Zhu Xiaohu: The model API business is difficult to maintain high gross margins in the long term

Zhu Xiaohu, founding partner of Jinsha River Ventures, stated: The model API business is difficult to maintain high gross margins in the long term. Anthropic currently has a gross margin of about 60%, primarily due to the strong intelligence capabilities of the model, especially in coding scenarios where it had a significant advantage, but that advantage has begun to cross a critical point; when other models catch up, the price may only be one-tenth of its current level, making it hard to maintain a high premium. Even if it remains ahead in the next few years, if the intelligence is only 10% stronger but the price is 50% higher, users may not continue to pay. Intelligence will ultimately become more like electricity and water, and the long-term gross margin for model APIs may be between 10% and 20%. Once capabilities converge, price wars will be hard to avoid, and Anthropic's urgency to go public may also be related to the window period.

Zhu Xiaohu mentioned that the AI office market is definitely much larger than AI coding. AI coding was the first to emerge because engineers have a high acceptance of AI, tasks are clear, and results are easy to verify, but it is only part of office work; the truly larger market is the white-collar market, namely AI office, which is essentially a general-purpose agent, similar to Microsoft Office. The competitiveness of AI office products lies in Model + Harness + Product + Context, and once harnesses converge, competition will shift to Context.

Regarding Tencent AI, Zhu Xiaohu stated that Tencent is a latecomer, and the Hunyuan model has squeezed into the top tier domestically, with WorkBuddy becoming the leading office agent application in the domestic top tier. Jinsha River Ventures has purchased a WorkBuddy enterprise account because it uses WeChat for Work, and the organizational relationships, historical communications, as well as Tencent Meeting and Tencent Docs constitute a ready-made Context.