A Zhipu equals ten MiniMaxes: the sheep remain, but the pigs are gone

Original Author: Xiaobing

On July 9, 2026, Hong Kong stocks experienced a moment of dark comedy.

MiniMax faced its first post-IPO lockup expiration, with its stock price plunging over 20% intraday, its market cap breaching the HK$100 billion mark and shrinking to around HK$90 billion. This represented 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 its lockup expiration. The market had braced for another sell-off, but instead, it opened 3% lower and then rallied throughout the day, closing up 13%. The next day, it rose another 11%, solidifying its market cap at HK$900 billion.

One Zhipu AI is now roughly equivalent to ten MiniMax.

Six months ago, this equation was reversed, or at least, the bets were placed in the opposite direction.

Initially, the Market Bet on MINIMAX

Let's rewind to January of this year.

On January 8, Zhipu AI went public, touting the title of "World's First Major AI Model Stock," but only managed a 13.17% gain on its first day, closing with a market cap of approximately HK$55.5 billion. The capital market gave it face, but not much.

The reason isn't hard to fathom. Before its IPO, 73.7% of Zhipu AI's revenue came from local deployment—essentially providing customized, private deployments for government and enterprise clients. This meant securing deals one by one in sectors like finance, energy, government affairs, and power grids, and handling every painstaking project phase: on-site work, debugging, operations, maintenance, and training.

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

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

Why did the market love it?

Because MiniMax told a story everyone understood and wanted to hear: C-end (consumer-facing), globalization, a super app.

Its prospectus showed that approximately 67% of its revenue came from consumer-facing AI-native products: overseas emotional companion app Talkie, domestic app Xingye, and video tool Hailuo AI. Three hundred million users, an overseas narrative, a full-stack multimodal suite. The capital market looked at Talkie and saw a TikTok documentary playing in its head.

By March, MiniMax's stock price hit HK$1,330, its market cap surpassing HK$410 billion, briefly exceeding that of Baidu.

At that point, it was clear who was the darling and who was overlooked.

Then, the script changed.

A Natural Controlled Experiment: Price Hikes

What may have widened the gap between the two companies was a textbook natural controlled experiment in the first half of 2026, with a single variable: price increases.

First, look at Zhipu AI's side.

On February 12, GLM-5 was released, with the Coding Plan package rising by 30%; on March 16, GLM-5-Turbo launched with another 20% increase; in April, GLM-5.1 saw a further 10% rise. Within one quarter, API pricing had accumulated an 83% increase.

In an industry context defined by price wars for market share, this was like going against the wind. So what happened?

Call volume didn't decrease; it increased by 400%. Market demand outstripped supply, leading to service queues and a public apology from the company, which subsequently launched a "Computing Power Partner" recruitment drive.

By the end of March, Zhipu AI's API platform annual recurring revenue reached RMB 1.7 billion, skyrocketing 60 times in one year. Nine out of China's top ten internet companies were using GLM models extensively on a daily basis. CEO Zhang Peng's external communication boiled down to one sentence: The bottleneck is computing power, not customers.

Now look at MiniMax's side.

On June 1, the flagship model M3 was released, priced at roughly double the previous generation. On the same day, its long-standing per-call pricing was switched to per-token billing, and it also eliminated a 29-yuan monthly subscription package for some users. The developer community called this a "stab in the back."

Market validation was swift. About a week later, M3 announced a permanent 50% price cut, bringing its price directly in line with DeepSeek. A supposedly next-generation flagship product couldn't sustain its premium for even seven days.

On June 12, JPMorgan released a research report that read like a verdict on this experiment: maintaining an "Overweight" rating on Zhipu AI, downgrading MiniMax to "Neutral," and slashing the target price from HK$1,100 to HK$400—a 64% cut.

The core logic of the report was simple: in a market where AI demand still exceeds inference supply, price cuts aren't proactive concessions; they're self-identification of insufficient competitiveness. The ability to raise prices without losing orders is the only hard indicator that a model's capabilities have been validated by the market.

A product that can raise prices by 83% and still have queues, versus a product that returns to its original price after seven days—the difference between these two is the market cap difference.

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 side pays for the other.

The core business model of that era boiled down to one sentence: users get it for free, advertisers pay, and the platform counts the money.

User attention was the advertising inventory, DAU was the speed of the money printer. More users meant marginal costs approached zero, and the revenue curve curved exponentially upwards. For two decades, all the giant myths of China's internet were variations on this single sentence.

Then the AI era arrived, and this sentence stopped working. It failed so completely that even a giant like ByteDance couldn't withstand it.

Look at Doubao (the AI assistant). It's the highest monthly active consumer AI application in China, with 345 million MAUs and daily token calls exceeding 180 trillion. Sounds like Douyin back in the day, right? But behind these numbers is a daily computing cost in the hundreds of millions of RMB, totaling tens of billions annually. So, on June 24 of this year, Doubao officially launched paid subscriptions, with three tiers priced at 68, 200, and 500 RMB per month.

The godfather of the free era personally turned off the lights for the free era.

The change happening here deserves to be stated very, very slowly:

In the mobile internet era, ordinary people's attention was advertising inventory. In the large model era, ordinary people's attention is inference cost.

Before, every extra minute a user spent meant more inventory the platform could sell. Now, every extra sentence a user types means more GPU costs burning for the platform. Before, it was "more users equals more profit." Now, it's "more users equals a thicker bill."

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

So, the existential question for consumer-facing big models has never been "do you have users," but rather, can you convert their attention into any of these: paid subscriptions, enterprise efficiency, transaction commissions, workflow entry points, or business decision power. If you can't convert it, DAU isn't an asset, it's a liability. MAU isn't a moat, it's a burn rate meter.

MiniMax's financials demonstrate this point with brutal clarity.

Its overall gross margin for the consumer business is only 4.7%. In the tech industry, this number is practically charity. Talkie's ARPPU is only $5, relying mainly on ad revenue. In contrast, its subscription-based sibling Hailuo AI boasts an ARPPU of $56. More critically, the combined monthly active user churn rate for Talkie and Xingye climbed to roughly 60% in the last quarter, and Talkie even faced removals and rectifications in some overseas markets.

300 million users, 4.7% gross margin. Revenue of approximately $79 million in 2025, with an adjusted net loss of about $250 million, and R&D expenses accounting for over 70% of revenue. Every user falling in love with a virtual girlfriend on Talkie is consuming real GPU resources, while contributing only $5 worth of ad inventory. The market has started asking the long-overdue question: Can a virtual companion really support large-model R&D?

This isn't a case of MiniMax not trying. It used the most standard playbook of the mobile internet era—making hit apps, chasing user scale, telling an overseas expansion story—and charged headfirst into an era whose rules had already been rewritten.

Every minute of companionship on Talkie was inventory in the old era, but a cost in the new era. It did everything right for the previous era, and then lost to the era itself.

What Did Zhipu AI Do Right?

Looking back at Zhipu AI now, many attribute its victory to "choosing the B2B path." That assessment is only half right.

Zhipu AI didn't ignore the consumer side. It launched Zhipu Qingyan in August 2023, one of the first batch of registered large model products in China. Yet, by November 2025, its combined monthly active users across the app and web browser was under 3 million. By consumer-grade metrics, that's a failing grade. But precisely because it struggled on the consumer front, Zhipu AI was forced to answer the question MiniMax was only compelled 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. Thanks to its coding capabilities, Anthropic's valuation surged past $900 billion in the first half of the year, making the entire industry realize that "models that can write code are the most valuable." People who write code are willing to pay for efficiency because AI code generation directly corresponds to quantifiable productivity. What Zhipu AI did was bring this proven pricing logic to the Chinese capital market for the first time and complete its local validation with an 83% price increase and a 400% surge in call volume.

Consider its previously derided "unsexy" government and enterprise work: building a power grid large model for 27 provincial subsidiaries of State Grid Corporation of China, developing an industry-wide full-process model for PetroChina, collaborating with Huawei on Ascend integrated machines, and adapting the GLM architecture to over 40 different types of domestic chips.

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

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

Zhipu AI's story is one sentence: "China's Anthropic."

MiniMax's story is four sentences: multimodal, consumer social, video generation, global expansion. None of these four lines is inherently wrong, but none is strong enough to command an independent valuation on its own. The market doesn't know whether to value it as a model company, an application company, or an overseas expansion company, so it ends up valuing it at the cheapest of the three.

Morgan Stanley projects a 2027 price-to-sales ratio of 57 times for Zhipu AI, and 29 times for MiniMax. In the same track, the valuation multiple is double. The difference isn't just in technology; it's also in the sharpness of the narrative.

Don't Rush to Crown Zhipu AI

Having said all this, concluding that "Zhipu AI has won it all" would be replacing one form of naivety with another.

First, let's say a fair word for MiniMax.

Founder Yan Junjie has believed from day one that multimodality is the endgame—text, voice, vision, and video will eventually converge, and a complete product ecosystem is the real moat. Over a five or ten-year horizon, this judgment might not be wrong.

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

It still has cards to play: its B2B open platform revenue grew nearly 198% annually, with a gross margin of about 70%. Management has stated a goal of reaching $1 billion in ARR by the end of 2026, and preparations for a return to the A-share market have begun. The steering wheel is turning; the question is just how fast it can turn and how much fuel is left in the tank.

Now, let's throw some cold water on Zhipu AI. Its 2025 revenue was RMB 724 million, with an annual loss of RMB 4.718 billion, and a market cap of HK$900 billion. This price-to-sales ratio would require teacher supervision in any securities analysis textbook. Moreover, there's a dark line buried beneath the cheers: the comprehensive gross margin shrank from 56.3% in 2024 to 41%. Zhipu AI hasn't built its own large-scale computing infrastructure. Under this asset-light model, computing costs rise linearly with token call volume. The price hike isn't just pricing confidence; it's a necessary action to maintain a self-consistent business logic. Half of its pricing power comes from model capability, and the other half from the current seller's market where computing demand outstrips supply. Once the computing power gap eases, no one can guarantee that half of this card will still be in play.

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

But regardless of what happens next for these two companies, the divergence over the past six months has already written an era-defining judgment onto the K-line chart, using a market cap difference of over HK$800 billion: The bible of the mobile internet era is obsolete.

The premise of "one side pays for the other" was that user attention could be wholesaled to the other party at near-zero cost. In the large model era, every unit of attention must be priced per token and depreciated against GPU costs.

The real change isn't that the sheep have lost their value. It's that the sheep's attention no longer inherently holds value. The sheep's intentions, tasks, transactions, and productivity—those are what hold value.

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

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