广告
加载中

智谱和MiniMax 又撞车了

kiki 2026-07-15 13:50
kiki 2026/07/15 13:50

邦小白快读

EN
全文速览

本文核心梳理了中国两家头部上市大模型企业智谱与MiniMax的最新发展动态,干货信息如下:

1. 两家企业近期撞车发布内部全员信,智谱提出「Touch High摸高」计划,明确长期聚焦AGI研发,不追求短期商业变现;MiniMax创始人宣布在公司实现AGI前不再领取薪酬,风格偏向务实,迎合资本市场期待。

2. 两家当前都面临两大共同考验:一是迎来上市后的首次解禁大考,基石投资者的去留会直接影响股价和市场信心;二是上市募资消耗速度远超预期,智谱已消耗93%募资,MiniMax消耗了77%,都在通过回A+港股再融资拓展资金渠道。

3. 行业当前共同面临「Token不经济」的问题,两家都针对性调整研发方向,重点提升模型的Coding能力和长程任务能力,解决成本高、交付不达预期的痛点。

本文披露了大模型行业的最新发展逻辑和趋势,能给布局AI的品牌商提供多维度参考:

1. 行业定价逻辑已经发生改变,原来以Token消耗规模为核心的大模型定价逻辑已经失效,Token成本攀升但交付效果不及预期的「Token不经济」问题凸显,品牌商在引入大模型服务时,要更关注模型解决实际业务场景问题的能力,而非单纯看Token规模。

2. 当前头部大模型厂商发展路线出现明显分化,智谱走全栈AI基础设施生态扩张路线,侧重投资并购布局生态;MiniMax走AI原生应用整合路线,侧重全球商业化落地,品牌商可以根据自身的需求,选择匹配的合作方。

3. 未来大模型技术会围绕企业需求升级,Coding能力和长程任务能力提升后,会产出更多可落地的Agent解决方案,品牌商可提前关注技术落地进展,为自身业务升级做准备。

本文梳理了大模型行业的最新变化,能给AI赛道相关卖家提供机会参考和风险提示:

1. 行业变化带来的新机会:当前行业已经从早期的技术叙事转向落地验证,原有靠Token规模讲故事获客的模式已经行不通,倒逼整个行业转向实际交付,To B类AI卖家可以抓住行业转型的机会,聚焦提升自身落地服务能力,抢占市场份额。

2. 头部厂商的扩张带来合作机会:智谱正在进行生态扩张,布局产业链上下游,MiniMax加大AI原生应用的全球商业化开发,都会释放大量的细分领域合作机会,卖家可以对接头部厂商的生态需求,获得增长空间。

3. 风险提示:当前整个行业都还处于亏损状态,资本市场对大模型项目的要求越来越挑剔,卖家要平衡长期研发和短期商业化,明确自身增长逻辑,提升增长确定性,才能获得资本和市场的认可。

本文披露的大模型行业动向,能给想要推进数字化转型的工厂提供诸多启示:

1. 未来大模型技术会更适配工厂的生产需求,当前行业正在解决Token成本高、交付效果差的痛点,后续重点提升模型的长程任务能力,刚好适配工厂生产规划、产品设计、供应链管理这类跨周期的复杂任务,能更好帮助工厂解决实际问题。

2. 后续工厂接入AI的成本会逐步降低,头部大模型厂商智谱正在联合国内芯片企业开发定制化AI推理芯片,降低对海外芯片的依赖,随着本土AI基础设施逐步完善,AI算力的成本会下降,工厂数字化转型的门槛会进一步降低。

3. 当前大模型厂商都在加速商业化落地,开放了Agent相关的封装能力,工厂可以提前对接相关能力,探索AI在生产设计、需求预测等环节的应用,抓住数字化升级的商业机会。

本文分析了大模型行业的核心痛点和发展趋势,能给AI相关服务商提供清晰的发展参考:

1. 当前行业的核心客户痛点已经明确,就是「Token不经济」:客户支付了越来越高的Token成本,但是大模型交付的效果达不到预期,客户对大模型的信任度下降,这是当前服务商拓展市场需要解决的核心问题。

2. 行业发展的核心方向已经清晰,头部厂商都将研发重心转向提升Coding能力和长程任务能力,通过把模型底层的编程、工具调用、多步执行能力封装成可调用可编排的Agent工作流,为客户提供更落地的解决方案,服务商可以跟随这个方向打磨自身服务能力。

3. 目前行业已经出现两种成熟的发展路线,分别是全栈生态扩张路线和应用整合路线,服务商可以根据自身的定位选择合作路径,对接头部厂商的生态资源,落地更多客户场景。

本文梳理了头部大模型平台的最新动态,能给AI平台商的运营和发展提供参考:

1. 当前市场对大模型平台的价值要求已经改变,原来靠Token规模做估值叙事的逻辑已经不被资本市场认可,市场更关注实际交付效果和增长确定性,平台商需要调整自身运营策略,强化落地能力,给市场传递清晰的增长预期。

2. 头部平台面对资金压力和解禁考验的做法可参考:面对上市募资消耗过快的问题,头部厂商都采用「回A申报+港股再融资」的组合方式拓展资金渠道,储备研发资源;发展路线上要么做全栈基础设施生态,要么聚焦应用整合落地,平台商可以结合自身情况选择适配路线。

3. 风险规避提示:平台商要平衡长期研发投入和短期商业化,提前规划资金储备,控制研发投入节奏,同时要给市场明确的增长确定性,避免因预期差导致估值大幅波动。

本文披露了中国大模型产业上市后的最新发展动向,暴露出行业新问题,提供了鲜活的研究样本,干货如下:

1. 产业新动向:两家头部独立大模型企业完成港股上市后,都面临解禁大考和资金短缺的问题,均启动了回AIPO+港股再融资的组合方案拓展资金,整体行业仍然高度依赖资本投入支持AGI研发;头部企业发展路线已经出现明显分化,智谱走全栈AI基础设施生态扩张路线,MiniMax走BC多元化布局的应用整合路线。

2. 行业新问题:行业原本共识的「价值与Token规模正相关」的逻辑已经失效,「Token不经济」成为全行业共同面对的问题,即Token成本攀升但交付效果不达预期,行业亟需建立新的价值定价逻辑。

3. 商业模式层面,当前两种路线都尚未实现盈利,都还在探索阶段,短期分别需要解决增长确定性和技术优势持续性的问题,长期行业竞争尚未定论,为产业研究提供了典型的对比研究样本。

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 summarizes the latest developments of two leading publicly traded Chinese large language model (LLM) companies, Zhipu AI and MiniMax, with key takeaways as follows:

1. The two companies recently released internal all-staff letters around the same time. Zhipu AI launched its "Touch High" initiative, clarifying its long-term focus on AGI research and development (R&D) rather than pursuing short-term commercial monetization. MiniMax's founder announced that he will forgo his salary until the company achieves AGI, a pragmatic move aligned with capital market expectations.

2. Both companies currently face two shared major challenges. First, they are confronting the first post-IPO lock-up expiration test, where the holding decisions of cornerstone investors will directly impact their share prices and market confidence. Second, they are burning through IPO proceeds much faster than expected: Zhipu AI has spent 93% of its raised capital, while MiniMax has used 77%. Both are expanding funding channels through a combination of a secondary A-share listing and follow-on financing in Hong Kong.

3. The entire industry currently faces the common problem of "Token uneconomics," so both companies have adjusted their R&D priorities accordingly. They are now focusing on improving models' coding capabilities and long-context task performance to address the pain points of high costs and underperforming delivery.

This article outlines the latest development logic and trends in the LLM industry, offering multi-dimensional insights for brands deploying AI:

1. The industry's pricing logic has shifted. The original pricing framework centered on token consumption volume is no longer valid, and "token uneconomics"—rising token costs paired with underwhelming delivery results—has emerged as a prominent issue. When adopting LLM services, brands should prioritize a model's ability to solve practical business problems over raw token volume.

2. Leading LLM players have clearly diverged in their development strategies. Zhipu AI pursues a full-stack AI infrastructure ecosystem expansion route, focused on building out its ecosystem via investment and acquisitions. MiniMax follows an AI-native application integration route focused on global commercial落地. Brands can select partners that align with their specific needs.

3. Future LLM technology will evolve around enterprise demand. Improvements in coding and long-context capabilities will enable more deployable Agent solutions, and brands can monitor technology progress in advance to prepare for their own business upgrades.

This article summarizes the latest changes in the LLM industry, providing opportunity references and risk warnings for sellers active in the AI sector:

1. Industry shifts have created new opportunities. The sector has moved from early-stage technology storytelling to落地 validation, and the old customer acquisition model that relied on token volume storytelling is no longer viable, forcing the entire industry to shift to actual delivery. To B AI sellers can capitalize on this industry transition, focus on improving their own delivery capabilities, and capture market share.

2. Expansion by leading players has opened up partnership opportunities. Zhipu AI is expanding its ecosystem across the upstream and downstream of the industrial chain, while MiniMax is accelerating the global commercial development of AI-native applications. Both are releasing substantial partnership opportunities in niche segments, and sellers can align with leading players' ecosystem needs to unlock growth.

3. Risk warning: The entire industry is still unprofitable, and capital markets have become far more selective about LLM projects. Sellers must balance long-term R&D and short-term commercialization, clarify their growth logic, and improve growth certainty to earn recognition from capital and the market.

The LLM industry trends outlined in this article offer multiple insights for factories looking to advance digital transformation:

1. Future LLM technology will become better adapted to factories' production needs. The industry is currently addressing the pain points of high token costs and poor delivery performance, and its subsequent focus on improving long-context task capabilities will fit perfectly with cross-cycle complex tasks in manufacturing such as production planning, product design and supply chain management, better helping factories solve practical problems.

2. The cost of AI access for factories will gradually decline. Leading LLM provider Zhipu AI is co-developing customized AI inference chips with domestic chip companies to reduce dependence on overseas chips. As local AI infrastructure matures, AI computing costs will fall, further lowering the barrier to digital transformation for factories.

3. LLM providers are currently accelerating commercial落地 and have opened up packaged Agent-related capabilities. Factories can connect with these capabilities early to explore AI applications in production design, demand forecasting and other links, and seize the commercial opportunities of digital upgrading.

This article analyzes the core pain points and development trends of the LLM industry, providing clear development guidance for AI-related service providers:

1. The core customer pain point in the current industry is clear: "token uneconomics." Customers are paying ever-higher token costs, but LLM delivery fails to meet expectations, dragging down customer trust in LLMs. This is the core problem service providers need to solve to expand the market.

2. The core development direction of the industry is now clear. Leading players have shifted their R&D focus to improving coding capabilities and long-context task capabilities, packaging the model's underlying programming, tool calling and multi-step execution capabilities into callable, orchestratable Agent workflows to deliver more practical solutions for clients. Service providers can build out their service capabilities along this direction.

3. Two mature development routes have now emerged in the industry: full-stack ecosystem expansion and application integration. Service providers can choose a partnership path based on their own positioning, access ecosystem resources from leading LLM players, and deliver solutions for more customer scenarios.

This article summarizes the latest developments of leading LLM platforms, offering references for the operation and development of AI platform operators:

1. The market's requirements for the value of LLM platforms have changed. The original valuation narrative based on token volume is no longer accepted by capital markets, which now prioritize actual delivery performance and growth certainty. Platform operators need to adjust their operation strategies, strengthen delivery capabilities, and communicate clear growth expectations to the market.

2. The approaches of leading platforms to funding pressure and lock-up expiration tests offer a useful reference. Faced with faster-than-expected burn of IPO proceeds, leading players have adopted the combined strategy of A-share listing application plus Hong Kong follow-on financing to expand funding channels and reserve R&D resources. On the strategic front, players either build a full-stack infrastructure ecosystem or focus on application integration and delivery, and platforms can choose a suitable route based on their own circumstances.

3. Risk mitigation warning: Platforms need to balance long-term R&D investment and short-term commercialization, plan funding reserves in advance, control the pace of R&D investment, and communicate clear growth certainty to the market to avoid large valuation volatility caused by expectation gaps.

This article discloses the latest post-IPO developments in China's LLM industry, exposes new industry problems, and provides a fresh research sample with key findings as follows:

1. New industry developments: After completing their Hong Kong IPOs, the two leading independent LLM companies both face the lock-up expiration test and funding shortages, and have launched a combined strategy of A-share IPO plus Hong Kong follow-on financing to expand capital. The overall industry still relies heavily on capital input to support AGI R&D. Leading players have already diverged clearly in their development routes: Zhipu AI pursues full-stack AI infrastructure ecosystem expansion, while MiniMax follows an application integration route with diversified B2B and B2C布局.

2. New industry problems: The industry's original consensus that "value is positively correlated with token volume" is no longer valid, and "token uneconomics" has become an industry-wide problem: token costs are rising while delivery fails to meet expectations, and the industry urgently needs to establish a new value pricing logic.

3. In terms of business models, neither of the two existing routes has achieved profitability yet, and both remain in the exploration stage. In the short term, they need to address growth certainty and sustained technological advantage respectively, and long-term industry competition remains unresolved. This provides a typical comparative research sample for industrial research.

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.

上市不意味着上岸,往往需要承载资本市场更挑剔的情绪和目光。

这一点,正发生在中国大模型双雄智谱(02513.HK)与MiniMax(00100.HK)身上。

继今年3月公布首份年度财报、默契提出相同的AGI公式后,两家中国大模型企业发布全员信的节点也撞车了。

智谱创始人唐杰在内部信中阐释了AGI的理解,智谱将开启「Touch High(摸高)计划」,即继续聚焦于AGI研究,而不是短期商业变现。

就在几天前,MiniMax创始人闫俊杰在内部信中宣布,从即日起,直到公司实现AGI的那一天,他将不再从公司领取任何薪酬。

仅从内部信形式来看,智谱偏向长期技术叙事,MiniMax则用上了资本市场最爱的「创始人不领薪酬」,前者理想、后者务实,但折射出的是两家公司当前面临的困境:

当市场意识到「Token不经济」,人们对大模型价值的定价逻辑正在发生改变。

智谱内部信的标题为《巨浪已来》,MiniMax则是《向天空的尽头》,光从标题来看,一个瞄向大海、一个望向天空,两家大模型公司都在技术层表达了对AGI的无尽探索。

但跳出技术叙事,两封内部信发布的节点很值得考究。

第一,两家大模型公司都迎来了上市后的首次「解禁大考」。

解禁不意味着抛售,可能是公司价值的直接体现。

本质上,解禁给了基石投资者兑现窗口,他们可根据公司的技术迭代速度、商业化增长、模型训练效率等因素选择是否继续持有,而基石投资者的选择不可避免会对股价、市场信心等方面产生影响。

第二,两家大模型公司都需更多的「粮草弹药」。

MiniMax港股IPO时,募资净额约54.61亿港元,上市约半年里,已使用约77%的募资用于研发与基建,公司称消耗速度超预期。

另一边的智谱上市前约50亿港元的募资总额,同期也已消耗募集资金的93%。

花钱速度远超预期,两家企业都在寻求更多的资金渠道。

一方面是,从港股回A。此前MiniMax和智谱都曾表示,已启动A股IPO申报或上市辅导。

另一方面,是港股再融资。

智谱宣布配售H股,拟募资约314.1亿港元,是年初IPO募资的6倍;MiniMax采用了更灵活的组合融资方式,以「配售新股+可转债」合计募资160.4亿港元,是年初IPO募资的仅3倍。

从募资用途来看,双方花钱的手段有相似性,都明确指向了基模迭代、算力基础设施建设以及商业化拓展。

不同的点在于,智谱偏向生态扩张,MiniMax专注在应用整合。

智谱提到了其资金用于战略投资及并购,这显示出智谱更强的生态扩张意图。

智谱曾多次下场做LP。据其2025年财报,智谱参与成立的星连资本出手30多个企业,整体价值为13.027亿元,增幅达186.26%,作为LP的智谱收获的财务账面权益值约2.79亿元。

MiniMax则提到资金将用于Harness类AI原生产品的全球商业化和开发上。

这和MiniMax此前因Agent收获的高光时刻有关,OpenClaw的火热下,MiniMax凭借M2.5和M2.7模型的53天高频迭代,抓住了Agent渗透的红利。

但很快,两家大公司都撞上了一面相同的墙。

这面墙叫做:Token不经济。

将时间拨回今年年初,在首份年报披露时,智谱和MiniMax给出了类似的公式。

智谱说:“AGI商业价值=智能上界×Token消耗规模”。MiniMax说:“平台价值=智能密度×Token吞吐量。”

两家大模型的共识是智能上限和Token消耗相关,这也是当时行业的普遍认知。

但很快,庞大的Token消耗订单撑不起实际交付成果,「Token越来越贵」「token plan的价格争议」一时间成为讨论的焦点。

这也是所谓的「Token不经济」:Token量是在激增、模型厂商为了覆盖成本提升API定价,但下游使用者承担了攀升的成本,一旦交付效果不及预期,很容易直接影响人们对模型能力的质疑。

这其实是大模型行业里的「灰姑娘水晶鞋效应」,用户对AI模型的期待就像灰姑娘的脚,永远等待一双尺寸合适的水晶鞋,想要锁定忠实用户,提升留存率,就必须要用SOTA模型的技术能力帮助他们解决实际业务场景的问题。

意识到这一点的智谱和MiniMax也开始掉转车头,集中解决「Token不经济」的问题。

在模型侧,重点提升的是两大能力:Coding能力和长程任务能力。

具体来说,Coding能力解决的Token定价问题,长程任务解决的是Token的交付效果。

一方面,中国大模型厂商从Anthropic里学到的启示是,Coding不仅是实现token商业化变现的核心场景,很容易实现溢价,而且能避免陷入「做模型、还是做产品」的创新者窘境。

MiniMax Agent首席架构师阿岛在采访中提到:“今天做通用Agent应用比较悲哀的一点是,你的创新,最终都会被模型内化掉,甚至你写的很多Skill,未来可能都不再需要了。”

但Coding场景很好地实现了「模型即产品」,编程能力领先的模型公司在看到被解锁的应用场景和用法后,往往会迅速将它们吸收、内化到下一代模型和Agent产品中,快速让忠诚度较高的企业用户和开发者直接体验到。

而长程任务能力则是解决模型的实际交付成果。

唐杰在内部信中也提到,今天模型令人兴奋的突破就在于,模型能学会完成一个极长的任务,不是即时问答,而是跨越数周、数月乃至数年的规划与执行。

换言之,为了让模型完成更复杂、更长的任务,两家大模型厂商都在尝试将底层模型的编程、工具调用、深度搜索、多步任务执行等能力,封装为可调用、可编排、可交付的Agent工作流,比如MiniMax推出的MiniMax Agent。

内部信同期发布的节点,两家企业也自然被拿来反复比较。

最明显的差距在于市值。

截至发稿前,智谱市值超7000亿港元,MiniMax接近700亿港元,相差十倍的市值差距,也让投资者产生好奇,相似的独立大模型基因、同样尚在亏损,为何走出了不同的行情?

第一,智谱有参照物。

早期智谱曾将自己定位中国版OpenAI,早在业绩电话会上就花了大量篇幅去对标Anthropic,这次再内部信中有特别强调了「极致的安全治理」。

而因为长期押注Coding,Anthropic实现了对OpenAI的阶段性反超,比如公司估值、年度经常性收入(ARR)。

智谱较早押注Coding方向,为其提升B端收入的可见性埋下了伏笔,相比之下,MiniMax走的是多元化路径,BC两条腿走路,叠加原生多模态心智,相比智谱,缺少对标对象的MiniMax也因此少了确定性。

第二,智谱享受技术基础设施的溢价。

从内部信中提到的摸高计划来看,智谱提到一个关键,未来并不追求短期的应用变现。

此前,DeepSeek和智谱还被曝正在自主研发AI推理芯片,旨在降低对英伟达、华为等外部芯片供应商的依赖,智谱正和国内多家芯片设计企业展开初步磋商,计划为旗下GLM大模型开发定制化AI芯片。

这都反映出,智谱向资本市场展示的,不单单是一个模型平台商,而是一个全栈AI基础设施商,这无疑是一个更长期、更大的叙事,智谱也天然享受了技术基础设施带来的溢价。

但无论是智谱还是MiniMax,资本市场的短期反差并不是终点,内部信起到的作用既是对投资者的预期管理,也是自身的一次再激励。

谁被低估、谁被高估,这个问题也并不重要。短期来看,财报基本面就是韧性的体现,MiniMa需要回答多元化路线的增长确定性,智谱则需要回答Coding的短期先发优势能否转化为持续的动能。

长期来看,中国大模型的竞争也尚未结束,谁落后、谁领先,改变还在继续。

参考资料:

1、腾讯研究院:Token不经济

注:文/kiki,文章来源:硅基研究室,本文为作者独立观点,不代表亿邦动力立场。

文章来源:硅基研究室

广告
微信
朋友圈

FAQ回顾

大模型行业的Token不经济是什么意思?

Token不经济是当前大模型行业的普遍痛点,指大模型Token消耗量激增,厂商为覆盖成本提升API定价,下游使用者承担攀升的使用成本,若交付效果不及预期,很容易引发用户对模型能力的质疑。

智谱和MiniMax的募资资金主要用途是什么?

两家企业募资资金均重点投向基模迭代、算力基础设施建设以及商业化拓展,其中智谱偏向生态扩张,将部分资金用于战略投资及并购;MiniMax专注应用整合,重点投入Harness类AI原生产品的全球商业化和开发。

智谱市值远超MiniMax的核心原因有哪些?

主要有两大原因,一是智谱有明确对标对象,押注Coding方向的路径对标Anthropic,增长确定性更强;二是智谱定位全栈AI基础设施商,正在研发定制化AI推理芯片,享受技术基础设施的估值溢价。

大模型厂商解决Token不经济的核心方向是什么?

当前大模型厂商主要从两大维度优化,一是提升Coding能力,解决Token定价问题,实现商业化溢价;二是提升长程任务能力,优化Token交付效果,满足用户复杂场景的落地需求。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +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