The Paradigm Shift of Sovereign AI

Only 2.5 years after ChatGPT stunned the world, there are AI factories rising in Lagos. LLMs trained in Malay. GPUs powering homegrown AI products in the Persian Gulf. AI is diversifying, localizing — nationalizing. Building on previous decades of governmental research and funding, big techs and AI startups swept the world into the generative AI era, with little direct oversight or control by sovereign countries — a departure from Cold War-era, state-borne innovations like plastic, nuclear power and the Internet. But a course correction is taking place: away from corporate-driven AI to state-driven models tailored to specific cultures, languages and needs. What does it take for countries to build, own and manage sovereign AI ecosystems? And are they up to the task?

Declaring AI Independence

Though not the first to originate the idea, a key figure in popularizing sovereign AI is Jensen Huang, the CEO of Nvidia. As he declared in November 2023:

“Every country needs to build their sovereign AI that reflects their own language and culture... The second wave of AI is the expansion of generative AI all around the world.”
Jensen Huang - CEO, Nvidia

More than a year and a half later, Nvidia drives the proliferation — and bottlenecks — of sovereign AI across the world. For now, at least, Nvidia’s GPUs are the most sought-after resource for any homegrown AI system. At once, this leads countries to “beg Jensen Huang” for GPUs, as one expert put it, or search for alternatives.

All of this is happening as countries seek to reverse the early dependence on American AI startups and Big Techs that at once provide a Silicon Valley bias in AI systems and a shaky reliance on foreign entities amid Trump-era realpolitik.

David Shrier is a technologist and professor at Imperial College in the UK, and he is advising numerous governments across the world on creating their own sovereign AI system. Foreign governments began reaching out to him to advise on sovereign AI capabilities in the lead-up to US elections last year. Countries feared being squeezed out of critical access to GPU chips by a potential Trump Administration and a forthcoming “America First” approach. Now that Nvidia chips lie squarely in the middle of Trump’s raging trade wars — both as a stick and as a carrot — governments are turbocharging efforts to strive for AI independence. The outsized leverage of both Nvidia and Trump hangs over the international semiconductor landscape, as does the general concentration of AI capabilities among leading tech firms.

“The private sector’s dominance means rapid progress but also concentration of power, expertise and infrastructure in a handful of global firms. This disrupts the traditional role of states as technological gatekeepers and makes it harder for governments to steer innovation for public benefit. As a result, efforts at sovereign AI often involve public–private partnerships, but the bargaining power and capacity gaps between countries — and between governments and tech giants — are stark.”
Lucia Velasco - AI Policy Lead, UN Office for Digital and Emerging Technologies

The ensuing sovereign AI movement — where countries steer the means and design of homegrown AI systems — expands far beyond earlier data localization efforts. The focus, says Shrier, isn’t merely protection; it’s a matter of synergizing digital economies and developing strategic autonomy.

Source: Trusted AI Alliance

“With sovereign AI, we're seeing something that's a bit broader-based and uses more in the policy toolkit, because it's not only looking at restrictions, but much more importantly… governments trying to empower innovation ecosystems.”
David Shrier - Professor of Practice, AI & Innovation, Imperial Business School

Shrier offers a glimpse of what a sovereign AI-enabled financial services sector looks like, for instance. A nation’s financial services sector can utilize a sovereign AI data center with a “financialGPT library” built into it. “If you’re a fintech entrepreneur and decide to go to a Country X instead of Country Y,” he explained, “Country X can offer a library of tools to accelerate building your fintech that’s running on top of their national AI infrastructure.”

“You'll be able to go to the sovereign GPT and say, ‘I want to build a neobank that has these offerings to it.’ As long as you can supply the regulatory capital, you can start your bank, and a lot of the mess will be dealt with by the automated systems.”
David Shrier - Professor of Practice, AI & Innovation, Imperial Business School

Under such scenarios, preexisting open banking and open data schemes serve as force multipliers. Sweden has plans to launch such a venture, along with countries in the MENA and ASEAN regions, according to Shrier.

Next in Line

Given sufficient capital resources and technical knowhow — no small feats amid AI’s high costs and insufficient human capital — countries must also attain sufficient hardware, software, regulatory and infrastructural capacities in a synchronized manner to build sovereign AI ecosystems. Not easy, and not cheap.

The task requires supplanting for domestic use the AI solutions released by US tech giants —who have already invested more than $300 billion into AI — or Chinese competitors of like DeepSeek, a startup that required a fraction of the cost as in the West to create an LLM that performs at similar levels as ChatGPT.

Across the board, government funding has played a critical role in AI development among leading countries; the UK stands out as one example of achieving the highest AI productivity per capita, according to Shrier, thanks to significant government funding to entities like the Alan Turing Institute. But consider that Google has published more research papers on AI than the U.S. and China combined. Notable earlier achievements in research included Google’s open sourcing of TensorFlow in 2016 and its transformative paper detailing BERT in 2018.

Source: Trusted AI Alliance

On the software side, building out a sovereign AI includes accumulating the vast data sets required to train LLMs on par with leading LLMs originating from Open AI, Google, Meta and the like — but attuned to local language and culture. Complex and expensive supercomputers are needed to provide the computational power needed to train these models. Nonetheless, a panoply of language- and culture-specific LLMs spanning French to Japanese, Arabic and dialects in between, are having releases across the globe.

In public-private partnerships, many countries are providing land and cheaper energy, fast-tracking permits and using funds to acquire chips, build AI data centers and the requisite infrastructure to make homegrown AI systems. Among a host of other countries, India has announced investments of $1.24 billion to build out sovereign AI capabilities; Brazil $4 billion; and Japan more than $740 million. The U.S. has announced The Stargate Project, which seeks $500 billion in AI infrastructure investments, and China launched a $47.5 billion semiconductor fund. Saudi Arabia committed $100 billion.

For those without the political connections and gobs of money, attaining scarce AI semiconductors remains a primary bottleneck for many countries. There are some competitors emerging to Nvidia’s market dominance, among them the GPU manufacturer AMD. But with Nvidia’s software layer forming the basis of many AI systems so far, the U.S. company’s market dominance still greatly influences which countries get their sovereign AI capabilities off and running — a dependence on one single corporate clearly at odds with the higher aspirations of sovereign AI.

However, stakeholders are confident the bottleneck won’t last forever. Ayotunde Coker is the CEO of Open Access Data Centers, a leading data center company in Africa building out AI-ready data centers across the Africa continent. He believes the shortage in GPUs is a “short to medium-term issue”. Rivals will vie for new, better, more affordable chips that replace current ones. Emerging is GPU-as-a-service (GPUaaS), a cloud computing model that provides on-demand access to GPUs over the Internet, allowing for users to pay for shared GPU resources according to their usage.

One potential rival to Nvidia, Groq, provides access to its proprietary Language Processing Units (LPUs) in a similar model that at once promises faster-processing and cheaper AI inference — potentially opening opportunities in emerging markets like Africa.

“It means that you don't have to think about getting the GPUs, but the endpoint agentic AI developers are able to get access to that GPU capacity as a service for them to deliver what they need to.”
Ayotunde Coker - CEO, Open Access Data Centers

Densifying Data Centers

For countries to build a self-sustaining sovereign AI ecosystem, they need a lot more than just chips. At this juncture, only 32 countries have AI-specialized data centers, and they are largely in the U.S., Europe and China.

Source: New York Times

OADC is among the leaders in building out the necessary infrastructure across the African continent. To do so, they must upgrade or create new data centers with the required densities in power to serve AI’s energy-intensive needs. OADC’s flagship data center in Lagos is upgrading to a max capacity of about 2 megawatts to 24 megawatts to facilitate sovereign AI capabilities, with similar projects across their core data centers in Durban, Johannesburg and Cape Town.

In building state-of-the-art data centers in a place like Africa, challenges certainly abound. AI-specialized data centers require large, stable supplies of energy to power data centers and sufficient water to cool facilities. OADC must contend with what Coker describes as the “90/90 challenge”: operating highly efficient data centers in conditions with 90% humidity, and 90 degrees Fahrenheit.

“We're having to secure power infrastructure commitments from utilities, and [in addition] we provide our own generation. We have locations [powered by] gas, for instance, some with hydro… we're ensuring that our core facilities are right now able to support the GPU demand that provides a sovereign AI infrastructure.”
Ayotunde Coker - CEO, Open Access Data Centers

Another crucial infrastructure task is building out the proper fiber optic networks necessary between core and edge data centers. Chris George is the Principal for SELF Infrastructure, which supports data center, submarine cable and network startups, after he spent ten years as a strategic negotiator for Google, building out Google’s network infrastructure. George notes the rising popularity of Managed Optical Fiber Networks (MOFNs) to handle the high-speed, low-latency and secure connectivity needed for AI data processing, training and deployment within country borders, and beyond them.

Talk To Your Telco

Possessing the requisite fiber optic networks, data centers, consumer trust and national footprint, telcos are often providing a crucial layer between state policy and AI infrastructure. 18 telcos across five continents have launched 18 AI factories powered by Nvidia technology.

Along this dimension, Singapore and its national telco, Singtel, stands out. Though disadvantaged in scale, Singapore’s government, academic research and private industry work in tight concert with one another to achieve ambitious AI goals. The UN’s Velasco applauds Singapore’s SEA-LION.AI, a family of open-source LLMs funded by the government and developed by researchers that “better understand Southeast Asia’s diverse contexts, languages and cultures”. Singtel has launched RE:AI, an AI cloud platform designed to democratize AI access across Singaporean industries, government agencies and academia. Singtel’s Paragon serves as an all-in-one aggregation and business orchestration platform that uses AI to help manage networks, clouds and multi-access edge computing infrastructure. Paragon is already being adopted by operators in Thailand, Indonesia, Spain and Taiwan, according to Keith Leong, managing director of Singtel. Fueling these advances in Singaporean AI is Nxera, Singtel’s data center arm that incorporates liquid cooling to support high density AI workloads in a scalable manner.

Source: Kearney

With government-funded scholarly research fueling the software, ready access to sufficient Nvidia chips, and valuable Big Tech partnerships along the way to accelerate AI development, Singapore’s software, hardware and infrastructural investments work in lockstep with the its declared National AI Strategy to complete the necessary components of a self-sustaining sovereign AI system.

“AI is the greatest disruptor of businesses today, and Singtel is in a position to help consumers and organizations benefit from it… We are putting AI at the center of what we do, from transforming our customer service channels to developing AI solutions to driving innovation.”
Keith Leong - Managing Director, Singtel

Such large investments towards AI and tight coordination between the public and private sector have positioned tiny Singapore among global AI leaders.

Striking a Balance

As countries build out their own AI ecosystems and patterns, calls for international harmonization collide with competing national interests. At the U.N., Velasco emphasizes an adherence to oversight, transparency and civil society involvement, while drawing “clear red lines” to “guard against abuse and uphold human rights”. At the same time, Velasco recognizes that countries have “the right to shape how AI is developed and deployed in ways that reflect their own values, cultures and legal frameworks.”

“The UN’s approach is to support both national sovereignty and global cooperation. The aim is not to impose one-size-fits-all solutions, but to promote interoperability, transparency and a common ground — safety, human rights and accountability. In practical terms, this means encouraging dialogue on socio-technical standards and governance options.”
Lucia Velasco - AI Policy Lead, UN Office for Digital and Emerging Technologies

Maintaining a balance between interoperability and tailorization, innovation and responsible governance — such dilemmas prove no less formidable in today’s fraught geopolitical terrain and hypercompetitive AI landscape. The dangers of autocratic manipulation and exploitation loom. It wasn’t long ago that the Internet and social media once unified people across borders in unprecedented ways — only for utopian visions to shatter in fragmented, polarized pieces.

Proper design and stewardship must be mission critical.

“There is the risk that we're going to wind up with a Tower of Babel. There is a risk that we're going to wind up with AI that makes it even easier to isolate people and to reinforce their worst tendencies. And this requires citizen advocates and a vigilant regulator and a literate population who is engaged in the democratic process.”
David Shrier - Professor of Practice, AI & Innovation, Imperial Business School

These challenges already exist at this current stage of sovereign AI development that Shrier likens to the Model T in car production: revolutionary, but “far from a Ferrari 458”. AI’s advances in the years ahead will completely transform how we conceive and carry out basic functions of commerce, communication and operations; agentic AI, or autonomously operating AI systems, are just around the corner.

There is a roadmap with clear parameters and models surfacing among responsible technologists and visionaries offering proper safeguards and material improvements if sovereign AI systems are implemented responsibly. How these powerful, evolving technologies are steered among sovereign countries squaring off in technopolitical gamesmanship — bankrolled by hundreds of billions of dollars — remains the existential question.