MiniMax, taking the blow for the entire AI industry

Both are AI companies, both facing the unlocking of restricted shares. 

Zhipu surged 13%, while MiniMax plummeted 24% in two days. 

In the same week, Zhipu and MiniMax delivered completely opposite market outcomes. 

This marks the first time the capital market has publicly declared with real money that the valuation honeymoon period for AI companies is completely over. 

From today onwards, whether it's foundational large models or consumer-facing applications, every player will face rigorous scrutiny one by one. 

MiniMax has become the first name called to stand trial. 

1. 48 Hours, Two Opposite Verdicts

This is how the events unfolded. 

On July 9, MiniMax saw its first batch of restricted shares unlocked since its listing. 153 million shares hit the market, accounting for nearly half of the total share capital. The stock plunged over 20% intraday, eventually closing down nearly 18%. 

Criticism flooded in immediately. 

The performance of its new-generation M3 model failed to meet market expectations, yet the billing model was forcibly switched from a subscription system to a token-based consumption system, effectively raising prices; the reputation of its Coding product collapsed. Its flagship consumer app Xingye was ordered to rectify operations, and coupled with the strictest AI regulations to date, its growth story could no longer hold up. 

The market reached a unified and swift conclusion: MiniMax is in danger.

Then, in less than 24 hours, the narrative completely flipped. 

On July 10, MiniMax announced the completion of a HKD 16 billion financing round, with participation from sovereign wealth funds, long-term capital funds, Chinese institutional investors, and multi-strategy funds spanning Asia-Pacific, Europe, and the United States. More than 100 institutions joined the deal, resulting in 7x oversubscription. 

The same day, founder Yan Junjie sent a company-wide letter, stating he would forgo all salaries starting immediately until the company achieves AGI. He also pledged 4% of his personal shares as team incentives and earmarked another 1% to establish an open-source fund. 

The same company, within 48 hours, received two completely different valuations from the market. 

While the secondary market was panic selling, over 100 institutions in the primary market were scrambling to pour in capital.

What on earth does this reveal? 

2. The Market Isn't Trading MiniMax Itself

Let's return to that most striking comparison. 

Zhipu's restricted share unlocking led to a rally; MiniMax's unlocking triggered a crash. 

To many, this is because MiniMax's products underperformed, its M3 pricing strategy backfired, and users abandoned the platform. 

This makes sense, but it only scratches the surface. 

The deeper shift is that capital has begun pricing AI companies using two entirely distinct logics. 

For one category of companies, the market trades on faith.

Zhipu is a typical example. Also listed and facing restricted share unlocking, it carries minimal baggage from consumer-facing products, and the market still values it as a model platform. Capital focuses on its technical moat, financing capabilities, and ecological positioning. 

Restricted share unlocking? It even turned into positive news. The fact that long-term institutions are willing to take up shares proves the market's long-term bullish confidence in the value of its underlying infrastructure. 

Its valuation is anchored on the premise that it will be worth far more in the future. Even though OpenAI and Anthropic are not listed, they follow the same logic. 

For the other category of companies, the market trades on financial statements.

Once a company falls into this category, everything changes immediately. What's its DAU? What's its ARPU? How are renewal rates and retention metrics? Why raise prices? What's the guidance for the next quarter? 

MiniMax happened to land exactly on this dividing line. 

It has Hailuo AI, Xingye, an open platform, as well as products, users, and revenue. These assets, which should have been its strengths, inadvertently gave capital a reason to evaluate it against the standards for internet product companies. 

More critically, a point many have overlooked: 

MiniMax is the first large model company to face real scrutiny in the secondary market.

Zhipu is also listed on the Hong Kong Stock Exchange, but Zhipu was never pressed by the market about its DAU sequential growth rate or commercialization conversion funnel. 

MiniMax is the first AI company to stand under the secondary market spotlight carrying the three mountains of products, users, and revenue, being reviewed line by line by the full set of public company disciplines: shareholders, restricted share unlocking, P/E ratios, financial reports, and institutional holdings. 

In the past, discussions about AI companies revolved around models, technology, financing, and valuations. 

Now, conversations about MiniMax have shifted to equity structure, overhang from tradable share unlocking, gross margins, commercialization paths, and growth sustainability. 

These two types of conversations belong to entirely different worlds. 

Today's MiniMax allows capital to measure an AI company for the first time using the full standards of a public listed firm. 

3. MiniMax's Real Challenge

MiniMax's current position is not the full story of this trial. 

To understand why this happened, we must first grasp what kind of company MiniMax really is. 

According to reports from Silicon Star People, an observer who has closely witnessed MiniMax's internal operations once documented this scene: 

On the eve of the M3 launch on June 1, dozens of engineers' cursors flickered on the same document, updating the latest benchmark data. Founder Yan Junjie (internal nickname IO) was also in the document. For most of the time, he only listened without directing, until enough information was gathered, then he spoke to make the final decision. 

This is a company that treats "context" as its operating system.

Whoever holds the relevant information joins the group chat. Information flows in real time and transparently, with almost no buffer layers. Colleagues from different departments argue openly in front of everyone, then shake hands and get back to work. Founders are directly challenged by employees. Two co-founders can engage in heated debates, yet resume normal collaboration in the next meeting. 

"At big tech firms, you worry about projects getting axed, blame shifting, or complaints from other teams. At MiniMax, as long as you figure out the right path and execute it correctly, you're fine." 

This culture delivers extremely high efficiency. After the M3 launch, its pricing plan sparked user fury, and the entire process from identifying the problem, internal public debate, plan revision, to issuing a public apology took only 24 hours. 

But the flip side of this efficiency is blind spots. The company's full attention has been placed entirely on model intelligence.

From internal KPIs to resource allocation, everything serves the single goal of delivering a more powerful model. Operational trivialities like how to smoothly migrate payment plans or communicate changes to new and existing users, which pure technical staff see as irrelevant to model technical metrics, were instinctively pushed to the bottom of the resource priority list. 

An insider involved in decision-making recalled: If we didn't make the change, users might not be able to fully utilize our new model. 

The logic wasn't wrong. But they skipped the step of explaining to users why the change was necessary.

Switching from subscription billing to token-based consumption came with no clear guidelines, and explanations on official pages were vague. Users only discovered the rules had changed after logging in, when their credit consumption surged sharply. 

This user-empathetic operational misstep quickly turned into a wave of refunds from B-end developers, amplified into a brand crisis on social media, and ultimately transmitted directly to the secondary market, materializing as a double-digit single-day drop in share price. 

The engineers underestimated their users.

Worse still, the colleague in charge of pricing had only joined the company two or three months earlier, with zero awareness of long-time users' historical sentiments. An insider said: What he lacked was context, and that's everyone's responsibility. 

Ironically, this company that emphasizes context more than any other lost the most critical context when facing its own users. 

Users were not in those internal group chats. They had no way to know the technical logic behind the scenes. Users only knew one thing: their money was being spent faster, and no one had told them why. 

The belief MiniMax has held since its founding can be summed up in one word: Scale.

Intelligence, models, foundational capabilities. M3 has a total parameter count of 428B, yet only activates 23B parameters per inference. Many were surprised its parameter count was smaller than expected, but this is exactly its philosophy: push intelligence to its peak with the most elegant architecture possible. 

"Mini" is the starting point, "Max" is the destination. This belief brought it this far, but also left it vulnerable to today's setbacks.

Capital has no patience to wait for a company to slowly turn scale into revenue, especially when that company is already publicly listed. 

4. An Entire Industry Is Rewriting Its Rules

MiniMax's predicament is not an isolated case—it is just the first one placed under the spotlight. 

Over the past three years, competition among AI companies was essentially a competition over models. Higher benchmark scores, larger parameter counts, newer architectures would guarantee capital inflows. 

In 2026, the game has changed. 

Today's AI competition requires three elements to hold true simultaneously: models, products, and capital narratives.

OpenAI's valuation has exceeded one trillion US dollars, not just because of its strong model capabilities, but also because ChatGPT boasts a user base of 1 billion monthly active users, stable enterprise-level paid revenue, and an increasingly mature commercialization closed loop. 

Anthropic continues to secure sky-high financing rounds. Not only because Claude leads in technical performance, but also because its reputation in the enterprise market is well established, with Amazon and Google competing to increase their stakes. 

Looking back at MiniMax: 

Its models are not underperforming. After adjusting the inference architecture, M3's TPS (tokens per second) was optimized from 30 at launch to 80, with its first-pack latency entering the global top tier. Core call volume metrics on its API open platform keep growing, and third-party enterprise clients that previously left due to performance issues are clearly returning. 

But these metrics are invisible to the secondary market, which is also unwilling to wait for them to materialize. 

All the secondary market sees is the pricing controversy, user churn, product rectification, and growth skepticism—each point deducts points against public company financial and compliance standards.

This marks the first time the AI industry has truly transitioned from model competition to corporate competition. 

In the past, the race was about who had smarter algorithms. Now, the competition is about who can translate technology into user retention and monetization, explain financial growth in a language the secondary market understands, and build a compliance moat amid tightening global regulatory trends. 

None of these are the strengths engineers are best at. Yet each of them is now starting to determine who gets to stay in the game. 

Notes beyond the text:

Many people interpret Yan Junjie's decision to forgo salaries and give away 5% of his shares as a gesture of tragic determination and crisis public relations. 

What's more noteworthy is another set of facts. 

MiniMax has not conducted layoffs, nor has it stopped model training. The HKD 16 billion financing with 7x oversubscription drew more than 100 institutions scrambling to participate. 

This means long-term capital in the primary market hasn't lost confidence in the fundamental logic of large models. Its direction of bet hasn't changed: betting on models, betting on intelligence, betting on scale. 

The real problem never lies with MiniMax alone. The real problem is: starting from the second half of 2026, the logic of capital investing in AI has completely transformed.

MiniMax did not lose on its model. It is simply the first AI company that capital demanded to prove itself as a real public listed enterprise. 

Today it's MiniMax. 

Tomorrow, it will be every AI company's turn. 

This article is from WeChat Official Account "Beyond the Layout", written by Huahua, republished with authorization from 36Kr.