OpenAI’s “Sovereign AI” Initiatives and the Global AI Landscape

I. OpenAI’s “Sovereign AI” Collaborations with Foreign Governments

OpenAI has disclosed a series of projects in the current year, collaborating with foreign governments to develop what it terms as their “sovereign AI” systems. The company contends that these agreements, some of which are coordinated with the US government, are integral to a broader endeavor aimed at enabling national leaders to exert greater control over a technology with the potential to reshape their economies.

A. The Emergence of “Sovereign AI” as a Buzzword

Over the past few months, “sovereign AI” has gained significant traction in both Washington and Silicon Valley. Proponents of this concept assert that it is of utmost importance for AI systems developed in democratic nations to be able to expand globally, especially as China intensifies its efforts to deploy its own AI technology overseas. As stated in the Trump administration’s AI Action Plan released in July, “The distribution and diffusion of American technology will prevent our strategic rivals from making our allies reliant on foreign adversary technology.”

B. OpenAI’s Partnerships with Non – Democratic Governments

At OpenAI, this movement has led to partnerships with countries such as the United Arab Emirates, a federation of monarchies. Jason Kwon, OpenAI’s chief strategy officer, posits that collaborating with non – democratic governments can facilitate their evolution towards greater liberalism. “There’s a wager that engagement is superior to containment,” Kwon remarked in an interview with WIRED last week at the Curve conference in Berkeley, California. “This approach has had varying degrees of success.”

Kwon’s reasoning mirrors statements made by some politicians regarding China over two decades ago. “We can either strive to guide China in the right direction or turn our backs, almost certainly pushing it in the wrong direction,” declared US president Bill Clinton in 2000, as China was preparing to join the World Trade Organization. Since then, numerous American companies have profited from trade with China, yet the Chinese government has become more authoritarian.

II. The Concept of “Sovereign AI”

A. Defining “Sovereign AI”

Some individuals argue that true sovereignty can only be attained if a government can inspect and, to a certain extent, control the AI model in question. “In my view, sovereignty is unattainable without open – source,” says Clément Delangue, the CEO of Hugging Face, a company that hosts open – source AI models. In this regard, China has taken the lead, as its open – source models are rapidly gaining global popularity.

Today’s sovereign AI projects vary in scope, from granting countries partial to full control over the entire tech stack, meaning the government manages all aspects of the AI infrastructure, from hardware to software. “The common underlying element for all these projects is the legal aspect—by tying at least part of the infrastructure to geographical boundaries, the design, development, and deployment will then comply with certain national laws,” explains Trisha Ray, an associate director at the Atlantic Council’s GeoTech Center.

B. The OpenAI – US – UAE Deal

The deal announced by OpenAI in partnership with the US government in the UAE encompasses a 5 – gigawatt data center cluster in Abu Dhabi (200 megawatts of the total planned capacity is scheduled to come online in 2026). The UAE is also rolling out ChatGPT nationwide. However, it appears that the UAE government will have no access to examine or modify the chatbot’s internal mechanisms.

III. Shifting Attitudes in Silicon Valley

Only a few years ago, the idea of constructing AI infrastructure in authoritarian countries might have triggered worker protests in Silicon Valley. In 2019, Google employees opposed the tech giant’s plan to deploy a censored search engine in China, ultimately succeeding in getting the project scrapped. “What’s occurring with some of these LLM projects is quite similar, but the backlash is not as intense,” Ray notes. “The concept that, ‘yes, if you operate within a country’s borders, you must abide by all local laws,’ has gradually become more normalized.”

A. OpenAI’s Stance on Censorship

Kwon is firm in stating that OpenAI will not censor information, even if requested by a foreign government. “We will not suppress informational resources,” he affirms. “We might add to them, but we will not eliminate them.”

IV. China’s Lead in Open – Source AI

A. Chinese Tech Companies’ Global Spread of Open – Source Models

While US AI firms are eager to partner with foreign leaders, Chinese tech companies are disseminating their open – source models globally. Giants like Alibaba and Tencent, along with startups such as DeepSeek, have released open – source foundation models that are on par with the capabilities of their US counterparts.

Alibaba reports that its Qwen family of AI models has been downloaded over 300 million times worldwide, and more than 100,000 derivative models have been developed using them. Startups in countries like Japan have widely adopted Qwen after finding that it excels at performing tasks in the local language. Last month, researchers in the UAE released a new state – of – the – art model based on Qwen2.5.

B. China’s Open – Source Advantage

Earlier this year, OpenAI released its first open – weight models since GPT – 2. According to WIRED’s previous reporting, while this project had been in progress for some time, OpenAI was also spurred by the popularity of models released by China’s DeepSeek, which made a significant impact in January.

Delangue contends that focusing on open – source models has enabled Chinese AI companies to iterate rapidly, as rival firms can quickly adopt each other’s training techniques. “They have progressed from being far behind five years ago to now being on par with the US and leading in open – source,” he observes. “It would not be unexpected if China takes the lead in AI overall next year.”

China’s open – source strategy, Delangue argues, also maximizes the efficiency of its AI infrastructure investments. “In the US, where most of the field is closed – source, one gigawatt means that every single lab is essentially conducting the same training,” he says. “In Europe or in China, due to the more open – source and open – science environment, the same gigawatt is actually distributed among the labs. One lab conducts the training and then releases its models, obviating the need for neighboring labs to repeat the same training.”

C. OpenAI’s View on Open and Closed Models

OpenAI maintains that both open and closed models can be part of the sovereign AI approach. “It doesn’t seem that we are on a path with only one model,” Kwon states. “When visiting various countries, it is evident that they desire to utilize both the best closed models and open models for different use cases.”

This is an edition of the Model Behavior newsletter. Read previous newsletters [here].