Redazione RHC : 16 October 2025 07:04
This year, OpenAI announced a series of projects with foreign governments to create “sovereign AI” systems. According to the company, some of these agreements are currently being negotiated with U.S. authorities. The goal is to give national leaders greater control over technology that has the potential to transform their economies.
In recent months, “sovereign AI” has become a buzzword in Washington and Silicon Valley . Advocates believe it’s crucial that systems developed in democratic countries be actively deployed globally, especially as Chinese models are increasingly being exported.
In its July AI Action Plan, the Trump administration framed this idea as follows: the spread of American technologies should prevent strategic rivals from making allies dependent on the decisions of “hostile” suppliers.
For OpenAI, moving in this direction means working beyond democracies. The company is collaborating with the United Arab Emirates, where power is vested in a federation of monarchies . OpenAI’s chief strategy officer, Jason Kwon, says that engaging with non-democratic regimes can push them toward greater openness. Kwon prioritizes inclusion over isolation, and this approach sometimes works, sometimes not.
Critics point out that similar arguments were raised against China twenty years ago. At the time, the United States was focusing on economic integration as a path to liberalization. As a result, many American companies profited from trade, while Beijing’s policies hardened.
The debate has also focused on what constitutes true sovereignty . Some experts believe that without the ability to inspect and, to some extent, control the model, sovereignty is impossible. Clement Delange, CEO of Hugging Face, says that “there is no sovereignty without open source.” In this area, China has already taken a leading role, with its open source models rapidly gaining popularity beyond its borders.
The term “sovereign AI” currently encompasses several architectures. In some projects, the state gains partial control over the stack, while in others it has complete control over the infrastructure, from hardware to software. The common denominator across all initiatives is legality. As Tricia Ray of the Atlantic Council’s GeoTech Center notes, constraining at least part of the infrastructure by geographic boundaries makes its design, development, and implementation subject to national law.
In the United Arab Emirates, a partnership between the United States and OpenAI is building a data center cluster with a total capacity of 5 GW . Approximately 200 MW of this capacity is expected to be operational by 2026. The country is also deploying ChatGPT for government services. However, there are no indications that the government will gain access to the internal components of the model or the right to modify its operation.
Just a few years ago, the idea of deploying AI infrastructure in authoritarian countries could have sparked protests in Silicon Valley. In 2019, Google employees managed to shut down a censored research project for China. Now, analysts note, the attitude has become more pragmatic. The “working in a country means obeying its laws” mentality has become significantly normalized, and thus there are few internal protests surrounding major LLM initiatives.
Kwon emphasizes that OpenAI will not delete information at the request of foreign authorities. The company may add local resources and features, but it doesn’t plan to “clean” the data.
While American companies are forging international alliances, Chinese companies are actively deploying open-source models around the world . Alibaba, Tencent, and startups like DeepSeek are publishing basic models with capabilities comparable to those of their Western counterparts. Alibaba claims to have downloaded over 300 million Qwen software programs and created over 100,000 derivative models. Qwen is making significant progress in Japan thanks to its high-quality local language support. Last month, researchers in the United Arab Emirates presented a new model based on Qwen2.5.
OpenAI also returned to an open-source format this spring, releasing the first open-weighted models since GPT-2. According to industry sources, the decision was influenced by the enormous popularity of DeepSeek’s open-source models earlier this year. Delang notes that focusing on open source accelerates progress. Companies quickly adopt each other’s successful training techniques, which is why, in just five years, Chinese teams have gone from lagging behind to comparable to their American counterparts, assuming a leading position in the open ecosystem. He estimates that China could assume leadership in AI as early as next year.
There’s also a pragmatic aspect. In a closed environment, the same energy-intensive training cycle is often repeated in parallel by different labs. In an open ecosystem, one center trains and publishes the model; neighboring centers don’t have to use the same resources, and capacity is distributed more efficiently.
OpenAI believes that sovereign AI is not a choice between “open” and “closed.” Different countries want to leverage the best of both worlds. Some tasks are more easily addressed with large-scale commercial models, while others are better based on open source solutions that can be tested and adapted to local legal and market requirements.
Redazione