The question of who controls vital infrastructure and the capabilities it enables has always loomed large in the technology sector. But now the AI age has supercharged those discussions, bringing them to the top of agendas in boardrooms and government alike.
Earlier this year, US export controls restricting foreign nationals’ access to Anthropic’s frontier Mythos and Fable models sparked a new wave of discourse around the role of AI models, and relatedly AI-ready compute, as a geostrategic asset. The US government’s reported concern was that safeguards in the models could be “jailbroken” or bypassed. Days after the June launch, the Department of Commerce applied export controls. Unable to verify nationality in real time, Anthropic suspended both models for all users, in the US and abroad, for almost three weeks.
While the US banned advanced chip sales to China in 2022, the move to block model use on the grounds of national security was unprecedented. It signaled that these advanced models and the capabilities they offer are increasingly being viewed as critical intelligence that should be ringfenced.
Now, export controls have been lifted. Anthropic restored global access to Fable 5 in early July. Mythos remains restricted to vetted organizations, largely cybersecurity vendors (Project Glasswing). But the move has already had wide ranging implications for data sovereignty in the age of AI. Access to critical infrastructure created or located in other nations is no longer guaranteed, even if that nation is a strategic partner.
One thing is clear: nations increasingly need greater control over, or assured access to, the data, models, compute, energy and skills on which critical AI capabilities depend. This presents a significant challenge, given the scale of investment, infrastructure, natural resources, and talent required to develop and deploy AI capabilities. Control in practice rarely means building everything domestically. It means knowing which dependencies matter most, and having a credible fallback when one of them is restricted.
A full-stack problem
For organizations and governments, AI sovereignty is a full-stack challenge. It affects every layer of the data center ecosystem; from data, models and compute power to the physical infrastructure that supports them. Three kinds of sovereignty are in play: data sovereignty (where data sits and who can access it), model sovereignty (who can use, adapt and switch off the models) and operational sovereignty (who builds, powers and runs the infrastructure). Each depends on six layers: data, models, compute, physical infrastructure and energy, semiconductors and supply chains, and talent.
As we saw with Anthropic, model access may become one of the most challenging factors for many nations seeking to leverage advanced AI domestically. Models like Fable 5 are closed-weight. For closed-weight models, access to certain capabilities or models can be blocked or restricted by the provider, depending on the user or jurisdiction.
In contrast, new Chinese models such as Kimi K3 promise leading benchmarks and a commitment to open weights. But concerns around censorship and sovereignty have surrounded such models since the launch of DeepSeek’s R1 model in 2025. Open weights improve portability, because a model can run on infrastructure the customer controls, but they do not settle questions of provenance, security assurance or fit with local regulation. Enterprises and governments will want those answered before placing such models in critical workflows.
AI talent is similarly concentrated in hubs in the US and China. The gravitational pull of those hubs draws in more talent, perpetuating their dominance in the space. Acquisitions further concentrate expertise. In December 2025, Meta agreed to acquire AI start-up Manus, though the acquisition was later blocked around concerns of transferring Chinese technology and talent to an American company. It is another example of AI expertise being treated not only as a commercial asset, but as a strategic national asset.
Data is another central consideration for any conversation surrounding digital sovereignty. Where data resides, where it is stored and processed, has long been subject to contestation, from the EU’s trailblazing GDPR to India banning Mastercard from issuing cards on the grounds it violated data storage rules back in 2021. It takes on new significance in the age of AI, however. Data is the lifeblood of large language models (LLMs), which require vast volumes of information for training, fine-tuning and evaluation. However, for organizations and governments, access to data is not simply a question of availability or volume. Sovereignty depends on the ability to confirm data was lawfully accessed, where data can be stored and processed, and the conditions under which it can be transferred across borders. In practice, this is driving demand for in-country data platforms, clear data lineage and governance, and the ability to keep sensitive data, and the inference that runs on it, inside a defined jurisdiction.
Data is the lifeblood of large language models (LLMs), which require vast volumes of information for training, fine-tuning and evaluation.
At the same time, physical resources are vital for data center operations. The computation that drives AI processes is resource intensive, drawing considerable energy from the grid. As workloads scale, managing the heat generated by AI infrastructure is becoming another consideration in where and how data centers are built. In a time of escalating geopolitical tensions and volatile energy costs, ensuring reliable, cost-effective compute at the necessary scale is no easy feat. Grid connection is often the first constraint: in many markets, securing power, and the transformers, switchgear and substations needed to deliver it, can take years, sometimes longer than building the facility itself.
AI-ready semiconductors and memory add another layer of complexity. The manufacturing bases for these components are limited, concentrated primarily in South Korea, Taiwan, China, and the US. There are other pockets of innovation – for example, the Netherlands leads the world in advanced lithography, which is necessary for many of the world’s most sophisticated microchips. But again, access and cost are limiting factors for new nations seeking to develop their own sovereign infrastructure.
An integrated approach
While most don’t have the resources or influence to singlehandedly meet the present challenges, there are still pathways to a sovereign, AI-ready future. To get there, nurturing new ecosystems and building strong partnerships across the data center, AI, energy and technology industries will be vital.
Leaders around the world are already concentrating resources towards AI sovereignty. The UK recently introduced a £500 million Sovereign AI Fund to provide British companies with capital, as well as access to national compute and research infrastructure. Meantime, IndiaAI’s Compute Capacity program aims to establish 10,000+ GPUs via public-private partnerships, to make compute available to India’s startups. The program aims to widen affordable domestic access to compute. Both initiatives address the same practical question: who gets access to compute, on what terms, and where it physically sits.
Malaysia’s initiative is a good example of a joined-up approach. The country boasts Huawei’s first AI incubation center outside of China. The government has also announced its intention to invest RM2 billion (approximately US $497 million) into the creation of a Sovereign AI Cloud intended to strengthen national data sovereignty. Adding to this, Malaysia’s Data Center Task Force was founded in February 2025 with a strategy for managing energy capacity and data center water usage.
As this example illustrates, ensuring sovereignty in the age of AI is a delicate balancing act. Multiple, potentially conflicting factors must be weighed against one another to ensure a robust, joined-up response to the present challenges.
What to ask before building
For operators, enterprises and governments planning AI infrastructure, sovereignty becomes manageable when it is reduced to a handful of practical questions. Where will the data and inference run, and who can access them? Can the site secure power, cooling and water at the scale planned, and on what timeline? And do we have the skills to build and run the facility locally, or the right partners to supplement them?
Where will the data and inference run, and who can access them?
What's clear is that sovereignty will be achieved incrementally through trade-offs and strategic investment. It’s a full-stack problem, spanning everything from talent to power availability to supply chain challenges. Meeting these challenges necessitates a joined-up approach. Partnerships between infrastructure operators, energy and technology providers and local governments will become increasingly crucial to ensure sovereignty across both the physical and digital infrastructure layers that underpin AI.