has set a new financing record, is still that technical idealist?
Stick to the Open Source Ecosystem
On October 17, DeepSeek officially announced that DeepSeek V4.1 Flash has been updated to the V4.1-Flash-1017 version, and the official web page, App and API models have been synchronously replaced with the new model.
The official statement says that the new version, trained through reinforcement learning based on self-feedback, improves the ability to build complete projects from scratch, supports more "wish-driven" programming tasks in scenarios such as game development and 3D modeling, and has carried out adaptive optimization for platform frameworks including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.
Recently, DeepSeek is not the only player that has updated its model: MiniMax open-sourced M3.1 Flash; Kimi released its flagship open-source model K3.1 Pro; Stepfun open-sourced Step 5 Preview; Zhipu AI is also stepping up the training of its next-generation flagship model, with the rumored parameter scale ranging from 2.5 trillion to 3.5 trillion, all targeting complex tasks and agent scenarios.
However, while peers continue to increase investment at the model layer, DeepSeek has done something beyond the industry's expectation. On October 20, DeepSeek directly open-sourced its entire internal inference framework DeepInfer and production-level codec component DeepSeek-recipe.
The inference framework is a software system that provides an efficient operating environment for large language models, responsible for coordinating and scheduling the huge computing power cluster, whose efficiency directly determines the computing power cost and power consumption of AI services.
DeepSeek officially announced that based on the DeepInfer framework, it has realized high-throughput, low-latency large-scale inference services for DeepSeek V4.1 Flash. The daily throughput of a single cluster is 155 trillion Tokens, the average input cache hit rate is 96%, and the revenue per card per hour calculated at valley electricity prices is 38 US dollars.
The reason why this entire set of frameworks can be open-sourced is explained by DeepSeek's official WeChat account with the quote from *Guanzi*: "Do good deeds with the public, share ingenuity with all people". DeepSeek's official WeChat account writes that sharing internal deployment experience is to help developers and enterprises optimize the throughput and latency of large-scale inference, efficiently build private and public cloud deployments, reduce computing power costs and energy consumption, "so that intelligence can reach a wider group of people at a lower price, and realize the true popularization of large models".
This is also the "original intention" that Liang Wenfeng has always emphasized. As early as in an interview in 2024, he asserted: "We will not close the source. We believe that building a strong technical ecosystem is more important than a closed business model."
At this year's closed-door investor meeting, he further shared his logic on open source: First, it is related to their vision, they "do this thing with very great kindness to the world, to be useful to humanity." Second, they believe that open source is beneficial for the commercialization of AI.
In Liang Wenfeng's view, unlike the past contradiction between open source and commercialization, AI is large enough that it may eventually account for 10% of human GDP. No company can monopolize it, and it must be shared with others. In this context, Liang Wenfeng believes that to achieve AGI, restraint is a must. "If you think that X percent of human GDP belongs to you, the more you think so, the less likely you are to succeed. The more restrained you are, the more likely you are to achieve AGI," Liang Wenfeng said.
On September 30, when DeepSeek announced the open-sourcing of infrastructure components for Huawei Ascend computing power platform, it stated that it will continue to promote technological innovation, and jointly build and progress with the community to form an open software ecosystem.
Some AI developers also said that DeepSeek itself is doing this thing with the idea of an open source ecosystem, and the creativity within its community is also the strongest in China's developer ecosystem.
Restrained Commercialization
Liang Wenfeng is not only driven by emotional narratives. Beyond open source and popularization, he has his own set of logic in commercialization.
At this year's closed-door investor meeting, he shared: "I think we have been doing commercialization all the time, but we do not take commercialization as the goal, but take AGI as the goal." In Liang Wenfeng's view, the DeepSeek team is conducting world-leading AGI research, and commercialization will happen at an appropriate time, but it is still far away from the point of fully shifting to commercialization. So they still hope to develop next-generation technologies to solve more existing problems. "This is also part of our restraint," Liang Wenfeng said.
He even hopes that other people can take advantage of this commercial opportunity: "We do not have enough energy, and the organization does not have enough people to do this thing." Up to now, DeepSeek has not directly sold subscriptions to C-end users, the only charging channel is selling APIs, which has led several rounds of API "price wars" in the industry.
"Our API pricing brings a reasonable profit, which means that if we purchase a batch of equipment in the market, we can recover the cost in ten months," Liang Wenfeng once explained this pricing standard to investors. In his view, if they pursue profit maximization, the price should be set higher, because user demand in this interval is inelastic, "if we double the price, the consumption of Tokens will not change much, and the total revenue can nearly double." But he did not do that.
The DeepSeek team also reached a consensus on this. "In the version of the model we released before, we were worried that the demand would be too high at the beginning, so we set the price relatively high, and the team members were not very happy. Later, we lowered the price to a quarter of the original, and everyone was happy." In his view, the reason why the team spent so much energy and carefully made the model well is to let everyone fully access very cheap and very effective models. This is the motivation and vision of the team, and also the internal consensus that the company can gather together.
Source: AI-generated
This set of pricing principles has not slowed down the company's revenue growth. According to the latest disclosure, DeepSeek's annualized recurring revenue (ARR) is growing extremely fast, from less than 500 million US dollars to 1 billion US dollars in only about three months, ranking in the first echelon of China's large model industry.
However, while revenue and call volume are rising rapidly, service stability is also under pressure. DeepSeek experienced two large-scale access anomalies this year, both of which took several hours to fix. The aforementioned AI developer also believes that DeepSeek's open Chatbot is more like a derivative, an additional service provided by the model capability in the process, which is not necessarily an active commercial choice. He therefore speculates that DeepSeek's future focus may fall more on B-end business, deploying models for different enterprises, such as integrating a model into a certain device, as well as a series of supporting optimization work.
DeepSeek is also trying to explore a more suitable commercialization method for itself. Along with the release of V4.1 Flash on October 17, the official announced the addition of Search and Fetch interface support, with the price of both interfaces at 5 yuan per thousand calls; when users use DeepSeek API in DeepSeek Harness, Claude Code and Codex, they can access these two interfaces through server-side tool calls, and the call fees will be charged per call from 00:00 on November 1, 2026.
This not only adds a new business line for the company, but also users' search data can feed back to model training, and the lower price can also make developers accustomed to using DeepSeek for search, retaining DeepSeek's search entry.
But in Liang Wenfeng's priority list, commercialization is never the first goal. He emphasized that the primary purpose of DeepSeek developing large models is to serve the iteration of internal AGI research, and accelerate the self-evolution of model capabilities through "self-use", that is, RSI (Recursive Self-Improvement), rather than simply pursuing C-end user experience or commercial monetization.
Beyond large language models, Liang Wenfeng also maintains a high degree of attention to embodied models. In August 2026, DeepSeek invested 141 million yuan in Unitree Robotics; in October, Liang Wenfeng plans to lead the investment in Winverse AI founded by Hu Han. In Liang Wenfeng's view, only when the model has the ability of continuous learning and self-development, can it finally move towards embodied intelligence and realize the complete closed loop of intelligence in the physical world.
This article is from the WeChat official account "China Entrepreneur Magazine" (ID: iceo-com-cn), Author: Kong Yuexin, Editor: Ma Jiying, released with authorization from 36Kr.