DeepSeek is reportedly pushing for an A-share IPO, with its proposed 480 billion yuan valuation sparking widespread industry debate

On July 15, as disclosed by media outlets including Shenzhen Business Daily and Chinese financial news agency, leading domestic large model developer DeepSeek has fully launched preparations for an A-share IPO, with plans to submit application materials within the year and complete its listing in 2027.

The company has just closed its first-round financing exceeding 500 billion yuan, with a pre-money valuation of 71 billion US dollars (approximately 4.8 trillion yuan). Tencent, CATL, and the National Artificial Intelligence Industry Fund have all participated in the investment.

Coupled with the market catalyst from the upcoming Hong Kong listings of Zhipu AI and MiniMax, the entire AI sector has rallied. This article breaks down its capital logic, technical strength, and deep-seated listing risks.

01 Frenzy of Capital

DeepSeek's capitalization path is a microcosm of the surging valuations in China's AI industry.

Since rumors of a 10-billion-US-dollar valuation first emerged in April 2026, its post-money valuation reached 52 billion US dollars after the first-round financing closed in June, and its pre-money valuation jumped to 71 billion US dollars during negotiations for the second round in July. In just three months, its valuation has multiplied sevenfold.

The total size of this first-round financing has exceeded 500 billion yuan. Founder Liang Wenfeng personally contributed 200 billion yuan to become the largest investor. Tencent put in 100 billion yuan, CATL invested 50 billion yuan, JD.com, NetEase, and IDG Capital each contributed 30 billion yuan, while the state-backed National Artificial Intelligence Industry Fund participated in the investment simultaneously. Industrial capital, financial capital, and policy capital have all gathered in its shareholder roster.

To maintain control over its core technology, DeepSeek has designed a differentiated investment structure. Apart from the National AI Industry Fund, other industrial investors are only entitled to profit-sharing rights, with no board seats or voting rights in business operations. All external shares are subject to a mandatory 5-year lock-up period with no transfer allowed. Liang Wenfeng holds nearly 78% of the control through his capital contribution and structural design, which also reflects that institutional investors prioritize long-term technological realization over short-term capital arbitrage.

For its listing destination, DeepSeek has abandoned the Hong Kong Stock Exchange route and chosen the Shanghai STAR Market, relying primarily on the fifth set of listing standards specifically for AI large models issued by the Shanghai Stock Exchange on June 17. This regulation does not require enterprises to meet traditional profitability indicators, only mandating core technological advantages, large-scale product implementation, complete compliance filings, and other rigid conditions. It opens a green listing channel for hard-tech AI enterprises still in the loss-making investment phase. The company has already initiated due diligence by accounting firms, striving to prepare a full set of financial application materials by the end of December.

In terms of market sentiment, the wealth effect from Zhipu AI and MiniMax's Hong Kong listings earlier this year has laid the groundwork for this IPO plan.

After its listing, Zhipu AI's share price peaked at HKD 2,980, with its total market capitalization once exceeding HKD 1.27 trillion. MiniMax surged 109% on its first trading day, reaching a peak market capitalization of HKD 410 billion.

Driven by DeepSeek's IPO news, AI-related stocks rallied collectively at the Hong Kong market close on July 15. MiniMax rose by over 13% in a single day, Zhipu AI climbed more than 6%, its investor Tencent Holdings jumped nearly 4%, and Alibaba also closed up over 2%. Capital continues to flow into the listing dividends of leading independent large model developers.

02 Technology as the Foundation

If capital is the "accelerator" for DeepSeek's IPO sprint, then technology is the "ballast" that underpins all of this.

DeepSeek originated from HiFunda, a leading domestic quantitative private equity fund, and officially became an independent operation in July 2023. In its early stage, its R&D was entirely funded by the stable cash flow generated from the parent company's 700-billion-yuan assets under management.

In 2025, HiFunda achieved a return rate of 56.55%, with annual revenue of approximately 8.6 billion yuan. Founder Liang Wenfeng holds 85% of the company's shares, receiving billions in dividends each year. This provides a steady stream of self-owned funds for the large model's high-intensity investment in computing power and talent, distinguishing it from most AI startups that rely on external capital burn.

In terms of technical architecture, DeepSeek focuses on the Mixture-of-Experts (MoE) model, positioning itself for low cost and high performance. It has iterated a full product matrix covering DeepSeek-V2, V3, R1, and V4, spanning general conversation, code generation, mathematical reasoning, multimodal interaction, and lightweight industrial deployment scenarios.

Among these, the DeepSeek-R1 model, which competes with OpenAI o1, had a training cost of only 5.576 million US dollars, far lower than its overseas counterparts. Its open-source version has seen a 300% increase in GitHub stars over the years, attracting massive global developers with its extreme cost-performance ratio, earning it the industry nickname "Pinduoduo of the AI world".

Its commercial layout has established a three-tier business system: the C-end offers lightweight AI assistants for individual users, the B-end focuses on private customized deployment for government and enterprise clients, and the underlying layer charges global developers for API access based on open-source technology.

Citing data from industry insiders, foreign media reported that DeepSeek's current annualized revenue reaches 400-500 million US dollars, with core revenue coming from paid API calls for its V4 series. This business segment has a gross profit margin exceeding 50%, making it the only mature revenue source at present.

However, the issues of a single revenue structure and lack of large enterprise clients have also laid the groundwork for valuation disputes.

To address the industry's pain point of independent innovation in computing power, DeepSeek plans to allocate a large portion of its financing to building GW-level self-operated intelligent computing centers and developing self-designed AI inference chips. It is implementing a dual-track strategy of "purchasing high-end overseas GPUs while adapting domestic chips". Its V4 model has already completed adaptation and deployment on Huawei chips, gradually reducing reliance on overseas supply chains and building a moat of independent computing power.

Looking at the broader industry landscape, China's general large model sector is shifting from a parameter race to a commercial implementation phase. Digital transformation for government and enterprises, intelligent manufacturing upgrades, and AI transformation of internet products are generating massive demand.

Industry research data shows that from 2026 to 2030, China's large model market will maintain a compound annual growth rate of over 45%, with model API calls, industry-specific customized solutions, and computing power leasing becoming the three core revenue growth drivers.

Currently, there are fewer than 10 independent large model enterprises in China that possess complete underlying self-development capabilities and have achieved large-scale commercial application. DeepSeek has secured a leading position in the track with its technological iteration speed and cost advantages. Tencent and CATL's investments are intended to open up two major implementation scenarios: the consumer internet and the new energy industry, forming a closed loop of industrial synergy.

03 The Critical Watershed

With the triple support of capital, technology, and policy, as DeepSeek advances toward its A-share IPO, whether its ultra-high valuation of 4.8 trillion yuan can be realized depends on its post-listing performance.

DeepSeek currently generates revenue only from scattered API orders and has not yet landed large benchmark projects for government and enterprise clients. Its management has also publicly stated that it will prioritize investment in cutting-edge technological R&D and defer short-term commercial monetization, making the timeline for profitability unclear.

General large models must strictly complete algorithm filing, training data copyright verification, and data security reviews. The fifth set of standards on the STAR Market imposes extremely strict requirements on information disclosure verification for AI enterprises, mandating full disclosure of algorithm architecture, data sources, and intellectual property ownership. Any compliance flaw will hinder the IPO process.

In addition, the continuous rise in salaries for high-end AI talent in the industry and the risk of core technical team attrition can easily cause disruptions in technological R&D.

Overall, DeepSeek is a benchmark asset in the capitalization era of domestic large models. It has captured the triple dividends of industrial upgrading, capital enthusiasm, and policy relaxation, with a high-quality listing narrative logic.

Subsequent investors need to focus on tracking three core indicators: API paid conversion rate, the progress of self-operated computing power deployment, and the scale of large government and enterprise contract signings. Only by completing the closed loop from technological R&D to sustained profitability can it break away from the attribute of concept speculation and truly grow into an AI leader with hard-core competitive barriers.

DeepSeek's IPO process will also become the critical watershed for the entire large model track to transition from capital frenzy to performance realization.

This article is from the WeChat official account "Insight IPO" (ID: dongcha-ipo), author: Sun Ningyu, editor: Jiang Zhou, published with authorization from 36Kr.