10月30日，据外媒援引消息人士称，DeepSeek第二轮融资拿到了1055亿元人民币（折合158亿美元），其中有近30家投资者共同出资，国有资本占比高达27%，风投占比26.5%，科技公司占比41.9%，成为迄今为止中国AI公司规模最大的风险投资，DeepSeek的估值因此也达到了6155亿人民币。
值得关注的是，第二轮融资的产业投资方名单堪称一份AI基建地图：上游有长鑫存储、中芯国际（存储芯片），吉利、美的、立讯精密、vivo（制造与终端），宁德时代（能源），以及腾讯、美团、小红书、携程（算力与场景）。从芯片到制造再到应用场景，这些科技公司各自出资10亿至30亿元人民币。
在此之前DeepSeek横空出世的两年多里，一直拒绝外部融资。
不过接受外部融资，并不代表DeepSeek会因此稀释太多股权。
今年6月底，DeepSeek首轮融资就有个极为特殊的安排：除国家AI基金外，其他投资者均需将资金投入梁文锋持有的实体，放弃投票权，并锁定股份五年。与此同时，工商信息显示梁文锋在融资前已将个人持股比例从1%增至34%，表决权达到100%。配合后续传出的科创板上市筹备、原高瓴合伙人出任CFO等动作，这是一套精心设计的融资但不失控的方案。
DeepSeek首轮融资超过500亿元，外媒指出，DeepSeek将最新一轮融资规模扩大一倍的原因在于两方面：
一是上游英伟达芯片的价格飙升。据悉，一台搭载八颗B300芯片的服务器价格从1500万元人民币飙升至约1900万元人民币，是同类服务器在大多数国家售价的五倍多。
二是为应对AI泡沫破裂可能带来的冲击而储备资金。此前首批上市的中国大模型厂商MiniMax和智谱AI的市值已大幅下跌，原因在于今年夏季上市前投资者认购限制解除。
不过从过去半年DeepSeek启动相关融资后一系列动作来看，缺钱可能还不是DeepSeek接下来最有可能遇到的问题，更主要的是如何应对算力制约，同时推进对国产算力产能瓶颈的突围。
以国产GPU厂商摩尔线程为例，在近期公布三季度财报中，已实现单季营收15.5亿元，同比暴增1773.9%，归母净利润4926万元，较原定目标提前扭亏为盈。财报指出其智算产品在AI创新企业加速落地，并已实现对DeepSeek、GLM、MiniMax等主流开源模型的Day-0支持。
2026年前七月DeepSeek营收约4.75亿元，毛利率44.6%，但净亏损仍超7亿元。同时，年化收入从不足5亿美元翻倍至10亿美元仅用了约三个月。但今年前七个月DeepSeek在AI infra相关支出也达到了110亿元。与业内许多大模型厂商一样，DeepSeek仍处于依靠API业务增收不增利的阶段。
结合近期国内业内人士也有不同判断：一方人士会更加关注DeepSeek的长期战略价值，能否搭建好算力基础设施，实现国产算力卡位，而另一方人士也会存有质疑，与许多模型厂商一样，DeepSeek同样商业化路径不够清晰、AI投资可能存在自循环等问题。
DeepSeek的融资故事讲到这里，真正的考验也才刚刚开始。（本文首发于钛媒体APP，文 | TechHorizon）
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On 30 October, according to external sources, the DeepSeek second round of financing received RMB 105.5 billion (equivalent to $15.8 billion), of which nearly 30 investors co-financed up to 27 per cent of State-owned capital, 26.5 per cent of windfall investment, 41.9 per cent of technology companies, and the largest risk investment ever made by China's AI company, and the DeepSeek valuation thus reached RMB 61.5 billion.
It is of concern that the list of industrial investors for the second round of financing is referred to as an AI infrastructure map: upstream storage of long snails, central cores International (storage chips), glean, American, well-advised, vivo (manufacturing and terminal), the Ningde era (energy), and Tencent, the United States Corps, REDnote, and transport (calculations and scenes). From chips to manufacturing to applications, these technology companies each contribute between 1 billion and 3 billion yuan yuan.
Before that, DeepSeek had refused external financing for more than two years.
The acceptance of external financing, however, does not mean that DeepSeek dilutes too much equity.
At the end of June, DeepSeek had a very special arrangement for the first round of financing: With the exception of the National AI Fund, all investors are required to invest their funds in the entities held by Liang Wenfeng, to renounce voting rights and to lock their shares for five years. At the same time, business information indicates that Liang Wenfeng has increased his shareholding from 1 per cent to 34 per cent and voting rights to 100 per cent prior to financing. This is a well-designed programme of financing but not out of control.
DeepSeek ' s first round of financing exceeds 50 billion yuan, and external sources point to two reasons why DeepSeek doubled the size of the latest round:
One is the price of the Upper British Wida chip. According to the information received, the price of a server carrying eight B300 chips rose sharply from RMB 15 million to about RMB 19 million, more than five times the price of a similar server in most countries.
The second is to reserve funds to respond to possible shocks from the bursting of the AI bubble. The market value of China's first LLM manufacturers, Mini Max and Zhipu AI, which were listed earlier, has fallen significantly, owing to the lifting of the pre-market investor subscription restrictions this summer.
However, in a series of actions following DeepSeek’s launch of related financing in the past six months, the lack of money may not be the most likely next problem for DeepSeek, but rather how to deal with the constraints of calculation, while pushing for a breakthrough in the domestic capacity bottlenecks.
For example, in the recently published three-quarter financial statements of the national GPU manufacturer, a single-season recovery of $1.55 billion was achieved, an increase of $1773.9 per cent over the previous year, resulting in a net profit of $4.926 million for the mother, which was in excess of the target. The financial paper noted that its smart products accelerated their landing in AI innovation enterprises and that Day-0 support for mainstream open-source models such as DeepSeek, GLM and Mini Max had been achieved.
In July 2026, the DeepSeek battalion collected approximately $475 million, representing a Maori rate of 44.6 per cent, but the net loss was still over $700 million. At the same time, annualized income doubled from less than $500 million to $1 billion for only about three months. In the first seven months of this year, DeepSeek spent $11 billion on AI infra. Like many LLM producers in the industry, DeepSeek is still at a stage of increasing revenue from API operations.
There is also a difference of opinion in the recent past: one side will be more concerned about the long-term strategic value of DeepSeek, whether it can build up a critical infrastructure and achieve a national credit card, while the other side will be skeptical, and, like many model manufacturers, DeepSeek’s commercialization path is not clear enough, and AI investment may be self-revolving.
DeepSeek's financing story is here, and the real test is just beginning. (This post was launched on titanium medium APP, Man, TechHorizon)
Auto-fetched and translated by AI Atlas. All rights belong to the original author.