DeepSeek Aims for AGI with $10 Billion on the Table: and Open-Source Remains Untouchable (at least for now)
Liang Wenfeng, founder of DeepSeek, has informed potential investors of the first external funding round of the lab about the intention to pursue AGI as a priority objective, continuing to release open-source models and forgoing short-term commercialization. The statement, reported by Bloomberg, emerged in the context of a fundraising of approximately 70 billion yuan, equivalent to 10 billion dollars, on a pre-money valuation estimated around 45 billion dollars.
This marks DeepSeek's first external capital ever. The Hangzhou lab, founded in 2023, has so far been entirely funded by High-Flyer Quant, Liang's own quantitative hedge fund, which closed in 2025 with an average return of 56.6%. That autonomy had been presented as a deliberate protection against pressure on the product roadmap. Today's turn reflects partly a matter of scale: the current size of DeepSeek's model training has exceeded what even a profitable hedge fund can sustain indefinitely.
Who is entering the round, and with what stakes
Among the likely investors is the National Artificial Intelligence Industry Investment Fund, a state vehicle that could contribute about 10 billion yuan, in addition to Tencent, IDG Capital, and Monolith Capital. Liang himself is expected to participate with a personal investment in the order of 20 billion yuan. Other details are currently unknown, and the identity of the investors, the final valuation, and the closing date have not yet been formalized, and the company has not released statements to Bloomberg.
The first rumors in April spoke of a fundraising of at least 300 million dollars on a valuation of 10 billion. Within a few weeks, the figures rose to 20 billion, then to 45 billion at the beginning of May after negotiations opened with the China Integrated Circuit Industry Investment Fund (the so-called "Big Fund"). Today's Bloomberg figure of 10 billion dollars as the fundraising size on a valuation of 45 billion confirms what emerged from the rumors.
AGI as a Framework, Open-Source as a Method
The investor presentation, as reconstructed by Bloomberg, leverages the research-oriented identity that produced the R1 model from January 2025. That release wiped out about 600 billion dollars of Nvidia's market capitalization in a single session, based on the idea that frontier reasoning models could be trained at a fraction of what U.S. labs were spending. Whether that price claim held up under verification has become a debate in itself, but the strategic point was made: a Chinese lab could keep pace with the frontier, and do so openly.
In April, DeepSeek released V4-Pro and V4-Flash, two Mixture-of-Experts with 1.6 trillion and 284 billion total parameters respectively, both under MIT license and with a context of 1 million tokens. The V4 family is optimized for Huawei Ascend and Cambricon, with DeepSeek deciding not to give Nvidia early access, confirming the not exactly relaxed atmosphere in China regarding AI models and U.S. technology.
The combination of open-source, AGI as a framework, and implications for the Chinese hardware ecosystem explains why a round of this scale attracts the state fund as a leading investor. This same combination should also capture the attention of Chinese regulators, who have been outlining how state oversight of foundation model developers should look for two years.