Pathways to Equitable AI Compute Access

Contents

Executive Summary

This multistakeholder dialogue, organized by Oxford Martin AI Governance Initiative, Columbia University, and the Quantum Hub, examines how countries—particularly in the Global South—can achieve equitable access to AI compute infrastructure. The discussion frames compute access not as a technical problem but as a political economy issue, exploring whether "shared compute hubs" (pooled international infrastructure) can address the concentration of AI power while navigating geopolitical tensions and national sovereignty concerns.

Key Takeaways

Compute is political, not just technical. The concentration of AI infrastructure determines geopolitical power, innovation access, and economic outcomes. Addressing it requires governance innovation, not just hardware deployment.
Shared compute hubs are viable but require deep trust. Success depends on multi-country governance, shared principles, and partnerships limited to nations with mutual trust. They cannot be universal solutions but can work within regional/alliance contexts.
Strategic agency beats self-sufficiency. Countries should focus on controlling/steering critical stack layers (safety, data practices, standards, inference optimization) and deliberately, confidently depending on partners for frontier models and chips—rather than attempting impossible autarky.
India's hybrid model demonstrates viability. Government-funded infrastructure with private-sector operation, aligned with local scale and pricing realities, can rapidly build compute capacity while maintaining democratic input and commercial incentives.
Inference and localization are underestimated levers. Most countries will unlock AI value not through frontier model development but through smart adoption, fine-tuning with local data, and deployment infrastructure—requiring far less compute and capital than training models from scratch.

Key Topics Covered

Compute as a Bottleneck: Why AI compute concentration limits innovation access, research agendas, and who gets to build AI applications globally
Shared Compute Hubs Model: Proposed mechanism where countries pool resources to build jointly-governed AI infrastructure with democratic access
India's Hybrid Approach: How India's government-industry partnership strategy differs from purely capitalist (US) or state-controlled (China) models
Digital Sovereignty vs. AI Sovereignty: Distinction between "control everything yourself" (impossible) vs. "strategic agency" (securing access and influence in key stack layers)
The AI Trilemma: The tension between maximizing speed of development, competitive advantage, and democratic control simultaneously
Geopolitical Risks: How international conflicts could undermine shared infrastructure trust; export controls and semiconductor restrictions as leverage
Open Source & Decentralization: Lessons from 1990s web battles (Linux, Apache, Mozilla) applied to current AI infrastructure challenges
Infrastructure Layers & Sovereignty: The distinction between ownership, control, meaningful operational autonomy, and reliable access
Talent Distribution: 90% of leading AI scientists globally are outside the US and China, yet many work within those countries
Energy, Training, & Inference: The need to balance compute investment across model training and deployment/inference phases

Key Points & Insights

Compute Equals Power in AI Era
- Control of compute infrastructure determines who can innovate, train models, set research priorities, and ultimately capture value. Concentration mirrors historical patterns (1990s web centralization) with serious long-term consequences for competition and innovation diversity.
Shared Compute Hubs: Addressing Sovereignty Without Going It Alone
- No country can independently build the entire AI stack (chips, power, talent, models, energy). Shared hubs allow countries to pool capital, negotiate better hardware prices, share governance structures, and reduce technical/financial barriers—while maintaining democratic input through institutional independence from individual political systems.
India's Hybrid Model as Proof of Concept
- In two years, India shifted from minimal domestic compute to aggregating 38,000+ GPUs through a government-industry partnership. Government funds GPU access and provides market certainty; private sector (Yota, others) builds and operates data centers. Result: rapid scaling aligned with India's "frugal economy" approach and ability to reach 1.4 billion people.
"Control, Steer, Depend" Framework Over Full Self-Sufficiency
- True digital sovereignty is not owning everything. Rather, it's exercising strategic agency by (a) controlling critical layers (e.g., safety standards, data practices), (b) steering markets through scale/procurement, and (c) deliberately depending on trusted partners for specialized capabilities. India exemplifies this in AI strategy.
Geopolitical Tensions Threaten Shared Infrastructure Trust
- Regional conflicts could weaponize shared compute access; export controls (e.g., US semiconductor restrictions) limit pluralism. Trust-building mechanisms must be robust: shared legal principles, multi-country governance structures, and partnerships limited to nations with existing mutual trust relationships.
Training vs. Inference: A Neglected Distinction
- Emphasis on frontier models (training) overshadows deployment (inference) opportunities. Adopting and fine-tuning existing models with local data, safety norms, and language adaptation is where most countries unlock transformative AI value for their populations. Inference-focused strategies reduce compute requirements.
Principles Over Products
- Governments pursuing shared compute must prioritize public interest principles (interoperability, openness, decentralization) over short-term "market fit" or winning an "AI race." Sustainable partnerships require agreed-upon governance principles, transparency requirements, and incentives for open-source contribution.
The $1 Billion Wall: Current AI Economics Are Unsustainable
- Frontier AI development requires unprecedented capital and infrastructure investment ("a wall of money"). Current models (largest, fastest, best) are economically unsustainable. Future landscape may shift toward smaller, specialized, decentralized AI—creating opportunities for alternative architectures.
Talent Is Globally Distributed; Access Is Not
- 90% of leading AI researchers are outside the US and China, yet many work there due to compute/capital access. Repatriating talent and creating local ecosystems requires simultaneous investment in compute, policy stability, and local AI applications (healthcare, agriculture, education).
Rule of Law & Meaningful Access as Infrastructure
- Beyond physical compute, equitable access requires robust legal frameworks guaranteeing access rights, transparent allocation criteria, and protection against arbitrary restrictions. Governance structures must reflect the principles of the "old world" (contract law, dispute resolution) applied to digital infrastructure.

Notable Quotes or Statements

Linda Griffin (Mozilla): "Compute equals power... if we don't distribute access then there's no competition without competition there is no innovation and we know that the market is just not incentivized to do this on its own."
Robert Trager (Oxford Martin AI Governance): "All countries in the world that are not the United States and that are not China seem to have a bit of a common interest in being sure that they have continued and certain access to these advanced technologies."
Sunil (Yota Data Services): "I think it's the most dangerous statement to give [to say] 'let private sector take over.' Government has done a great job in funding the training cycle of models. Let them take it through the inferencing cycle also."
Sangeita Gupta (NASSCOM): "The demand far outpaces [current supply]... there are a large number of startups [and] researchers who will need much more affordable access to compute."
Pente (Tony Blair Institute): "We want to look at AI in two different phases... training and inference... it's about drawing on the models to infer or to create outputs that are useful to your citizens so that we really unlock the transformative potential of AI."
Hector (Microsoft): "No state or no company can or must dominate every layer of the IT tech stack... the practical play here is to decide where you want to build durable influence."

Speakers & Organizations Mentioned

Speaker	Role/Title	Organization
Rohit Kumar	Moderator, Founding Partner	Quantum Hub
Robert Trager	Co-director	Oxford Martin AI Governance Initiative
Linda Griffin	Vice President, Global Policy	Mozilla Foundation
Sangeita Gupta	Senior VP & Chief Strategy Officer	NASSCOM (India)
Hector	Director, Responsible AI Public Policy	Microsoft
Pente	Senior Policy Adviser	Tony Blair Institute for Global Change
Sunil	Managing Director & CEO; "Data Center Man of India"	Yota Data Services
Khalik Kabali	Referenced	(India AI governance discussions)
Martin Tisday	Architect (referenced)	French AI Summit
Michael Katzios	Economist (referenced)	—

Institutions/Initiatives:

Oxford Martin AI Governance Initiative
Quantum Hub (Columbia University)
Committee of Global Thought, Columbia University
Mozilla Foundation
NASSCOM (National Association of Software and Service Companies, India)
Tony Blair Institute for Global Change
Microsoft AI and Research
Yota Data Services (India)

Technical Concepts & Resources

Key Models & Systems Referenced

Frontier Models: GPT (OpenAI), Gemini (Google), Claude (Anthropic), Mistral, Cohere, Deepseek
India's Homegrown Models: Mentioned as beating ChatGPT and Gemini in certain AI domains (discussed as part of India's AI initiative success)

Infrastructure & Compute Concepts

GPU Aggregation: 38,000+ GPUs aggregated in India through India Mission; target scaling to higher numbers
Data Center Architecture: GPU-specific data centers (25-50x denser power/cooling than traditional facilities); networking, energy optimization
Optical Fiber Latency: 250-millisecond transport capacity between continents for distributed compute coordination
Power Requirements: Scaling from 200 MW (current data center capacity) → 1.4 GW → 3-10 GW (projected with AI adoption)

Governance & Policy Frameworks

Shared Compute Hubs Model: Pooled international infrastructure with governance independent of individual political systems; capital contribution → proportional access rights
Control, Steer, Depend (CSD) Framework: Pente's framework for redefining digital sovereignty as strategic agency rather than autarky
- Control: Critical stack layers (safety, standards, data practices)
- Steer: Market influence through scale and procurement
- Depend: Reliable partnerships for specialized capabilities
Hiroshima Process Reporting Framework: Referenced as baseline for AI risk management, transparency, and code of conduct reporting (applicable to compute infrastructure)
Rule of Law: Access decisions grounded in obligations, objective standards, and legal protections

India-Specific Initiatives

India Mission: Government-industry hybrid model funding GPU access and providing compute to startups/researchers on consumption basis
AI Governance Guidelines: Released by India touching on themes of compute access, open source, competition
Frugal Economy Model: India's approach to rapid scaling with low unit costs, adapted from mobile/payment revolution successes
Digiatra Model: Referenced as seamless integration of state/central policy incentives (aspirational comparison for coordination)

Historical Precedents & Open-Source Lessons

1990s Web Battles: Linux (OS), Apache (web server), Mozilla (browser) emerged from open competition; demonstrated affordability, innovation enablement, and non-proprietary models at scale
Open Source AI: Decentralized, cross-country collaboration; emerging emphasis on small AI, specialized models

Theoretical Concepts

AI Trilemma (Prof. John Hui, Harvard Kennedy School): Difficulty maximizing all three of (1) speed of development, (2) competitive advantage, (3) democratic control/oversight simultaneously
Geopolitical Dynamics: Export controls (US semiconductor restrictions), FTA-equivalent AI trade agreements, regional conflict risks to shared infrastructure

Data & Metrics

AI Adoption Gap: 24.7% of working-age people in Global North use generative AI; 14.1% in Global South (widening gap)
Talent Distribution: 90% of leading AI scientists globally are outside US/China; many work within those countries
Countries Leading in AI Adoption: Norway, Singapore, UAE (despite smaller compute bases)

Energy & Sustainability

Renewable Energy: Yota data centers noted as renewable-powered; energy sustainability a critical variable in scaling compute
Energy Planning: Need for coordinated energy infrastructure planning alongside compute (not just power availability but predictable supply)

Document Context: This transcript is from the Oxford Martin AI Governance Initiative Impact Summit, a multistakeholder dialogue on equitable AI infrastructure access. The discussion intentionally shifts from technical solutions to political economy, emphasizing that compute access is fundamentally a governance and power-distribution issue, not merely an engineering challenge.