In May 2026, all eyes in China’s technology world are fixed on a single name — DeepSeek. This AI company, once defined by its self-imposed labels of “no fundraising, no commercialization,” detonated two explosive announcements within a single month. First came the April release of its next-generation flagship model DeepSeek-V4, followed in May by reports that it had initiated its first external fundraising round, with its valuation potentially breaking through USD 50 billion and its fundraising target ballooning to RMB 50 billion (≈ USD 6.94 billion). Liang Wenfeng — who once refused to meet with investors — has suddenly become the hottest pursuit target in the capital markets. From the “isolated island” of technological idealism to the “high ground” of competing capital, DeepSeek is undergoing a profound transformation. And behind this transformation lies a fundamental restructuring of the competitive landscape of China’s AI industry.
A Three-Stage Model Leap: From Matching GPT-5 to Defining a New Paradigm
To understand DeepSeek’s current valuation, one must first understand its position on the technology track. From R1 in early 2025 to V4 in April 2026, DeepSeek has maintained an extremely high density of technical achievement at an unusually low release frequency, completing the transition from pursuer to leader.
1) December 2025: V3.2 — Open-Source Model Matches GPT-5 for the First Time
On December 1, 2025, DeepSeek simultaneously released two official model versions: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. On public reasoning benchmarks, V3.2 reached the level of OpenAI’s GPT-5, released in August of the same year, falling only slightly short of Google’s Gemini-3.0-Pro. This was the first time a domestically produced open-source model had gone head-to-head with the world’s top closed-source models on reasoning capability.
The two models have clearly differentiated positioning. V3.2 is focused on strong cost-performance and everyday use, with reinforced Agent capabilities enabling it to autonomously complete complex work such as generating reports and writing code. The Speciale version, on the other hand, is a “lopsided genius” — it specializes in solving highly difficult mathematics problems and programming competitions, achieving gold-medal-level results across all four of the world’s top competitions: IMO 2025, CMO 2025, ICPC World Finals 2025, and IOI 2025. Its scores in ICPC and IOI correspond to the level of the second-place and tenth-place human competitors respectively.
The more critical technical breakthrough lies in V3.2’s integration of thinking into tool invocation for the first time. Previous large models could not invoke external tools while in thinking mode, but V3.2 supports tool invocation in both thinking and non-thinking modes, realizing a cyclical workflow of “think first, then invoke tools, then continue thinking based on results.” This reached the highest level among open-source models at the time for large-scale Agent tasks.
2) April 2026: V4 — An “Unannounced” Paradigm Leap
If V3.2 was about “matching,” then DeepSeek-V4, released on April 24, 2026, represents a genuine act of “defining.”
V4 contains two MoE model versions: V4 Pro and V4 Flash. The Pro version has total parameters reaching 1.6T, setting one-million-token context length as standard for the first time, while achieving computing power consumption per token at just 27% of V3.2 under the one-million-context condition, with KV cache occupancy at only 10%. In Agentic Coding evaluations, V4 Pro has reached the leading level among open-source models.
The most closely watched technical decision in V4 is its inclusion, for the first time in an official technical report, of Huawei’s Ascend 950 series chips (including the 950P/950PR and other models) alongside NVIDIA GPUs in its hardware validation list. This signals that DeepSeek has completed a full-stack migration and adaptation from NVIDIA’s CUDA ecosystem to domestically produced computing power. Huawei subsequently announced that its Ascend Ultra-Node products provide full support for DeepSeek V4.
This is not merely a display of technical muscle. The Ascend 950 series inference chips entered mass production in March 2026. Their single-card computing power is approximately 3 times higher than NVIDIA’s H20 — the version specially supplied to the Chinese market — while their procurement price is approximately one-third to one-quarter that of NVIDIA’s H200. In low-precision inference scenarios, domestically produced chips can demonstrate efficiency surpassing that of general-purpose GPUs through new homogeneous SIMD/SIMT designs and proprietary HBM. This comes precisely as the industry’s center of gravity shifts from “training dominance” to “inference as the decisive battleground,” and DeepSeek’s choice of domestic adaptation has landed precisely in step with the times.
A Fivefold Valuation Jump in 21 Days: An Unprecedented Fundraising Frenzy
If the technical breakthroughs represent DeepSeek’s fundamentals, then the fundraising frenzy of April to May 2026 is a story that has set the entire investment world’s pulse racing.
1) Four Valuation Rewrites
Since reports first emerged in mid-April that it had initiated its first external fundraising, DeepSeek’s valuation has been rewritten four times in just 21 days.
Early April: approximately USD 10 billion (≈ RMB 72 billion); DeepSeek initiates its first fundraising round, planning to raise no less than USD 300 million (≈ RMB 2.16 billion).
April 22: exceeding USD 20 billion (≈ RMB 144 billion); major internet giants including Tencent and Alibaba were reported to be in investment discussions.
May 6: approximately USD 45 billion (≈ RMB 324 billion); the National Integrated Circuit Industry Investment Fund (the “National Big Fund”) was reported to be in discussions to lead the investment.
Early May: the fundraising scale target rose to a maximum of RMB 50 billion (≈ USD 6.94 billion), with the post-investment valuation potentially surpassing USD 51.5 billion (≈ RMB 370.8 billion).
In RMB terms, the fundraising scale expanded from approximately RMB 2.2 billion (≈ USD 305.56 million) to RMB 50 billion (≈ USD 6.94 billion). Many industry insiders believe the initial valuation was excessively conservative and inconsistent with DeepSeek’s technical level and industry standing, but the speed and magnitude of the subsequent surge still exceeded most people’s expectations.
2) Who Is Chasing DeepSeek?
The lineup of capital in this fundraising round is formidable. According to the Financial Times, the National Big Fund is in discussions to lead DeepSeek’s first fundraising round, with other participants in negotiations including multiple internet giants and other state-backed funds. Tencent and other internet majors have also indicated investment interest.
Most notable is founder Liang Wenfeng’s own contribution. According to The Information, Liang Wenfeng plans to personally contribute up to RMB 20 billion (≈ USD 2.78 billion) in this fundraising round, representing a substantial proportion of the total amount to be raised. If the deal closes, it will break the record for a single fundraising round by a Chinese AI company.
3) The Deeper Significance of the National Big Fund’s Entry
The appearance of the National Big Fund is no coincidence. The fund’s first phase was established in 2014 with registered capital of RMB 98.72 billion (≈ USD 13.71 billion) and ultimately raised total funds of RMB 138.7 billion (≈ USD 19.26 billion), marking China’s first large-scale intervention in the semiconductor industry in the form of a national industrial fund. If it leads the DeepSeek investment, this will be the first publicly visible instance of the National Big Fund openly taking a position in a leading domestic large model company.
Some analysts have noted that this signals that national-level strategic capital is no longer content to be a bystander in the AI race. Considered alongside V4’s deep adaptation to domestically produced computing power, the Big Fund’s entry is simultaneously a capital action and a carrier of industrial strategy considerations.
4) Why Now? Three Layers of Real-World Pressure
Liang Wenfeng had cultivated an image defined by “no fundraising, no commercialization, no roadshows.” Backed by High-Flyer Quant — with an average return rate of 56.55% in 2025 and assets under management exceeding RMB 70 billion (≈ USD 9.72 billion) — DeepSeek was genuinely not short of money. Why then choose to seek capital at the most fiercely competitive moment?
First, computing power. The V4 series extended context to 1M tokens and began testing multimodal capabilities; the continuous iteration of frontier models keeps pushing computing power requirements upward. As industry insiders analyze, “frontier models are becoming less and less like writing a paper or training a model — reasoning capability, Agent capability, ultra-long context, and enterprise-grade stability will all continue to push computing power demands higher.”
Second, talent. According to disclosures in the V4 technical report, DeepSeek’s “research and engineering” team numbers approximately 300 people, with at least 10 confirmed departures, an attrition rate of approximately 3.3%. Although the percentage is not high, the confirmed departing core members include: Wang Bingxuan, a core author of the first-generation LLM (who joined Tencent); Guo Daya, a core researcher on R1 (who joined ByteDance); Luo Fuli, a core contributor to V2 and V3 (who joined Xiaomi); Wei Haoran, a core author of the OCR series; and Ruan Chong, a core contributor to multimodal work (who joined Yuanrong Qixing) — spanning core technical areas including base models, reasoning, OCR, and multimodal.
Third, and most critically, option pricing. For a company without external funding, stock options have only an internal book value that has not been verified by real money, lacking the liquidity and premium reference that comes with market pricing. Only by bringing in external funding can employee options obtain market pricing, making it possible to attract and retain talent. In other words, what Liang Wenfeng wants is not just money — it is a valuation anchor that makes “growing together with the company” concrete and tangible for the team.
A Changing Landscape: China’s AI Enters the “National Big Fund Era”
DeepSeek’s fundraising is not an isolated event. In the same period, Moonshot AI (Kimi) completed approximately USD 2 billion (≈ RMB 14.4 billion) in fundraising, with its post-investment valuation exceeding USD 20 billion (≈ RMB 144 billion); StepFun is valued at approximately USD 10 billion (≈ RMB 72 billion) and has dismantled its red-chip structure to accelerate a Hong Kong IPO. Together with Zhipu and MiniMax, which listed on the Hong Kong Stock Exchange on January 8 and January 9, 2026 respectively, four of the “top five large model companies” now have valuations exceeding RMB 100 billion.
This recalls the listing wave of the “AI Four Dragons” a few years ago. But today’s competitive landscape is entirely different. At the global level, OpenAI’s valuation stands at approximately USD 850 billion (early 2026 data), and Anthropic is in negotiations for a new round of fundraising at an approximately USD 900 billion valuation. The China-U.S. large model contest is rapidly converging toward the leading players.
China’s battleground is particularly distinctive. Internet giants such as ByteDance, Baidu, Alibaba, and Tencent continue to increase their investment, backed by capital, computing power, and ecosystem advantages. Leading unicorns such as Zhipu and MiniMax are competing in close proximity, and technological iteration and commercialization are both accelerating.
On the user side, according to QuestMobile’s “Q3 2025 China AI Application Market Report,” DeepSeek had 145 million monthly active users, second only to ByteDance’s Doubao at 172 million, firmly holding second place in domestic AI applications. March 2026 data showed DeepSeek’s monthly active users at 127 million, reflecting a certain downward trend, with ByteDance’s Doubao eating into the market at an even larger scale.
The National Big Fund’s entry adds a new variable to this competition. If DeepSeek’s first fundraising round closes smoothly, it will signify that national-level strategic capital has officially taken a seat at the table of frontier AI. This may have far-reaching implications for the capital structure, technical roadmap, and even the international competitive landscape of the entire industry.
Can V4’s technical breakthrough translate into commercial returns? Can the company in its post-funding form maintain its innovative vitality? These questions are far more complex than any funding press release.
From “no fundraising” to RMB 50 billion (≈ USD 6.94 billion), from a pure NVIDIA approach to Ascend adaptation, from silence to V4’s “unannounced” release — DeepSeek’s 2026 almost concentrates all the tensions of China’s AI industry: autonomy and openness, ideals and reality, technology and capital.
The DeepSeek team has quoted a passage from the Xunzi: “Not tempted by praise, not frightened by slander, following the right path, upright in oneself” — which can be read as a response to the world’s judgments of them. Yet when the funding closes, when more capital and expectation pour in, can this company maintain the steadfastness of “following the right path” under the spotlight?
This is a story about holding firm. But the next chapter of this story may no longer be about holding firm itself — it may be about how to find a new equilibrium in the midst of transformation.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.