DeepSeek raised more than 50 billion yuan, roughly $7.4 billion, in June in its first-ever round of outside financing, valuing the Chinese AI company at more than $50 billion and making it, by that measure, China’s most valuable AI startup. Until this round, DeepSeek had taken no external investment at all, funded entirely by founder Liang Wenfeng out of profits from High-Flyer, the quantitative hedge fund he also built, even through the January 2025 release of its R1 model that briefly rattled US tech stocks and forced a global reassessment of how far behind China’s AI labs actually were.
Liang’s path to DeepSeek runs entirely through that hedge fund. He co-founded High-Flyer in 2015 with two college classmates after studying at Zhejiang University, building it into a quantitative fund that peaked at roughly $14 billion in assets under management in 2021 and still manages around $8 billion today. In 2021, he began buying Nvidia chips for what was then an AI side project, and by his own account in a 2023 interview, High-Flyer had already acquired 10,000 Nvidia A100 GPUs before the US government’s export restrictions on advanced AI chips to China took effect, a stockpile that gave DeepSeek a computing head start most of its Chinese rivals didn’t have when it formally launched in July 2023. Liang personally holds only about 1 percent of DeepSeek directly; the remaining 99 percent sits inside a High-Flyer-controlled investment vehicle, a structure that made this June’s round, in effect, the first time DeepSeek’s ownership has been meaningfully diluted at all.
The scale of what R1 actually did to markets in January 2025 is worth remembering as context for why this funding round matters. When DeepSeek released R1 on January 20, 2025, claiming performance rivaling OpenAI’s o1 at a fraction of the training cost, Nvidia’s stock fell 17 percent in a single trading day one week later, erasing nearly $600 billion in market value, the largest one-day value loss for any company in history at that point. Part of the shock was pricing: DeepSeek’s API charged $0.55 per million input tokens and $2.19 per million output tokens, dramatically undercutting OpenAI’s equivalent rates of $15 and $60. That was the moment DeepSeek stopped being a curiosity and became a company US markets had to price in.
The round’s structure this time is unusual enough to be the real story. Most of the new capital flows through a limited partnership controlled by Liang himself, and investors buying into that vehicle receive no voting rights and are locked in for five years. Liang personally contributed the largest single check, roughly $3 billion, or 20 billion yuan. Tencent put in about $1.5 billion, CATL, the battery manufacturer, contributed roughly $740 million, and NetEase, JD.com, and IDG Capital rounded out the round’s corporate backers, all through the same non-voting structure.
There is exactly one exception. China’s state-backed National AI Industry Investment Fund invested directly into DeepSeek itself rather than through Liang’s limited partnership, which means it is the only outside party in the entire round with actual voting rights and no lock-up restriction. Every other investor, including Tencent, one of China’s largest technology companies, has capital at risk in DeepSeek with no formal say in how the company is run.
That structure sharpens a question that’s followed DeepSeek internationally since R1’s launch: how closely is the company’s incentive structure actually tied to Chinese state priorities. With the state investment fund holding the only meaningful governance stake among outside investors, the answer looks less ambiguous than DeepSeek’s public positioning as a scrappy, founder-led lab would suggest. Liang himself offered a version of this framing back in 2023, telling reporters at the time, “Money has never been the problem for us; bans on shipments of advanced chips are the problem,” a line that still holds: this round wasn’t raised because DeepSeek needed cash, it was raised to lock in a specific set of strategic backers.
Chip access has, if anything, become more openly contested since. A senior Trump administration official alleged in late February 2026 that DeepSeek had trained a newer model on Nvidia Blackwell chips at a data center in Inner Mongolia, chips explicitly banned from export to China, with the suspected route running through smuggling networks exploiting gaps in US export enforcement. DeepSeek has responded not by denying the pressure but by leaning into it, confirming plans to design and manufacture its own AI chips and recruiting semiconductor talent toward that goal, an unusually direct admission that reliable Nvidia chip access can no longer be assumed even with billions in fresh capital sitting in the bank.
The funding lands as DeepSeek’s open-weight models remain among the most widely used low-cost alternatives to Western frontier systems for developers who can’t or won’t pay Claude or GPT-class API pricing, a trend that’s accelerated further this year as other Chinese labs, notably Zhipu’s Z.ai with its GLM-5.2 model, have pushed the same cost-performance argument even harder. This round signals DeepSeek intends to keep releasing competitively rather than fade after its initial 2025 moment, and it now has real capital, if not a diversified investor base or guaranteed chip supply, behind that intention.
DeepSeek’s models remain genuinely useful and genuinely cheap for Philippine developers and startups building AI features on tight margins, and nothing about this funding round changes that technical reality. But the governance structure is a real due-diligence flag for any Philippine company in a compliance-sensitive category, BSP-regulated fintech, anything handling personal data under the Data Privacy Act, weighing whether to build core infrastructure on DeepSeek’s API or self-hosted models. The company’s only outside party with actual accountability and voting power is a Chinese state investment vehicle, not a normal cap table of shareholders with fiduciary duties to a diversified investor base, and that’s a different risk profile than building on Anthropic or OpenAI, whatever one thinks of those companies’ own governance, or their own current fight with Washington over chip supply and model review.
Share this article