预测：2027年前中国AI创业公司将迎来第二波淘汰潮
一条挺敢说的预测：到2027年底中国AI创业公司可能只剩两家活得滋润，DeepSeek排第一，智谱和月之暗面争第二，判断依据是GLM 5.5和K3.1的表现。
作者预测到2027年Q1中国AI创业公司将出现第二波倒闭潮，原因是算力资本开支呈指数增长、国内算力稀缺，以及'略不同但更慢更贵'的模型产品市场空间消失。作者预计到2027年底可能只剩两家仍在增长的头部创业公司，第一是DeepSeek（采用MIT开源旗舰模型策略），第二最可能是智谱Zhipu，其次是MoonshotAI，并认为GLM 5.5和K3.1的表现将决定两者排序。文中还提到阿里巴巴、京东、腾讯等有独立营收的大厂会维持内部实验室，而Meituan LongCat可能因主业亏损被认为成本过高。
Speaking of the ruthlessness of the sage I think that by Q1 2027 we'll see the second wave of Chinese AI startup die-off (since they're Chinese, they might keep resisting as they get Baichuanized, but…); more decent ones will go this time, ones I root for, even. Capex needs are growing exponentially and nation-scale compute scarcity bites in. PMF for "slightly different… but slightly slower and more expensive" is disappearing as models become more general. Xi is getting tougher on involution and wants international competitiveness. Industrial and financial capital is betting on winners. Major IT companies with independent revenue (Alibaba, JD, Tencent) will keep their internal labs afloat, depressing the state's willingness to support startups for anti-monopoly reasons; even then, something like Meituan LongCat might be deemed too costly as Meituan bleeds money from involutionary warfare on its main turf. Finally, DeepSeek's policy of MIT'ing its flagships and Apache'ing its codebases will discourage demand for an internal full-stack fallback at other companies with AGI ambitions (a big reason they even do this is lack of faith in third party providers).
Cynically, I expect that by the end of 2027, there may be as little as two major startups still growing, as Wenfeng deems "enough"; and the #1, I believe, will be DeepSeek, scaling towards gigawatts of capacity and serious, no-caveats competition with Western frontier on AGI and ASI. The most probable #2 is Zhipu, due to good product velocity, financials and, tbh, state ties. I'd rather it be MoonshotAI, which is the second most probable; I believe GLM 5.5 and K3.1 will strongly indicate who is who. It would be nice if the losers of the race gracefully "transform", as Wenfeng puts it, by joining the winners. I know that some of their staff are very much in awe of DeepSeek, which is understaffed, and perhaps a dignified consortium deal may be had. But I am already bittersweet for the era of Chinese Cambrian explosion of AI that I see coming to an end.
