India - AI Research Infrastructure

Overview

AIRAWAT, set out in an approach paper prepared by NITI Aayog, proposes a national AI-specific cloud computing infrastructure designed to supply India’s researchers, startups and public-good projects with specialised compute, storage and orchestration capabilities required for modern ML/DL workflows. The document explains why conventional HPC systems are not sufficient for many AI workloads, and why national-scale, GPU/TPU-enabled infrastructure can accelerate research, reduce costs, and strengthen data sovereignty. The approach paper benchmarks leading global systems (for example, Japan’s ABCI and U.S. supercomputing efforts) and outlines a modular, multi-tenant architecture, governance arrangements, financial models and phased implementation steps. The original approach paper is published by NITI Aayog and can be accessed at AIRAWAT: Establishing an AI Specific Cloud Computing Infrastructure for India (NITI Aayog).

Definitions

This approach paper defines AIRAWAT as an "AI-specific cloud computing infrastructure" that provides: (a) specialised compute (GPU/TPU-based nodes) for model training and inference; (b) multi-layer storage for high-throughput reads and bulk archival retention; (c) orchestration and resource partitioning to support multi-tenancy; and (d) data and model management services for sharing, reproducibility and governance. Key terms include "host institute" (the research or academic body that hosts hardware), "system integrator" (entity responsible for procurement, deployment and maintenance), "multi-tenant resource partitioning" (software/firmware-based isolation among projects), and "public-good workloads" (priority sector projects that receive concessional access).

Governance and Institutional Framework

The approach paper proposes a layered governance framework combining strategic oversight by NITI Aayog and a cross-ministerial Steering Committee with operational management by host institutes and a contracted system integrator. It recommends clearly delineated responsibilities: strategic policy and prioritisation by NITI Aayog, implementation and operations oversight by the host institution and systems integrator, and an independent technical advisory group for standards and benchmarking. The model emphasises transparent user-selection criteria, an allocation policy for compute resources, and mechanisms for stakeholder consultation. For implementation and policy alignment, the paper recommends coordination with central agencies including the Ministry of Electronics and Information Technology and domain ministries; see NITI Aayog's publication at NITI Aayog: AIRAWAT working paper page for governance detail.

Key Focus Areas

AIRAWAT identifies multiple technical and policy focus areas: architecture and hardware selection (GPU/TPU choice, networking, storage hierarchy), software tooling (orchestrators, container runtimes, ML lifecycle platforms), multi-tenancy and quota management, data ingestion and management pipelines, reproducibility and experiment tracking, security and privacy controls, and financial sustainability models. The paper also highlights sectoral focus for early prioritisation: healthcare (medical imaging, diagnostics), agriculture (precision agri and yield prediction), education (adaptive learning), smart cities and mobility (traffic modelling), climate and weather forecasting, and financial inclusion/fraud detection. Complementary operational matters addressed include user onboarding processes, monitoring and telemetry, SLAs for availability and maintenance, and policies for backups and disaster recovery. The approach stresses that an indigenous platform reduces dependence on commercial hyperscalers and can help enforce data sovereignty and localisation where necessary.

Implementation Framework

The recommended implementation follows a phased and modular model: a pilot phase to validate architecture and governance; a scale-up phase adding capacity and geographic distribution; and a mature phase offering federated regional nodes. Key operational recommendations include appointing a systems integrator under a clear AMC (annual maintenance contract), defining host-institute responsibilities (power, cooling, safety, physical security), and establishing APIs and SLAs for containerised workloads. The paper recommends a mixture of capital funding and operational charges, with concessional or subsidised access for academic research and socially important projects. It also proposes capacity partitioning policies to balance commercial and public-good uses while maintaining transparent selection and cost-recovery mechanisms.

Monitoring and Evaluation

AIRAWAT proposes continuous monitoring of technical and programmatic metrics: uptime and availability, job throughput, GPU-hours consumed, storage utilisation, energy efficiency (PUE), number of research outputs supported (papers, prototypes), startup engagements, and impact in priority sectors. It recommends establishing a reporting cadence to the Steering Committee, independent audits for procurement and operations, and periodic public reporting on allocation, usage and outcomes to ensure accountability and to attract stakeholder trust. The approach also supports building telemetry and observability for security incident detection and forensic analysis.

Penalties, Liability, and Appeals

As an approach document rather than primary statute, AIRAWAT does not itself create criminal penalties; instead it defines contractual and administrative remedies: performance-based contract provisions with the systems integrator (including liquidated damages for SLA breaches), termination clauses for misuse of the facility, and indemnity obligations for users. The paper recommends transparent appeal mechanisms for allocation disputes and a grievance redressal process managed by the governance board. It also highlights liability allocation for data loss or operational failures within contractual agreements between host institutes, system integrators and users.

Relationship to Other Instruments

AIRAWAT is proposed as a practical implementation instrument arising from the National Strategy for Artificial Intelligence (NSAI, NITI Aayog, 2018). It is intended to align with existing governmental initiatives on digital infrastructure, data governance, and research funding. The approach paper recommends compatibility with national data protection norms and operational coordination with the Ministry of Electronics and Information Technology and agencies running national portals and research infrastructure. The NITI approach paper should be read alongside other policy documents on Responsible AI and sectoral strategies for health, agriculture and education.

International Alignment

The paper benchmarks AIRAWAT against international high-performance AI facilities (for example ABCI in Japan) and suggests adopting international best practices in hardware selection, energy efficiency, and open standards for interoperability. It recommends collaborations and learning exchanges with international research consortia and supercomputing centers while prioritising domestic capability development. The approach emphasises the need for standards that allow portability of workloads and reproducibility of experiments to support international research collaboration while safeguarding data where national policies require localisation.

Implementation Timeline

Sources and References