Hello everyone, welcome to Best Partner. Today, we will deeply dissect an interview with Arthur Mensch, the CEO of Mistral, a leading AI company in Europe. As a European representative competing with Chinese and American giants in the AI arena, Mistral's perspective can be said to combine technical depth with business realism. Arthur Mensch's sharing directly reveals the core shift in the AI industry from pursuing the illusion of omnipotence to returning to the essence of business. This interview covers ten key topics including foundation model trends, enterprise application landing, open-source vs. closed-source, and sovereign AI strategy. Every viewpoint could rewrite our understanding of the future of AI. Next, let's follow Arthur Mensch's thinking and see what underlying changes are happening in the AI industry.

To start, Arthur presented a surprising conclusion: foundation models are becoming commoditized, and this is an inevitable trend. Many people may wonder, in 2023 we were still amazed by the capabilities of GPT-4, in 2024 open-source models began to catch up, but by the end of 2025, companies like Google had already caught up with OpenAI's leading-edge models. Why has the script of "the strong always win" suddenly failed?

Arthur gave three key reasons. The first reason is that the threshold for core technology is not as high as imagined. Arthur mentioned that there are only about ten laboratories in the world that master the core technology of AI foundation models. The data acquired by these laboratories is highly similar, and the algorithms and technical solutions they follow are largely the same. More importantly, these core algorithms are actually very streamlined. This means that building an AI foundation model is not a unique skill, but a set of standardized processes that can be replicated and learned.

The second reason is that the speed of knowledge diffusion is far beyond expectations. The key knowledge and technical details required to train models are spreading rapidly within the industry, making it difficult to form lasting intellectual property barriers. When everyone is doing the same thing and can access similar data and algorithms, it becomes extremely difficult to achieve a breakthrough in leadership and significantly surpass competitors.

The third reason is that the model itself is becoming a commodity, like oil or steel. The performance differences between foundation models are shrinking, making it difficult for leaders to maintain their technological advantage for a long time, and the speed of asset depreciation is far beyond imagination. This leads to a key question: when foundation models are no longer scarce, where will the value of the AI industry flow? How will the profit model change?

Arthur bluntly pointed out that some competitors are now investing hundreds of billions of dollars to create general assets that depreciate rapidly, and this model poses a huge risk. A typical example is OpenAI's plan to invest $1.4 trillion in infrastructure. But if the performance of the next generation of models is quickly caught up by competitors, it is questionable whether this astronomical investment will bring corresponding strategic returns. Mistral's choice is to turn to downstream applications and return to pragmatism.

Arthur emphasized that the biggest pain point in the AI industry is that we heard countless grand promises a few years ago, but until now, most companies have not truly benefited from AI. The core problem lies in insufficient customization. Many companies first have an AI solution and then look for problems, rather than focusing on their core pain points first. Therefore, Mistral's approach is to help customers find the right application scenarios and provide deep customization. For example, a supply chain scheduling process that originally required twenty people to collaborate can be streamlined to two people through in-depth AI optimization. This tangible efficiency improvement is the true embodiment of AI value.

Arthur emphasized that all investments must ultimately be paid for by the free cash flow and value creation generated downstream. The AI industry must prove that the value growth brought by technology is sufficient to cover the huge amount of capital invested by society. As foundation models become commoditized, the industry's narrative focus will inevitably shift from building a god-like AGI to building business applications that solve specific problems. Many people may remember that AGI was the hottest topic in the industry a few years ago, and everyone fantasized about a panacea that could solve all the world's problems. But now, even giants like OpenAI are turning to enterprise-level applications and business building. Is this shift driven by a sense of crisis brought about by technological commoditization?

Arthur's answer was very direct. He said that AGI is too abstract for businesses, and can even be said to be an illusion. In reality, there is no panacea that can solve all problems. Just as no one can be proficient in all professions, solving specific problems requires specialized capabilities. We are returning to systems thinking from illusory thinking, no longer pursuing a versatile model, but starting to think about which data can significantly improve the efficiency of specific tasks, and how to build a flywheel to continuously optimize the application experience through human-computer interaction.

Arthur particularly emphasized that AGI is more like a North Star guiding the continuous evolution of the system, an unachieved goal, rather than a solution that can be landed now. Another realistic reason for the change in industry narrative is that AI companies find it difficult to explain to investors why their technology will not be surpassed by competitors. In this case, companies must delve into the details of their business and prove their value by solving real-world problems.

Arthur mentioned that AI will eventually become decentralized, and due to the limitations of data volume and Scaling Law, the demand for customization capabilities will become increasingly strong. Therefore, Mistral's goal from the beginning has been to provide enterprises with this specialized customization capability, rather than pursuing a versatile AGI system. This return to the essence of business is becoming a consensus among more and more AI companies. When the core of AI shifts from making models smarter to building systems that solve problems, a new architectural logic emerges: models are just components, and the orchestration layer is the core.

Arthur broke it down into two complementary core parts: static definition and dynamic components. The so-called static definition is the workflow rules and system behavior guidelines set by humans, equivalent to the manual information we use to define the system, such as the company's approval processes, business rules, decision trees, etc. These are relatively fixed and are the guardrails of the system. The dynamic components connect the model to various tools and give instructions. The model can autonomously call tools and decide on execution logic. This part is flexible and variable.

Arthur particularly reminded that it is overly idealistic to think that a dynamic system without human guidance can solve all problems. Over the past three years, with the enhancement of model reasoning capabilities, the increase in the types of callable tools, and the improvement of programming capabilities, the dynamic part has indeed been growing. However, the static part remains extremely important. It is the combination of static guardrails and dynamic autonomous decision-making that can create better and more reliable systems to solve complex problems that were previously unsolvable. The combination of the two will occupy the core position of AI systems for a long time, they are not substitutes for each other, but complement each other, jointly promoting us to solve more complex business challenges.

This systems thinking also directly affects the reconstruction of enterprise software by AI. Future enterprise software will no longer be a fragmented collection of tools, but will be centered on a context engine, eliminating bloated middleware, and allowing the front-end interface to be generated on demand. This may be the most profound transformation of enterprise operations by AI.