Liang Wenfeng Pours ¥20 Billion Into DeepSeek: Locking In Talent, Locking In the Ecosystem, and Going All In on AGI

The Founder's ¥20 Billion Self-Investment: A Fundraise That Rewrites the Rules

When news broke that Liang Wenfeng had personally injected ¥20 billion into DeepSeek, the investment world fell into a momentary stunned silence. This wasn't ¥2 million or ¥20 million — it was a founder using his own money to fund his own company. According to information compiled by creators on Bilibili, the investor lineup is nothing short of star-studded: Tencent contributed roughly ¥10 billion, CATL ¥5 billion, JD.com, NetEase, and IDG each ¥3 billion, and the National AI Industry Investment Fund ¥1 billion — yet even all of these giants combined still fall short of Liang's personal contribution.

Liang's financial firepower traces back to High-Flyer (幻方量化), the quantitative hedge fund he founded — one of China's leading quant funds, which at its peak managed over ¥100 billion in assets. Quantitative trading relies on mathematical models and computer algorithms to execute automated trades in financial markets, with massive GPU compute power at the core for processing enormous volumes of market data. Because of this inherent demand for GPU compute, High-Flyer began large-scale procurement of NVIDIA GPU clusters around 2019, accumulating deep operational expertise in AI infrastructure. This explains why DeepSeek, from its very inception, possessed far greater compute resources and financial backing than a typical startup — it essentially grew out of a fintech foundation.

Even more intriguing is the structure of the deal. According to reports, investors receive virtually no voting rights and face a five-year lock-up period. The only party with voting rights is the national industry fund — a privilege granted specifically because of its government-backed nature. All other institutions play the role of passive financial investors.

If traditional fundraising means a company ceding equity to outsiders in exchange for capital, this round looks more like a redefinition of the rules entirely. Liang Wenfeng is making a statement through his actions: the power to set DeepSeek's direction stays firmly in his hands.

The Real Purpose Behind DeepSeek's Fundraise: Not Capital, But Retention

Anyone familiar with DeepSeek knows the company launched with a founding motto: "no fundraising, no capital needs, no IPO." Liang Wenfeng comes from the quant trading world, sits on ample cash reserves, and has never hesitated to spend on GPUs and servers. So why break his own rule?

The answer comes down to two words: talent retention.

Over the past two years, DeepSeek has become a prime poaching target for China's tech giants. According to leaked information, the lead authors of DeepSeek's first-generation large language model were recruited away by Tencent; a researcher known in the industry as "Rooftop Girl" (罗福莉) was reportedly courted personally by Xiaomi's Lei Jun at a premium price; and another key team member jumped to ByteDance earlier this year. The offers weren't modest bumps in pay — they were outright multiples of existing compensation, with total packages routinely hitting eight figures, no negotiation needed.

The core issue is options pricing. Stock options are the most common talent incentive tool at startups, granting employees the right to purchase company shares at a pre-agreed price in the future. Their value depends entirely on how much the company's valuation grows. In Silicon Valley and China's tech industry, options have been the key mechanism for attracting top talent to early-stage companies — early employees at Facebook, Google, and ByteDance all built substantial wealth through options. But options have a fatal weakness: liquidity. If a company never raises funding and never goes public, options are just numbers on paper — impossible to cash out. DeepSeek's salaries are not low, but as a company that had never raised external capital, the question of what those options were actually worth had no answer. They could be worthless scraps, or they could be a goldmine — entirely dependent on an unknowable future. And young people can't afford to wait: mortgages, living costs, and concrete high-paying offers from major tech companies make the rational choice obvious.

Liang Wenfeng understood this clearly. Only by bringing in external investors could the company's valuation be established by the market. Once a valuation is set, core employees' options immediately become quantifiable in value. He reportedly also laid down a hard rule for investors: no poaching DeepSeek staff, and no encouraging team members to go off and start their own companies. Cross the line and you're out. This fundraising round is fundamentally a lock on the core team — and Liang Wenfeng holds the only key.

The Four-Hour Closed-Door Meeting: DeepSeek Is Going All In on AGI

What truly ignited discussion across the internet was a reportedly four-hour recording of a closed-door meeting. In it, Liang Wenfeng clearly defined DeepSeek's boundaries: no multimodal AI, no world models, no consumer-facing products, and even popular B2B verticals like finance and healthcare are not high priorities.

So what does DeepSeek actually want to build? The answer: Artificial General Intelligence (AGI), via the path of open source.

AGI refers to an AI system with generalized cognitive capabilities on par with or exceeding humans — one that can autonomously learn, reason, and solve problems across any intellectual task, rather than excelling only in specific domains like chess or image recognition. Today's large language models — GPT-4, Claude, and the like — are classified as "narrow AI." Despite their impressive capabilities, they are fundamentally systems based on statistical pattern matching, lacking genuine understanding, planning, and autonomous goal-setting. AGI is regarded by industry leaders like OpenAI's Sam Altman and DeepMind's Demis Hassabis as the ultimate destination of AI development — though the timeline for achieving it remains hotly debated. By targeting AGI directly rather than commercial applications, DeepSeek is positioning itself as a foundational research institution rather than a product company.

What does open source actually mean? Consider this analogy: a large AI model is like a master who has trained for years, with hundreds of billions of "moves" (parameters) stored in their mind. Under a closed-source model, if you want to use their capabilities, you follow their rules — and pay for the privilege. Open source, on the other hand, is like that master writing their life's knowledge into a manual and posting it online for anyone to download, modify, and use commercially.

In the recording, Liang Wenfeng directly called out ByteDance and other companies employing closed-source strategies, saying bluntly: "I don't see what's good about being closed source." Behind this statement lies a strategic calculation rooted in the broader US-China AI competition.

The Compute Constraint and Open-Source Strategy: DeepSeek's Ecosystem Play

Liang Wenfeng's read on the US-China AI competition is sharp: at the talent level, the gap is almost nonexistent — the same cohorts of Tsinghua and Peking University graduates circulate between Silicon Valley and Chinese teams. The real gap is in compute, specifically in chips.

According to the source material, domestic chips are actually adequate in performance — roughly four Huawei cards can match one NVIDIA card — but the bottleneck is production capacity. Huawei's annual chip supply to major internet companies is around 100,000 units, and DeepSeek only receives approximately 16,000 — far too few for large-scale model training, and money can't solve the scarcity.

Given constrained compute, using every chip wisely becomes critical. Under a closed-source model, a company must burn enormous compute on inference and deployment for users, plus build internal teams to handle optimization — burning through precious GPU resources on service operations. Under an open-source model, the global developer community will spontaneously handle inference optimization, engineering adaptation, and other unglamorous heavy lifting. The company's own compute can then be concentrated entirely on training the next generation of models.

The deeper calculation involves ecosystem lock-in. The global AI development landscape is currently deeply locked into NVIDIA's CUDA ecosystem. CUDA (Compute Unified Device Architecture), launched by NVIDIA in 2006, has over nearly 20 years become a massive and difficult-to-displace software ecosystem: virtually all mainstream deep learning frameworks (PyTorch, TensorFlow), scientific computing libraries, and AI inference engines are built on top of CUDA. This deep entrenchment means that even as domestic chips like Huawei's Ascend series gradually close the hardware performance gap with NVIDIA, developers face enormous code migration costs — large volumes of existing AI code simply cannot run in non-CUDA environments.

According to the source material, Liang Wenfeng's team is working to rebuild a CUDA-like development environment on Huawei hardware using the Triton language. Triton is a GPU programming language open-sourced by OpenAI, designed to provide a high-performance compute abstraction layer that doesn't depend on any specific hardware vendor. Once DeepSeek's open-source user base grows, Huawei will have a strong incentive to supply more chips and resources — a mutually reinforcing "ecosystem co-building" dynamic. By tying its fate to domestic chips, DeepSeek also carves out its own space to survive and thrive.

Technical Vision: From Parameter Stacking to AI Self-Evolution

On the technical roadmap, Liang Wenfeng has articulated a conviction that transcends today's mainstream paradigm: the next breakthrough won't come from continuing to scale up parameter counts, but from enabling AI to continuously learn, self-reflect, and self-evolve.

This view is set against the backdrop of the Scaling Law — the dominant breakthrough paradigm of the past several years in AI. Proposed by OpenAI researcher Jared Kaplan and colleagues in 2020, this theory established that model performance follows a predictable power-law relationship with parameter count, data volume, and compute — in short, "bigger is better." This law drove every leap from GPT-2 to GPT-4 and fueled the global compute arms race. But since the second half of 2024, the field has been grappling with the idea that Scaling Law is hitting a ceiling — the marginal returns from simply adding more parameters are diminishing, while training costs continue to rise exponentially.

The direction Liang Wenfeng points to closely aligns with several frontiers being explored in academia: test-time compute (letting models do more "thinking" during inference), self-play (self-improvement through competition, similar to AlphaGo), and self-improvement loops based on reinforcement learning. DeepSeek's previously released R1 model is a prime example of this direction in practice — it introduced a "Chain of Thought" mechanism at the inference stage, representing an early proof-of-concept for this technical approach.

This is a road the entire global field is exploring, with no settled answers yet. And until that destination is reached, the most pragmatic strategy right now is to reduce costs, improve results, and accelerate iteration. Open source is seen as the optimal path to achieve this — generating revenue isn't the primary concern (the money is already there). Rather than converting compute into short-term profits, the better play is to "make broad alliances" — invite the global developer community to participate together and accelerate the arrival of AGI.

Lock In Talent Through Funding, Lock In the Ecosystem Through Open Source: A Formidable Grand Strategy

Taken together, these moves form a self-consistent logical loop: fundraising to lock in talent, open source to lock in the ecosystem. Liang Wenfeng's goal is not to compete for market share or users — his sights are set on AGI itself. In his game, money, chips, and people are all instruments and pieces on the board.

One detail worth noting: Tencent, CATL, JD.com, and NetEase — all calculating, strategically-minded industrial investors — were willing to accept a five-year lock-up and surrender their voting rights. That alone is a powerful signal. Each company has clear strategic motivations: Tencent is embedding AI capabilities across WeChat, gaming, and cloud services, and investing in DeepSeek secures a priority channel to frontier model capabilities; CATL, the world's largest battery manufacturer, is actively applying AI to battery materials research, smart manufacturing, and autonomous driving; JD.com and NetEase have strong needs in e-commerce logistics intelligence and game AI respectively. Notably, this "strategic investment with minimal control rights" model is extremely rare in Chinese tech investment — investors typically demand board seats or veto rights. The willingness of these giants to forgo control reflects both DeepSeek's scarcity value and the urgency of "picking a side" in today's Chinese AI competitive landscape. What they're betting on isn't DeepSeek's next product — it's Liang Wenfeng the person, and the question of whether China can carve out a path on the AGI track.

Of course, as interpretive content originating from social media platforms, some of the specific figures and details above — such as exact investment amounts, options valuations, and the contents of the closed-door meeting recording — have yet to receive official confirmation. Readers should maintain independent judgment. But regardless of how the details shake out, DeepSeek's relative silence in recent months may not signal a retreat — it may simply be gathering momentum for the next technological leap. The outcome of this high-stakes bet still awaits the verdict of time.