Key Takeaways from Liang Wenfeng’s 2026 Investor Meeting on Large‑Model Strategies

During the investor communication on May 20 (recorded) and compiled on July 16, 2026, Liang Wenfeng focused on large‑model technology routes, compute supply, and commercialization speed. He argued that the industry should avoid blindly chasing ever‑larger parameter models and instead prioritize real‑world effectiveness, algorithmic architecture optimization, and inference efficiency. He highlighted current constraints in high‑end compute, the progress of domestic compute infrastructure, and cost‑control strategies, while acknowledging remaining challenges in hardware‑software integration.

On the product side, the team concentrates on practical models for industry scenarios, balancing general capabilities with domain‑specific customization to avoid pure, application‑agnostic technical stacking. Strategically, the company emphasizes scenario‑driven revenue validation over uncontrolled expansion, focusing on government and enterprise projects, and predicts that AI competition will shift from parameter battles to engineering excellence and sustainable business models.

The article also references a curated collection of 52 related analyses that delve into topics such as DeepSeek V4 inference cost comparisons (H100 vs. Ascend 950PR/910C), domestic AI chip adaptations, supernode architecture debates (Ethernet, InfiniBand, NVLink), PCIe 8.0 reaching 1 TB/s, memory‑pooling techniques for trillion‑parameter training, and various supernode hardware evaluations (ScaleX640, Atlas 950/960, etc.). These resources provide deeper technical context for the meeting’s insights.