One Zhipu Equals Ten MiniMaxes: The Sheep Are Still Here, the Pigs Are Gone
TechFlow Selected TechFlow Selected
One Zhipu Equals Ten MiniMaxes: The Sheep Are Still Here, the Pigs Are Gone
Attention is no longer valuable.
Navigating Web3 tides with focused insightsBy: Xiao Bing
On July 9, 2026, HK stocks staged a piece of black humor.
MiniMax faced its first lock-up expiration since listing. The stock price fell over 20% during the session, market cap broke through the 100 billion HKD threshold, shrinking to near 90 billion. Compared to the historical high of 410 billion in March, 320 billion evaporated.
Just the day before, Zhipu also faced lock-up expiration. The market was ready for a second crash, but after opening 3% lower, it rallied all the way, closing up 13%, and rising another 11% the next day, with market cap steadying at 900 billion HKD.
One Zhipu equals about ten MiniMax.
Half a year ago, this equation was written in reverse, or at least bet on in reverse.
Opening, the Market Voted for MINIMAX
Turn the clock back to January this year.
On January 8, Zhipu rang the bell, bearing the title of "Global Large Model First Stock". On the first day, it only rose 13.17%, closing with a market cap of about 55.5 billion HKD. The capital market gave face, but not much.
The reason is easy to understand. Before listing, 73.7% of Zhipu's revenue came from localized deployment. Translated into human language, it means doing customized private deployments for government and enterprise clients. Finance, energy, government, power grid, deal by deal, project by project. On-site, debugging, operations, training, a full set of hard and tired work.
In the eyes of investors used to exponential growth curves, the portrait of this business equals "AI Old Timer": High marginal costs, difficult scaling, long payment collection cycles. Not sexy enough, not OpenAI enough.
The next day, January 9, MiniMax listed. First day surge of 109%, market cap broke through 105 billion HKD, almost twice that of Zhipu.
Why does the market love it?
Because MiniMax tells a story everyone understands and wants to hear most: C-end, Globalization, Super App.
The prospectus shows about 67% of its revenue comes from C-end AI native products, overseas emotional companion app Talkie, domestic Xingye, video tool Hailuo AI. 300 million users, overseas narrative, multimodal suite. The capital market looked at Talkie and played TikTok documentaries in their minds.
By March, MiniMax stock price hit 1,330 HKD, market cap broke 410 billion, briefly surpassing Baidu.
At that time, who was the darling and who was the has-been was clear at a glance.
Then, the script changed.
A Natural Control Experiment: Price Hike
What opened the gap between the two companies was perhaps a textbook natural control experiment in the first half of 2026. There was only one experimental variable: price hike.
First, look at the Zhipu group.
February 12, GLM-5 released, Coding Plan package price up 30%; March 16, GLM-5-Turbo launched, up another 20%; April GLM-5.1 up another 10%. Within one quarter, API pricing cumulative increase of 83%.
In the context of the whole industry trading price wars for market, this is going against the wind. The result?
Call volume did not decrease but increased by 400%, market demand exceeded supply, services queued, the company apologized publicly, and subsequently launched "Compute Partner" recruitment.
By the end of March, Zhipu API platform annual recurring revenue reached 1.7 billion RMB, a 60x explosion in one year; among China's top ten internet companies, nine were deeply calling GLM every day. CEO Zhang Peng's statement to the outside was only one sentence: The bottleneck is compute, not customers.
Then look at the MiniMax group.
June 1, flagship model M3 released, pricing about twice the previous generation, changed the long-used per-request billing to per-Token billing on the same day, and conveniently cancelled a batch of users' 29 RMB monthly subscription plans. The developer community called this "Backstab".
Market validation came extremely fast. About a week later, M3 announced a permanent price cut of 50%, price directly fell back to the same range as DeepSeek. A flagship product claimed to be a generation change, premium did not last seven days.
June 12, JPMorgan released a research report, writing the verdict for this experiment: Maintain Zhipu "Overweight", downgrade MiniMax to "Neutral", target price slashed from 1,100 HKD to 400 HKD, a cut of 64%.
The core logic of the report is simple. In a market where AI demand still exceeds inference supply, price cuts are not active concessions, but self-certification of insufficient competitiveness; being able to raise prices without losing orders is the only hard indicator of model capability validated by the market.
One product with an 83% price hike still queuing, and one product with a price hike beaten back to original form in seven days, the gap in between is the market cap gap.
The Death of Attention
To understand the result of this experiment, one must first understand the greatest invention of the mobile internet era: Wool comes from the pig.
The core of the business model of that era can be summed up in one sentence: Users use for free, advertisers pay, the platform counts money.
User attention is ad inventory, DAU is the speed of the money printing machine. More users, marginal cost approaches zero, revenue curve bends exponentially upward. For twenty years, all the giant myths of the Chinese internet are variations of this sentence.
Then the AI era came, and this sentence became invalid. Invalid very thoroughly, thoroughly enough that even the Universe Factory couldn't hold it.
Look at Doubao. The domestic C-end AI application with the highest monthly active users, 345 million MAU, daily Token call volume broke 180 trillion. Sounds a lot like Douyin back in the day? But behind these numbers is daily compute cost on the scale of hundreds of millions of RMB, 10 billion scale annually. So on June 24 this year, Doubao officially launched paid subscriptions, three tiers priced at 68, 200, 500 RMB per month.
The godfather of the free era personally turned off the lights for the free era.
The change happening here is worth saying once at the slowest pace:
In the mobile internet era, ordinary people's attention is ad inventory; in the large model era, ordinary people's attention is inference cost.
Before, every extra minute a user scrolled, the platform had one more piece of inventory to sell; now, every extra sentence a user chats, the platform burns one more portion of GPU. Before it was "more users equals more money", now it is "more users, thicker bill".
The sheep are still that group of sheep, but the sheep's attention is no longer naturally valuable. Only the sheep's intent, tasks, transactions, and productivity are valuable.
So the life or death question for C-end large models has never been "do you have users", but whether you can convert user attention into any one of paid subscriptions, enterprise efficiency, transaction commissions, workflow entry points, or commercial decision rights. If you can't convert, DAU is not an asset, it's a liability; MAU is not a moat, it's a money-burning speedometer.
MiniMax's ledger demonstrates this sentence to a cruel degree.
Its C-end business overall gross margin is only 4.7%. In the tech industry, this number equals charity. Talkie's ARPPU (Average Revenue Per Paying User) is only 5 USD, mainly relying on ad monetization; while Hailuo AI from the same stable walking the subscription model has an ARPPU of 56 USD. More critically, last quarter Talkie and Xingye MAU churn rate climbed to about 60%, and some overseas markets were removed for rectification.
300 million users, 4.7% gross margin. 2025 revenue about 79 million USD, adjusted net loss about 250 million USD, R&D expenditure accounts for more than 70% of revenue. Every user dating Talkie is consuming GPU in real money, and what they contribute is 5 USD of ad inventory. The market finally started asking that overdue question: Can virtual lovers actually drive large model R&D?
This is not MiniMax not working hard. It used the most standard posture of the mobile internet era, making blockbuster Apps, rushing user scale, telling overseas stories, and ran into an era where the rules had already been rewritten.
Every minute of Talkie's companionship, in the old era was inventory, in the new era is cost. It did all the right moves for the previous era, then lost to the era itself.
What Did Zhipu Do Right?
Looking back at Zhipu now, many people summarize Zhipu's victory as "chose B-end". This summary is half right.
Zhipu did not not do C-end. Zhipu Qingyan launched in August 2023, it was among the first batch of filed large model products in China. But by November 2025, App plus web end MAU combined was less than 3 million. By C-end scorecard, this is failing, but precisely because C-end couldn't be moved, Zhipu was forced to answer early that question MiniMax was forced to answer only at lock-up expiration: If attention isn't valuable, what is?
Its answer is: Productivity. Specifically, code.
This path was not invented by Zhipu, it originated from Anthropic. Relying on coding capability, Anthropic's valuation rushed past 900 billion USD in the first half of the year, letting the whole industry see that "models that can write code are the sexiest". People who write code are willing to pay for efficiency because AI writing code directly corresponds to priced productivity. What Zhipu did was move this validated pricing logic into the Chinese capital market for the first time, and used an 83% price hike amplitude and 400% call volume growth to complete local validation.
Look again at those government and enterprise hard tasks once mocked as "not sexy": Making power large models for 27 provincial companies of State Grid, making industry models covering the full process for PetroChina, doing Ascend All-in-One with Huawei, GLM architecture adapting 40+ domestic chips.
In the old narrative, this is called customized outsourcing; in the new narrative, this is called "Base Model + Domestic Compute + Developer Platform", every word steps on the G-spot of the current capital market.
The capital market always pays a premium for clear narratives.
Zhipu's story is one sentence, "Chinese Anthropic".
MiniMax's story is four sentences, Multimodal, C-end Social, Video Generation, Global Overseas. None of the four lines are wrong, but none are strong enough to price independently. The market doesn't know whether to value it as a model company, application company, or overseas company, finally valuing it at the cheapest one.
Morgan Stanley gives Zhipu 2027 expected PSR 57x, gives MiniMax 29x. Same track, valuation multiple difference is a full one times. The difference is not just technology, but narrative focus.
Don't Rush to Deify Zhipu
Speaking to this point, if one concludes "Zhipu won big" from this, that is replacing one naivety with another.
First, speak a fair word for MiniMax.
Founder Yan Junjie believed from day one that multimodal is the endgame. Text, voice, visual, video will eventually fuse, complete product ecosystem is the moat. This judgment on a five or ten-year scale, may not be wrong.
The problem is, endgame correct and midgame making money are two different things: You can be right on the endgame, but there is also probability of burning yourself to death in the midgame.
It still has cards in hand: B-end open platform revenue annual growth near 198%, gross margin about 70%, management promised 2026 end rush 1 billion USD ARR, back to A-share tutoring has also started. The steering wheel is turning, the question is just whether the turn is fast enough, and how much fuel is left in the tank.
Then splash a basin of cold water on Zhipu. 2025 revenue 724 million RMB, year loss 4.718 billion RMB, market cap 900 billion HKD. This PSR placed in any security analysis textbook belongs to content requiring teacher accompaniment. And there is a dark line in the financial report covered by cheers: Comprehensive gross margin compressed from 56.3% in 2024 to 41%. Zhipu did not large-scale self-build compute infrastructure, asset-light model, compute cost rises linearly with Token call volume. Price hike is less pricing confidence, more a necessary move to maintain business logic consistency. Its pricing power, half comes from model capability, half comes from the current compute supply shortage seller's market; when compute gap eases, whether this half card remains, no one dares to guarantee.
It won the relative race, not absolute valuation.
But regardless of how the two follow up, this half year's divergence has already written an era-level judgment on the K-line with 800+ billion HKD market cap difference: The Bible of the Mobile Internet is invalid.
The premise of "Wool comes from the pig" is that sheep's attention can be wholesaled to pigs at near zero cost; while in the large model era, every portion of attention must be priced by Token, depreciated by GPU.
The real change is not that sheep aren't valuable, it's that sheep's attention is no longer naturally valuable. Sheep's intent, sheep's tasks, sheep's transactions, sheep's productivity, are what's valuable.
In the previous era, who gathered more sheep, who wins.
In this era, who figures out what on the sheep is valuable first, who survives.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
