GLM 5.2: The Chinese Breakthrough Redefining AI in Business

A Dependency Revealed by Political Decisions

For several years, the AI strategy of European companies relied on a simple approach: using APIs provided by American tech giants. This comfortable model was abruptly called into question when the U.S. government decided to restrict access to Anthropic's models. This decision highlighted the fragility of this dependency, as it revealed that the AI strategy of many companies was vulnerable to political decisions made thousands of miles away.

This situation forced companies to reassess their priorities. The crucial question is no longer which model performs best, but rather how to ensure continuous and reliable access to these models for their teams. This new perspective has radically changed the game.

GLM 5.2: An Unexpected Turning Point

The emergence of the GLM 5.2 model, of Chinese origin, has surprised the market. This open model now competes with the best proprietary models while being much more cost-effective, with an inference cost about ten times lower. Moreover, it offers unprecedented flexibility, capable of running on European servers such as those from Scaleway in Paris or OVHcloud in Gravelines, and even on companies' internal infrastructures.

This trend is confirmed by the saturation of GPU instances on European clouds, not due to academic research, but because companies are bringing their inference workloads back in-house. This shift towards local AI, once considered a whim, is becoming the default choice for organizations planning for the long term.

The irony of the situation lies in the fact that the open AI revolution, initially launched by American labs, is now being led by Chinese players who seem poised to win this race.

The Importance of Tooling Around Models

Companies must understand that true value no longer lies solely in the AI model itself, but in the environment surrounding it. A raw model, no matter how sophisticated, only produces plausible text. In contrast, a well-framed model, with clear business rules, safeguards, and controlled access to company systems, can produce genuinely useful work.

This concept, known as the "execution harness," is essential. It is a software layer that transforms generic intelligence into a specialized collaborator. A well-designed harness allows for increasingly precise rules to be added to an open model, making the model interchangeable, but not the harness.

This phenomenon has already occurred in the history of computing: value has shifted from infrastructure to usage. Processors and storage have become commoditized, but not software or data. Today, it's the turn of AI models to follow this evolution.

The Value of Internal Data

The real strategic news for companies in 2026 lies in their own data. Every SME over ten years old possesses an untapped treasure: CRM histories, support tickets, business emails, application logs, and production data. Individually, this data may seem insignificant, but when structured and analyzed by a local and cost-effective model, it becomes a valuable source of insights.

AI models will continue to improve, but without quality data, their potential remains limited. Structuring and leveraging internal data thus becomes the top priority for companies.

Refocusing the Company on Essentials

This paradigm shift offers companies the opportunity to reinvent their operational model. When AI and data take over repetitive tasks such as qualification, reporting, or follow-ups, only two essential functions remain for the company: acquiring customers and delivering what has been promised. Everything else becomes support, and support can be automated.

The most successful organizations have already made this distinction. They have not simply added AI to an unchanged organizational chart, but have redesigned their structure around the crucial question: where is the human irreplaceable? The answer to this question defines key positions, while the rest is destined for automation.

Actions to Take

For the next six months, three concrete actions are necessary:

• First, audit your dependency by identifying every process that could be interrupted if your American API provider cut access.

• Second, inventory your dormant data, such as logs, CRM, and tickets, and assess what they could reveal if analyzed by a model.

• Third, test an open model on a European instance with a real use case and a strict rules framework.

The cost of these experiments is now negligible, while ignorance can be costly. The current revolution belongs neither to those who possess the largest models nor to those who shout the loudest, but to those who master their data, their tooling, and know where the human makes the difference.