Strategic Studies National Security Defence Technology

AI in National Security: A Practitioner's View

Dipankar Sarkar
Dipankar Sarkar · · 5 min read

This is a practitioner’s view on AI in national security — written by someone who has built production AI systems for 18+ years and studied strategic studies at The Takshashila Institution. It is a synthesis of the Substrate Pattern, the Tiered Governance Model, the Federated Learning work, and the Indo-Pacific technology stack. It is also a candid assessment of where India stands.

The thesis

AI is the next general-purpose technology. The country that wins AI wins the next century. This is the framing used by the US National Security Commission on AI, the EU AI Act, the China AI 2.0 plan, and the IndiaAI Mission. The framing is correct. The implication is that AI is not just a commercial technology — it is a strategic one. The same way nuclear energy, semiconductors, and biotech are strategic. The country that loses AI loses the next century.

What AI does for national security

Five categories, in order of maturity:

  1. ISR (Intelligence, Surveillance, Reconnaissance). Satellite imagery analysis, signals intelligence, pattern-of-life analysis. The most mature AI-for-security use case. Both the US and China have deployed these at scale. India is behind but catching up.
  2. Cyber. Defensive and offensive cyber operations. AI is changing both. The most acute near-term risk: AI-generated phishing, deepfakes for social engineering, AI-enhanced vulnerability discovery. The most acute near-term opportunity: AI for threat detection, anomaly detection, automated incident response.
  3. Command and control. AI-assisted decision-making for military commanders. The “AI co-pilot” pattern, applied to warfighting. Both the US (Project Maven) and China are deploying these. India is in early stages.
  4. Autonomous systems. Drones, autonomous vehicles, autonomous weapons. The most controversial category. The killer robots debate is real but secondary to the operational reality: the US, China, Russia, Israel, and others are deploying autonomous systems today.
  5. Decision support at the strategic level. AI for strategic-level analysis, scenario planning, red-teaming. The most nascent category, but the most strategic. This is where I expect the most impact in the next 10 years.

Where India stands

India is in the middle. India has the talent (the IIT system produces world-class AI researchers), the data (the IndiaAI dataset initiative), and the demand (the Indian military is large and modernising). India lacks:

  • Indigenous compute. The IndiaAI compute cluster (18,693 GPUs) is a start but it is small relative to what is needed. India is still dependent on NVIDIA.
  • Indigenous models. The Indic language models (BharatGPT, Airavat) are in early stages. India is still dependent on US and Chinese foundation models.
  • Defence procurement speed. The Indian defence procurement system is slow. By the time an AI system is procured, the technology has moved. The iDEX (Innovations for Defence Excellence) programme is trying to fix this.
  • Talent retention. The best Indian AI researchers go to the US. The brain drain is real.

The catch-up is possible. The IndiaAI Mission, the iDEX programme, the National Quantum Mission, and the India Semiconductor Mission are all serious. But the time horizon is long. India is not winning the AI race today. India could win it by 2035 if the right bets are made now.

The four bets India should make

  1. Sovereign AI infrastructure. India needs its own training compute, its own inference infrastructure, and its own foundation models. The IndiaAI Mission is a start. The bet should be on sovereign AI for the public sector, defence, and strategic industries (semiconductors, biotech, finance, energy). Commercial AI can continue to use US and Chinese models.
  2. Defence-tech innovation. The iDEX programme should be expanded. The bet should be on dual-use startups — companies that can sell to both defence and commercial customers. The capital is there; the procurement is the bottleneck.
  3. Federated learning and privacy-preserving AI. For defence and intelligence, data sharing is critical but legally constrained. Federated learning, differential privacy, secure multi-party computation — these are the technologies that enable data sharing without compromising sovereignty. India should be a leader here, not a follower.
  4. AI safety and governance. India is well-positioned to lead the global conversation on AI safety and governance. The Tiered Governance Model (4 tiers, T0-T3) is one example. India should publish frameworks, lead in standards bodies, and build the regulatory infrastructure for the rest of the world.

The dual-use dilemma

The most acute strategic question: how does India balance the dual-use nature of AI? The same AI that improves logistics can improve targeting. The same AI that detects fraud can detect fraud in tax collection. The same AI that translates languages can generate disinformation.

The answer is not to suppress the dual-use. The answer is to build the governance infrastructure that makes the dual-use safe. Defence in depth, kill switches, audit logs, runtime policy gates — these are the technical primitives. The Tiered Governance Model is the policy primitive. The Substrate Pattern is the engineering primitive. Together, they are the answer.

What I am working on

I work on AI agent infrastructure. The agent runtime, safety, observability, and deployment layers. This work is dual-use by design — the same runtime that serves a financial-services customer can serve a defence customer. The Tiered Governance Model is the policy layer; the Substrate Pattern is the engineering layer; the runtime is the substrate.

My conviction: the strategic question for India is not “should we build AI?” (yes, obviously). It is “how do we build AI in a way that preserves sovereignty, ensures safety, and serves the national interest?” The frameworks I am building are my answer.

— Dipankar Sarkar IIT Delhi alumnus · Takshashila Institution alumnus · Founder, Neul Labs

This is a 2026 essay drawing on production AI work, 18+ years of building, and 8 years of strategic studies at Takshashila. The views here are my own.