Who controls intelligence? Open-weight models, institutional AI

Nvidia Ceo Jensen Huang used his first post on X not to promote a product, but to raise a question: Who should control advanced artificial intelligence? Sharing a letter titled “Open Weights and American AI Leadership,” Huang wrote in his post that open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.

The sovereignty issue Huang highlighted is also critically important for Türkiye. The central question is how institutions can use artificial intelligence without sending their data outside their own environments or becoming dependent on a single provider. Today, open-weight models offer one of the most realistic paths toward that goal. Türkiye took early steps in this direction. But real progress will come when national models and data infrastructures evolve into systems that continuously learn from each institution’s own data.

So why choose open-weight models? They expand access and competition while enabling customization, independent security research, and technological sovereignty. They allow institutions to adapt models while keeping sensitive data within their own environments, retain ownership of accumulated knowledge, and avoid dependence on a single provider.

Hidden cost of closed models

There is no disputing that closed models currently offer frontier-level performance, managed infrastructure, and technical support. But every institution using these models sends its data to the provider’s servers. The provider determines pricing, terms of use, and when a model version will change or be discontinued. More importantly, access itself depends on the political decisions of another country.

This summer, we witnessed a concrete example. Due to export controls imposed by the U.S. Department of Commerce, access to Anthropic’s Fable 5 and Mythos 5 models was briefly suspended worldwide on June 12, 2026. Access that can be cut off overnight by another country’s decision cannot be the sole foundation for critical operations such as public services, defense, health care or finance. Moving workloads that process sensitive data and require uninterrupted availability to open-weight models is therefore not a choice, but a necessity.

Dependence has a quieter face: shadow AI. Without secure, capable internal AI, employees turn to closed-source chatbots through personal accounts. Contracts, customer records, source code, and internal correspondence can reach servers abroad without the institution knowing. Bans do not solve the problem; they merely hide it. The remedy is an internal alternative at least as useful as external tools.

The need for open-weight models in defense has also been demonstrated in practice. In July, an OpenAI agent undergoing testing escaped its controlled environment and infiltrated Hugging Face’s systems. According to Nvidia, when the safety filters of closed models could not distinguish the attacker from the defender, Hugging Face ran the open-weight GLM 5.2 model on its own infrastructure, analyzed more than 17,000 actions, and stopped the attack. In a crisis, a model operating in your own environment and under your own rules may prove more valuable than the best closed model.

What should institutions choose?

The right choice depends on the capabilities required, the sensitivity of the data, in-house technical expertise, risks, licensing terms, and the importance of avoiding dependence on a single provider. But on-premises AI is not inherently secure. Hardware investment, operating model-serving infrastructure, managing updates and security, and monitoring performance all require substantial in-house expertise in large language model operations (LLMOps). Such expertise cannot be built through a patchwork of tools; it requires an end-to-end platform.

An institutional AI platform needs four layers: the first is a data pipeline to collect, clean, and prepare training data; second is a model pipeline to fine-tune, compress, and evaluate open-weight models; third is a serving layer offering models through standard internal interfaces, with access controlled by keys and quotas; and lastly, agents that work with institutional documents, cite sources, and access databases or other tools as needed. Sensitive data and models must remain within the institutional network, with every model traceable from training data to response endpoint. Today, even 120-billion-parameter open-weight models can run on a single datacenter GPU, while 30-billion-parameter models fit on a powerful personal computer. This architecture is accessible not only to technology giants but also to medium-sized public institutions. Institutions can therefore keep sensitive-data agents on their own servers while using controlled cloud access for lower-risk tasks.

An often-overlooked component is the verification layer. A large language model generates answers, but smaller, specialized decision models can check whether they are source-supported, whether retrieved documents answer the question, and whether training data is correctly labeled. This layer also safeguards the quality of data entering the learning loop.

Türkiye’s priority

Türkiye has taken an important step in this direction. Bilge, a family of Turkish large language models developed by TÜBITAK Bilgem, was unveiled at the Türkiye AI Summit on June 13, 2026, alongside the country’s 2026-2030 AI Action Plan. Yet, developing national models is only the beginning. Their long-term value will depend on institutions’ ability to adapt and continuously improve them using their own data and feedback.

The first link in this transformation is data infrastructure. No model can learn from real-world operations without infrastructure that collects, cleans, labels, and securely stores usage data, examples of errors, and expert corrections from the field. In November 2022, before ChatGPT introduced the world to generative AI, Türkiye’s Presidency of Defense Industries unveiled the Defense Industry Artificial Intelligence Platform. The platform brought data, model training and computing power under one roof. Evren, opened for general access in September 2026, took this structure a step further. Data is processed on infrastructure located in Türkiye, 11 open-weight language models are offered through a single interface, and the shared knowledge pool grows as users upload and label data. This vision, which began in the defense industry, needs to spread to every sector and public institution, starting with health care, energy, finance, and agriculture.

Türkiye’s approach to developing and fine-tuning sector-specific AI models should be strengthened by a feedback mechanism that continuously channels data from real-world applications into model development. Models would then cease to be one-off projects and become institutional assets that learn from practice and improve over time. Open-weight models are the essential tool for enabling this cycle while keeping data within the institution. This approach also aligns with the goals for the data ecosystem, computing infrastructure, and domestically developed models set out in Türkiye’s AI Action Plan, which came into effect on Aug. 18, 2026. The crucial question is whether data from the field, examples of errors, and expert corrections will be integrated into an institutional learning loop. If that loop is established, Türkiye can develop a body of capabilities that becomes increasingly differentiated through its own sector-specific data and cannot easily be replicated from outside. Ultimately, model sovereignty is less about owning a model than about owning this learning loop.

Institutional memory

AI sovereignty is the ability of an institution or a state to operate, manage and control AI systems within its own legal, physical and operational boundaries. Where models run, how data flows, who holds the weights, who can access the capabilities, and which country’s laws apply are all elements of this capacity. Model sovereignty lies at the heart of this structure, because it is precisely at this layer that an institution’s knowledge accumulates over time.

As an institution adapts an AI model to its documents, terminology, workflows, and decision-making processes, the system gradually becomes part of its institutional memory. Methodological know-how, customer experience, and operational intuition built up over years are then stored not only in employees’ minds and archives, but also in the weights of a fine-tuned model. This accumulated knowledge is among the hardest institutional assets to replicate, because it cannot be purchased; it can only be developed through experience. By retaining control over the systems behind critical decisions, institutions can make AI an integral part of their own intelligence.