Most AI companies talk about changing the world and then quietly optimize for quarterly revenue. DeepSeek, the AI lab that emerged from a Chinese hedge fund, is making a different bet: it wants to actually build artificial general intelligence, and it says the money is just fuel for that mission.
Founder Liang Wenfeng told investors in May 2026 that DeepSeek’s north star is fundamental AI research, not commercialization. The company plans to keep its best models open-source, a posture that costs revenue but wins something arguably more valuable right now: credibility.
A $10 billion vote of confidence
DeepSeek is in the middle of a funding round estimated at roughly 70 billion yuan, or about $10 billion, which would push the company’s valuation to approximately $45 billion.
The backing comes from China’s state AI fund and High-Flyer, the quantitative hedge fund where Liang built his fortune before pivoting to AI research.
Open-source as a strategy, not a charity
DeepSeek has already released models including R1 and V4 variants, at price points that made Western competitors visibly uncomfortable when the benchmarks dropped. Its mixture-of-experts architecture scales to 1.6 trillion parameters, which is a technical way of saying the model is very large and very capable.
Liang’s commitment to keeping future top models open-source is strategically interesting. Open-source AI creates a gravitational pull. Developers build on it, researchers cite it, and companies embed it into their products. Most Western AI labs have moved in the opposite direction, locking down their most capable models behind API walls and subscription tiers.
What this means for the AI competitive landscape
DeepSeek represents a distinct model of AI development: state-adjacent backing, research-first culture, and open-source distribution as a moat-building strategy rather than a monetization shortcut.
Western AI labs now face a rival that is not trying to beat them at their own game of enterprise contracts and API pricing. DeepSeek is playing a different game, one where the scoreboard is research citations, parameter counts, and community adoption.