Chinese AI companies are attempting to forge a path distinct from Silicon Valley. MiniMax founder, chairman, and CEO Yan Junjie (IO), in his first English-language podcast interview, stated that the company will remain committed to open source until achieving artificial general intelligence (AGI), and that the key to profitability is not raising model prices, but continuously driving down inference costs through technological innovation.
In his view, the U.S. AI ecosystem relies on massive capital and computing resources to push breakthroughs in model capabilities, while Chinese companies must improve efficiency under limited computing conditions. MiniMax hopes to demonstrate that even without access to the same scale of top-tier chips, it can narrow the gap with leading U.S. frontier models through model architecture, infrastructure, and algorithmic optimization.
This approach is already reflected in the business model. Yan Junjie revealed that as of August, MiniMax's annualized run rate (ARR) for annual recurring revenue had exceeded $800 million, with enterprise business accounting for more than 80%. He believes that low inference costs not only allow customers to use AI at lower prices, but also enable the company to conduct reinforcement learning and post-training at lower costs, creating a virtuous cycle where commercial growth and model capability improvement reinforce each other.
Meanwhile, MiniMax is advancing the integration of language models and video generation models, hoping to evolve AI from generating text and video clips into intelligent agents capable of planning, executing, and completing complex projects. Yan Junjie emphasized that AGI should not belong to any single company, country, or organization, but to everyone.
Key takeaways from Wall Street CN are as follows:
1. MiniMax will remain committed to open source until AGI is achieved. Yan Junjie believes open source is not an alternative to commercialization, but a catalyst for both commercialization and technological innovation. Developers can build new products based on the model, forming a feedback loop that helps the model iterate faster.
2. Low inference costs are MiniMax's core moat in its business model. Yan Junjie believes AI profitability should not rely on price increases, but on reducing inference costs through technological innovation. Lower costs attract more customers and allow the company to expand post-training scale at lower expense.
3. The AI ecosystems in China and the U.S. are forming different development paths. The U.S. relies on capital, chips, and computing resources to advance frontier models, while Chinese companies place greater emphasis on the efficiency of each unit of computing power. Resource constraints have actually driven Chinese AI labs to continuously innovate in architecture, infrastructure, and algorithms.
4. MiniMax is betting on the convergence of language models and video models. The company currently has the M-series language models and H-series multimodal models, and hopes to integrate both into a unified system in the future, upgrading AI from content generation to completing complex tasks in advertising, film, and design.
5. The next phase of AI Agents is not just executing tasks, but autonomously setting goals and verifying results. Yan Junjie believes that truly valuable agents should possess autonomy in long-term tasks.
6. MiniMax's competitive advantage lies not only in technology, but also in the founder's capital allocation ability. Scottish Mortgage investment manager Lawrence Burns believes Yan Junjie can reduce ineffective experiments, make key technology choices, and achieve better R&D results with fewer resources.
7. Profitability still faces two major challenges: price competition and rising frontier R&D costs. Lawrence Burns noted that China's AI market is highly competitive, and MiniMax still needs continuous investment to stay at the technological frontier. While the market space is enormous, profitability and profit margins still need to be validated.
From Resource Constraints to Efficiency Advantage: MiniMax Chooses a Different AI Path
For MiniMax, resource constraints are not a short-term problem that can be easily eliminated, but a reality it has had to face since its founding.
Leading U.S. AI companies have more abundant capital, chips, and computing resources, allowing them to find optimal technology paths through extensive experimentation. Chinese companies must use every unit of computing power more carefully. This difference has gradually shaped MiniMax's organizational culture and technology strategy.
Yan Junjie summarizes the company's goal as: minimizing inference costs while making frontier intelligence accessible to everyone.
He believes that the ultimate competition in the AI industry is not just about model performance, but also about how much intelligence can be created per unit of computing power.
Open Source Is Not Abandoning Commercialization, but Building a Larger Innovation Network
In the development paths of Chinese and U.S. AI companies, open source versus closed source has become an important dividing line.
U.S. companies like OpenAI and Anthropic mainly adopt closed-source models, commercializing by controlling model access and API services. Chinese AI companies are more actively adopting open-weight models, allowing developers to download models and deploy and modify them on their own hardware.
MiniMax has chosen to stand with the latter. Yan Junjie stated that open source is not an alternative to commercialization, but rather a catalyst for commercialization and further innovation.
He cited the company's H3 model as an example. According to his account, in the first two weeks after H3's release, approximately 300 open models built on the model had already appeared.
Among them, a U.S. company called fal.ai conducted post-training based on H3 and developed H3 Max, which can generate 5 seconds of video in 3 seconds, achieving near-real-time video generation.
This means the value of open-source models is not limited to the original model itself. External developers can optimize on top of it and develop applications that were never originally envisioned.
Yan Junjie gave an example of developers who have used the technology to build AI live-streaming systems that run 24 hours a day, with content generated in real time based on user needs. Similar capabilities could extend to real-time gaming, education, and other interactive experiences.
For enterprise and government customers, open source has another advantage: models can be deployed on their own hardware, better protecting data and privacy. Therefore, MiniMax hopes to allow developers, users, and the model company to form a mutually reinforcing ecosystem by opening model weights.
Yan Junjie even explicitly stated that the company plans to remain open source until AGI is achieved.
This does not mean the open-source model has no commercialization challenges. Model downloads and developer numbers ultimately still need to translate into API revenue, enterprise customers, and actual usage. But MiniMax's judgment is that an open ecosystem can expand the model's application scope and allow the company to obtain more technical feedback at lower cost.
Profitability Does Not Rely on Price Increases: Low Inference Costs Determine Both Revenue and R&D Efficiency
Compared to the open-source strategy, Yan Junjie's explanation of MiniMax's business model is more noteworthy.
Currently, the AI industry generally faces high R&D and computing investment, with many companies still operating at a loss. For MiniMax, how to achieve profitability while maintaining model capability growth is one of investors' biggest concerns.
Yan Junjie's answer is: AI profits are not obtained by raising prices, but by reducing inference costs through innovation.
He believes that low inference costs simultaneously constitute the company's revenue moat and profit moat.
On one hand, lower inference costs mean MiniMax can provide services to customers at lower prices, expanding market coverage. Not all enterprises need the world's most advanced models; many daily office tasks, programming, and business processes only require models that are good enough in performance and reasonably priced.
If model performance meets business needs while the price is significantly lower than competitors, the return on investment for enterprises adopting AI will increase.
On the other hand, low inference costs also affect the R&D efficiency of the model itself.
Yan Junjie particularly emphasized the importance of post-training. After a model completes initial training, it still needs to continuously practice, correct errors, and improve complex task handling through supervised fine-tuning, reinforcement learning, and data synthesis.
These processes all require substantial inference computing. Therefore, the lower the inference cost, the more capable the company is of expanding post-training scale, giving the model more practice opportunities.
Lawrence Burns offered a golf analogy: two golfers both need continuous practice to improve, but if one can buy practice balls at a lower cost, then with the same budget, he can practice more and is ultimately more likely to become the better golfer.
MiniMax's advantage lies in having cheaper "practice balls." This means low cost is not just a pricing advantage, but can also translate into an advantage in improving model capabilities.
Yan Junjie revealed that as of August, MiniMax's ARR had exceeded $800 million, with B2B business accounting for more than 80%.
However, ARR is an annualized run rate calculated based on current revenue levels and does not equal full-year revenue already realized.
For MiniMax, the ideal development path is: inference costs continue to decline, attracting more customers, expanding revenue scale, then reinvesting resources into post-training and model R&D, ultimately driving revenue growth faster than cost growth. This is the growth flywheel the company hopes to build.
Another Competitive Approach for Chinese AI Companies: Creating More Intelligence per Unit of Computing Power
When discussing China-U.S. AI industry competition, Yan Junjie believes the two ecosystems are forming different development models.
U.S. AI companies have more capital, computing power, and chip resources, enabling them to push model capability breakthroughs through large-scale computation. Chinese companies need to place greater emphasis on efficiency per unit of computing power.
He stated that the U.S. AI ecosystem is building a closed system supported by massive capital, while China's AI ecosystem is building an efficiency-driven collaborative system.
He does not believe the two ecosystems will ultimately converge on the same path. This difference is reflected not only in business models but also in technical architecture and R&D methods.
MiniMax focuses its investment in three areas: model architecture, infrastructure, and optimization technology. Yan Junjie stated that from the beginning, the company wanted its models to run well on different chips rather than relying entirely on a single type of hardware.
In his view, the more constrained the computing power, the more important algorithm and inference efficiency become.
These resource constraints have actually made the company's engineers and researchers better. This also explains why MiniMax summarizes its corporate culture as flat, focused, and frugal.
The company currently has only about 500 employees, so it must concentrate resources on the most important things.
Yan Junjie proposed two R&D principles: first, keep things as simple as possible; second, always build infrastructure. He believes that launching products directly may be faster, but in the long run, infrastructure is the foundation for achieving scale.
Scottish Mortgage investment manager Lawrence Burns believes that one of Yan Junjie's standout strengths is his ability to simultaneously play the role of both an excellent scientist and a capital allocator.
In many U.S. AI labs, researchers can conduct extensive experiments with abundant capital and chip resources. But MiniMax does not have such conditions, so it must reduce ineffective attempts and quickly judge which technology paths are worth investing in.
Burns believes Yan Junjie can more accurately judge technology direction and achieve better results with fewer experimental choices. This not only improves capital efficiency but also helps attract and retain top Chinese AI talent.
From Video Generation to Complete Tasks: The Convergence of Language Models and Video Models
MiniMax does not intend to only do language models or video generation models. The company currently has two main model series: the M-series language models and the H-series multimodal models.
The M-series primarily targets programming, agents, and productivity applications; the H-series focuses on video generation and can also handle images and audio.
Yan Junjie stated that the company hopes to integrate the two series into a unified model in the future.
Behind this strategy is his judgment about the changing forms of AI applications.
Current video generation models can generate video clips based on prompts, but they still fall short in understanding complex instructions, modifying content based on feedback, and completing continuous tasks.
For example, in advertising, film, and design work, users not only need to generate a video but also need to continuously adjust shots, modify content, understand creative intent, and ultimately complete an entire project.
This requires combining the understanding, planning, and execution capabilities of language models with the generation capabilities of video models. Lawrence Burns believes this convergence could open up a larger market than simply generating video clips.
When AI can upgrade from generating content to completing entire projects, the fees enterprises are willing to pay may also increase significantly.
At the same time, MiniMax is also betting on AI Agents. Yan Junjie believes the competitive focus of the next generation of agents is not just whether they can complete complex tasks, but whether they can autonomously set goals and verify task results.
Currently, MiniMax has made programming and Agent capabilities core directions for its M-series models. He also specifically mentioned that he was concentrating his efforts on developing the next-generation model M3 Pro and believes it could become an important breakthrough.
Profitability Still Faces Two Major Challenges: Price Wars and Frontier R&D Costs
Although MiniMax has shown rapid revenue growth, profitability remains an issue the company must face.
Lawrence Burns pointed out that in the first half of this year, MiniMax's revenue nearly quadrupled, but its reported loss also grew to about three times the original size.
He believes the company faces two main challenges in achieving profitability.
First, price competition in China's AI market is extremely fierce.
MiniMax needs to compete with rivals like DeepSeek for customers, and intense price competition could drive down service prices across the entire industry.
Second, the cost of maintaining frontier model capabilities continues to rise.
As AI models continue to advance, the resources required for training and inference also increase. This is one reason MiniMax raised $2 billion in July.
In Burns's view, the company needs to rely on a lean cost structure, correct technology choices, and a market that can be monetized at scale to address these challenges.
But he also noted that some positive changes have already emerged: R&D spending growth is beginning to fall below sales revenue growth, and enterprise business revenue is also expanding rapidly.
According to his account, enterprise business revenue grew about eightfold and currently accounts for 63% of total revenue. Yan Junjie's figure of more than 80% is calculated on an annualized basis.
The increase in enterprise customers helps improve commercialization capability, because as model performance improves, enterprises can use AI to complete more work and are more willing to pay for higher productivity.
However, Burns emphasized that MiniMax remains an early-stage investment, and the final outcome carries significant uncertainty. This is one reason the fund controls the size of its portfolio position.
AGI Does Not Belong to Any Single Company: MiniMax's Ultimate Vision
At the end of the interview, Lawrence Burns asked Yan Junjie what the world would look like if MiniMax achieved its goals.
Yan Junjie's answer was: the real question is not whether AGI can be achieved, but who AGI will ultimately benefit.
In his view, AGI should not belong to any single company, country, or organization, but to everyone.
He believes that the people who use the models, the developers who build applications based on the models, and the researchers who contribute to the models should all be part of the AGI ecosystem.
This is also why MiniMax chose the slogan "Intelligence with everyone" rather than "Intelligence for everyone."
The two may seem to differ by only one word, but the meaning behind them is different.
The former emphasizes that intelligence should not only serve everyone but should be jointly participated in and owned by everyone.
For MiniMax, open source is not just a technology strategy but also a choice about who AI belongs to in the future.
For investors, what ultimately needs to be validated is another matter: whether this open and low-cost approach can generate sustainable commercial returns in fierce industry competition.
Comments