The digital development community has make this mistake before. We chase politically appealing solutions that sound like empowerment but actually deepen dependency on the very systems we claim to escape.
Now we’re doing it again with sovereign AI.
The latest fashion in tech policy circles is pushing countries to demand sovereign Generative AI solutions – the idea that countries need their own domestically controlled artificial intelligence systems rather than relying on global models from American and Chinese companies.
From Canada’s $2 billion Sovereign AI Compute Strategy to India’s $1.25 billion IndiaAI Mission, governments are racing to build their own AI capabilities. The rhetoric is seductive: technological independence, cultural preservation, national security.
This is digital nationalism wrapped in development language, and it threatens to leave most LMICs further behind than ever.
Sign Up Now for more critiques of digital development activities
Crazy AI Development Costs
Let me start with some numbers that should terrify any finance minister. Training GPT-4 reportedly cost more than $100 million, with Google’s Gemini Ultra estimated at $191 million in compute costs alone. In just the first six months of 2024, Microsoft, Amazon, Google, and Meta collectively spent more than $100 billion on AI and cloud infrastructure.
Now compare this to the sovereign AI budgets that governments are announcing with such fanfare.
The EuroHPC project, with the largest sovereign AI budget at roughly $2 billion, represents just 2 percent of what four American companies spent in six months. France’s Jean Zay upgrade involving 1,456 Nvidia H100 GPUs is less than 6 percent of the new 25,000 advanced GPUs that Microsoft plans to install in its French data centers by end of 2025.
The reality: most countries don’t have the compute infrastructure, capital, or technical talent to compete at this level. Pretending otherwise is actively harmful to their digital development prospects.
Sovereign AI Is Data Localization 2.0
This feels familiar because we’ve seen this movie before. As I detailed in previous analysis of digital sovereignty and development, these policies follow the same flawed logic that drives data localization requirements.
Like data localization, sovereign AI is primarily a political decision, not a technological one. The motivations are identical: government control, economic protectionism, national security theater, and international relations posturing. The technological benefits are questionable at best.
Data localization laws have already demonstrated how these policies harm cloud computing efficiency, scalability, and disaster recovery capabilities. Sovereign AI requirements will create the same problems for AI deployment – forcing inefficient local infrastructure, preventing optimization across regions, and ultimately delivering inferior service to citizens who deserve better.
The 10% vs. 90% Problem
Here’s what the sovereign AI advocates won’t tell you: the vast majority of AI use cases don’t require sovereign control. Yes, there are genuine edge cases – maybe 10% of applications where national security, military operations, or truly sensitive government functions legitimately need air-gapped, locally controlled AI systems.
But for the other 90% – healthcare diagnostics, agricultural optimization, education systems, financial services, transportation management – global models are not just adequate, they’re superior. At Intelehealth, our telemedicine platform reduced time to diagnosis by 50% in LMICs using generic global models finely tuned to local contexts.
The technical superiority of global models isn’t accidental. These models benefit from massive scale, network effects, deep research investments, and the ability to spread R&D costs across worldwide operations. No individual country can replicate this infrastructure and expertise at comparable cost or quality.
Application Layer Opportunity
Rather than chasing impossible dreams of AI independence, developing countries should focus on what they can actually control: the application layer. This means becoming world-leading at deploying, customizing, and building solutions on top of global models.
Three strategies make infinitely more sense than sovereign AI:
1. Dominate specific use cases.
Instead of building general-purpose models, focus on solving particular problems exceptionally well. Using already developed LLMs while developing the necessary infrastructure and programming talent for localization can leapfrog the time, expense, and technical capabilities required to create models anew.
2. Leverage open-source models strategically.
Open Source Generative AI systems can allow for greater access by sidestepping the steep development costs associated with building computing capacity and initially training a model. Countries can fine-tune existing open models for local languages and cultural contexts without rebuilding everything from scratch.
3. Build regulatory and deployment excellence.
Create frameworks that ensure AI systems serve local populations while maintaining interoperability with global platforms. This approach delivers sovereignty benefits without the impossible economics of full technological independence.
Digital Resilience Over Digital Sovereignty
The development community should advocate for digital resilience and rights-based governance rather than chasing sovereignty mirages. This means building interoperable systems that work across borders while maintaining local control over critical decisions.
For LMICs specifically, this translates to practical priorities: investing in digital literacy, strengthening governance frameworks, creating regional collaboration mechanisms, and developing local technical capacity for AI deployment and customization.
The path to technological empowerment doesn’t run through expensive infrastructure projects that recreate existing capabilities poorly. It runs through smart integration with global systems while building genuine local capacity where it matters most.
