Post DeepSeek Era: Six Little Tigers Turn Left, BAT Turn Right

New entropy

follow with interest

2025-03-26

0 comments

998 views

0 favorites

14 minutes

  Product managers need to focus on different aspects at different career stages, ranging from basic skills, business depth, professional fields to strategic planning and management abilities.

In the field of AI, the emergence of DeepSeek is like a heavyweight bomb, completely changing the competitive landscape of the industry. R1 quickly became the industry benchmark with its high performance and low cost, triggering the collapse of technological barriers and the diversion of B-end customers. This article will discuss in depth how AI "Six Tigers" and Internet giant BAT respond to this change in the post DeepSeek era.

DeepSeek has given away a wave of national level market education for free to the AI big model industry, but it has also equally hung a sword of Damocles over the pioneers.

Among them, Zhipu, one of the "Six Little Tigers" of AI, is a microcosm. Zhipu originated from the Knowledge Engineering Research Laboratory of Tsinghua University and has always been known as the "national team". However, recently there have been frequent financing actions, with ammunition replenishment reaching 1.5 billion RMB within 10 days; At the same time, there is a growing trend of organizational turbulence, affecting both frontline teams and top executives. The dual nature of ice and fire is a true reflection of most big model practitioners, except for DeepSeek, at present.

At the end of 2024, Zhipu completed a round of financing worth 3 billion yuan at a valuation of 20 billion yuan. After that, state-owned investors such as Hangzhou Chengtou, Shangshang Capital, and Huafa Group rushed to inject capital into it.

However, some venture capitalists have analyzed the "new entropy" and found that DeepSeek's emergence has had a certain negative impact on the valuation of Zhipu. The quick money may also be to grab the relatively high price that has already reached an upward bottleneck.

In contrast to the opening of the cash granary, Zhipu has shown a shrinking trend in team size and external investment. During the peak period, a team of over a thousand people continued to undergo a certain proportion of elimination in the delivery and sales sectors; Among the executive team, Li Huizi, Vice President of Applied AI and Cooperation, Zhang Kuo, Chief Strategy Officer, and Qu Teng, Vice President, have successively withdrawn.

What kind of product manager does B-end need?

B-end product managers aim to improve the efficiency of the supply side, so B-end demand is mainly guided by business issues. This is an important aspect of B2B products, which serve a subject

View details>

Besides Zhipu, other members of the "Six Little Tigers" are also quietly or openly digesting the pros and cons caused by DeepSeek, BAT、 Bytes and other Internet giants have also experienced differentiation in route selection, making this spring also imply a taste of returning to winter.

The DeepSeek era has arrived after 01

DeepSeek is like a giant rock thrown into a calm lake, causing a huge wave in the AI industry. It can even be exaggerated to say that it directly rewrote the competition rules of domestic large models.

The open-source big model DeepSeeker R1 quickly became the industry benchmark with its performance comparable to GPT-4o and cost only one tenth, leading to two chain reactions - the collapse of technical barriers and the diversion of B-end customers. Many enterprises have started to deploy independently, and small and medium-sized hardware manufacturers have seized the opportunity to quickly launch the "DeepSeek all-in-one machine", directly impacting the market for private deployment of large models represented by Zhipu.

Although Zhipu's current business direction focuses on both B-end and C-end, as CEO, Zhang Peng has firmly leaned towards the former. He admitted that the main purpose of promoting C-end products is to showcase their capabilities to B-end customers. To this end, Zhipu deliberately imitated cloud manufacturers to establish a regional sales team, and recruited Wu Weijie, the chief commercial officer of the former Flying Book of ByteDance. In addition, Chen Xuesong, who serves as Vice President, has seamlessly connected a large number of sales talents to the Zhipu team through his past experience at Megvii Technology, one of the "Four Little Dragons" of AI.

By cultivating in the blue ocean market, the Zhipu MaaS platform has accumulated 700000 enterprise and developer users, with overall commercial revenue exceeding 100% by 2024; According to incomplete statistics, the number of successful bids for To B and To G projects throughout the year reached 32, with a disclosed amount of approximately 130 million yuan, second only to iFlytek and Baidu.

However, the AI equality brought by DeepSeek unexpectedly arrived. From the launch of 70 "AI intelligent employees" based on the DeepSeek big model in Futian District, Shenzhen, to the announcement by Shanghai Jiao Tong University that the full range of DeepSeek models have been locally deployed on a computing power cluster composed of domestic chips, government and enterprise units have been the first to be swept by this AI wave. According to statistics, as of the end of February this year, 45% of the 98 state-owned enterprises in China have completed the deployment of the DeepSeek model.

At the forefront of crazily grabbing orders for government and enterprise big models are not pioneers such as Zhipu, but more traditional data equipment vendors, computing power service providers, and IT integrators such as Lenovo, ZTE, Inspur, Shenzhou, and so on. Even small and medium-sized businesses in Huaqiangbei, Shenzhen, are eager to try.

New Entropy learned from industry insiders that some government and enterprise orders have explicitly requested support for the DeepSeek interface. Compared to big companies like BAT, which can unload their psychological burden and choose to walk on both self-developed and open-source paths, the "Six Tigers" represented by Zhipu must bet on the underlying technical capabilities of their own large models. Once left behind too much by DeepSeek, it will lose its core competitive advantage, which is why the capital market has begun to re-examine the value of closed source big models.

02 Differentiation of AI Six Tigers

The competition in the AI industry is extremely fierce, and it is already a rapidly changing track. DeepSeek's strong debut further exacerbates the industry's crisis awareness. The intelligence spectrum of the "Six Little Tigers" is still at the forefront, and some backward enterprises have been greatly impacted than the industry dividends, gradually showing a clear differentiation trend.

According to statistics, except for Jieyue Xingchen, which is also known as "North Zhipu, South Leap" along with Zhipu, there has been no news of executive resignations yet. The other "Six Little Tigers" have had several high-level changes in the past year, especially in the post DeepSeek era, where the wave of resignations is accelerating.

For example, Jiao Ke, co-founder of Baichuan Intelligence, has recently officially resigned, and another co-founder, Chen Weipeng, is also about to leave and is currently processing relevant procedures. It is reported that both of them have entered the AI entrepreneurship track. Wei Wei, Vice President and Head of Open Platforms at MiniMax, who had extensive experience in enterprise services at Tencent Cloud, resigned at the beginning of the year. There has been no further confirmation yet.

The adjustment of the talent team is often accompanied by changes in business direction. Baichuan Intelligent has now chosen to shift towards the vertical medical field, perhaps based on a deep exploration of its own technological advantages and market demand, attempting to cultivate and refine in specific niche areas to avoid direct competition with powerful competitors such as DeepSeek. However, this also means that its competitiveness in the general AI big model market will further weaken, and its market share is at risk of being encroached upon.

Zero One Thing is also being forced to overturn its business model. CEO Li Kaifu had previously reluctantly announced the abandonment of training super large models. "Looking up at the stars to pursue AGI requires sufficient and even low-cost ammunition reserves. Looking at it from a down-to-earth perspective, our highest priority at this stage is to consolidate the strength of obtaining ammunition first." Currently, Zero One Thing is "making blood" through the splitting business, and Li Kaifu's attitude is "to talk to investors first and see if anyone is willing to invest.

However, in the context of rapid iteration of big model technology, excessive focus on the monetization of the application layer, even in the case of self redemption under financing difficulties, will inevitably lead to a gap with industry leaders in terms of technological core competitiveness, which is somewhat like drinking poison to quench thirst.

The dark side of last year's industry star month is also considered one of the big model companies that has been greatly affected in this wave. Compared to last year's large-scale investment campaign for Kimi across all online channels, it has been revealed that the investment strategy has been adjusted, including suspending investment in multiple Android channels and giving the top spot in the app to Yuanbao, who has been crazily investing recently.

On the other hand, in the open source wave sparked by DeepSeek, Zhipu, MinMax, and Step Leap Star all have varying degrees of open source foundation and have the advantage of keeping up with the pace. However, from the publicly available code repository of DeepSeek, it can be found that it is more compatible with Nvidia's CUDA framework due to its hardware model selection. Considering various factors, it is highly likely that there will be a wave of domestic chip substitution in the future. Therefore, on the path of open source, the "Six Tigers" have not missed any opportunities.

03 Large factory launches self rescue operation

With the release of the latest financial reports of Baidu, Alibaba and Tencent and the holding of performance teleconferences, AI has become an Internet giant who wants to be labeled all over the body.

Tencent, which was once one step behind the starting line of entering the big model, became the first representative to let go of its arrogance and fully integrate into DeepSeek. Apart from Yuanbao taking over the overflow of traffic from DeepSeek's official website server capacity, Tencent's C-end products such as WeChat, Tencent Docs, QQ Browser, and Tencent Music Player quickly switched sides, completely ignoring the feelings of their own hybrid models.

The B-end business represented by cloud services is more suitable for accepting DeepSeek's "transformation". In addition to Tencent Cloud, Baidu Cloud, Alibaba Cloud, Huawei Cloud and the ByteDance volcano engine all worked overtime during the Spring Festival holiday to launch a version of DeepSeek.

However, after "borrowing" algorithms optimized by multiple models to achieve comprehensive cost reduction and efficiency improvement, the cloud war between major manufacturers will inevitably increase their price power again. According to a research report by Guotai Junan Securities, the cost of training Deepseek V3 (671B parameter full blood version) is 5.576 million US dollars (about 40.7 million RMB), which is only 7% of Meta's large language model Llama 3. The cost of training ChatGPT-4o with OpenAI is as high as 78 million US dollars or even 100 million US dollars, and it also requires tens of thousands of H100 chips. And Deepseek-V3 is trained using a specially designed AI chip with performance castrated - the NVIDIA H800.

A salesperson from a cloud computing service provider revealed to "New Entropy" that the price war that swept across the industry last year was ostensibly led by several large companies such as ByteDance and Alibaba. However, starting from the second half of the year, DeepSeek has released ultra-low prices in a small range, which has made it even more difficult for everyone to relax and breathe.

In addition, it is worth noting that after the logic advocated by foreign AI giants was shattered, major companies have gradually returned to the curve overtaking route brought by DeepSeek. Whether it is the old giant BAT or the rising star ByteDance, joining the ranks of the big companies indicates that they have already gained an absolute advantage in their respective business territories. Fully utilizing this high ground to charge towards the AI mountain is obviously more efficient and effective.

For example, Doubao, which is backed by Tiktok traffic, has been quite confident in not integrating with DeepSeek. And even so, it has not lost too many users. According to QuestMobile data, as of February 9, 2025, the average daily active users of Doubao were 18.45 million, second only to DeepSeek and higher than BAT's main AI applications.

For big companies, whether it is open source or closed source, free or paid, it will not cause too much burden on the cost side. Both sides can choose to bet with a higher winning rate. On the other hand, Zhipu's open source strategy and "wintering" approach may become a symbol of the domestic AI big model race moving from fanaticism to rationality.

The catfish effect of DeepSeek has forced the entire industry to recalibrate the triangular relationship between technology, capital, and markets. Open source is no longer an idealistic slogan, but a necessity for survival; Commercialization cannot only rely on government orders or capital foam, but needs to take root in the real scene. After enduring winter, spring will not be far away.

Author | Fushen Editor | Siyuan Editor | Jiuli

This article is authored by Renren who is a product manager [New Entropy]. The WeChat official account is [New Entropy]. The original/authorized release on Renren who is a product manager is prohibited from reprinting without permission.

The image is from Pixabay and is based on the CC0 protocol.

后DeepSeek时代：六小虎向左，BAT向右

新熵关注

2025-03-26

0 评论998 浏览0 收藏14 分钟

在AI领域，DeepSeek的出现如同一颗重磅炸弹，彻底改变了行业的竞争格局。R1以其高性能和低成本迅速成为行业基准线，引发了技术门槛的崩塌和B端客户的分流。本文将深入探讨后DeepSeek时代，AI“六小虎”和互联网大厂BAT等如何应对这一变革。

DeepSeek给AI大模型行业，免费赠送了一波国民级别的市场教育，却也平等地在先行者们头上，悬起了一把达摩克利斯之剑。

其中，AI“六小虎”之中的智谱就是一个缩影，智谱脱胎于清华大学知识工程研究室，素来有“国家队”之称。然而就在最近开始频频出现融资动作，10天之内补充弹药达15亿人民币；可与此同时，组织震荡颇有加剧之势，从一线团队到高管大牛皆有波及。冰火两重天的态势，可谓是目前除了DeepSeek之外，大多数大模型从业者们，真实写照的一个缩影。

2024年底，智谱曾以200亿元的估值，完成一轮30亿元人民币的融资，在这之后，包括杭州城投、上乘资本、华发集团等国资背景的资方快马加鞭地赶到为其注资。

不过，也有风投人士对「新熵」分析，DeepSeek的横空出世还是对智谱的估值造成了一定负面影响，快速拿钱也可能是为了抢下已经出现上涨瓶颈的相对高价。

与大开现金粮仓之门形成反差的是，智谱在团队规模和对外投资上呈现出收缩之势。高峰期阶段的一千余人团队，在交付和销售板块持续进行一定比例的淘汰；高管队伍中，则有应用AI与合作副总裁李惠子、首席战略官张阔，以及副总裁曲滕等先后退出。

B端究竟需要什么样的产品经理？

B端产品经理都是以提升供应侧的工作效率为目的，所以B端需求主要是以业务问题为导向。 这个是B端产品比较重要的一点，B端产品是服务于一个主体 ...

查看详情 >

智谱之外，其他“六小虎”成员也都各自或默默或高调地消化着DeepSeek引起的利弊影响，BAT、字节等互联网大厂同样出现了路线选择上的分化，让这个春天也隐含了一丝返冬的味道。

01 后DeepSeek时代来了

DeepSeek犹如一颗投入平静湖面的巨石，在AI行业掀起了滔天的波澜，甚至可以夸张点说，其直接改写了国内大模型的竞争规则。

开源大模型DeepSeek-R1以性能比肩GPT-4o、成本仅为十分之一的优势，迅速成为行业基准线，进而产生出两个连锁反应——技术门槛崩塌与B端客户分流。不少企业开始独立部署，以及中小硬件厂商抓住机会，快速推出“DeepSeek一体机”，直接冲击了以智谱为代表的大模型To B私有化部署市场。

智谱目前的业务方向虽然对于B端和C端各有侧重，但作为CEO的张鹏曾坚定地偏向前者，他坦言称推C端产品的主要目的，就是为了向B端客户展示能力。为此，智谱刻意模仿云厂商建立了一支以区域划分的销售团队，并且挖来了前字节跳动旗下飞书的首席商业官吴玮杰。另外，担任副总裁的陈雪松，凭借过去在AI“四小龙”之一的旷视科技任职经历，带动了一大批销售人才无缝衔接到智谱团队。

在蓝海市场的耕耘，使得智谱MaaS平台积累下70万企业和开发者用户，2024年的整体商业化收入超过100%；据不完全统计，全年的To B及To G项目的中标数量达32个，披露金额约1.3亿元，仅次于科大讯飞和百度。

然而，DeepSeek带来的AI平权，猝不及防地杀到。从深圳市福田区推出70名基于DeepSeek大模型的“AI数智员工”，到上海交通大学宣布在国产芯片组成的算力集群上，完成了DeepSeek全系列模型的本地部署，首当其冲被这场AI浪潮席卷的就是政企单位。有统计数据显示，截至今年2月底，国内98家央企中的45%，皆已经完成了DeepSeek模型的部署。

冲在疯抢政企大模型订单一线的，不是智谱等大模型先行者，而是更传统的数据设备商、算力服务商，以及IT集成商，诸如联想、中兴、浪潮、神州等等。甚至有深圳华强北的中小商家，也开始跃跃欲试。

「新熵」从行业人士处了解到，部分政企订单已明确要求“支持DeepSeek接口”。相比BAT等大厂可以卸下心理包袱，选择自研与开源两条腿走路的路线，以智谱为代表的“六小虎”们，则必须押注在自家大模型的底层技术能力上。一旦被DeepSeek甩开太多，将失去核心竞争优势，这也是为什么资本市场已经开始重新审视闭源大模型的价值。

02 AI六小虎的分化

AI行业的竞争异常激烈，本就是一个日新月异的赛道，DeepSeek的强势登场则进一步加剧了业内的危机意识。“六小虎”中的智谱尚处于前列，一些落后企业受到的冲击则远大于行业红利，逐渐呈现出明显的分化态势。

据统计，除了与智谱并称为“北智谱，南阶跃”的阶越星辰，目前尚未传出高管离职的消息，其他“六小虎”近一年内都已经有多位高层变动，尤其在进入后DeepSeek时代，离职潮隐隐有着加速的趋势。

比如，百川智能的联合创始人焦可近期已正式离职，另一位联合创始人陈炜鹏也即将离开，目前正在办理相关手续。有消息称，两人均已投身AI创业赛道。MiniMax的副总裁、开放平台负责人魏伟，曾在腾讯云拥有丰富的企业服务经验，也于年初离职，暂未有更进一步的确认消息。

人才队伍的调整，往往伴随着业务方向的变化。百川智能如今选择转向垂类医疗领域，或许正是基于对自身技术优势与市场需求的深度挖掘，试图在特定细分领域深耕细作，以避开与DeepSeek等强大对手的正面竞争。然而，这也意味着其在通用AI大模型市场的竞争力会进一步减弱，市场份额面临被侵占的风险。

同样被逼推倒业务模式的还有零一万物。CEO李开复此前无奈宣布放弃训练超大模型，“仰望星空追寻 AGI 需要充足甚至不计代价的弹药储备，脚踏实地来看，我们现阶段的最高优先级是先巩固拿到弹药的实力。”目前，零一万物正通过拆分业务“造血”，李开复的态度是“先去找投资人聊，看有没有人愿意投”。

不过，在大模型技术快速迭代的背景下，过度专注于应用层的变现，即使是融资困境下的自我救赎，仍不可避免会在技术核心竞争力上与行业领先者逐渐拉开差距，多少有些饮鸩止渴的意思。

去年的行业明星月之暗面，也被认为是这波受影响较大的大模型企业。相比于去年在全网渠道为Kimi进行的大规模投流，目前被爆出已调整投流战略，包括暂停多个安卓渠道的投放，将APP榜首的让给了近期疯狂投流的元宝。

另一边，由DeepSeek掀起的开源浪潮中，智谱、MinMax和阶跃星辰皆有不同程度的开源基础，更有跟上脚步的优势。但是，从DeepSeek公开的代码库内容可以发现，其出于硬件型号上的选择，更适配英伟达的CUDA框架。而从综合因素考量，未来极大概率会出现一波国产芯片的替代潮，因此在开源这条路上，“六小虎”们并非没有一点机会。

03 大厂开启自救行动

随着百度、阿里巴巴、腾讯的最新财报陆续发布和业绩电话会议的召开，AI成为互联网巨头们恨不得贴满全身的标签。

曾经在入局大模型的起跑线上慢人一步的腾讯，成为第一个放下矜持全线接入DeepSeek的代表。除了元宝承接了DeepSeek官网服务器容量不足外溢出的流量外，腾讯旗下的微信、腾讯文档、QQ浏览、腾讯音乐器等C端产品，都迅速地倒戈，丝毫不顾及自家混元大模型的感受。

以云服务为代表的大厂B端业务，更适合接受DeepSeek的“改造”。除了腾讯云，还有百度云、阿里云、华为云和字节跳动的火山引擎，均在春节假期里加班加点上线了接入DeepSeek的版本。

不过，在“借鉴”了多模型优化的算法实现全面降本增效后，大厂之间的云战争必然会再度加大价格战力度。国泰君安证券的研报显示，训练Deepseek V3（671B参数满血版）的成本为557.6万美元（约合4070万人民币），仅为Meta公司的大语言模型Llama 3的7%，而OpenAI训练ChatGPT-4o所花费的成本高达7800万美元甚至是1亿美元，并且还需要上万张H100芯片。而Deepseek-V3训练时使用的是性能被阉割过的特供AI芯片——英伟达H800。

有云计算服务商的销售人员对「新熵」透露，去年全行业掀起的价格战，明面上都以为是字节、阿里几个大厂在打头阵，实际上从下半年开始，DeepSeek在小范围内释放出的超低价，更让大家不能放松喘气。

另外，值得注意的是，在由国外AI巨头曾经倡导的高举高打逻辑破灭后，大厂们也逐渐回归到DeepSeek所带来的弯道超车路线。不管是老巨头BAT，还是新贵字节，跻身大厂之列就说明已然在一方业务地盘上占据绝对的优势，充分利用这一高地再向AI山头发起冲锋，显然更能事半功倍。

比如背靠抖音流量的豆包，就相当有自信地一直坚持未与DeepSeek融合。并且即便如此，也没有丢掉太多用户，据QuestMobile数据显示，截至2025年2月9日，豆包的周日均活跃用户数为1845万，仅次于DeepSeek，高于BAT旗下的主力AI应用。

对于大厂们来说，开源还是闭源，免费还是收费，在成本侧都不会造成太大的负担，两边全都要或者选择胜率较大的押注即可。而智谱的一边开源一边“过冬”，或许会成为国内AI大模型赛道从狂热走向理性的标志。

DeepSeek的鲶鱼效应迫使全行业重新校准技术、资本与市场之间的三角关系。开源不再是理想主义的口号，而是生存的必需品；商业化也不能仅依赖政府订单或资本泡沫，而需扎根真实场景。当熬过了冬天，春天就不会远了。

作者丨茯神 编辑丨思原   主编丨九黎

本文由人人都是产品经理作者【新熵】，微信公众号：【新熵】，原创/授权 发布于人人都是产品经理，未经许可，禁止转载。

题图来自 Pixabay，基于 CC0 协议。