Written by: Xiaobing

On July 9, 2026, the Hong Kong stock market put on a black humor show.

MiniMax faced its first round of unlock after going public, with stock prices dropping over 20% during the day, and market value crashing through the 100 billion HKD barrier, shrinking to around 90 billion, evaporating a total of 320 billion from the historical high of 410 billion in March.

Just the day before, Zhipu also experienced an unlock, and the market was ready for a second stampede. However, it opened down 3% and then surged, closing up 13%, and rising another 11% the next day, firmly establishing a market value of 900 billion HKD.

One Zhipu is roughly equivalent to ten MiniMax.

And six months ago, this equation was written the other way around, at least in terms of bets placed.

At the start, the market votes for MiniMax

Let's rewind to January of this year.

On January 8, Zhipu rang the bell, carrying the title of "the first global large model stock", but it only rose 13.17% on its first day, closing with a market value of about 55.5 billion HKD. The capital market showed some respect, but not much.

The reasons are not hard to understand. Before going public, Zhipu derived 73.7% of its income from localized deployments, which in plain language means providing customized private deployments for government and enterprise clients. Financial services, energy, government, electric grid, negotiating one deal at a time, completing one project after another, on-site work, debugging, operations, and training — a full set of tough and tedious tasks.

In the eyes of investors accustomed to exponential growth curves, this business representation is roughly equivalent to "AI Old Deng": high marginal cost, difficult to scale, long payment cycles. Not attractive enough, not OpenAI enough.

The next day, January 9, MiniMax went public. It surged 109% on the first day, with its market value suddenly exceeding 105 billion HKD, almost double that of Zhipu.

Why does the market love it?

Because MiniMax tells a story that everyone understands and wants to hear: Consumer end, globalization, super applications.

The prospectus shows that about 67% of its revenue comes from consumer-end AI native products, such as the overseas emotional companionship application Talkie, domestically from Xingye, and video tool Haizhuo AI. 300 million users, overseas narratives, and a multimodal comprehensive package. The capital market looks at Talkie and thinks of the TikTok documentary.

By March, MiniMax's stock price soared to 1330 HKD, and its market value surpassed 410 billion, briefly exceeding Baidu.

At that time, who was the little sweetie and who was Mrs. Niu was very clear.

Then, the script changed.

A natural comparison experiment: Price Increase

What perhaps widened the gap between the two companies was a textbook natural comparison experiment in the first half of 2026, with only one experimental variable: price increase.

First, let’s look at Zhipu’s group.

On February 12, GLM-5 was released, and the Coding Plan package increased by 30%; on March 16, GLM-5-Turbo was launched, increasing by another 20%; in April, GLM-5.1 increased by another 10%. In one quarter, API pricing was cumulatively raised by 83%.

In the context of the entire industry engaging in a price war for market share, this was an act of defiance against the wind; what was the result?

The call volume not only didn’t decrease but instead increased by 400%, market supply couldn’t meet demand, services queued, and the company publicly apologized, then launched the “computational power partner” recruitment program.

By the end of March, Zhipu’s API platform's annual recurring revenue reached 1.7 billion, a staggering increase of 60 times; nine out of China’s top ten internet companies were deeply calling GLM every day. The CEO, Zhang Peng, had only one statement to the outside: The bottleneck is in computing power, not in customers.

Now let’s look at MiniMax’s group.

On June 1, the flagship model M3 was released, priced at about twice that of its predecessor, and on the same day changed the longstanding per-use billing to token-based fees, also canceling a batch of users’ 29 HKD monthly lease package, which the developer community called “backstabbing.”

The market’s validation came quickly; about a week later, M3 announced a permanent price reduction of 50%, bringing the price back to the same range as DeepSeek. A flagship product claimed to be next-generation could not maintain its premium for even seven days.

On June 12, JPMorgan published a research report, writing a judgment for this experiment: Upgrading Zhipu to “buy” and downgrading MiniMax to “neutral”, cutting its target price from 1100 HKD to 400 HKD, a cut of 64%.

The core logic of the report is very simple: in a market where AI demand still exceeds reasoning supply, price reduction is not an active concession but rather a self-certification of insufficient competitiveness; being able to increase prices without losing orders is the only hard indicator of the model's capability being validated by the market.

A product that increased prices by 83% and still has to queue, compared to a product that was hit back to its original form within seven days, illustrates the difference in market value.

The Death of Attention

To understand the results of this experiment, one must first understand the greatest invention of the mobile internet era: The wool comes from the pig.

The core of the business model of that era was simple: users use it for free, advertisers pay, and platforms count their money.

The attention of users is the advertising inventory, and DAU is the speed of the money printing machine. The more users there are, the marginal cost approaches zero, and the revenue curve rises exponentially. All the myths of the giants in China's internet over the past twenty years are variations of this concept.

Then the AI era arrived, rendering this concept ineffective. It became obsolete to the point where even the universe's giants couldn't bear it.

Look at Doubao. The highest monthly active AI application in the domestic consumer end, with 345 million monthly active users, daily token call volume soaring over 180 trillion — doesn’t this sound a lot like the Douyin of those days? But behind these numbers are daily operating costs of over 100 million, scaling to tens of billions a year. Thus, on June 24 of this year, Doubao officially launched paid subscriptions, with three pricing tiers: 68, 200, and 500 HKD per month.

The founder of the free era has personally turned off the lights on that era.

The change happening here is worth stating slowly:

In the mobile internet era, ordinary people's attention is advertising inventory; in the large model era, ordinary people's attention is reasoning costs.

Previously, the longer users swiped, the more inventory platforms could sell; now, the more users chat, the more GPU the platform consumes. It was once “the more users, the more profit”, now it’s “the more users, the thicker the bills”.

The sheep are still the same flock of sheep, but their attention is no longer inherently valuable; only the intentions, tasks, transactions, and productivity of the sheep are valuable.

Therefore, the life-and-death question for large consumer models has never been about “whether there are users”, but whether you can convert users' attention into any of the following: paid subscriptions, enterprise efficiency, transaction commissions, workflow access, and business decision-making power. If it can’t be converted, DAU is not an asset but a liability; monthly active users are not a moat but a speedometer for burning money.

MiniMax’s ledger demonstrates this point to a brutal degree.

Its consumer-end business has an overall gross margin of only 4.7%, which is roughly equivalent to charity in the technology industry. Talkie's ARPPU (average revenue per paid user) is only 5 USD, mainly relying on advertising monetization; while its partner, subscription-based Haizhuo AI, has an ARPPU of 56 USD. Worse still, by the end of last year, Talkie and Xingye's monthly active user churn rate had climbed to about 60%, with some overseas markets even facing removal and rectification.

300 million users, 4.7% gross margin. Projected revenue for 2025 is about 79 million USD, with an adjusted net loss of about 250 million USD, and R&D expenses accounting for over 70% of revenue. Every user that becomes romantically involved with Talkie is consuming GPU in real monetary terms, while they contribute merely 5 dollars' worth of advertising inventory. The market has finally begun to ask that overdue question: Do virtual partners really support large model development?

This is not because MiniMax isn't trying; it is applying the most standard approach from the mobile internet era, creating blockbuster apps, boosting user scale, and telling overseas stories, diving into an era where rules have already been rewritten.

Every minute of companionship from Talkie, in the old era, represented inventory; in the new era, it represents cost. It did all the right things from the previous era and then succumbed to the era itself.

What did Zhipu do right?

Looking back at Zhipu now, many people sum up Zhipu’s victory as “choosing the B-end”, which is half correct.

Zhipu didn't shy away from the C-end; it launched Qingyan in August 2023, being among the first batch of large model products registered domestically, yet by November 2025, the total monthly active users for the app and web version was less than 3 million. By C-end metrics, this is failing; but precisely because the C-end couldn't gain traction, Zhipu was forced to answer the question MiniMax only reluctantly faced on the unlock day: If attention is not valuable, what is?

Its answer is: productivity. More specifically, it is code.

This path was not invented by Zhipu; it originated from Anthropic. Relying on programming capabilities, Anthropic’s valuation shot past 900 billion USD in the first half of the year, making the entire industry realize that “models capable of writing code are the most desirable”, and that those who write code are willing to pay for efficiency, as AI writing code directly corresponds to measurable productivity. What Zhipu did was to bring this proven pricing logic into the Chinese capital market for the first time, completing the local verification with an 83% price increase and a 400% increase in call volume.

Next, consider its often ridiculed "unsexy" hard work for government and enterprises: creating large model systems for 27 provincial companies of the State Grid, developing comprehensive industry models for PetroChina, collaborating with Huawei on Ascend integrated machines, and adapting GLM architecture for over 40 domestic chips.

In the old narrative, this was called customized outsourcing; in the new narrative, it is called “foundation model + domestic computing power + developer platforms”, with every word hitting the G-spot of the current capital market.

The capital market always pays a premium for clear narratives.

Zhipu's story is simply stated as “China's Anthropic”.

MiniMax’s story contains four sentences: multimodal, C-end social, video generation, and global expansion. Each line isn't necessarily wrong, but none is strong enough to command independent pricing; the market doesn't know whether to value it as a model company, an application company, or an overseas company, ultimately opting for the cheapest option.

Morgan Stanley gives Zhipu a projected sales multiple of 57 times in 2027, while MiniMax receives 29 times, a full doubling in valuation multiples on the same playing field, where the differences lie not just in technology but in narrative focus.

Don't rush to canonize Zhipu

If we reach the conclusion that “Zhipu has won thoroughly” from this point, it is merely replacing one kind of naivety with another.

First, let’s give MiniMax a fair word.

Founder Yan Junjie has believed since day one that multimodality is the endgame; text, voice, vision, and video will eventually converge, and a complete product ecosystem will serve as a barrier. This judgment isn't necessarily wrong when considering the scale of five or ten years.

The issue is that being correct in the endgame and making money in the mid-game are two different things: you can be right in the endgame but also risk burning out in the mid-game.

It still has cards to play: the B-end open platform revenue is growing nearly 198% annually, with a gross margin of about 70%; management has announced intentions to reach 1 billion USD ARR by the end of 2026, and return-A counseling has already commenced. The steering wheel is being turned; the only question is whether the turn can be fast enough and how much fuel remains.

Now let’s splash some cold water on Zhipu. In 2025, revenue was 724 million RMB, with a loss of 4.718 billion RMB for the year, and a market value of 900 billion HKD; this sales multiple would be considered a topic in any securities analysis textbook that requires a teacher to read alongside the students. Moreover, there is a line in the financial report that was drowned out by applause: the overall gross margin compressed from 56.3% in 2024 to 41%. Zhipu has not built a large infrastructure for computational power; under a light asset model, computing costs rise linearly with token call volume. This makes price increases less about pricing confidence and more about a necessary action to maintain the consistency of business logic. Its pricing power comes partly from model capabilities, and partly from a seller's market driven by the current shortfall in computing power; once the gap in computing power narrows, whether this half of the card still holds is an uncertain bet.

It wins in relative races, not absolute valuations.

But regardless of how both companies fare going forward, the differentiation over the past six months has written a generational judgment in market value differences over 800 billion HKD on the K-line: The mobile internet’s bible has become ineffective.

The premise of “the wool comes from the pig” is that sheep's attention can be almost zero-cost wholesaled to pigs; but in the age of large models, every ounce of attention must be charged per token and depreciated based on GPU usage.

The real change is not that sheep are no longer valuable, but that the attention of sheep is no longer inherently valuable. The intentions of the sheep, the tasks of the sheep, the transactions of the sheep, and the productivity of the sheep are what truly hold value.

In the last era, those who round up the most sheep win.

In this era, those who first understand what is valuable about the sheep will survive.