【徐瑾】梁文锋的商业逻辑与矽谷叙事截然不同。科技圈有人将他比作张小龙，但他骨子里更像一位务实的南方商人——重视成本、愿意分利、看得见大局。他明确表示API定价标准是伺服器十个月回本，融资所得优先全部换成显卡，并预判国内基座厂商最终将收敛至三至四家。这套在资源约束下的生存逻辑，强调「活下去比讲故事更重要」，在国内AI圈比任何愿景都更具分量。他对开源的态度同样值得关注：最强模型完整开源，自用版与开源版保持一致，不担心第三方分流，因为算力与成本才是真正的壁垒。用开放换生态、用成本优势守住核心，这是商人的算盘，而非布道者的愿景。

梁文錋对AGI经济规模的判断具有深刻洞见。他认为AGI最终将占整体经济的10%，而推动这一切的核心变量是成本。这一判断与互联网扩张史高度吻合——从伺服器到频宽再到云计算，每一次成本的数量级下降都带来一轮新的应用爆发，AI亦不会例外。他进一步指出，互联网格局将在AI时代重演：头部聚合平台吃掉大部分价值，长尾创作者分走剩余，被淘汰的是中间那批不上不下的企业。国内基座厂商收敛至三至四家，说的正是同一件事——行业洗牌后，活下来的人靠的是清晰的成本与规模逻辑。

在技术路线上，梁文錋展现出罕见的克制。他明确指出持续学习是当前核心瓶颈，而多模态和世界模型仅是配套组件，不值得投入核心资源。他认为AGI是渐进过程，不存在突变奇点，模型大约每两至三个月更新一版。这种节奏感本身带著历史感，是长期主义者才会有的表述。Naval Ravikant将AI定义为这个时代的杠杆工具，而杠杆的关键从来不在工具本身，而在于谁能以最低成本持续运转。梁文錋比大多数人更早想清楚了这件事，尽管风险同样真实存在。（关键数字： three months）

Liang Wenfeng's business logic stands in sharp contrast to Silicon Valley narratives. Some in the tech community compare him to Zhang Xiaolong, yet he is fundamentally a pragmatic southern Chinese businessman—cost-obsessed, willing to share profits, and able to see the bigger picture. He has stated explicitly that API pricing targets a server payback period of ten months, that funds raised are prioritised entirely for purchasing GPUs, and that domestic foundation-model vendors will ultimately consolidate to three or four. This survival logic under resource constraints, emphasising that "staying alive matters more than storytelling," carries more weight in China's AI sector than any grand vision. His stance on open source is equally notable: the strongest model is fully open-sourced, the internal and open-source versions remain identical, and he does not fear third-party diversion because compute power and cost constitute the real moat. Trading openness for ecosystem and defending the core through cost advantage—this is a businessman's calculus, not an evangelist's vision.

Liang Wenfeng's assessment of AGI's economic scale offers profound insight. He believes AGI will eventually account for 10% of the overall economy, with cost as the core variable driving this transformation. This judgement aligns closely with the history of internet expansion—from servers to bandwidth to cloud computing, each order-of-magnitude reduction in cost triggered a new wave of application proliferation, and AI will be no exception. He further argues that the internet's structural pattern will repeat in the AI era: dominant aggregation platforms will capture the majority of value, long-tail creators will claim the remainder, and the companies eliminated will be those stuck in the middle. The consolidation of domestic foundation-model vendors to three or four speaks to the same dynamic—after the industry shakeout, survivors will be those with clear cost and scale logic.

On technical direction, Liang Wenfeng displays a rare restraint. He identifies continual learning as the current core bottleneck, while classifying multimodal capabilities and world models as merely supporting components unworthy of core resource investment. He views AGI as a gradual process with no sudden singularity, with models updated roughly every two to three months. This sense of pacing itself carries a historical sensibility—a framing that only a long-term thinker would adopt. Naval Ravikant defines AI as the leverage tool of this era, yet the key to leverage has never been the tool itself but rather who can operate it continuously at the lowest cost. Liang Wenfeng figured this out earlier than most, though the risks remain equally real.