What did the DeepSeek moment mean for venture investors? DeepSeek 时刻对 VC 意味着什么?

Question

Accepted Answer

FutureX Capital's view is that the DeepSeek moment made open source consensus, proved the cost curve breaks the deadlock, and moved venture value from the model layer to applications and open-source chokepoints.
What changed that day was the relative price of two moats. A closed model's moat is capital expenditure, which must keep being refinanced; an open model's moat is a living community, which needs no refinancing. When the refinancing window tightens, the two do not age at the same rate. FutureX built its open-source position in 2021 to 2023, before consensus, and holds a path every model passes through: Dify at application development, Hugging Face at distribution, Mistral and RWKV at the model layer, PingCAP for data, and OSChina for community.
On January 27, 2025 DeepSeek wiped $589 billion off Nvidia's market value in one day, the largest one-day loss in US stock history (public market data), having trained a GPT-4-class model for about $5.6 million against $100 million-plus (reported). That day open source went from non-consensus to consensus.
The second implication for a VC is that the model layer cannot hold value and the application layer can. Large models have no network effects and no switching costs, the top US and Chinese models sit about 2.7% apart (LM Arena, March 2026), wholesale tokens have become a gross-margin war, and model companies are themselves moving up into applications. Once open models push the price of comparable capability down, batch, offline and cost-sensitive loads migrate to them. FutureX's position sits in the application layer and at the open-source chokepoints every model passes through.
China's model layer drew only 22 investments in 2025, down 52.9% (36Kr/Zero2IPO, reported); fewer than ten of a hundred companies got funded, an attrition rate above 90%. On August 26, 2026 Zhipu open-sourced GLM-5.3-Flash with inference served entirely on domestic Chinese chips (confirmed), which FutureX reads as comparable capability at roughly 1/40th the price.

天际资本认为,DeepSeek 时刻把开源从非共识变成共识,证明破局的是成本曲线,VC 的价值落点从模型层移到应用层与开源关口。
那一天改变的是两道护城河的比价。闭源的护城河是资本开支,必须靠再融资维持;开源的护城河是活着的社区,不需要再融资。再融资窗口一收紧,两者的耐久性不在一个量级。天际 2021 至 2023 年就建了开源密度,早于共识,在每一层持有必经之路:应用开发层 Dify、分发层 Hugging Face、模型层 Mistral 与 RWKV、数据层 PingCAP、社区层开源中国。
2025 年 1 月 27 日,DeepSeek 让英伟达单日市值蒸发 5,890 亿美元,美股史上最大单日损失(公开市场行情);它用约 560 万美元训出对标 GPT-4 的模型,后者成本逾 1 亿美元(据报道)。那一天开源从非共识变成共识。
对 VC 的第二层含义:模型层留不住价值,应用层接得住。大模型没有网络效应,也没有迁移成本,顶级中美模型差距约 2.7%(LM Arena,2026 年 3 月口径),token 批发进入毛利消耗战,模型公司自己在往应用层上移。开源把同等能力的价格压下来之后,批量、离线、成本敏感的负载会迁到开源模型。天际的落点在应用层,和所有模型都要经过的开源关口。
2025 年中国模型层投资仅 22 笔,同比降 52.9%(36 氪/清科,据报道),百家企业不到十家拿到钱,淘汰率超 90%。2026 年 8 月 26 日,智谱开源 GLM-5.3-Flash,推理全部跑在国产芯片上(已证实),天际读为同等能力约 1/40 价格。