AI, Sovereignty, and Cooperation: Shaping the Future Across the Global Sout

Contents

Executive Summary

This panel discussion explores how nations—particularly in the Global South—can build sovereign AI capabilities while maintaining meaningful international collaboration. The speakers argue that AI sovereignty doesn't mean complete self-reliance in building every component of the stack, but rather developing local, context-specific models and governance frameworks while leveraging global open-source tools and infrastructure. The conversation frames AI resilience as analogous to climate resilience: both require integrating fragmented data systems, building redundancy, and viewing intelligence not merely as prediction but as adaptability to inevitable unpredictability.

Key Takeaways

Don't Build Everything Yourself: Sovereignty in AI means building domain-specific, locally contextualized models and governance—not replicating the full tech stack. Leverage open-source foundations and partnerships strategically.
Compute-as-a-Service, Not Compute-as-Ownership: Governments should provide subsidized, managed compute infrastructure via portals (as India does) rather than letting each institution buy expensive hardware. This scales efficiently and maintains sovereign control.
Interoperability Over Isolation: The future is modular, open data exchange (like DPI principles applied to AI). Nations should design systems that orchestrate data and tool use while maintaining oversight—not hoard data in silos or depend on single vendors.
Data Governance Is Now Existential: For the Global South, establishing consent-driven, transparent data frameworks before external actors mine domestic datasets is critical. Once data is consumed by a model, it's permanently gone.
Governance Diversity Is Fine, But Benchmarks Matter: Multiple models with different alignments are legitimate. What matters is shared safety, reliability, and transparency standards—not model monoculture. Ensemble resilience beats single-point-of-failure.

Key Topics Covered

AI Stack Sovereignty: The five-layer model (energy, chips, datasets/data centers, models, applications) and which layers nations can realistically own
India's AI Mission: Compute-as-a-service model, 12 foundation models, 38,000 committed GPUs, and the rationale for sovereign infrastructure
Global South Data Exploitation: The paradox that Global South nations have abundant data ("oil") but lack the infrastructure ("Ferrari") to use it; risk of data extraction without local benefit
Domain-Specific vs. Frontier Models: Why most use cases don't need state-of-the-art large language models; SLMs (Small Language Models) for practical applications
Digital Public Infrastructure (DPI) Parallels: How DPI principles (modularity, interoperability) apply to AI governance and data exchange
Climate-AI Nexus: Topology similarities between complex adaptive systems; AI for climate prediction and resilience; digital twins of Earth
Data Interoperability & Governance: Technical orchestration of data flow; trust-building across fragmented systems; consent-driven, bias-free AI
Multi-Model Ecosystems: Benefits and governance challenges of multiple competing models (12 Indian models, Chinese models, French models, etc.)
Compute Accessibility for Small Businesses & Startups: Democratization challenges; existing but underutilized programs
International AI Alliances: Potential for regional clustering and knowledge-sharing among smaller nations

Key Points & Insights

Sovereignty ≠ Complete Self-Reliance: Only very large countries (China, India) can realistically build full-stack AI independently. Most nations should focus on domain-specific models, local data governance, and leveraging open-source foundations rather than recreating frontier models.
Compute-as-a-Service Model Works: India's approach of offering subsidized, government-managed GPU access via a portal with 40% subsidy to users demonstrates how to democratize compute at scale while maintaining sovereign control. This is more efficient than nations buying compute outright.
The Data Paradox of the Global South: These regions have abundant data (the "gas"), but lack infrastructure and governance frameworks (the "Ferrari") to extract value. Without sovereign frameworks, external actors will mine and exploit this data, and once consumed by models, it's gone permanently.
Topology of Climate = Topology of AI: Both are complex adaptive systems exhibiting nonlinear, exponential behavior and phase transitions. Both require resilience (not just prediction) to handle inevitable unpredictability. Both are fragmented across siloed systems. Data interoperability is key to solving both.
Multiple Models, Multiple Alignments ≠ Fragmentation: Having 12 Indian foundation models with different alignments reflects legitimate diversity. Governance should focus on safety/reliability benchmarks rather than model monoculture. The principle: "Let a million flowers blossom" on diverse models, but within a streamlined governance structure.
SLMs (Small Language Models) Are the Practical Future: Most government and commercial problems don't require frontier LLMs; they require efficient, domain-specific SLMs. Four of India's 12 foundation models are SLMs. SLMs also reduce energy footprint, helping with climate concerns.
DPI-Inspired Orchestration for AI: Modularity and interoperability—proven with Digital Public Infrastructure—apply to AI. Data exchange architectures should focus on controlling how data flows and which tools are invoked, not on restricting access. This lets nations use best-of-world components while maintaining sovereignty.
Consent, Transparency, and Bias Mitigation Are Governance Pillars: India's approach combines a Data Protection Act (consent), AI Governance Guidelines (safety/reliability), and transparent data tracking (AI Datastore to detect bias). These provide a replicable template for other nations.
Resilience > Prediction in Uncertain Times: As unpredictability becomes the norm (due to climate tipping points, geopolitical shifts, technological phase transitions), the capacity to withstand shocks matters more than forecasting. Cohesion and integration drive resilience.
Energy and Legal/Governance Capacity Are Overlooked Constraints: Beyond compute and models, nations need energy infrastructure (India's Shakti Bill), institutional capacity to implement and sustain AI services, and legal frameworks. Without these, even available compute remains unutilized.

Notable Quotes or Statements

Rahul Mathan (Trilegal): "Sovereignty and collaboration are not contradictory—they are complementary. Sovereign people across the planet can collaborate to achieve the progress of humanity."
Shilpa Pakurka (Nvidia): "The future of AI is global, but it's going to be deeply local as well... sovereignty is about domain-specific models that solve the unique problems of a nation using local datasets and cultural nuances."
Sudeepa G (Ministry of Electronics & IT, India): "Only strong sovereign nations can do collaborations for the benefit of humanity... we want the entire global south to benefit from this AI revolution."
Daniel Abadi (CDPI): "Bureaucracy creates inequality, inflation, and poverty because it separates people. AI is a tool to kill bureaucracy and make it more efficient... governments need to have their house in order before scaling AI" (the "Ferrari without gas tank" analogy).
Tulio Vigevani (COP 30 Presidency): "The topology of climate change and AI is the same—both are complex systems with nonlinear, exponential behavior. Intelligence has two legs: prediction AND resilience. When unpredictability is the norm, resilience matters more... nations must become collective intelligence ecosystems."
On SLMs and Efficiency: "Small language models are going to play a very critical role... most of our problems, day-to-day issues are domain specific" (Sudeepa). And Daniel: "Governments only need a small language model. You are not doing rocket science."
On Data as Non-Renewable: The implicit but powerful framing that once a nation's data is ingested by an external model, "it's gone"—this is a one-time extraction, unlike traditional resources.

Speakers & Organizations Mentioned

Speaker	Organization	Role
Rahul Mathan	Trilegal	Moderator; legal expert; raises sovereignty questions
Shilpa Pakurka	Nvidia	Hardware/compute infrastructure; global AI partnerships
Sudeepa G (Sud)	Ministry of Electronics & IT, Government of India	India AI Mission lead; sovereign stack architect
Daniel Abadi	CDPI (Center for Digital Public Infrastructure)	DPI expert; works with Latin America, Africa, Asia; governance design
Tulio Vigevani	COP 30 Presidency (Brazil)	Climate-AI nexus; complex systems resilience; "planetary intelligence"
Mahir (implied)	Summit organizer	Opening logistics

Institutions/Initiatives Referenced:

India AI Mission (March 2024 launch; compute-as-service, 12 foundation models)
Nvidia Inception Program (virtual incubator for startups)
Ministry of Electronics & IT (India)
COP 30 (Brazil, climate/AI integration)
CDPI (DPI frameworks for Global South)
IITs (Indian Institute of Technology)
Digital Public Infrastructure (DPI) movement (emerging from G20)

Technical Concepts & Resources

AI/ML Concepts

Large Language Models (LLMs) vs. Small Language Models (SLMs): SLMs highlighted as more practical, efficient, and domain-appropriate for most use cases
Domain-Specific Models: Fine-tuned, locally contextualized versions of foundation models
Foundation Models: Base models (12 in India's mission) that can be adapted for specific sectors
Model Alignment: Value/ethical alignment embedded by model builders; challenge of governance across multiple models with different alignments

Infrastructure & Architecture

Five-Layer AI Stack (per India's framing):
1. Energy
2. Chips/Hardware
3. Datasets & Data Centers
4. Models
5. Applications
Compute-as-a-Service (CaaS): Government-managed, subsidized GPU access via portal; 38,000 GPUs committed (India phase 1), 20,000–25,000 additional (phase 2)
Digital Twins: Simulations of real systems (e.g., digital twin of Earth for climate modeling)
Open-Source Models: Leveraged by nations to avoid reinventing; locally fine-tuned for sovereignty

Data & Governance

Data Interoperability: Breaking siloed data systems to enable cross-regional, cross-sector intelligence (both climate and AI contexts)
Orchestration Layer: Controls how data flows, which models/tools are invoked—keeps sovereignty while using diverse components
Consent-Driven AI: Data Protection Act (India) ensures data cannot be used without explicit consent
Bias Detection: Transparent data tracking (e.g., India's "AI Datastore") to identify and flag biases
Benchmarks & Standards: Shared safety/reliability standards across models, but allowing diverse implementations

Climate-AI Integration

Complex Systems Theory: Nonlinear dynamics, exponential behavior, phase transitions, cascading tipping points
Resilience Engineering: Capacity to withstand unpredictable shocks, not just predict known futures
Cohesion & Integration: Key drivers of systemic resilience in the context of climate/geopolitical uncertainty

Governance Models

India AI Governance Guidelines (November 2025): Trust, innovation, safety, reliability framework
Shakti Bill (India, recent): Addresses energy infrastructure for AI data centers
Data Protection Act (India): Consent-based data usage
PLI (Production Linked Incentive), DLI, Electronics Component Manufacturing schemes (India): Chip/hardware development

Programs & Initiatives

Nvidia Inception Program: Virtual incubator offering GPU access to startups
India AI Mission: Empaneled 14 Cloud Service Providers; selection underway for more foundation model developers
Digital Public Infrastructure (DPI): Modular, interoperable systems (UPI, digital identity, payment systems) now being extended to AI
Planetary Intelligence Initiative (emerging from Brazil-India partnership): Open regional AI frameworks for the Global South

Potential Limitations & Gaps in the Transcript

Incomplete speaker attribution: Some responses lack clear attribution; tone suggests informal panel but editing would help clarity
No specific technical metrics: Discussion of "state-of-the-art" models (e.g., Mistral) is referenced but not detailed
Limited Q&A depth: Audience questions are mentioned but truncated; full responses not always captured
Policy documents referenced but not quoted in detail: India's AI Governance Guidelines and Data Protection Act are cited but specific provisions are paraphrased
Energy requirements vaguely mentioned: SLMs reduce compute/energy, but no specific figures (kWh, carbon footprint) are provided

Document Version: Conference talk summary
Confidence Level: High (direct transcript quotes and clear speaker statements)
Intended Audience: AI policy makers, researchers, government digital officers, development practitioners in the Global South