广告
加载中

她把公司卖给智谱

陈佳 2026-07-24 12:05
陈佳 2026/07/24 12:05

邦小白快读

EN
全文速览

本文核心内容是智谱完成对AI基础设施企业中科加禾的战略收购,这是智谱近年最大收购案例之一,以下是核心干货:

1. 收购标的核心情况:中科加禾成立于2023年,创始人是清华出身、中科院计算所的女研究员崔慧敏,核心团队是有20年以上经验的编译器团队,曾参与龙芯、华为昇腾等多款国产芯片的编译器研发,目标是打造跨不同AI芯片的软件底座,降低模型迁移适配成本。

2. 收购的核心逻辑:智谱发力云端大模型服务,需要自行承担算力成本,已经建成1GW级国产芯片数据中心,需要编译器技术提升单芯片利用率,收购前双方已经合作解决了GLM-5大模型推理的异常问题,补齐了智谱的算力释放短板。

3. 行业背景:AI编译器是连接大模型和芯片的核心“翻译官”,国产AI芯片目前软件生态薄弱,这块是国产AI发展的关键缺口。

本次收购事件为AI领域品牌商传递了清晰的产业趋势和布局参考,核心干货如下:

1. 产业发展趋势:当前AI产业竞争已经从大模型训练、芯片制造,向上游底层基础软件环节延伸,AI编译器这类中间基础设施的价值正在快速凸显,国产芯片生态完善过程中催生了大量刚需机会,AI相关品牌可提前布局底层软件赛道,建立差异化优势。

2. 市场需求变化:大模型商业化落地正逐步转向云端按使用量持续付费的模式,这种模式对模型运行稳定性、算力效率的要求远高于本地化部署,To B客户对底层能力的关注度持续提升,品牌布局AI业务必须重视底层基础设施优化。

3. 资本层面参考:智谱收购补齐核心技术短板的消息放出后,在股价回撤70%的情况下单日反弹超40%,说明布局核心技术能够有效获得资本市场认可,为AI品牌的技术投入方向提供了明确参考。

本次收购事件为AI领域相关卖家揭示了行业新变化、机会与风险,核心干货整理如下:

1. 新增增长机会:当前大模型商业化进入深水区,云端服务模式普及带动底层算力优化需求爆发,AI编译器、异构算力适配、推理性能优化相关服务已经实现商业化落地,中科加禾成立不到一年就获得数千万元订单,相关卖家可切入国产芯片生态适配这个缺口。

2. 风险提示:目前国产AI芯片软件生态碎片化,不同芯片的编程框架、软件体系互相独立,做AI应用、模型服务的卖家如果布局多芯片方案,必须提前解决适配优化问题,否则容易出现模型运行异常,影响用户体验和订单转化。

3. 创业退出参考:硬核技术创业团队可以依托科研积累,先通过和上下游合作验证产品价值,获得融资后快速发展,最终被头部企业收购实现退出,这种路径对底层技术创业团队来说可行性很高。

本次收购事件为国内AI芯片、算力硬件相关生产企业传递了新的需求和发展启示,核心干货如下:

1. 产品需求变化:当前国产AI芯片已经实现大规模应用,下游客户已经从追求芯片数量,转向追求芯片实际利用效率,配套基础软件优化已经成为客户的核心需求,AI芯片生产企业必须重视配套软件生态建设,否则硬件性能无法充分释放,会直接影响产品市场竞争力。

2. 潜在商业机会:跨多品牌芯片的通用编译底座是当前行业明确缺口,芯片生产企业可以和专业编译器团队开展合作,共同优化产品性能,拓展下游大模型、智算中心客户,目前行业已经有数十家厂商达成相关合作,市场空间广阔。

3. 数字化转型启示:底层核心技术是AI产业的核心壁垒,拥有核心技术的主体更容易获得资本和市场的认可,硬件生产企业推进智能化转型不能只停留在应用层面,要重视底层核心技术的积累,才能建立长期竞争力。

本次收购事件为AI基础设施相关服务商明确了行业趋势、客户痛点和可行方向,核心干货整理如下:

1. 行业发展趋势:随着大模型产业落地加速和国产AI芯片渗透率不断提升,AI编译器这类连接芯片与大模型的中间层基础设施已经成为明确刚需,市场需求正在快速增长,未来随着国产芯片替代推进,通用编译软件的市场空间会持续扩大,是值得布局的赛道。

2. 核心客户痛点:目前下游大模型厂商转向云端服务模式后,需要自行承担高额算力成本,而现有国产芯片软件生态不统一,模型跨芯片适配成本高,芯片利用率偏低,推高了厂商的运营成本,高并发场景下还容易出现运行异常,影响用户体验,这是客户普遍存在的痛点。

3. 可行解决方案方向:参考中科加禾的路径,打造横跨多品牌、多型号AI芯片的异构编译底座,搭配推理引擎、微调引擎、算子自动转译工具,可以有效帮助客户降低适配成本,提升芯片利用率,解决高并发场景的运行问题,是清晰可行的商业化方向。

本次收购事件为大模型、算力相关平台商揭示了行业需求、可借鉴做法和风险规避方向,核心干货如下:

1. 当前行业需求:大模型平台转向云端持续付费模式后,对算力能力的要求发生了变化,平台不仅需要拿到足够数量的芯片解决算力供给问题,还需要底层编译优化技术提升单芯片的利用率,才能支撑大规模用户的高并发调用,避免出现算力紧张、用户体验下降的问题。

2. 可借鉴的布局做法:智谱的路径是一方面自建1GW级的国产芯片大型数据中心解决算力供给问题,另一方面收购核心编译器团队解决算力释放问题,这种软硬结合的布局方式,为大模型平台的算力建设提供了清晰参考,平台可通过收购核心技术团队快速补齐自身短板。

3. 风险规避方向:大模型平台的企业估值受业务进展和技术布局影响很大,股价和市场信心波动较高,平台需要及时披露核心技术布局和业务进展,传递企业价值,稳定市场信心,同时要提前规划算力建设,避免业务增长后算力跟不上影响运营。

本次收购事件反映了国内AI产业的最新动向、新问题与新的商业模式,对产业研究有较高价值,核心干货整理如下:

1. 产业最新动向:当前国内AI产业竞争正在从前端的大模型训练、中端的芯片制造,向后端上游的基础软件环节延伸,AI编译器已经成为资本和头部企业布局的热点,硬核底层技术团队的商业价值得到行业认可,中科加禾成立不到两年完成五轮累计1.3亿元融资,最终被头部大模型企业收购,印证了这个趋势。

2. 产业待解决的新问题:当前国产AI芯片存在软件生态碎片化的问题,不同芯片厂商拥有独立的软件栈和编程体系,模型跨芯片迁移的适配成本很高,制约了国产芯片的大规模推广应用;同时大模型商业化转向云端模式后,算力成本成为大模型厂商的核心负担,提升算力效率是产业当前需要解决的核心问题。

3. 商业模式创新观察:依托科研机构技术积累的底层技术创业团队,先通过和上下游合作验证产品价值,获得多轮融资后最终被头部企业收购实现退出,是适合高壁垒底层技术创业的可行商业模式,也为科研成果转化提供了新的路径。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

This article centers on Zhipu AI’s strategic acquisition of Zhongke Jiahe, an AI infrastructure firm, one of Zhipu’s largest acquisitions in recent years. Key takeaways are as follows:

1. Profile of the acquisition target: Founded in 2023, Zhongke Jiahe was founded by Huimin Cui, a Tsinghua alumna and female researcher from the Institute of Computing Technology, Chinese Academy of Sciences. Its core team is a compiler group with over 20 years of experience that has participated in compiler R&D for multiple domestic chips including Loongson and Huawei Ascend. The company aims to build a cross-architecture software base for diverse AI chips to cut model migration and adaptation costs.

2. Rationale for the acquisition: As Zhipu scales up its cloud-based large model services and bears its own computing power costs, the company has built a 1GW domestic chip-based data center. It needs compiler technology to boost per-chip utilization. Prior to the acquisition, the two parties had already cooperated to resolve inference anomalies with the GLM-5 large model, and the acquisition fills the gap in Zhipu’s computing power release capabilities.

3. Industry context: AI compilers act as the core "translator" connecting large models and chips. Domestic AI chips currently suffer from weak software ecosystems, making this segment a critical bottleneck for China’s AI development.

This acquisition offers clear industry trend insights and strategic references for AI brands. Key takeaways are as follows:

1. Industry development trend: AI industry competition has now shifted upstream from large model training and chip manufacturing to foundational software. The value of intermediate infrastructure such as AI compilers is growing rapidly. The ongoing improvement of domestic chip ecosystems has created large numbers of high-demand opportunities. AI brands can get ahead of the curve by entering the foundational software track to build differentiated competitive advantages.

2. Shifting market demand: Large model commercialization is gradually moving toward a cloud-based pay-as-you-go model, which imposes far higher requirements for model operation stability and computing efficiency than on-premise deployment. B2B customers are increasingly focusing on underlying capabilities, so brands building AI businesses must prioritize optimization of foundational infrastructure.

3. Capital market reference: Following the announcement of Zhipu’s acquisition to fill its core technology gap, Zhipu’s share price rebounded more than 40% in a single day after a 70% drawdown. This demonstrates that investments in core technology gain clear recognition from capital markets, providing a clear reference for AI brands’ technology investment direction.

This acquisition reveals new industry shifts, opportunities and risks for AI-related sellers. Key takeaways are as follows:

1. New growth opportunities: As large model commercialization enters a deeper stage of development, the spread of cloud-based service models has spiked demand for underlying computing power optimization. Services related to AI compilation, heterogeneous computing adaptation, and inference performance optimization have already achieved commercialization. Zhongke Jiahe secured tens of millions of yuan in orders less than a year after its founding, so sellers can enter the gap of domestic chip ecosystem adaptation.

2. Risk warning: The software ecosystem of domestic AI chips is currently fragmented, with independent programming frameworks and software systems across different chips. If sellers of AI applications and model services pursue a multi-chip strategy, they must resolve adaptation and optimization issues in advance; otherwise, operational anomalies are likely, hurting user experience and order conversion.

3. Startup exit reference: Hard tech startup teams can leverage existing research expertise, validate product value via cooperation with upstream and downstream partners, scale quickly after securing funding, and eventually exit via acquisition by a leading enterprise. This path is highly feasible for bottom-layer technology startup teams.

This acquisition delivers new demand signals and development insights for domestic AI chip and computing hardware manufacturers. Key takeaways are as follows:

1. Changing product demand: Domestic AI chips have now achieved large-scale adoption, and downstream customers have shifted their focus from chip volume to actual chip utilization efficiency. Supporting foundational software optimization has become a core customer demand. AI chip manufacturers must prioritize supporting software ecosystem building; otherwise, hardware performance cannot be fully unlocked, which will directly harm product competitiveness.

2. Potential business opportunities: A general-purpose compilation base supporting cross-brand chips is a clear industry gap. Chip manufacturers can partner with professional compiler teams to jointly optimize product performance and expand downstream customers including large model developers and AI computing centers. Dozens of manufacturers have already reached relevant cooperation, and the market opportunity is substantial.

3. Insights for digital transformation: Core bottom-layer technology is the key competitive moat for the AI industry. Entities holding core technology gain far more recognition from capital markets and customers. When advancing intelligent transformation, hardware manufacturers should not stop at the application layer. They must prioritize accumulating core bottom-layer technology to build long-term competitiveness.

This acquisition clarifies industry trends, customer pain points and viable strategic directions for AI infrastructure service providers. Key takeaways are as follows:

1. Industry development trend: With the accelerated commercialization of large models and rising penetration of domestic AI chips, intermediate infrastructure such as AI compilers that connect chips and large models has become a clear, high-demand market. Market demand is growing rapidly, and as domestic chip substitution progresses, the market size of general-purpose compilation software will continue to expand, making it a track worth entering.

2. Core customer pain points: After downstream large model developers shifted to cloud-based service models, they must bear high computing costs on their own. The fragmented software ecosystem of current domestic chips drives up cross-chip model adaptation costs, keeps chip utilization low, pushes up developer operating costs, and often causes operational anomalies in high-concurrency scenarios that hurt user experience — this is a widespread pain point for customers.

3. Viable solution direction: Following Zhongke Jiahe’s path, building a heterogeneous compilation base compatible with multiple brands and models of AI chips, paired with inference engines, fine-tuning engines, and automatic operator translation tools, can effectively help customers cut adaptation costs, boost chip utilization, and resolve operational issues in high-concurrency scenarios — making this a clear and viable commercial direction.

This acquisition reveals industry demand, replicable strategies and risk mitigation directions for large model and computing power platform operators. Key takeaways are as follows:

1. Current industry demand: After large model platforms shifted to the cloud-based recurring revenue model, their requirements for computing power have changed. Platforms not only need to secure sufficient chip volume to meet computing supply, but also require bottom-layer compilation optimization to improve per-chip utilization, to support high-concurrency access from large-scale user bases and avoid computing shortages and degraded user experience.

2. Replicable strategic approach: Zhipu’s strategy combines building a 1GW domestic chip large-scale data center in-house to meet computing supply, and acquiring a core compiler team to unlock greater computing efficiency. This integrated hardware-software layout provides a clear reference for large model platforms’ computing infrastructure development, and platforms can quickly fill capability gaps by acquiring core technology teams.

3. Risk mitigation: The valuation of large model platforms is highly sensitive to business progress and technology layout, leading to high volatility in share prices and market confidence. Platforms should disclose core technology layouts and business progress in a timely manner to communicate corporate value and stabilize market confidence. They should also plan computing infrastructure development in advance to avoid operational disruptions from insufficient computing capacity amid business growth.

This acquisition reflects the latest trends, emerging problems and new business models in China’s AI industry, offering high value for industrial research. Key takeaways are as follows:

1. Latest industry trends: Competition in China’s AI industry is shifting upstream from front-end large model training and mid-end chip manufacturing to back-end foundational software. AI compilers have become a hot area of investment and布局 for both capital and leading enterprises, and the commercial value of hard-core bottom-layer technology teams has gained industry-wide recognition. Less than two years after its founding, Zhongke Jiahe completed five rounds of financing totaling 130 million yuan before being acquired by a leading large model enterprise, confirming this trend.

2. Unresolved industry challenges: Domestic AI chips currently face fragmented software ecosystems: different chip vendors maintain independent software stacks and programming systems, leading to high adaptation costs for cross-chip model migration that constrains large-scale adoption of domestic chips. Meanwhile, after large model commercialization shifted to the cloud model, computing costs have become the core burden for large model developers, and improving computing efficiency is the core problem the industry must currently resolve.

3. Observations on business model innovation: Bottom-layer technology startup teams built on research institution expertise can validate product value via cooperation with upstream and downstream partners, secure multiple rounds of financing, and eventually exit via acquisition by a leading enterprise. This is a viable business model for high-moat bottom-layer technology startups, and also offers a new path for commercializing academic research outcomes.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

双赢。

作者/陈佳

报道/投资界PEdaily

智谱买下一个团队。

投资界从投资方处获悉,智谱已完成对AI基础设施企业中科加禾(XCore Sigma)的战略收购,缔造智谱近年最大的收购案例之一。

这次被智谱买下的是一支编译器团队。中科加禾成立于2023年,创始人崔慧敏是清华计算机系毕业生、中科院计算所研究员、博士生导师,也是这一硬核技术领域为数不多走向创业一线的女性学者。

过去三年,她带领团队完成五轮融资,身后集结不少我们熟悉的投资机构。但随着交易完成,中科加禾近日发生工商变更,投资方股东退出,崔慧敏持股降至11%,智谱一跃成为最大股东。

清华学姐

集结一支罕见团队

先从崔慧敏的经历说起。

她1997年进入清华大学计算机科学与技术系,本科毕业后继续在清华攻读硕士,随后进入中国科学院计算技术研究所攻读博士。长期以来,她聚焦异构编程模型、异构编译优化以及数据中心编程与编译技术研究,曾任中科院计算所编译团队负责人。

放眼AI产业,大模型和芯片占据了很多关注。但还有一个隐藏在两者之间的基础软件环节,正在决定算力能否真正释放——编译器。

通俗来说,编译器是连接AI模型与芯片的“翻译官”。AI模型要在不同硬件平台上高效运行,需要编译器将模型计算任务转换为适配目标芯片的执行代码,并通过算子优化、计算调度等方式释放硬件性能。

对比国内外产业,这一环节的分量会更清晰。英伟达的优势不仅来自GPU硬件,也来自围绕CUDA构建的软件生态。大量AI开发工具、算法框架和应用能够基于CUDA进行优化,从而降低开发者适配不同GPU产品的成本。

相比之下,国产AI芯片软件生态建设还有不足。不同芯片厂商通常拥有独立的软件栈、编程框架和优化体系,模型迁移过程中往往需要进行算子适配、编译优化和性能调优。如果软件生态跟不上,芯片硬件性能也难以充分释放。

而崔慧敏团队长期关注的,便是异构计算环境下的编程模型与编译优化问题。

正因为这项技术难度极高、周期漫长,中科加禾团队的分量也就更容易理解。官网显示,中科加禾团队具备20年以上的编译器开发经验,主持或参与研发过包括龙芯、神威、寒武纪、华为昇腾等多款芯片的编译器,以中科院计算所作为技术源头。

这次创业始于2023年5月,崔慧敏带队尝试将长期积累的编译与芯片工具软件能力做成商业化产品。公司的目标是打造一套横跨不同品牌、不同型号AI芯片的软件底座,降低模型在不同国产芯片之间的适配和迁移成本。

2024年7月,中科加禾发布异构原生算力产品矩阵,包括异构原生推理引擎SigInfer、异构原生微调引擎SigFT,以及算子自动生成和转译工具SigTrans。其愿景是,通过搭载中科加禾工具软件包,让每颗国产芯片/英伟达受限版芯片(以A800和H800为代表的功能受限芯片)都能融入生态,在大模型时代定义编译层通用国产基础软件生态标准。

如今回过头看,在国产AI芯片软件生态薄弱的背景下,中科加禾汇集了一支难以在短期内复制的编译器团队。人才就是壁垒,这就不难理解智谱为何将中科加禾团队收入囊中。

身后投资人体面退出

中科加禾曾留给创投圈深刻印象。

早在2023年8月,公司完成数千万元种子轮融资,由BV百度风投领投,晨山资本、陆石投资、奇绩创坛、寒武纪、天创资本等跟投。

寒武纪此后披露的上市公司文件显示,截至2025年3月31日,寒武纪通过全资子公司南京显生间接持有中科加禾0.76%的股权,期末账面余额为200万元。

很快,中科加禾再获一轮融资。

那是2024年3月,中科加禾完成数千万元天使轮融资,由元禾原点领投,新尚资本、中科院创投、晨山资本和BV百度风投跟投。此时,中科加禾成立约八个月,已经开始与上下游生态厂商开展产业合作,产品也在部分头部客户中形成商业化订单。

2024年6月,中科加禾完成天使+轮融资,国中资本入局。

半年后,公司又获得北京市人工智能产业投资基金领投的数千万元Pre-A1轮融资。彼时,中科加禾已与数十家智算中心、芯片厂商、服务器厂商、运营商及互联网公司达成合作,并获得数千万元商业订单及意向订单。

2025年4月,中科加禾完成Pre-A+轮融资,投资方为深高新投。

至此,中科加禾在不到两年时间里完成五轮融资。投资界向中科加禾方面核实,此前融资累计金额约1.3亿元。但交易金额、交易方式、原有股东安排及团队后续分工的相关细节未获得披露。

最新消息,企查查显示中科加禾近日发生工商变更,三亚百川致新私募股权投资基金、北京市人工智能产业投资基金等15名股东退出,智谱关联主体Z.AI成为新增大股东,持股约60%。

智谱下一程

正如外界好奇:智谱为什么要收购中科加禾?

答案可能藏在算力里。

The Information报道,智谱正在增加面向企业客户的云端模型服务收入,降低对本地化部署项目的依赖。两种模式的差别在于:本地化部署是把模型安装在客户自己的服务器上,算力开支主要由客户承担,智谱获得的通常是一次性的部署和项目收入;云端服务则由智谱提供算力,客户通过API调用模型,并按照Token用量持续付费。

于智谱而言,云端模式能够带来更稳定、可持续的收入,但也意味着模型运行的算力成本从客户转移到了智谱。调用量越大,智谱需要承担的芯片采购、数据中心建设和推理成本就越高。

今年1月,智谱公告称,GLM-4.7上线后,GLM Coding Plan用户数高速增长,导致算力资源阶段性紧张。尽管公司已经启动扩容,仍一度将每日可销售量降至调整前的20%,优先保障现有用户。

于是,智谱开始大规模补充算力。

彭博社援引知情人士报道,智谱已经建成一座全部采用国产芯片的大型数据中心,并开始部分运行。这座数据中心按照1GW级功率设计,理论耗电规模相当于约75万户家庭的即时用电量。知情人士还称,智谱已建成或运营多个由超过1万颗芯片组成的计算集群,这些算力将主要用于GLM系列模型的训练和开发。

但算力的问题,并不只是芯片数量的问题。把大量AI芯片放进数据中心,并不等于自动获得同等规模的有效算力。

调用量快速增长后,智谱不仅需要更多芯片,还需要进一步提高每块芯片的利用效率。GLM-5发布后,智谱的Coding Agent每日调用量达到数亿次。但在高并发、长上下文场景下,部分用户开始反馈模型出现乱码、重复回答和生僻字等异常。

4月下旬,智谱在官方技术博客《Scaling Pain:超大规模Coding Agent推理实践》中复盘了这组问题,并定位到推理系统中的多个底层竞态Bug。值得一提的是,文章末尾专门感谢中科加禾和中国科学院计算技术研究所处理器芯片全国重点实验室团队的合作与支持。

在SGLang开源社区,相关代码提交显示这次修复由智谱与中科加禾共同完成,合作覆盖根因分析、验证和上线准备。这意味着,在收购消息浮出水面之前,中科加禾已经参与智谱GLM-5推理基础设施的实际优化工作。

至此,智谱近期的两项动作也可以串联起来:1GW数据中心解决的是算力供给问题,中科加禾解决的是算力释放问题。

而这笔收购发生时,智谱经历了一轮剧烈的股价波动。

6月22日,智谱股价盘中一度触及2980港元,市值达到约1.33万亿港元。此后股价迅速回落,7月17日和7月20日分别下跌28.49%和19.56%,最低较高点回撤约70%。

7月21日,智谱完成对中科加禾收购、落地1GW级国产AI算力数据中心的消息同时传出。当日智谱股价盘中最高上涨超过40%,总市值回升至5676亿港元,单日增加超过1500亿港元。

如此,效应开始显现。

注:文/陈佳,文章来源:投资界(公众号ID:pedaily2012),本文为作者独立观点,不代表亿邦动力立场。

文章来源:投资界

广告
微信
朋友圈

FAQ回顾

智谱为什么要收购中科加禾?

智谱正发力云端模型服务,需承担高额算力成本,单纯堆叠AI芯片无法直接获得有效算力。中科加禾团队有20年以上编译器开发经验,收购后可解决算力释放问题,提升芯片利用效率,支撑GLM系列大模型的训练和推理需求。

中科加禾是做什么的?

中科加禾是2023年成立的AI基础设施企业,核心团队有20年以上编译器开发经验,曾主导多款国产芯片的编译器研发,主打异构原生算力产品矩阵,目标是打造跨不同AI芯片的软件底座,降低大模型在国产芯片间的适配迁移成本。

AI编译器的作用是什么?

AI编译器是连接AI模型与芯片的“翻译官”,可将模型计算任务转换为适配目标芯片的执行代码,通过算子优化、计算调度等方式释放硬件性能,能解决国产AI芯片软件生态薄弱导致的硬件性能难以充分释放的问题。

智谱收购中科加禾带来了哪些市场反应?

2025年7月21日,智谱完成对中科加禾收购、落地1GW级国产AI算力数据中心的消息传出后,当日股价盘中最高上涨超过40%,总市值回升至5676亿港元,单日市值增加超过1500亿港元。

这么好看,分享一下?

朋友圈 分享

APP内打开

+1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0