Building Public Interest AI: Catalytic Funding for Equitable Compute Access

Contents

Executive Summary

This panel discussion at an AI summit centers on democratizing compute infrastructure globally, with India's 38,000-GPU public compute initiative serving as a flagship model. The conversation challenges narrow interpretations of "democratization" and "sovereignty," arguing instead for integrated ecosystems that combine compute with data governance, talent development, and local agency—moving from theoretical principles to operational, scalable implementation across the Global South.

Key Takeaways

"Democratization is not theoretical—it is operational": AI access for the Global South requires building integrated ecosystems (compute + data + governance + talent + institutions), not just distributing GPUs. This is already underway in India but requires acceleration and replication.
Challenge the Compute-First Narrative: While 38,000 GPUs is significant, the real bottlenecks are often energy, governance, skilled talent, quality datasets, and institutional capacity. Philanthropies and governments should distribute resources across the full stack, not overinvest in hardware alone.
Measure, Don't Assume Demand: Use frameworks like the "Compute Demand Index" and "Investment Readiness Index" to quantify actual GPU-hour needs and countries' operational capacity before committing infrastructure. This prevents waste and enables honest South–South negotiations.
Redefine Sovereignty as Agency: Move away from territorial control narratives toward relational sovereignty—ensuring communities and nations have decision-making authority over data, models, and AI systems that affect them, and access to open, interoperable tools.
Build Intermediaries and New Institutions: Governments, philanthropies, and private actors cannot solve this alone. Support organizations like Kalpa Impact that sit at intersections, connecting policy, technical capacity, and community needs at scale.

Key Topics Covered

Compute as Public Infrastructure: India's AI mission and the concept of compute as a digital public good
Democratization vs. Agency: Distinction between access to technology and genuine decision-making power over AI systems
The Compute Bottleneck Myth: Questioning whether compute is the primary constraint; emphasis on energy, models, data, and skills
Data Governance & Privacy: Unlocking datasets for training while respecting privacy and sovereignty
Global South–South Partnerships: India–Africa and Asia–Asia collaboration models; practical GPU hour metrics and demand frameworks
Institutional Innovation: New intermediaries needed to bridge government, philanthropy, private sector, and frontline communities
Sovereignty Reconceptualized: From territorial control to relational agency and indigenous data sovereignty principles
Energy & Hardware Constraints: Physical limitations of latency, power availability, and infrastructure readiness
Multistakeholder Platforms: The MAITRI initiative (Multi-stakeholder AI for Trusted and Resilient Infrastructure) as a collaborative framework
Investment Readiness vs. Compute Demand: Assessing countries' ability to operationalize compute (talent, governance, use cases) before infrastructure deployment

Key Points & Insights

Compute is Not the Primary Bottleneck: Multiple panelists (Martin Vélias, Sean Siao) argue that compute capacity alone won't solve AI democratization; energy costs, hardware access, skills gaps, and governance frameworks are equally or more critical constraints.
The "Compute Demand Index" Approach: Dr. Shiku Gatau's framework quantifies GPU hour needs per continent (Africa needs 2.5–7.5 million GPU hours annually) rather than relying on arbitrary gigawatt targets, enabling practical South–South negotiation.
Investment Readiness ≠ Compute Supply: Countries receiving GPU infrastructure without power, talent, data, or governance capacity risk "white elephant" data centers—resources wasted without operational capacity or local use cases.
Data Stewardship as Critical Infrastructure: Martin Vélias emphasizes that data innovation has lagged compute innovation; enterprise users have privacy-preserving technical safeguards that individual/citizen users lack. Governance models like data trusts remain unproven at scale.
Sovereignty as Relational, Not Territorial: The panel reframes "sovereignty" from Westphalian territorial control to indigenous data sovereignty concepts—about agency, authority, and choice over what data represents a people, not merely owning servers on your land.
Interdependence Over Independence: Villis Dhar argues against competitive "sovereignty" narratives; instead, interconnected infrastructure allows mutual value exchange—countries with compute capacity provide services; recipients build applications and return value through local innovation and markets.
MAITRI Platform as Digital Public Good: Proposed multi-stakeholder infrastructure designed as customizable, non-binding, modular—allowing countries to adopt compute, data, and governance mechanisms aligned with local contexts without vendor lock-in.
Three Foundational Pillars Beyond Compute: Capability (skills diffusion, joint research), collaboration (open platforms, mutual learning), and compliance (governance frameworks flexible enough for cultural/social diversity).
Philanthropy's Catalytic Role: Institutions must subsidize compute costs, resource open-source dependencies, fund intermediaries (e.g., Kalpa Impact), and support institutional capacity—not just donate hardware.
Use Cases Drive Infrastructure Decisions: AI infrastructure should be built in service of specific development outcomes (health, agriculture, education), not deployed speculatively; this reverses the typical hierarchy and makes compute demand measurable and justified.

Notable Quotes or Statements

Deepali Anand (Rockefeller Foundation): "The digital divide is rapidly becoming a compute divide. AI today is not constrained by imagination. It is constrained by infrastructure… Democratization in this context is not about catching up. It is about expanding who gets to lead."
Dr. Sorav Gar: "Most countries are not really seeking only access to AI but also seeking agency in AI… AI systems need to reflect each country's own development priorities, languages and social contexts."
Vishal Sikka (referenced by Dr. Gar): "Unlike compute infrastructure measured in gigawatts, a human being requires only 2,000 calories—not more than a 100-watt bulb. Are we missing something?" (Provocation challenging excessive compute assumptions.)
Villis Dhar: "AI diffusion is a passive concept… If we made the rich as rich as possible, somehow the benefits would filter down to everybody else… that sounds like trickle-down economics. But there's an alternate model… It requires us to step in and build institutional layers."
Martin Vélias: "There's a complete tragedy that we haven't seen [data innovation] anywhere close to as much innovation when it comes to compute… If we could harness even 20% of the brain capacity going into compute and apply it to data, we'd be in a very different place."
Dr. Shiku Gatau: "Africa needs 2.5 million GPU hours a year, 7.5 million for the next three years to be able to start competing… It's not just about facilitating GPU; it's what is the GPU in service of—solving for health, education, agriculture."
Sean Siao: "The ownership of compute [is] actually a bit overrated… The issue is actually deeper than just ownership—it's more access… The question is how do we make it more accessible for startups and impact organizations."
Villis Dhar (closing): "The response to overdependence isn't internal dependence. It's interdependence… allowing for centers of excellence that allow local capacity and local competence to drive what gets built."

Speakers & Organizations Mentioned

Government & Policy:

Dr. Sorav Gar – Secretary, Ministry of Statistics and Program Implementation, Government of India; chair, Democratizing AI Resources Working Group
Abishek Singh – Indian government official (acknowledged but absent from session)

Philanthropy & Foundations:

Deepali Anand – Rockefeller Foundation
Andrew VP – Rockefeller Foundation (moderator)
Villis Dhar – President, Patrick J. McGovern Foundation; UN Secretary General's AI Advisory Board
Sean Siao – CEO, Philanthropy Asia Alliance (80 member organizations across Asia)

Technology & AI Leadership:

Martin Vélias – Founder, Current AI; Public Interest Envoy, France's AI Action Summit; 15+ years on multistakeholder initiatives
Dr. Shiku Gatau – Founder/CEO, Kalpa AI (Kenya); established Safaricom Alpha; AFDB Digital Transformation Lead
Vishal Sikka – Referenced speaker (previous day remarks on compute efficiency vs. human energy)

Supporting Organizations:

Kalpa Impact – Intermediary organization connecting government, technology, and communities
Current AI – Open hardware and linguistic diversity initiatives
Safaricom Alpha – African AI infrastructure program
Roots (Robust Open Source Trust & Safety) – Funding open-source dependencies
MERI/MAITRI – Proposed multi-stakeholder AI infrastructure platform
UN Development Programme (UNDP) – Mentioned compute capacity offerings

Countries/Regions Referenced:

India (primary model: 38,000 GPU India AI Mission)
Kenya, Nigeria, Burundi (Africa demand examples)
Indonesia (Southeast Asia; data residency challenges)
Egypt (co-chair of Democratizing AI Resources Working Group)
New Zealand (Māori data sovereignty example)

Technical Concepts & Resources

Infrastructure & Architecture:

GPUs (Graphics Processing Units) – Primary compute bottleneck; measured in GPU hours or GPU clusters
Data Centers – Physical facilities for compute; latency challenges across geographies (50–100ms between India–Indonesia)
Energy/Power – Fundamental constraint; hydro, solar, wind cost reductions enabling accessibility in Asia
Gigawatt/Megawatt/Kilowatt – Units of power; discussions of appropriate scale per country

AI Governance & Frameworks:

MAITRI (Multi-stakeholder AI for Trusted and Resilient Infrastructure) – Proposed platform for shared compute, data, and governance; non-binding, modular, customizable digital public good
Six Foundational Pillars: Compute, Capability, Collaboration, Connectivity, Compliance, Context
Data Trusts/Data Stewardship – Governance mechanisms for privacy-preserving data sharing (noted as unproven at scale)
Indigenous Data Sovereignty – Relational, not territorial; authority over data representing a people (Māori example cited)

Measurement & Planning Tools:

Compute Demand Index – Framework quantifying GPU-hour needs by country/region (e.g., Africa: 2.5–7.5M hours/year)
AI Investment Readiness Index – Assessment of infrastructure readiness: power, talent, governance, data, use cases

AI Models & Systems:

Open-Source AI Stacks – Mentioned as critical but under-resourced; dependencies often volunteer-maintained
Domain-Specific/Niche Models – Alternative to large general-purpose models consuming gigawatts; potential future direction
Contextual/Localized AI – Models trained on local data, languages, and cultural diversity

Related Initiatives/Programs:

India AI Mission – 38,000 GPU public compute capacity plan
Aadhaar – Referenced as transformative tech stack example (mentioned in context of India's capability)
Open Government Partnership – Multistakeholder governance model (Martin Vélias's background)
Jensen Huang's AI Stack – Framework: Energy → Hardware → Compute → Models → Applications (Sean Siao reference)

Geographic/Regional Considerations:

Data Residency Requirements – Laws preventing data export; enable local data center build-out
Latency Limitations – Physical distance (10,000+ km) creates 50–100ms latency, making real-time compute sharing impractical
Power Availability Variance – Determines where compute infrastructure should be sited

Document Status: Working version released; feedback/comments accepted until March 31st.