DeepSeek’s Liang Wenfeng built a reputation as the quiet force behind some of the most efficient AI models to emerge from China. Shy. Nerdy. More focused on code than cameras. Yet last week the executive found himself at the center of a storm after a nearly four-hour investor call from May leaked across Chinese social platforms and beyond.
The recording captured candid remarks on compute shortages, Huawei chips, Nvidia’s vulnerabilities and the narrow gap separating Chinese labs from their American counterparts. Within hours links vanished. An unverified PDF surfaced on GitHub. And reports emerged that the normally reserved founder was spiraling. Frustrated. Furious at the breach.
But the leak did more than expose private talk. It pulled back the curtain on a company that turned down outside money for years, poured resources into open-source models and now sits atop a $50 billion-plus valuation after its first major fundraising. The episode highlights tensions at the heart of China’s AI push. Talent wars. Export controls. The pressure to commercialize without losing focus on long-term research.
The Funding Structure That Locked In Control
DeepSeek closed that record round in recent weeks. Roughly $7.4 billion at a valuation exceeding $50 billion, according to multiple accounts. The deal carried unusual terms. Investors funneled money through a limited partnership managed by Liang himself. They received no voting rights. A five-year lockup applied to shares. And every participant underwent personal vetting by the founder’s team.
Liang put in $2.8 billion of his own money from his hedge fund High-Flyer Capital. Tencent committed $1.4 billion. CATL, NetEase, IDG Capital and others followed. Only China’s National AI Industry Investment Fund gained direct equity and board influence. The structure preserved Liang’s grip even as the company prepared for an IPO that could file as soon as late 2026.
Those terms reflected lessons from earlier losses. Key researchers had left for rivals. One contributor to the V3 model joined Xiaomi. Others moved to ByteDance and Tencent. In a May virtual meeting with prospective backers, Liang laid down a firm condition. No poaching DeepSeek staff. No encouragement to spin out new companies. The demand appeared non-negotiable.
CNBC reported the details in June. https://www.cnbc.com/2026/06/18/no-poaching-our-people-chinas-deepseek-reportedly-tells-investors.html. The article noted DeepSeek had shunned external capital until talent departures forced its hand. Research remained the priority. Commercialization came second.
Yet the funding success proved short-lived in momentum. Bloomberg News revealed days ago that DeepSeek told prospective investors it would pause new fundraising. The decision came after the leak. No paperwork had been signed for a second round. Talks that could have valued the firm near $71 billion went quiet. Reuters covered the shift on July 14. https://www.reuters.com/world/china/chinas-deepseek-considers-new-fundraising-after-first-round-ft-reports-2026-07-14/.
The timing raised eyebrows. DeepSeek had just secured its biggest check. Plans for a hiring spree and infrastructure build-out were public. Financial Times stories detailed the escalation in China’s AI talent war and the company’s interest in fresh capital one month after closing the first round. But the leak changed calculations.
Gizmodo broke the story of Liang’s reported distress on July 25. https://gizmodo.com/deepseeks-normally-boring-ceo-is-reportedly-spiraling-after-supposed-leak-2000790888. The piece described a man unaccustomed to public scrutiny suddenly facing viral quotes that struck some as impolitic. Analysts familiar with the technical details grew more convinced the transcript was genuine. References to exact Huawei Ascend 950 counts, compiler workarounds and internal assessments matched known realities too closely to dismiss.
And what exactly did the recording reveal? A blunt assessment of the AI race. China trails the United States mainly in resources, not talent. DeepSeek operates at tens of billions of activated parameters while top American models reach hundreds of billions. The company holds about 16,000 Huawei 950 chips and a handful of Nvidia equivalents for a total around 20,000 H100-class GPUs. Training frontier systems would require 200,000 such cards. Huawei’s production capacity remains constrained for years.
Liang reportedly told investors Nvidia was digging its own grave. CUDA’s dominance would erode as Chinese teams built alternatives like their own compiler and TileLang. A Huawei 950 supernode could eventually replace Nvidia’s GB200 or GB300 at a 4-to-1 ratio. The ecosystem gap might close inside a year. These points circulated rapidly on X after the leak. One detailed thread by @LuizaJarovsky on July 23 laid out 11 excerpts. Another post summarized the roadmap: language models to chain-of-thought reasoning, then agents, continuous learning and finally embodied intelligence.
The candor extended to strategy. DeepSeek pursues reasonable returns only. Profit maximization invites trouble. Open-source serves as a core tactic for reaching artificial general intelligence. Half the team still views OpenAI’s offerings as superior. Hallucinations yield to better post-training. Capital structure limits expensive labeling approaches common in the U.S. The gap with American labs sits at roughly two years but achieved with one-twentieth the compute. Restraint defines the approach. Avoid super apps. Skip video generation or world models if they distract from core goals. Focus talent on data. No overtime. No KPIs. Vision drives the organization.
Transcripts posted to Substack and Reddit captured the full scope. One on AI Proem dated July 23 linked to the now-removed WeChat original and provided an AI-translated version. https://aiproem.substack.com/p/must-read-deepseek-liang-wenfeng. Another analysis on Hello China Tech dissected the theory that compute explains nearly every shortfall in talent, models and applications. https://hellochinatech.com/p/deepseek-liang-wenfeng-transcript. The leak’s speed and scale surprised observers. Similar breaches had not occurred before. Speculation swirled over whether government pressure or private request prompted the rapid takedowns.
Liang’s personal trajectory adds color. Once ranked among China’s richest after DeepSeek’s rise, he sits above Dario Amodei on some lists. The hedge-fund background shaped his insistence on control. High-Flyer Capital funded early efforts without diluting authority. That same discipline now faces tests. DeepSeek eyes an IPO on the mainland. Preparations could lead to a filing before year-end and debut in 2027, Yahoo Finance reported on July 14. https://finance.yahoo.com/technology/ai/articles/china-ai-leader-deepseek-file-144602998.html.
Recent models keep the pressure on. V4 arrived in April with upgrades in reasoning, agentic capabilities and token efficiency. CNN noted the preview’s promise to rival OpenAI, Anthropic and Google. The company stresses long-term leadership over short-term monetization. Liang told investors in a later meeting that artificial general intelligence and open-source remain priorities despite computing limits. Storyboard18 covered those remarks three days ago. https://www.storyboard18.com/digital/deepseek-says-long-term-ai-leadership-matters-more-than-short-term-monetisation-105299.htm.
But the leak exposed vulnerabilities. Exact hardware counts. Internal doubts. Pricing math that favors volume over margins. Critics and supporters alike parsed every line. Some saw validation of Beijing’s efficiency focus. Others detected overconfidence in hardware substitution timelines. The founder’s frustration reportedly grew as online commentary spread. A man who avoided the spotlight now watched his words fuel debate from Paris AI summits he skipped to domestic WeChat groups.
So what happens next? DeepSeek may resume fundraising once dust settles. Hiring continues. Model development presses forward. The episode serves as reminder. In a sector defined by secrecy and speed, even careful founders cannot fully control the narrative once audio escapes. Liang’s team built models that shook markets and erased billions in Nvidia value on prior releases. Now the man behind them must manage fallout from his own voice.
China’s AI sector watches closely. Talent retention. Compute access. The balance between openness and protection. DeepSeek positioned itself as the low-cost, high-volume producer in a global race. The leaked call reinforced that thesis with data points few expected to hear publicly. Whether the pause proves temporary or signals deeper caution remains unclear. One thing is certain. The quiet CEO no longer operates in silence.
WebProNews is an iEntry Publication