Summary of the Key Points
The global AI large-model industry is currently making a collective push towards General Artificial Intelligence (AGI). Major overseas manufacturers released new models with significantly improved performance at the beginning of September, even claiming that we have entered the AGI era. MiniMax, one of the earliest domestic startups in the field of large-model development, initially went public with the aura of a "full-modal approach and early-mover advantage," reaching a market value of up to 350 billion RMB. However, due to its technology not being at the industry's forefront, it had to rely on low cost and value for market competitiveness. Its business strategy shifted from targeting ordinary users to corporate clients, resulting in continuous losses. Its stock price plummeted by 85%, and it is now at an awkward crossroads of trying to catch up with AGI through cost-effectiveness.
---
Detailed Analysis
1. The AI industry is buzzing in September; is AGI really near?
The recent activity in the AI industry is as intense as the start of the school season: in the first three days of September, overseas giants released groundbreaking models one after another—Anthropic, Meta, and Google, with OpenAI finally unveiling GPT-6 Astra and declaring, "Welcome to the AGI era."
Many may not understand what AGI is. Simply put, traditional AI was like workers who could only use specific software; you had to clearly define your needs and feed them through official interfaces for them to perform tasks. GPT-6 Astra, on the other hand, can operate any software on your computer screen just like a human, from coming up with ideas, writing code, to producing results—all without your direct instruction.
Now, all major AI companies worldwide consider achieving AGI as their ultimate goal. Domestic companies were a bit slower to update in September. In contrast, Saudi Arabia released the world's first Arabic-language large model, which was based on MiniMax's open-source model. Founded in 2021, MiniMax released its first text model in April 2022, seven months before ChatGPT.
2. Lacking industry-leading technology, relying on "cost-effectiveness"
MiniMax took a path few startups dare to take: while others focused on text models or video generation, it invested in all four areas—text, voice, video, and music models—hoping to reach AGI through integrated approaches quickly.
However, in reality, none of its technologies stands out among its peers. It ranks 17th globally in text models and 9th in China, failing to make it into the top tier. Its H3 video generation model performs best, with the best results for image-to-video conversion, and it also ranks in the top three for text-to-video and video editing. It even solved the problem of garbled text in AI-generated videos. Previous AI could only randomly generate text; MiniMax's model allows text to move with the camera and fit the perspective of objects, such as text on a milk tea cup that doesn't distort when rotated.
Yet, even its strongest video model doesn't win based on quality but on price. While other companies charge 1-3 yuan per second for 1080P video generation, MiniMax charges only 0.8 yuan per second for 2K video, delivering results comparable to those of more expensive models. Domestic short-video producers often use MiniMax for the most critical scenes and generate the remaining 90% of the content with it. Overseas users praise it for being fast and affordable, the most reliable option at its price point.
3. Shifting from user-funded models to corporate revenue
MiniMax's commercial strategy has changed dramatically in the past two years. In 2024, 70% of its revenue came from individual users. Its product Talkie is like an AI version of a facial customization game, where users can customize the appearance, voice, and personality of AI characters and interact with them. Revenue came from user subscriptions and character card purchases, similar to traditional internet social games.
This approach is no longer viable: there are too many AI chat apps, and users can easily switch without significant loss of convenience, making it hard to retain them and increase payment rates. Now, over 60% of its revenue comes from corporate clients. It rents its AI capabilities by the use, similar to charging for shared power banks by the hour.
Its main competitive advantage is price: while top domestic models cost around 13 yuan per complex task, MiniMax charges only 3.4 yuan. However, since it operates in all four areas, its computational costs are higher than those of companies focusing on text models. The cheaper it sells, the more it uses, and the greater its losses. It is losing billions annually, and all analysts predict it will continue to lose money until at least 2028.
4. Market value plummeted from 350 billion to over 50 billion RMB
MiniMax's stock price has been extremely volatile: it went public on the Hong Kong Stock Exchange in January at HKD 165, doubling on the first day and reaching a peak of HKD 1330 in March, with a total market value of over 350 billion RMB. The market was excited for several reasons: it was a domestic pioneer in large-model development, its full-modal approach seemed most promising for reaching AGI first, and over 60% of its revenue came from overseas customers with higher payment habits, increasing the likelihood of successful commercialization.
However, the stock price plummeted to HKD 186, a 85% drop. Two events triggered this: in June, when it released the new M3 text model without prior notice, many developers found their account balances depleted, and the model's performance fell short of expectations, causing a 15% drop. In July, a large number of early shareholders sold their restricted shares, and the market, already skeptical of its technology, lost confidence in its valuation. When the well-received H3 video model was released, the stock price slowly recovered to around HKD 300.
5. Narrowing the path to catch up with AGI through cost-effectiveness
MiniMax faces a tough situation: its all-modal approach is only affordable for giants like Google and ByteDance. Even OpenAI had to shut down its leading Sora video model this year. Startups like MiniMax have very limited room for error on this path.
The technical barriers are being raised by overseas giants, and domestic competitors are also competing fiercely on both technology and price. The cost-effective model approach can never generate high profits due to the premium of advanced technology. Even if a model is three times more expensive, users with high requirements will still use it. MiniMax can only take on the lower-end orders, making it difficult to cover high R&D and computational costs.
The founder has admitted that MiniMax is now only in the first tier in China but was never the leader. He set the goal of AI contributing 2% to GDP, expecting to achieve this in two to three years. Time is running out for this early mover to catch up with the technology leaders.