India's AI summit: big signal, real wins, and critical gaps
The India AI Impact Summit (16-20 February 2026, New Delhi) was the largest AI gathering yet: 100,000+ attendees, 400+ sessions, and multiple heads of state. The message was clear: AI governance isn't a club for Western capitals; it's a public issue with Global South weight behind it.
The New Delhi Declaration on AI Impact was non-binding and aspirational, but 91 governments and international bodies signed on, including the US, China, and Russia. Unlike in Paris 2025, Washington stayed in the tent. The summit avoided military AI, continued the series trend, and India formally joined the US-led Pax Silica initiative on strategic tech, minerals, and semiconductors.
Big firms arrived with cash and deals. OpenAI and Anthropic announced major India partnerships, while Google's philanthropic arm committed US$60 million to public services and research. China sent no official delegation. The US team, led by Michael Kratsios, argued that "complete technological self-containment" is unrealistic-and that buying American hardware can itself be a form of sovereignty. With tighter US visa policies for Indian engineers, building capacity in India now makes more business sense.
What the summit got right
Delhi's tone was pragmatic. Government, industry, and civil society aligned on using AI to drive development-especially in healthcare, agriculture, and public services. For practitioners and policymakers, that's the right near-term focus. See also: AI for Government.
India leaned on genuine strengths in digital public infrastructure. Aadhaar's biometric ID and the Unified Payments Interface are proven, population-scale systems. For background, see UIDAI (Aadhaar).
Privacy was treated as a design constraint, not an afterthought. After the Supreme Court's Puttaswamy ruling established privacy as a fundamental right, the Digital Personal Data Protection Act (2023) set a consent-first, lighter-weight compliance approach rolling out through 2026-27. Official text: DPDP Act, 2023.
New rules kicked in on 20 February requiring permanent metadata and visible labels on AI-altered images and video-among the most comprehensive mandates anywhere. The open question: can platforms and tools enforce this at scale without breaking creative and enterprise workflows?
What the agenda missed
Agentic AI-the shift to systems that take multi-step actions with minimal supervision-was underplayed. That's a disconnect with how the industry has moved since late 2025. For hands-on resources, see AI Agents & Automation.
Tech leaders on stage compressed their own timelines for AI that can do "any cognitive work humans can do." If even roughly right, many voluntary commitments from Delhi will lag reality before they're implemented. The risk isn't theoretical: early agent deployments are already re-scoping jobs and product roadmaps.
Why this matters for India-and for builders everywhere
India is the world's largest exporter of IT services. That makes it directly exposed if a meaningful slice of coding, testing, integration, and support is automated by credible agents. We're already seeing stress in software stocks tied to partial automation threats.
There's a bigger pivot hiding in plain sight: software advantage won't be enough. Progress depends on the physical stack-energy, silicon, fabs, inference compute, robotics, and advanced manufacturing. That's where the next bottlenecks-and leverage-sit. Delhi emphasized software deployment and services; the harder questions are infrastructure, capacity, and supply-chain resilience before Geneva 2027.
Signals to watch
- Whether the New Delhi Declaration catalyzes concrete procurement, safety, or data-sharing projects across signatories-or stays aspirational.
- How India enforces permanent watermarking/metadata at internet scale without collateral damage to benign creative uses and open-source tooling.
- DPDP Act enforcement timelines and the regulator's posture on cross-border data flows, sensitive-sector carve-outs, and SME compliance.
- Actual capital formation in compute, energy, and semiconductor packaging versus MOU-heavy announcements.
- US-India tech ties under tighter US visa rules: do firms accelerate India-based engineering and data-center builds?
- Whether future summits finally address military AI or keep that track separate.
Action guide for governments, IT leaders, and developers
For governments
- Budget for the bottlenecks: grid upgrades, data-center power, domestic/ally fabs, advanced packaging, and inference clusters with clear access policies.
- Issue practical agentic-AI guidance: tool-use permissions, audit trails, escalation rules, and red-teaming standards for public deployments.
- Move from pilots to procurement in healthcare, agriculture, and service delivery. Tie funding to measurable outcomes and open benchmarks.
- Operationalize DPDP: consent flows, DPIAs for high-impact use cases, and streamlined SME compliance sandboxes.
- Make watermarking rules testable: reference codecs and open tests; phase enforcement; publish false-positive/negative targets.
For IT services and product companies
- Build an "agent-ops" stack: secure tool catalogs, least-privilege access, policy engines, and run-time monitors. Treat agents like junior staff with logs.
- Productize domain expertise: vertical agents for claims, loan ops, field service, QA, and DevOps. Ship outcome SLAs, not just hours.
- Diversify into hardware-adjacent services: model-serving optimization, inference cost engineering, on-prem clusters, and energy-aware scheduling.
- Automate your own delivery pipeline first: codegen, testgen, CI/CD approvals, L4 support triage, and contract review.
- Re-skill at scale: agent frameworks, tool-calling, retrieval, evals, and safety. Make every team "automation-native."
For developers and data teams
- Learn agent patterns: planning, tool-use, memory, and recovery. Prioritize determinism and observability over raw "wow."
- Ship guardrails: constrained functions, schema validation, HIPAA/GDPR/DPDP-aware data guards, and cost ceilings per task.
- Measure with evals that mirror jobs-to-be-done, not just benchmarks. Track regressions with every model update.
- Prepare for infra diversity: CPU/GPU mix, on-device inference, and regional deployment for data residency.
- Design for handoff: agents that escalate early with clean context save time and reduce risk.
The bottom line
Delhi proved AI is a public project, not a niche debate. But the hard work now shifts to agents, infrastructure, and enforcement-the places where progress stalls or compounds. Those who build capacity there will set the terms for everyone else.
Your membership also unlocks:
