Liang Wenfeng's 4-hour investment meeting speech leaked. This 34,000-word transcript suggests a sense of urgency or 'guilt' because his weakness was precisely targeted: DeepSeek's (especially R1) most proud achievements are open source and architectural innovation (like MoE/MLA/V3/R1's extremely low inference costs). But open source is a double-edged sword: once you open-source top-tier ideas, competitors (like Moonshot Kimi, Alibaba Qwen, Tencent, etc.) can copy and tune them within a month. The 'technical breakthrough' achieved with heavy investment might only have a 3-month lead. Without continuous leadership in the next generation, the previous advantage will sink into a swamp of mediocre homogenization.
[Transcript Summary]:
Welcome investors. We didn't start this company to maximize profit or go public. We started with goodwill toward the world, believing this is useful for humanity beyond money. Our vision isn't a slogan on the wall; it's how we actually operate. We are vision-driven, not KPI-driven. We don't have a formal organization; we are organized by vision. This has pros and cons, but it is our characteristic.
Why do we insist on open source? Because our vision requires it. Unlike others who might open-source out of necessity, for us, it is our original intention. We believe that for AI to succeed commercially, open source is beneficial. AI is so large—potentially 10% of human GDP—that no one can monopolize it. If you try to monopolize it, you will be abandoned by history. You need restraint.
We are a group of ordinary people who achieved extraordinary things through restraint. Open source is part of that restraint. We only seek reasonable profit. Our API pricing is based on a 10-month payback period for hardware costs. If we wanted to maximize profit, we could double the price without losing demand, but we chose to make it affordable. When we lowered prices, our employees cheered because they wanted the models to be useful to everyone.
Restraint is a strategy. By giving up some short-term gains, we increase the probability of achieving AGI. We don't want to be the next ByteDance or Tencent; we want AGI. Last year's C-end traffic and this year's B-end revenue are just by-products of our journey toward AGI. We don't worry about others deploying our open-source models to compete; the market is big enough.
On the technical roadmap: We believe the path is CoT (Chain of Thought) -> Agents -> Continuous Learning -> AGI Singularity -> Embodied AI. The next bottleneck is 'Continuous Learning'—making models learn like humans over time rather than just through static training. Once a model can learn continuously, it can iterate on itself, leading to the singularity.
On competition and resources: The gap with the US is about resources (chips), not talent. We have about 20,000 H-equivalent GPUs. We want to buy as many as possible. Domestic chips (Huawei) have a historical opportunity. The CUDA moat is being dismantled by AI-assisted coding and new languages like our TileLang. Within a year, the perception that domestic chip ecosystems are 'unusable' will be overturned. We are working closely with Huawei, though their capacity is currently limited.
Our core interest is team stability. As long as the team stays, we will achieve AGI. Everything else is secondary. We don't want to compete with big tech; we want to empower them. We focus on the 'main line' of AGI and ignore distractions like video generation or 3D, which we view as commercial products rather than intelligence breakthroughs.
