DeepSeek Founder Liang Wenfeng’s Leaked Investor Meeting: Nvidia’s CUDA Moat, China’s Chip Future, and an AGI-First Strategy

Nvidia's CUDA Moat Is 'Crumbling'

Liang argued forcefully that Nvidia's software ecosystem—long considered an unbreachable competitive advantage—is being rapidly dismantled. He stated that AI itself now makes it far easier to write code for non-Nvidia hardware, collapsing the lock-in advantage . He specifically cited tools like TileLang and AI-powered code generation that can translate CUDA-optimized workloads to run on competing chips .

Liang characterized Nvidia's dominance as self-undermining: "Nvidia is digging its own grave," he said, because the very rise of AI lowers the switching cost away from CUDA . He predicted that meaningful cracks in Nvidia's ecosystem will appear within about a year . DeepSeek itself trained its V3 model on Nvidia hardware without relying on Nvidia's software ecosystem, instead using its own high-level compiler, TileLang, for nearly everything above the silicon .

China's Domestic Chip Ecosystem: 'Historic Window'

Liang was emphatic that China's domestic AI chips—from Huawei, Alibaba, Moore Threads, and others—face no fundamental hardware or ecosystem adaptation obstacles. He repeatedly stated that the only real constraint is insufficient production capacity .

He predicted that within one year, real-world deployment would overturn the market's perception that domestic chips are "unusable or hard to use" . Even a 1–2× price premium versus Nvidia would be acceptable, he said . Liang framed this as a "historic win" opportunity, arguing that U.S. chip export controls have forced the Chinese ecosystem to build alternatives, creating a unique window .

DeepSeek's Hardware Constraints and Huawei Partnership

Liang openly acknowledged that the main gap between U.S. and Chinese AI labs is compute access, not algorithm or talent. Chinese labs face far tighter GPU supply due to U.S. export restrictions .

However, he disclosed that DeepSeek has already collaborated closely with Huawei, running its production ecosystem on tens of thousands of Huawei 910B and follow-on "950" chips. He said the performance-per-price ratio of these Huawei clusters can already "flat-replace" high-end Nvidia products like the GB200 and GB300 . DeepSeek secured roughly 16,000 Huawei 950 chips for this effort .

Liang predicted that chip supply should no longer be a binding constraint for DeepSeek within roughly 12 months . The company also plans to build its own large AI computing clusters in the future .

AGI-First Strategy: No Profit Targets, No Distractions

Perhaps the most striking part of the meeting was Liang's articulation of DeepSeek's strategy. He said the company was never founded to "make big money, go public, or cash out" . Instead, achieving AGI is the singular goal, prioritized above user growth and short-term revenue .

Liang stated that DeepSeek will likely keep its most advanced models open-source, arguing that open development and commercial monetization are not contradictory . "Restraint is a strategy used to increase the probability of achieving AGI," he said .

He explicitly stated that DeepSeek will not invest resources into "3D, video generation, world modeling," consumer apps, or enterprise products at this stage. Products exist only as stepping stones toward AGI . The current focus is on a Coding Agent and continuous learning capabilities, with the next phase aiming to achieve autonomous model iteration .

The company operates without revenue maximization goals, without near-term IPO plans, and without conventional business KPIs—a "KPI-free, vision-driven" organization . The entire investor meeting was designed to align new investors with this non-commercial, AGI-first philosophy during the $10 billion fundraise .

Key Caveats

The transcript was leaked without DeepSeek's authorization and has not been officially verified by the company . Reuters and Bloomberg confirmed its authenticity through sources familiar with the matter, but attribution carries that standard caveat . As with any leaked internal communication, the content should be read as a single snapshot of the founder's thinking at a specific point in time, not as a formal company statement.