跌破千亿!MiniMax为何惨败至此?Breaking the 1000-day mark! Why did MiniMax suffer such a defeat?


AI free token/AI免费积分

2026年7月8日到9日,港股大模型双雄迎来首轮限售股解禁。

智谱7月8日解禁,2568万股,市值超400亿港元。当天股价低开后掉头向上,盘中一度飙涨19%,收盘涨13.35%,市值突破8000亿港元。

MiniMax7月9日解禁,1.46亿股,占总股本46.44%,解禁市值超500亿港元。当天盘中一度跌超20%,股价跌破290港元,市值跌破千亿。

同一场考试,一个涨,一个跌。

有人说是因为智谱解禁比例小(5.76%),MiniMax解禁比例大(46.44%)。有人说是因为智谱国资股东多,MiniMax财务投资者多。

都对,但都是表象。

真正的原因是,市场开始算账了。

过去靠想象力撑着的估值,现在要面对一个残酷的问题,你的估值到底应该怎么计算?

图片

01 MiniMax为什么输了

半年前,情况完全相反。

2026年1月上市首日,智谱市值约579亿港元,MiniMax约1067亿港元,后者几乎是前者的两倍。

3月,MiniMax股价冲到1330港元,市值4100亿港元。

同月,智谱还在爬坡。

半年后,MiniMax市值跌去77%,只剩智谱的14%。一个智谱能买七个MiniMax。

发生了什么呢?MiniMax为什么败下阵来?

关键原因在于收入结构。

我们来看智谱,它2025年总营收7.24亿元,B端业务贡献了99%。企业级通用大模型3.66亿,开放平台及API1.90亿,企业级智能体1.66亿。MaaS API平台的ARR(年度经常性收入)达到17亿元,同比增长60倍。

MiniMax2025年总营收7903.8万美元(约5.6亿人民币)。C端AI原生产品收入5307.5万美元,占67.2%;B端开放平台及企业服务收入2596.3万美元,占32.8%。

看到没有,差距就在这里。

智谱有体量巨大的B端业务,且整个公司业务聚焦在B端,而MiniMax业务分散,有通用大语言模型,有编程模型,有生图模型,有语音模型,什么都做,业务也没有以B端为主。

这个收入结构,决定了市场对这两家公司不同的评值。

02 大模型只有两种产品能收到B端的钱

整个大模型行业,目前能系统性从B端收到钱的,只有两类产品。

第一类,编码,也就是当下的coding.

AI编程已经从一个程序员尝鲜的玩具,变成了企业软件工程的基础设施。

Stack Overflow 2025年调查显示,84%的开发者已在开发流程中使用或计划使用AI工具,47.1%每天使用。

而据Gartner预测,到2028年,90%的企业软件工程师将使用AI代码助手。

驱动这个变革的,是应用创新。

Claude Code横空出世,他是一个命令行编程工具,可以由开发者在终端用自然语言里直接跟AI协作写代码。

就是这么一个看起来毫不起眼的产品,到2026年2月,年化收入已经做到了25亿美元。

而Anthropic整体的年化营收轨迹更吓人,2024年1月还只有8700万美元,到2024年12月达到10亿美元,2025年底攀升至90亿美元,2026年4月达到300亿美元,到H轮融资时已突破470亿美元。

三个月,从100亿飙到440亿,每天新增约9600万美元。

图片

从未见过这样的增速。

这就叫作爆炸性增长!

Anthropic它的用户量远不及OpenAI,但估值已经反超,接近万亿美元。

世界财富十强企业中80%已成为其付费客户。

国内也一样。智谱GLM-5.2的软件工程能力逼近Claude Opus 4.8,Coding Plan的价格只有Claude的七分之一。市场买单的不是它7.24亿的营收和47亿的亏损,而是中国版Anthropic的预期。

字节也在转向这条路。豆包2.1 Pro补上了Coding和Agent能力。字节高层到访Anthropic后,内部启动AI资源重整,把更多资源从豆包这类大众产品挪向企业服务和编程模型。

大模型数据审核团队从1500人扩到3000多人,专门为编程模型清洗数据。

第二类:视频生产。

字节的Seedance跑通了这一商业模式。

据业内人士爆料,Seedance当前年化收入(ARR)已达20亿美元(约143亿元),单月营收超10亿元,绝大部分来自企业客户,而这一产品的毛利率达70%。

20亿美元是什么概念?智谱2025年全年营收7.24亿人民币,Seedance一个季度可能就超过了。

客户是谁?AI短剧、漫剧、广告、电商素材、游戏宣发。

这些场景的共同特点是,需求高频、付费明确、反馈快速。企业算账很直接,如果AI生成的视频如果比人便宜、比旧流程快,同时产出算得过来账,那就值得付费。

2026年上半年,国内AI短剧市场规模已突破110亿元,预计全年将攀升至350亿元以上。做一部80集AI短剧,每集1分钟,光视频生成就要三四万。

这些钱最终流向了谁?卖算力、卖模型、卖工具的人。

快手可灵也证明了这条路:近70%的收入来自专业创作者,每月666元的钻石会员成为摇钱树。

03 其他产品为什么收不到钱

除了编码和视频生产,其他大模型产品在B端基本收不到稳定的钱。

聊天机器人?企业凭什么付钱?一个能聊天的AI,和客服系统里已有的FAQ有什么区别?

AI写作?一篇稿子几百字,企业愿意为一个不确定质量的结果付多少钱?

AI绘画?一张图几块钱,但企业要的是可复用的、风格统一的、能放进商业流程的素材,不是随便画个图。

现在大量的企业愿意在这上面花的钱,都是小钱。

而在编程模型上花的钱,一些大企业,一年消费已达几千万美元。普通的IT企业月消费十几万、几十万,都不算大支出,甚至有游戏公司暴出一晚上就消费了二百万。

C端场景、或B端低价值场景面临用户付费意愿低与推理成本倒挂的双重瓶颈。

不付费的用户不再是企业的资产,而是负资产,用规模换利润的互联网铁律在大模型时代彻底失效了。

字节已经想明白了。未来C端豆包主要承担模型打磨、数据迭代与品牌曝光功能,B端才是核心营收与利润来源。

连字节都在掉头,其他人还在犹豫什么?

04 卖铲子的人

大模型厂商本质上是卖铲子的。

谁家的铲子好用,挖矿的人就用谁的。企业客户不关心你的模型参数有多少、榜单排第几,只关心一件事,用你的铲子能不能更快更便宜地挖到矿。

编码铲子的客户是程序员和企业研发部门。他们要的是代码质量、调试效率、工程化能力。

视频铲子的客户是短剧公司、广告商、内容生产者。他们要的是生成速度、画面质量、成本控制。

这两个场景有一个共同特征,结果可量化、ROI可计算。也就是说可以进行经济计算。

一个程序员用AI编程工具,每天多写200行代码,少花2小时调试,这个账能算。一个短剧公司用AI视频生成,每集省下5000元制作费,一个月省15万,这个账也能算。

而一旦跑通了ROI,那么,不用大模型的企业,就是傻子,因为明明能省钱,明明能提高效率,你却不用,这不是傻是什么?

说不清ROI的东西,企业不会掏钱。算得清ROI的东西,企业就会毫不犹豫地使用。

05 铲子的生产工具才是真正的赢家

当然,最后真正的赢家,还是产业里那些更重要的角色。

不管哪家大模型厂商赢,三星、英伟达这些生产铲子的人都是赢家。

大模型的竞争越激烈,算力的消耗越大。视频生成模型的算力需求远超文本和图片,需要同时建模空间、时间、运动轨迹、主体一致性、音画同步和物理合理性。编码模型的推理需求也在指数级增长。

谁在生产这些算力?英伟达的GPU、三星的HBM、台积电的CoWoS封装。

大模型厂商在互相厮杀,但每一场战斗都在给算力供应商送钱。

这就是基础设施的商业模式——不管谁赢,我都赢。

图片

06 生死战才刚刚开始

2026年是大模型商业化的分水岭。

行业已经告别了单纯的C端流量竞争与价格战,商业化锚点正式切换至B端企业生产力兑现周期。比起继续追用户规模,找到愿意持续付费的开发者和企业,正在变得更重要。

智谱和MiniMax都宣称2026年底实现10亿美元ARR。从5.6亿人民币到70多亿人民币,一年翻十几倍,这个目标能不能实现,将决定两家公司的生死。

但方向已经明确了。只有B端业务才是真正的矿石厂。 只有编码和视频生产这两类产品能系统性地从企业口袋里掏出钱来。

那些还在追求C端用户规模、还在做免费聊天、还在靠想象力撑估值的大模型厂商——留给他们的时间不多了。

产业虽然巨大,但谁能活下来,现在说都还早。但有一点可以确定,人工智能产业商业化的路径已经清晰,市场不会再听概念了,他们要听到的是收入。

 

On July 8 and 9, 2026, the restricted shares of two Hong Kong-listed AI heavyweights, Zhipu and MiniMax, entered their first lock-up expiration period.

Zhipu unlocked 25.68 million shares on July 8, worth over HK$400 million. The stock opened lower but reversed course with a 19% intraday jump, closing up 13.35% to nearly HK$800 billion in market cap.

MiniMax, unlocking 146 million shares (46.44% of total) on July 9, saw a massive 20% intraday drop to below HK$290 per share, collapsing to under HK$1000 billion in market value.

Two companies, one exam — Zhipu thrived, MiniMax failed.

Some blame the large unlock proportion (46.44% vs Zhipu's 5.76%), others point to institutional ownership or speculative froth. But the real reason lies deeper: the market is waking up to valuation reality.

都对 , 但都是表象。

Zhipu's model focuses on enterprise AI services — from general-purpose language models to specialized APIs and agents — generating 99% of its 2025 revenue from B2B clients. Its MaaS API alone boasts a $1.7 billion annualized revenue run rate, growing 60x year-on-year.

MiniMax, by contrast, relies heavily on consumer-facing AI products (entertainment, chat, image generation), which account for 67% of its revenue. While fun, consumer products don’t typically command high enterprise multiples or sticky subscriptions.

图片

01 MiniMax 为什么输了

Video AI — used for short films, ads, or training materials — is another gold mine. Seedance (ByteDance’s AI video arm) now generates over $2 billion ARR with 70% profit margins, largely from corporate clients producing marketing content.

Other AI offerings — like chatbots, writing assistants, or general image generators — struggle to find stable corporate customers due to low ROI or lack of clear integration into workflows.

The lesson: AI startups must pivot from flashy demos to solving real business problems. Zhipu’s success stems from its tight alignment with industrial automation, software development, and data analytics sectors. MiniMax’s fall suggests its consumer-first approach won’t cut it in the long run.

同月,智谱还在爬坡。

Meanwhile, the real titans — NVIDIA, Samsung, TSMC — are laughing all the way to the bank as demand for their computing chips skyrockets with every new model launch.

The moral? In the AI game, “B2B” is the new black. If you're not helping companies make more money or save more time, you're just background noise.

The key reason lies in the revenue structure.

Let's look at Zhipu. In 2025, its total revenue was 724 million yuan, with the B-end business contributing 99% of that. Enterprise-level general models accounted for 366 million, the open platform and API for 190 million, and enterprise-level agents for 166 million. The ARR (Annual Recurring Revenue) of the MaaS API platform reached 1.7 billion yuan, showing a 60-fold year-on-year increase.

MiniMax's total revenue in 2025 was 79.038 million USD (approximately 560 million RMB). The C-end AI native products generated an income of 53.075 million USD, accounting for 67.2%, while the B-end open platform and enterprise services generated 25.963 million USD, accounting for 32.8%.

Can you see the difference?

Zhipu has a massive B-end business and is a fully B-end company. In contrast, MiniMax's business is fragmented, spanning general language models, programming models, image generation models, and speech models. They are doing a bit of everything but haven't focused their operations on the B-end.

This revenue structure dictates how the market values these two companies.

02 Only Two Types of Products Can Generate Revenue from the B-End

In the entire large model industry, there are only two categories of products that can systematically generate money from the B-end.

The first category is coding.

AI programming has evolved from a toy for programmers to experiment with into the fundamental infrastructure of enterprise software engineering.

The Stack Overflow 2025 survey shows that 84% of developers use or plan to use AI tools in their workflow, with 47.1% using them daily.

Gartner predicts that by 2028, 90% of enterprise software engineers will use AI code assistants.

This shift is driven by application innovation.

Claude Code emerged as a command-line programming tool that allows developers to directly collaborate with AI using natural language in the terminal.

Despite its seemingly humble appearance, this product achieved an annualized revenue of 2.5 billion USD by February 2026.

Anthropic's overall revenue trajectory is even more impressive. It went from 87 million USD in January 2024 to 1 billion USD in December 2024, then skyrocketed to 9 billion USD by the end of 2025, and surpassed 47 billion USD by the time of its H-round financing.

This represents a leap from 10 billion to 440 billion within three months, effectively adding 96 million USD per day.

图片

This is explosive growth.

Anthropic's user base is much smaller than OpenAI's, but its valuation has surpassed it, approaching 1 trillion USD.

Eighty percent of the world's top ten wealthiest enterprises are now its paying customers.

The same trend is visible in China. Although Zhipu's loss was 4.7 billion, the market is betting on its potential to become the Chinese version of Anthropic.

ByteDance is also pivoting to this direction. Doubao 2.1 Pro now includes Coding and Agent capabilities. After ByteDance's leadership visited Anthropic, they initiated a reorganization of AI resources, moving them from consumer products like Doubao to enterprise services and programming models.

The data review team for the large model has been expanded from 1500 to over 3000 people, specifically to clean data for programming models.

The large model data auditing team grew from 1500 to over 3000 people, specifically to clean data for programming models.

第二类:视频生产。

ByteDance's Seedance has already validated this business model.

According to industry insiders, Seedance's current annualized revenue is 2 billion USD (approx. 14.3 billion RMB), with monthly revenue exceeding 1 billion RMB. The vast majority of this comes from corporate clients, and the product boasts a gross margin of 70%.

To put this into perspective, Zhipu's total revenue for 2025 was 724 million RMB. Seedance is outperforming that in a single quarter.

Who are the clients? AI short dramas, manga, advertising, e-commerce content, and game promotion.

What they all have in common is high-frequency demand, clear pricing, and fast feedback. Companies perform a simple cost-benefit analysis: if AI-generated videos are cheaper and faster than human-produced ones, and the output is marketable, then the payment is justified.

In the first half of 2026, the domestic AI short drama market exceeded 11 billion RMB and is expected to reach over 35 billion RMB for the full year. Producing an 80-episode AI short drama would cost around 34 thousand RMB just in video generation.

Where does this money go? To those selling computing power, models, and tools.

Kuaishou's Keling also proves this point: nearly 70% of its revenue comes from professional creators, with the Diamond membership at 666 RMB per month being a significant cash cow.

03 Why Other Products Cannot Generate Revenue

Outside of coding and video production, other large model products cannot generate stable revenue from the B-end.

Chatbots? Why would a company pay for a chatbot? It's essentially a more advanced FAQ.

AI writing? Companies aren't willing to pay for a piece of content with uncertain quality, especially when it might only be worth a couple of hundred words.

AI painting? While a single image might cost a few bucks, businesses need reusable, consistent-style assets that fit into their commercial workflow, not random doodles.

Currently, the money businesses are willing to spend on these other areas is minimal.

However, some large enterprises are spending tens of millions of dollars per year on programming models. Regular IT companies are shelling out tens or hundreds of thousands monthly, and there are reports of a gaming company spending 2 million RMB in a single night.

The consumer market and low-value B-end segments face dual bottlenecks: low user payment intention and excessive inference costs.

In this context, non-paying users are no longer assets; they are liabilities. The internet logic of trading scale for profit has failed in the era of large models.

ByteDance has figured this out. In the future, Doubao will primarily be used for model tuning, data iteration, and brand exposure, while the B-end will drive core revenue and profit.

If even ByteDance is making this shift, what is stopping the rest of the industry?

04 卖铲子的人

Large model vendors are essentially selling shovels.

The better the shovel, the more gold miners will buy it. Business clients don't care about parameter size or leaderboard rankings; they care about whether your shovel helps them dig faster and cheaper.

The customers for coding shovels are programmers and R&D departments who value code quality and debugging efficiency.

The customers for video shovels are short drama companies, advertisers, and content producers who value generation speed, visual quality, and cost control.

Both of these markets have one thing in common: the ROI is measurable.

If a company can calculate that using a specific AI tool saves them money or increases productivity, they will adopt it. If not, they will stick to traditional methods.

Once the ROI is proven, not using these tools becomes a competitive disadvantage.

For things where ROI is unclear, businesses won’t pay. For things where ROI is clearly calculable, businesses will use them without hesitation.

05 The True Winners are the Shovel Makers

Of course, the real winners in the long run are the manufacturers of the "shovels" themselves.

Regardless of which large model company dominates, Samsung and NVIDIA will profit.

As the competition between large models intensifies, the demand for computing power skyrockets. Video generation models, in particular, require massive amounts of processing power because they have to model spatial relationships, time sequences, motion, consistency, and audio synchronization.

While the model companies fight each other, the hardware providers are raking in the benefits.

This is the nature of the infrastructure business model: everyone wins when the industry grows.

This is the business model of infrastructure — no matter who wins, I win.

图片

06 生死战才刚刚开始

2026 marks a watershed year for large model commercialization.

The industry is moving away from a simple "race to the bottom" on price and consumer engagement. The focus is shifting from chasing users to finding paying developers and enterprise customers.

Zhipu and MiniMax have both set a goal of achieving 1 billion USD in ARR by the end of 2026. This represents a massive jump from their current RMB equivalents (roughly 5.6 billion RMB total). Whether they can achieve this will determine their viability.

The direction is clear: B-end businesses are the mines, and only coding and video production products can reliably extract value from them.

Companies that are still focused solely on consumer-facing apps, free chats, or speculative valuations are running out of time.

The industry is vast, and the ultimate survivors are still unknown, but one thing is certain: the path to monetization is through enterprises, and the market is demanding revenue, not just concepts.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry