Digital Public Infrastructure

Overview

Digital Public Infrastructure has emerged as the defining strategic context for AI deployment at population scale—not as a prerequisite to be completed before AI work begins, but as a force multiplier that determines whether AI benefits compound or concentrate 1. India's layered stack of identity (Aadhaar), payments (UPI), health identifiers (ABHA), and data-sharing frameworks (DEPA) now constitutes a proven template that other nations are actively studying and adopting 23. The central question at this summit was no longer whether to integrate AI with DPI, but under whose terms, with what safeguards, and at what cost to sovereignty and public interest 4. Speakers across sectors—agriculture, health, energy, education, disaster management, financial services—converged on a sobering consensus: the binding constraints are governance architecture and data interoperability, not model sophistication. The next 12–18 months represent a genuine window of strategic choice before global AI dominance patterns ossify 56.

---

Key Insights

• DPI is a multiplier, not a launchpad. Countries need not complete DPI before starting AI work, but the two must be deliberately coordinated. The worst outcomes come from AI built in parallel silos atop public infrastructure, creating new layers of gatekeeping on top of open foundations 17. The risk of "reverse colonization"—Indian data training foreign models sold back to India at premium—is structural, not hypothetical 5.

• Governance architecture matters more than model choice. Across agriculture 89, health 10, disaster management 11, and energy 12, speakers consistently framed the core challenge as designing accountability structures—who owns data, who manages models, who is liable when systems fail—rather than selecting the right algorithm. India's Agristack intelligence layer, where AI will operate atop farmer registries, is the current live test of whether this design can be done right 8.

• Data interoperability is the actual bottleneck. Approximately 80% of AI pilots fail at scale due to data silos and governance gaps, not algorithmic limitations 13. Solving fragmented, non-standardized data across federal systems—28 state governments in the disaster management context alone 11—requires unglamorous foundational work: standardized formats, metadata governance, cross-agency sharing protocols, and machine-readable glossaries 1415.

• Open standards and modularity determine whether benefits compound or concentrate. Closed, proprietary AI systems built atop public DPI create new forms of exclusion. The same logic that made TCP/IP generative applies: open agent protocols (MCP, A2A) and interoperable data exchange standards are as strategically important for AI as they were for the internet 1616. Vendor lock-in is a sovereignty risk, not merely a procurement inconvenience 17.

• Sovereignty is multidimensional and cannot be solved by data residency alone. True digital sovereignty requires control of the control plane, operational autonomy, domestic developer talent with deep platform knowledge, dedicated national budgets independent of donor cycles, and freedom from single-vendor dependency 41819. Nations that outsource these layers cannot expect to shape outcomes.

• Inclusion must be engineered, not assumed. Voice-first, multilingual interfaces are not accessibility features—they are the primary interface for the next billion users 22021. Systems designed for the last person (illiterate farmers with feature phones in remote areas) produce more robust systems for everyone 22. Women's absence from farmer registries is not a gap to be noted; it produces algorithmic exclusion at scale 2223.

• Procurement and regulatory capacity are as binding as compute. Technical excellence dies in bureaucracy without reformed procurement (favoring outcomes over lowest cost), trained civil servants, and regulators who can understand and oversee AI systems 2425. Investment in regulatory capacity—including AI safety institutes localized to the Global South—is as urgent as investment in GPU clusters 2326.

• Sandboxes are governance tools, not just technical ones. Regulatory sandboxes surface tradeoffs, test safeguards, and generate evidence on rights and accountability before population-scale rollout. UIDAI's sandbox for 20+ entities and its SITHA grant program exemplify how to build institutional learning capacity alongside technical innovation 272528.

• Frugal AI is strategically distinct from inferior AI. Optimization at hardware, model, and inference layers enables AI at 1–2 cents per transaction, unlocking markets invisible to Western enterprise AI 2. Small language models tuned to local contexts—agriculture in Swahili, governance in Telugu—outperform general-purpose LLMs for most developing-world applications 2930. The "bazaar" of networked small agents will outcompete centralized mega-models in the long run 5.

• Post-quantum cryptography is an urgent, non-optional DPI concern. Harvest-now-decrypt-later attacks mean current cryptographic identities (RSA, elliptic curve) are vulnerable within 10–20 years. Adoption of NIST-standardized lattice-based cryptography should begin immediately across identity infrastructure 27.

---

Recurring Themes

• Trust as the binding constraint on speed. Across at least a dozen sessions spanning disaster management, agriculture, health, financial inclusion, and identity, speakers independently arrived at the same conclusion: the critical gap is not compute or model quality but citizen trust, civil society confidence, and institutional credibility. Trust is built through transparency, participation, and demonstrated results—not through education campaigns or policy documents 2531323334. Failed government AI erodes trust across all digital services simultaneously 7.

• The human factor cannot be automated away. Whether framed as "AI + HI" hybrid governance 30, the irreducible role of frontline workers in agricultural AI adoption 93521, the non-delegability of ethical judgment in public services 3033, or the warning that 26% of implementers understand ethical frameworks 30, speakers consistently rejected the notion that institutional capacity-building is secondary to technical deployment. Change management, workforce retraining, and addressing job security fears are prerequisites for adoption, not afterthoughts 1310.

• The DPI playbook is India's most exportable asset. MOSIP is live in 35 countries 3; the agricultural DPI model (Mahavistar's 2.5 million users, planned Bharat Vistar national expansion) is explicitly designed for South-South replication 22; the UPI model for payments is being adapted in Indonesia and across the Global South 36. Speakers from Africa, APAC, and Latin America repeatedly cited India's DPI experience—not its AI models—as the primary reference architecture 243735.

• South-South cooperation over waiting for global consensus. Rather than awaiting UN consensus or defaulting to US-China framework choices, speakers from multiple regions called for active bilateral and regional cooperation: Smart Africa's AI Council, BRICS AI task forces, India-Brazil digital partnerships, and structured peer-learning pathways that transfer implementation knowledge without requiring full infrastructure rewrites 437357. India's linguistic and agroclimatic diversity positions it as a natural convener because its solutions already carry Global South relevance.

• Speed mismatches between technology and governance create systemic risk. The observation that technology advancement is outpacing institutional and social capacity was made across AI safety 266, health 10, identity 27, education 38, and enterprise AI 39 sessions. The response is not to slow deployment but to build agile governance institutions, pre-position safety standards, and treat data governance as a design-stage requirement rather than a compliance layer added afterward 6231.

---

Open Challenges & Tensions

• Inclusion versus speed in scaling. There is genuine tension between the urgency arguments (millions die without better health and agriculture systems, children lack personalized education) and the inclusion-by-design arguments (rushing to deploy without co-design, local language support, and offline capability replicates existing inequalities at AI scale) 242322. No speaker resolved this cleanly. The closest thing to consensus was "start with the last-mile user as the design constraint, not as an afterthought"—but this slows deployment timelines in ways that urgency advocates resist.

• Federated versus centralized data architecture. The energy 12, agriculture 219, and health 10 sectors all face the same unresolved architectural question: how to preserve departmental or state-level data ownership while creating unified intelligence that requires cross-system data flows. Federated learning, edge AI, and consent-based data exchange (DEPA model) are proposed solutions, but implementation at the scale of 28 Indian states with heterogeneous legacy systems remains largely unproven 116.

• Who governs the AI layer atop public DPI? Several speakers flagged that India has built the DPI foundations but has not yet designed the "intelligence layer"—where AI models operate, who manages them, who audits them, and who is accountable when they fail 81. This is the open governance question. The risk is that private sector actors fill the vacuum with proprietary, non-auditable systems built atop public infrastructure, recreating the very concentration that DPI was designed to prevent.

• Compute sovereignty versus regional cooperation. There is an unresolved debate between building dedicated national AI compute infrastructure (to avoid dependency) and pooling compute regionally to achieve economies of scale and reduce per-country costs 40. Both positions have merit and neither is obviously correct; the decision involves political economy as much as technical architecture. Demand guarantees from governments could bridge the gap, but no clear multilateral mechanism for this exists yet.

• Standards setting in a fast-moving environment. The window for establishing open, interoperable AI standards is closing as frontier model companies establish de facto proprietary standards through sheer deployment scale 1626. Formal standards bodies move more slowly than industry; the process recommendations (standardizing how organizations identify and manage risk rather than prescribing specific controls) may be the right long-term answer but offer little protection in the near term 26. India's position as a significant voice in ITU, ISO, and bilateral standard-setting forums is an asset that speakers agreed is currently underutilized.

---

Notable Examples

• Aadhaar and UIDAI's innovation ecosystem. UIDAI has enrolled 1.4+ billion people, opened a sandbox for 20+ entities, launched the SITHA grant program for academic research, and run hackathons targeting liveness detection, de-duplication, and presentation attack detection. The system's security architecture (liveness detection, renewable biometrics, cryptographic derivation) was designed in 2009–2010 rather than patched after breaches—cited as the model for AI security-by-design 2728.

• Mahavistar and Bharat Vistar agricultural DPI. Maharashtra's Mahavistar platform has reached 2.5 million users by combining farmer IDs, data exchanges, and shared protocols in an open, interoperable architecture. Its planned national expansion as Bharat Vistar is explicitly designed to be replicated in Africa, Southeast Asia, and Latin America—with India's linguistic and agroclimatic diversity serving as a built-in stress test for Global South conditions 22.

• MOSIP's 35-country deployment. The Modular Open Source Identity Platform, developed in India, is now operational in 35 countries as open, IP-free digital identity infrastructure. Cited repeatedly as evidence that India's DPI approach creates genuine alternatives to proprietary identity systems—and that the model can scale without creating new dependencies 318.

• Negotiate COP: open-source AI for climate diplomacy. Funded under the EU's NextGen EU program, Negotiate COP provides small delegations with real-time AI-assisted access to negotiation baselines and document synthesis, narrowing the information asymmetry between large and small country delegations at COP. Cited as a proof-of-concept for the digital public goods model in high-stakes multilateral settings 41.

• TGDEX and e-sahamati as data-sharing models. Telangana's TGDEX platform hosts 1,100+ government datasets as anonymized open data, enabling startups to build applications without rebuilding data infrastructure from scratch. The e-sahamati model for consent-based financial data sharing demonstrates that data portability can be architecturally enforced rather than left to goodwill—with direct relevance to agricultural, health, and education data ecosystems 36.