一个智谱等于十个MiniMax：羊还在，猪没了

Original Author: Xiaobing

On July 9, 2026, Hong Kong stocks staged a dark comedy.

MiniMax faced its first post-IPO lockup expiration. Its share price plunged over 20% intraday, and its market cap fell below the HK$100 billion mark, shrinking to around HK$90 billion. This represents an evaporation of HK$320 billion from its all-time high of HK$410 billion in March.

Just the day before, Zhipu AI also faced a lockup expiration. The market braced for another sell-off. Instead, it opened 3% lower, then rallied throughout the day to close up 13%. The following day, it gained another 11%, stabilizing its market cap at HK$900 billion.

One Zhipu AI is now roughly equivalent to ten MiniMax's.

Six months ago, the equation was reversed – or at least, that's where the bets were placed.

Initially, the Market Bet on MINIMAX

Let's rewind to January this year.

On January 8th, Zhipu AI rang the IPO bell. Debuting with the title "World's First Major AI Model Stock," it only gained 13.17% on its first day, closing with a market cap of around HK$55.5 billion. The capital market gave it face, but not much.

The reason wasn't hard to understand. Before its IPO, 73.7% of Zhipu AI's revenue came from local deployment. In plain English, this means providing customized, private deployments for government and enterprise clients. Sectors like finance, energy, government affairs, and the power grid – it's about negotiating deals one by one, executing projects individually, involving on-site work, debugging, operations, maintenance, and training. A full set of tough, labor-intensive tasks.

In the eyes of investors accustomed to exponential growth curves, this business profile looked like an "AI old-timer": high marginal costs, difficult scalability, and long payment cycles. Not sexy enough, not OpenAI enough.

The next day, January 9th, MiniMax went public. It surged 109% on its debut, pushing its market cap past HK$105 billion – nearly double that of Zhipu AI.

Why did the market love it?

Because MiniMax told a story everyone could understand and was most willing to hear: Consumer-facing, globalized, super-app.

Its prospectus showed that about 67% of its revenue came from C-end AI-native products, including the overseas emotional companion app Talkie, China's Xingye (Star Wilds), and the video generation tool Hailuo AI. 300 million users, an overseas narrative, a multi-modal full suite. The capital market looked at Talkie and imagined a TikTok documentary playing in their heads.

By March, MiniMax's stock price hit HK$1,330, pushing its market cap past HK$410 billion, temporarily surpassing even Baidu.

At that time, who was the sweetheart and who was the wallflower was crystal clear.

Then, the script flipped.

A Natural Controlled Experiment: Raising Prices

The divergence between the two companies was perhaps driven by a textbook natural experiment in the first half of 2026. The only variable in this experiment was: raising prices.

Let's look at the Zhipu AI group first.

On February 12th, GLM-5 was released, with the Coding Plan package rising by 30%. On March 16th, GLM-5-Turbo launched with another 20% increase. In April, GLM-5.1 followed with a 10% rise. Within a single quarter, API pricing was raised by a cumulative 83%.

In an industry context where everyone is competing for market share through price wars, this was blatant defiance. And the result?

API call volumes didn't drop; they increased by 400%. Market demand outstripped supply, services had queues, the company issued a public apology, and subsequently launched a "Computing Power Partner" recruitment drive.

By the end of March, Zhipu AI's API platform had reached an annual recurring revenue (ARR) of RMB 1.7 billion, a staggering 60-fold increase in one year. Nine out of China's top ten internet companies were making deep daily calls to GLM. CEO Zhang Peng's only external statement was: The bottleneck is computing power, not customers.

Now let's look at the MiniMax group.

On June 1st, the flagship model M3 was released, priced roughly double its predecessor. On the same day, the long-standing per-use billing was changed to per-Token billing. It also quietly canceled the RMB 29 monthly subscription package for a group of users. The developer community called this a "betrayal."

The market's verdict came swiftly. About a week later, M3 announced a permanent 50% price cut, bringing its price back down to the same range as DeepSeek. A flagship product purportedly representing a generational upgrade couldn't sustain its premium pricing for even seven days.

On June 12th, JPMorgan released a research report, essentially delivering a verdict on this experiment. They maintained an "Overweight" rating for Zhipu AI while downgrading MiniMax to "Neutral," slashing the target price from HK$1,100 to HK$400 – a cut of 64%.

The core logic of the report was simple: In a market where AI demand still exceeds inference supply, lowering prices isn't proactive concession; it's self-certification of insufficient competitiveness. Being able to raise prices without losing business volume is the only hard indicator that validates a model's capability in the market.

A product that can raise prices by 83% and still have queues, versus a product whose price increase is reversed within a week – the gap between these is the market cap gap.

The Death of Attention

To understand the outcome of this experiment, you first need to understand the greatest invention of the mobile internet era: One thing is paid for by another.

The core business model of that era could be summed up in one phrase: Users use it for free, advertisers pay for it, and the platform counts the money.

Users' attention was the advertising inventory; DAU was the speed of the printing press. More users meant marginal costs approaching zero and an exponentially rising revenue curve. For two decades, all the myths of China's internet giants were variations on this single phrase.

Then the AI era arrived, and this phrase stopped working. It stopped working so completely that even the giant couldn't withstand it.

Look at Doubao. It's the highest MAU AI application in China's consumer market, with 345 million monthly active users and daily Token call volume exceeding 180 trillion. Sounds a lot like Douyin back in the day, right? But behind these numbers are daily computing costs in the hundreds of millions of RMB, totaling tens of billions annually. So, on June 24th this year, Doubao officially launched paid subscriptions, with three tiers priced at RMB 68, 200, and 500 per month.

The founding father of the free era personally turned off the lights for the free era.

The change happening here deserves to be stated as slowly and clearly as possible:

In the mobile internet era, ordinary people's attention was advertising inventory; in the era of large models, ordinary people's attention is inference cost.

Before, every extra minute a user spent on the platform meant more inventory for the platform to sell. Now, every extra sentence a user types means more GPU power burned by the platform. Before, it was "more users, more profit." Now, it's "more users, a thicker bill."

The sheep are still the same sheep, but their attention is no longer inherently valuable. Only the sheep's intentions, tasks, transactions, and productivity hold value.

So, the life-or-death question for C-end large models has never been "do you have users?" but rather, can you convert users' attention into any of the following: paid subscriptions, enterprise efficiency, transaction commissions, workflow gateways, or business decision-making power. If you can't, DAU isn't an asset; it's a liability. MAU isn't a moat; it's a burn rate meter.

MiniMax's financial statements demonstrate this point with brutal clarity.

Its overall C-end business gross margin is only 4.7%. In the tech industry, this number is essentially a charitable endeavor. Talkie's ARPPU (Average Revenue Per Paying User) is only $5, primarily relying on ad monetization. In contrast, its sibling product Hailuo AI, which uses a subscription model, has an ARPPU of $56. More critically, the monthly active user churn rate for Talkie and Xingye surged to about 60% in the last quarter, and some overseas markets faced delisting and rectification.

300 million users, 4.7% gross margin. Annual revenue of approximately $79 million in 2025, an adjusted net loss of around $250 million, and R&D expenses accounting for over 70% of revenue. Every user chatting with their virtual companion on Talkie is burning GPU power with real money, contributing only $5 worth of ad inventory. The market finally started asking the belated question: Can a virtual girlfriend really support the R&D of a large model?

This isn't because MiniMax didn't try hard. It used the most standard playbook of the mobile internet era – creating hit apps, chasing user scale, telling a global expansion story – but charged into an era where the rules had already been rewritten.

Every minute of companionship on Talkie was inventory in the old era but is a cost in the new era. It did everything right according to the previous era's playbook, but lost because the era itself changed.

What Did Zhipu AI Do Right?

Looking back at Zhipu AI now, many people summarize its victory as "choosing the B2B path." This summary is only half correct.

It's not that Zhipu AI never tried the C-end market. Zhipu Qingyan was launched in August 2023, one of the first batch of registered large model products in China. Yet, by November 2025, its combined monthly active users for the app and web version were less than 3 million. By C-end metrics, this is a failing grade. But precisely because it failed in the C-end market, Zhipu AI was forced to answer early the question that MiniMax only had to confront on its lockup expiration day: If attention isn't valuable, what is?

Its answer was: productivity. Specifically, code.

Zhipu AI didn't invent this path; it originated with Anthropic. By leveraging coding capabilities, Anthropic's valuation surged past $900 billion in the first half of the year, making the entire industry realize that "models capable of writing code are the most valuable." People who write code are willing to pay for efficiency because AI coding directly translates into quantifiable productivity. What Zhipu AI did was bring this already validated pricing logic into the Chinese capital market for the first time, completing its local validation through an 83% price increase and a 400% surge in call volume.

Now, consider those "unsexy" government and enterprise tasks it was once ridiculed for: building power models for 27 provincial subsidiaries of State Grid, creating industry-wide models covering the entire workflow for PetroChina, collaborating with Huawei on Ascend all-in-one machines, and adapting the GLM architecture to over 40 types of domestic chips.

In the old narrative, this was called customized outsourcing. In the new narrative, it's called a "foundation model + domestic computing power + developer platform." Every word hits a sweet spot for today's capital market.

The capital market always pays a premium for a clear narrative.

Zhipu AI's story is simple: the "Anthropic of China."

MiniMax's story is more complex: multi-modal, C-end social, video generation, global expansion. None of these four lines is wrong, but none is strong enough to stand alone for valuation. The market doesn't know whether to value it as a model company, an app company, or a global expansion company. Ultimately, it gets valued according to the cheapest metric.

Morgan Stanley gives Zhipu AI a projected 2027 price-to-sales ratio of 57x, but gives MiniMax only 29x. On the same track, the valuation multiple difference is a full double. The gap isn't just in technology; it's also in narrative focus.

Don't Rush to Crown Zhipu AI

At this point, concluding "Zhipu AI has completely won" would be replacing one naivety with another.

Let's first say a fair word for MiniMax.

Founder Yan Junjie believed from day one that multi-modality is the endgame – that text, voice, vision, and video will eventually converge, and a complete product ecosystem is the true moat. This judgment, viewed on a five or ten-year horizon, might not be wrong.

The problem is that being right about the endgame and making money in the middle game are two different things. You can be right about the endgame but also burn yourself out in the middle game.

MiniMax still has cards to play: its B-end open platform revenue grew nearly 198% year-on-year with a gross margin around 70%. Management has stated a goal of reaching $1 billion ARR by the end of 2026, and has initiated preparations for a return to the A-share market. The steering wheel is turning; the only questions are how fast the turn can be and how much fuel is left in the tank.

Now, a dose of cold water for Zhipu AI. It had a revenue of RMB 724 million and a net loss of RMB 4.718 billion for the year 2025. Its market cap is HK$900 billion. This price-to-sales ratio, by any textbook on securities analysis, warrants a chaperone for reading. Moreover, a worrying trend hidden beneath the cheers in the financial report: the consolidated gross margin compressed from 56.3% in 2024 to 41%. Zhipu AI hasn't built large-scale computing infrastructure on its own. Under this asset-light model, computing costs rise linearly with token call volumes. The price increase is less a sign of pricing confidence and more a necessary move to maintain logical business coherence. Half of its pricing power comes from model capability, the other half from the current seller's market where computing power supply can't meet demand. Once the computing power gap eases, no one can guarantee that this half of the pricing power will remain.

It's winning a relative race, not an absolute valuation race.

But regardless of how either company evolves going forward, the divergence of the past six months has inscribed an era-defining judgment on the K-line chart, backed by a market cap difference of over HK$800 billion: The bible of the mobile internet is obsolete.

The premise of "one thing paid for by another" was that users' attention could be wholesaled to the paying party at near-zero cost. In the era of large models, every unit of attention must be priced per Token and depreciated per GPU.

The real change isn't that the sheep aren't valuable anymore. It's that the sheep's attention is no longer inherently valuable. What holds value now is the sheep's intentions, the sheep's tasks, the sheep's transactions, and the sheep's productivity.

In the previous era, whoever herded the most sheep won.

In this era, whoever figures out first what part of the sheep is truly valuable, survives.