Global Tech, Local Impact: Governing AI Where It’s Deployed

Contents

Executive Summary

This panel discussion examines "AI sovereignty"—the strategic capacity of countries, particularly middle powers and Global South nations, to maintain meaningful control over AI infrastructure, data, and governance within their borders. Rather than advocating for complete digital independence, speakers present sovereignty as a risk-management strategy involving selective autonomy in critical areas, often requiring partnerships with foreign technology companies. The conversation emphasizes that effective AI governance must balance geopolitical resilience, local implementation needs, economic development, and individual user agency.

Key Takeaways

Sovereignty is Strategic Risk Management, Not Isolation: Countries should identify critical data, applications, and infrastructure layers (e.g., public services, healthcare, defense) and apply targeted controls there, while accepting necessary foreign partnerships elsewhere. Complete autarky is neither feasible nor optimal.
The Global South Must Build Compute Infrastructure Today or Face Permanent Dependency: With 1.2 billion people entering working age in next decade, absence of local AI infrastructure now guarantees digital colonialism. Regional collaboration (hub-and-spoke models, fractional compute sharing) offers viable scaling path without requiring individual country to build gigawatt-scale data centers.
Open Source Wins in Global South—Chinese Models Now Leading: Meta's retreat from aggressive open-source development ceded ground to Chinese models (Qwen) demonstrating superior performance. Open source + performance = inevitable technology choice for resource-constrained developers; Western closed-source models losing market share despite geopolitical backing.
Regulation Must Adapt to Avoid Stifling Innovation or Missing Risk: Neither rigid pre-regulation nor post-hoc intervention will work. Outcome-focused frameworks with innovation sandboxes allow governments and companies to collaborate on solving specific problems (fairness, privacy, security) without prescribing technology.
Address Societal Risk (Inequality) Alongside Existential Risk (AGI): The urgent risk for most of world is not killer robots but growing AI-enabled inequality between tech-literate elites and excluded populations. Sovereignty frameworks must ensure local capacity to serve underserved populations (farmers, healthcare patients, unbanked citizens) in local languages and contexts.

Key Topics Covered

AI Sovereignty Definition & Scope: Moving beyond the concept as an all-or-nothing proposition; understanding it as degrees of control and risk management across the AI stack
Geopolitical Context: How digital dependency on foreign platforms creates vulnerability (vaccine access, conflict narratives, export controls)
The AI Stack: Infrastructure, model development, data projects, and compute capacity as distinct sovereignty vectors
Global South Implementation: Practical constraints and opportunities for countries with limited compute resources, fragmented data ecosystems, and diverse language populations
Open Source Models: Role of open-source frameworks (Llama, Qwen) in democratizing AI development; popularity of Chinese models in Africa
Regulation & Competition Law: Limitations of traditional competition law; trend toward outcome-based regulation rather than prescriptive rules
Infrastructure Strategy: Frugal/lightweight compute approaches, micro data centers, edge computing, and regional collaboration models
Risk Considerations: Existential risk vs. societal risk (inequality, digital colonialism); individual consumer trust and agency
Business vs. Public Applications: Distinction between commercial AI use cases and public/defense applications requiring sovereignty
Digital Colonialism: Pattern where economic value from technology is extracted and concentrated outside the Global South

Key Points & Insights

Sovereignty is a Spectrum, Not a Binary: The vast majority of public-sector sovereign AI projects tracked globally include partnerships with foreign technology companies. True isolation is neither feasible nor desirable; countries instead strategically manage which parts of the stack they prioritize.
India's Multi-Layered Motivation for Sovereignty:
- Geopolitical resilience (avoiding foreign policy dependencies as demonstrated by platform censorship during conflicts)
- Supply chain security (semiconductor vulnerabilities)
- Cultural and linguistic diversity preservation (100+ Indian languages/dialects underrepresented in globally-trained models)
Rapid Growth of Sovereign AI Projects: Documented increase from 40 projects across 30 countries (2024) to roughly 130 projects across 50+ countries by early 2026. India uniquely operates across all three stack layers (infrastructure, models, data).
Global South Faces Structural Compute Inequality: Africa holds less than 1% of global compute capacity despite housing 18% of world population. This creates unavoidable dependency on external infrastructure, but also opportunity for regional collaboration models (e.g., ECOWAS hub-and-spoke approach).
Open Source as Democratizing Force: ~70% of sovereign AI model projects tracked are released as open source. Chinese open-source models (Qwen) gaining popularity in Africa due to superior performance and accessibility compared to U.S. closed-source alternatives—geopolitics inseparable from technical choice.
Regulation Must Focus on Outcomes, Not Technology Prescriptions: Given AI's rapid evolution, regulators cannot keep pace with technology-specific rules. Shift toward outcome-based approaches (fairness, transparency) using innovation sandboxes and trusted industry collaboration rather than adversarial legal frameworks.
Digital Colonialism Pattern Repeating: Parallels to historical resource extraction: foreign platforms (WhatsApp, Facebook, Instagram) generate economic value locally but concentrate profits globally. AI infrastructure follows similar pattern; governments awakening to demand for local investment and value capture.
Risk Landscape Misaligned with Implementation Priorities: Early AI discourse fixated on existential/safety risk; now pendulum has swung to downplaying risk entirely. In reality, the urgent risk for Global South is societal risk—widening inequality between tech-savvy urban elite and vast populations with limited AI literacy and access.
Performance + Access Drive Technology Choice: Developers in Global South prioritize open-source models offering both performance benchmarks and accessibility over ideological alignment with Western platforms. Tinymachinelearning movement in Africa demonstrates sophisticated local innovation on edge computing.
Individual vs. National Sovereignty Tension Unresolved: Cumulative individual user autonomy doesn't necessarily translate to national strategic advantage; defense, healthcare, and public services require national-level controls that individual privacy choices cannot substitute for.

Notable Quotes or Statements

"Sovereignty to us is not just about trying to beat the race between the US and China but it is more about de-risking geopolitically and it goes across the entire stack all the way from the chips all the way to the platforms." — Shashi (on India's sovereignty rationale)

"2.2 billion people that are unconnected. That means a lot of their data is actually not represented. Their culture, their languages, their knowledge systems are not represented and integrated in the large language models that are the most widely used today." — Kate (on underrepresentation in global models)

"Sovereign AI will involve partnerships with third parties outside of the country and it's just going to be a matter of managing those relationships." — Pablo (on realistic sovereignty expectations)

"We can't decouple AI from geopolitics or technology from geopolitics... open source is playing a large role into enabling communities who historically have been marginalized." — Kate (on geopolitics and open source in Global South)

"We're about to enter a similar digital colonialism as we're getting into AI because the truth is most of the compute infrastructure sits outside of our countries today." — Kate (on the colonialism analogy)

"It's not necessarily a given that what's right for an individual is right for the country or vice versa." — Natalie (on tension between individual and national sovereignty)

"The bigger concern is not so much about the risk of existential issues but more about the divide." — Shashi (on India's prioritized risk: inequality)

"Outcome-based regulation... you need industry to come to you and say we would like to do this but we're not quite sure how you're going to feel about it." — Natalie (on regulatory approach)

Speakers & Organizations Mentioned

Speaker	Organization	Role/Expertise
Akash Kapoor	New America, NYU Gov Lab, Princeton University (visiting scholar)	Moderator; focuses on geopolitics of AI, sovereignty
Shashi	(India-based; formerly CEO of All India Radio)	Indian AI policy, cultural/linguistic diversity in AI
Pablo	Center for New American Security (CNAS)	Tracking sovereign AI projects globally; released "Sovereign AI Index"
Natalie Black	Ofcom (UK communications regulator), executive board member	Infrastructure, connectivity, AI implications in regulatory context; trade expertise
Kate	(Global South-focused company, name not fully specified)	Edge computing, frugal infrastructure, distributed compute in Africa, Caribbean, Latin America, Southeast Asia
Josh Tan	(mentioned as listener)	EU-focused shared compute infrastructure work
Gordon	(moderator/organizer)	Panel coordinator
Dario Amodei	(referenced, not present)	Author of essay on AI risk cited by moderator

Institutions/Bodies Referenced:

Ofcom (UK communications regulator)
New America
Princeton University
NYU Gov Lab
Center for New American Security (CNAS)
All India Radio
Indian School of Public Policy
Anoha Parika Institute for Defense Studies and Analysis
IMT Ghaziabad
Meta (Open AI's partner; context on open-source model retreat)
Microsoft
OpenAI

Technical Concepts & Resources

AI Models & Frameworks

Llama (Meta): Open-source foundational model; losing popularity to Chinese alternatives
Qwen (Chinese): Open-source model gaining adoption in Africa; cited as superior performance + accessibility
Gemini (Google): Globally-trained model; underrepresents regional cultural context
ChatGPT (OpenAI): Example of models missing local nuance

Infrastructure & Compute Concepts

Micro data centers: Small-scale, renewable-powered computing infrastructure for data residency
Edge computing: Intelligence deployed closer to data source (devices, local centers) vs. centralized cloud
D2M (Direct-to-Mobile Broadcasting): Standard (ATSC3) for secure model/dataset distribution to edge devices
Fractional compute: India's approach of purchasing time-shared access to hyperscaler GPUs rather than building dedicated infrastructure
Hub-and-spoke model: ECOWAS regional collaboration approach—larger compute in energy-rich country funded by region; smaller local compute for data sovereignty
GPU access: Hyperscalers (Microsoft, Google, etc.) providing compute via time-share arrangements

Policy & Governance Frameworks

Outcome-based regulation: Focus on fairness, transparency, security outcomes vs. technology-specific rules
Innovation sandboxes: Regulatory environments allowing controlled experimentation with specific use cases
Competition law: Traditional approach deemed inadequate for AI oligopolies; alternatives needed
Export controls: U.S. chip export restrictions; mentioned as external constraint on sovereignty efforts
U.S. Cloud Act: Legislation allowing U.S. government access to data on U.S. cloud platforms globally

Data & Language Concepts

Digital footprint disparity: ~100+ Indian languages/dialects; many lack written digital representation
Spoken interface vs. UI interface: User preference difference in India (populations comfortable with voice but not graphical interfaces)
Use-case localization: UPI (Unified Payments Interface) as India-specific payment modality requiring localized AI
Data residency: Requirement that critical data remain within national borders
Cultural nuance in training data: Problem of globally-trained models missing regional context, cultural references, language subtleties

Relevant Statistics & Data Points (from Sovereign AI Index)

2024: 40 projects across 30 countries
Early 2026: ~130 projects across 50+ countries
~40% of sovereign model projects based on fine-tuning Llama/open-source models (increasingly Qwen)
~70% of sovereign AI model projects released as open source
17 of 20 G20 countries engaged in sovereign AI projects
India: operates across all 3 stack layers (model, infrastructure, data)
Africa: <1% of global compute capacity for 18% of global population
India: ~1 billion mobile devices; average user consumes 20-30GB data/month (video-heavy)
UK: only 28% of population comfortable using AI
Africa: 2,500 languages; majority unwritten/spoken-only
Global South: 1.2 billion people entering working age within next decade

Research/Documents Referenced

Pablo's forthcoming work: "Sovereign AI Index" (Center for New American Security)
Dario Amodei essay: On AI risk (cited by moderator as framework for reconsidering risk discussion)
Tinymachinelearning movement: Edge AI community; Kenya was 3rd largest chapter globally after US/Europe

Note: This transcript represents a live panel discussion with informal speaking patterns, overlapping commentary, and incomplete speaker attributions in places. Direct quotes are preserved as spoken; summary synthesizes loosely organized remarks into thematic structure.