China's Big Fund Leads DeepSeek's First Funding Round at $45 Billion Pre-Money Valuation

(Source: Caijing AI Pai)

DeepSeek's model iteration requires continuous investment in large-scale training compute, and sufficient funding is a crucial source to support this long-cycle investment.

By Caijing reporter Zou Lu

Edited by Liu Yiqin

Caijing has learned that AI large model company DeepSeek's first round of financing is nearing completion. A person involved in the deal negotiations said this round is led by the National Integrated Circuit Industry Investment Fund (known as the "Big Fund"), with several market-oriented investment institutions also on the negotiation list.

Since April this year, the primary market has been buzzing with excitement, and DeepSeek has become a red-hot investment target. A person close to the deal said, "Everyone wants to get in touch with DeepSeek, but no one can guarantee they'll get in." Until the last moment, both the investors and the investment amount remain uncertain. However, the source recently told Caijing that the first-round investor list has been largely finalized by the end of this month.

This is the first time three-year-old DeepSeek has opened up to outside funding. A month ago, DeepSeek's valuation was still $20 billion. According to media reports on May 22, DeepSeek's latest funding round has risen to 70 billion yuan (approximately $10 billion), with a pre-money valuation of $45 billion. If this round is completed successfully, DeepSeek will become the company with the largest first-round financing in the history of China's AI large model industry.

Who wants in?

An investor familiar with DeepSeek's senior management initially understood that this round would be led by state capital and didn't consider investing. "The pie is already very big. If market-oriented institutions go in, the return on investment may not be very high." In financial circles, "white horse stocks" usually refer to large-cap companies with relatively stable performance. He judged that DeepSeek might be a "big white horse."

The turning point came after the May Day holiday this year. He noticed that DeepSeek was opening up opportunities to market capital. Several well-known investment institutions, including Monolith Capital and Hillhouse Capital, were intensively contacting DeepSeek. He quickly changed his stance, and his firm subsequently joined the negotiations.

In the VC circle, DeepSeek shares are hard to come by. There was once a rumor that Liang Wenfeng set up an email address requiring institutions to introduce themselves and explain their understanding of AI. The rumor has been debunked. The investor familiar with DeepSeek's senior management said that funds that don't know Liang Wenfeng personally or lack reputation basically can't get in.

Caijing has learned that in this round, some investors are attempting to use SPVs (Special Purpose Vehicles) to package assets into tradable securities and bring in new LPs (Limited Partners). A person close to the deal told Caijing that some lesser-known LPs who normally couldn't invest in DeepSeek could enter through such channels.

Investing through SPV penetration structures requires additional management and channel fees, so top funds rarely enter the financing pool this way. In recent years, as the global AI investment boom has risen, SPV penetration structures have been widely used in primary market transactions.

A deal person told Caijing that such operations only occur with particularly hot deals. "Even after paying a layer of channel fees, at DeepSeek's current valuation, you'll still make money when it goes public."

However, SPV transactions carry risks due to potential investment fraud. Last week, US AI unicorn Anthropic upgraded its ban on unauthorized share transactions, explicitly prohibiting SPV investments in Anthropic's past or future funding rounds, stating that transactions through unauthorized channels would be deemed invalid. This move shook the private equity secondary market, with some publicly traded funds holding SPV shares plunging in price.

Contrary to the market's overall optimism, some investors who never showed interest believe DeepSeek's current valuation is too high and "unfathomable." Participating in negotiations requires significant time and money, and the chances of actually getting in are too low. Considering the return on investment, they don't see this deal as worthwhile.

Chen Shi, a venture partner at FreeS Fund, told Caijing that investing in DeepSeek would certainly be good, but getting a tiny allocation through "particularly indirect" and uncertain relationships doesn't mean much. "Investing now probably won't result in losses, but I personally think it's more valuable to invest in opportunities you've discovered yourself, projects where you can secure larger stakes."

Several major tech companies also appear on the previously reported negotiation list. According to media reports, energy giant CATL is planning to enter the deal, strategically targeting the energy variables behind computing infrastructure. Media also previously reported that Tencent and Alibaba are negotiating with DeepSeek. Alibaba has publicly denied this, saying it has little interest. As for whether Tencent is still in contact, Caijing sought confirmation from Tencent but received no response.

Multiple AI investors analyzed that between Tencent and Alibaba, Tencent has more incentive to get in, as its Hunyuan model still hasn't caught up with the first tier, and a strategic partnership with DeepSeek would make sense. Alibaba, on the other hand, has built its own ecosystem from models to chips, so its strategic need to invest in DeepSeek is less compelling.

DeepSeek's Crossroads

Since 2025, several AI large model companies that survived the "hundred-model war" have been accelerating their financing and commercialization processes. In January this year, Zhipu Technology and MiniMax, both part of the AI "Six Little Dragons," were listed on the Hong Kong Stock Exchange. Zhipu's latest market cap is 6,353.26 billion Hong Kong dollars (approximately $81.093 billion), and MiniMax's latest market cap is 266.59 billion Hong Kong dollars (approximately $34.028 billion).

Moonshot AI recently completed a $2 billion funding round at a valuation exceeding $20 billion, and is dismantling its VIE and red-chip structures in preparation for an IPO. Among the leading AI large model startups, DeepSeek started fundraising latest, but its current $10 billion first-round financing far exceeds its competitors.

Regarding DeepSeek's rising valuation, many investors believe it remains within a reasonable range with room for growth. In terms of model inference efficiency, the industry widely recognizes DeepSeek as being at the forefront, maintaining low cost and high performance even after releasing new models. Taking the V4-Pro released in April as an example, the company disclosed that in a 1-million-token long-context scenario, single-token inference FLOPs are only 27% of the previous generation V3.2. This efficiency advantage is directly reflected in API pricing: V4-Pro costs 0.025 yuan per 1,000 tokens, among the lowest levels of mainstream global models.

How to commercialize is the critical question hanging over DeepSeek's high valuation. According to Caijing's previous report, DeepSeek executives revealed at a meeting in March that the outside world is closely watching DeepSeek's business model implementation and technological progress. DeepSeek has been making many efforts and attempts and has preliminarily verified some paths.

Multiple investment professionals in the AI field told Caijing that DeepSeek chose to open up financing for two main reasons: first, to let employee stock options be validated by the market to retain talent; second, the funds needed for model training will only increase, and the self-owned assets previously generated by High-Flyer Quant may not be able to meet R&D needs in the future.

Additionally, a person close to DeepSeek said that DeepSeek employees also want the company to go public.

Regarding whether DeepSeek should build products, an investor close to Liang Wenfeng told Caijing that Liang Wenfeng realized the importance of productization around the end of last year. "He wants to build products." The source said DeepSeek's product team is still small and has struggled to find a suitable leader. The company has struggled with its product direction. "Product matters — Liang Wenfeng needs to figure that out himself."

According to media reports from May 22, Liang Wenfeng stated at an investor meeting that the company will continue to advance its open-source AI models with the goal of achieving Artificial General Intelligence (AGI). In his view, technological breakthroughs are DeepSeek's core mission — pushing the boundaries of technology rather than pursuing profits.

The investor close to Liang Wenfeng believes that the technical ideal of achieving AGI and commercialization are not in conflict. A more practical consideration is that DeepSeek's model iteration requires continuous investment in large-scale training compute, and sufficient funding is essential to support this long-cycle investment.

Chen Shi believes that leading large model companies like Anthropic and OpenAI have top-tier business models in the industry, but they also have "the most high-pressure business models." This is because the Scaling Law is still at play — each generation of models costs more to train than the last, requiring continuous investment in compute, data, and talent. Once you fall off the table, the market forgets you. For AI large model startups, today's primary market can no longer meet the funding needed for model iteration. State funds can support part of the journey, but the next step can only be the secondary market.

DeepSeek has not previously focused on product development, with its focus remaining on model training. But today's market environment has changed — Coding and Agent applications all require substantial high-quality token consumption. "If DeepSeek decides to invest energy in commercialization, its revenue should also increase rapidly and won't be lower than its current competitors," Chen Shi said.

The state team leading this round is the most notable aspect of this financing. Song Xiangqing, vice president of the China Society of Commercial Economics, told Caijing that the Big Fund leading DeepSeek's first-round financing marks the first time in the fund's 12-year history that it has cross-sector invested in a pure large model company. The AI industry is entering a phase of "state team leadership, market capital following." He believes DeepSeek is rising to become a national-level AI strategic platform.

AI and computing power have been upgraded to a national strategy, which is also reshaping the capital structure of leading model companies. DeepSeek's deep integration with the domestic chip ecosystem is one of the backgrounds behind the state team's entry. The DeepSeek-V4 released in late April was the first to support Huawei's new-generation AI chip, the Ascend 950PR, meaning model training will no longer be entirely dependent on NVIDIA's CUDA chip ecosystem.

Changes in capital structure do not necessarily mean changes in direction. "Liang Wenfeng is a relatively pure person," a source close to him told Caijing. Even with the state team leading the way to the secondary market, DeepSeek's daily operations and technical direction won't necessarily be tied down by the capital market. Going public is just a capital operation, not an inflection point that changes direction.

However, as DeepSeek steps out of its "self-sufficient" technological utopia and faces the capital market, it cannot avoid the question that every tech startup must confront: how to balance technical ideals with commercial returns over the long term.