White Paper: Democratising Access to AI Infrastructure

The paper belongs to the PSA's White Paper Series, which the Office describes as explanatory briefs on specific policy issues, drawn from multi-stakeholder consultations, bilateral and multilateral AI policy engagements and expert review. The series says the papers are intended solely as explanatory documents and should not be construed as formal policy positions of the Office.¹ This paper was prepared by Animesh Jain of the Office, with advice and review from Dr. Preeti Banzal of the Office, Debjani Ghosh of NITI Aayog and colleagues from the EkStep Foundation.¹

The PSA publications page lists three white papers in the series: Advancing Indigenous Foundation Models (March 2026), Democratising Access to AI Infrastructure (December 2025) and Strengthening AI Governance Through Techno-Legal Framework (January 2026).² The date printed on this paper, 29 December 2025, matches that listing. A Pacta source repository gives the date as December 2026, which is after today's date and is not supported by either document.

The paper defines democratising access to AI infrastructure as making compute, datasets and the model ecosystem available and affordable so that a wide set of users can benefit from AI. It says these resources are currently concentrated in a handful of global firms and urban hubs, which limits equitable participation, and argues that India should treat them as shared resources.¹ It points to the India AI Governance Guidelines, which name three enablers of AI adoption: expanding access to high-quality and representative datasets, providing affordable and reliable access to computing resources, and integrating AI with digital public infrastructure. It also cites the United States National AI Research Resource as a federated example.¹

The paper's central proposal is to treat the building blocks of AI as digital public goods, positioned as shared public utilities and not proprietary assets. It recalls that the 2018 National Strategy for AI proposed an AI commons, a shared public digital infrastructure with standardised datasets, tools and platforms.¹

The paper is a policy brief, so its recommendations are framed as considerations, not obligations.

A DPI approach for AI. The paper says DPI for AI should be understood as a set of modular public-good enablers addressing specific coordination gaps, not as a single platform. It should evolve in phases. Early phases would use lighter elements such as directories, metadata standards, access protocols and registries. Later phases could add federated data access systems, consent-based data flows or coordinated compute-exchange mechanisms. A key next step is to examine technical pathways to an integrated stack connecting compute and data across distributed providers.¹

Limits of the approach. It depends on consistent technical standards, high-quality metadata and sustained institutional capacity for governance, oversight and auditability across multiple custodians. Privacy safeguards, cybersecurity and clear accountability matter more as access scales, and badly designed access layers could create new bottlenecks or exclusion for smaller institutions. The paper adds that democratisation cannot come from state-owned infrastructure alone.¹

Considerations. The paper highlights public-private partnerships for regional AI data centres, with edge facilities planned in Jaipur, Coimbatore and Chandigarh; incentives for private entities to contribute to AIKosh, the Open Government Data Platform and the National Data and Analytics Platform; resource efficiency, since scaling AI data centres needs an additional 45 to 50 million square feet of real estate by 2030 and data centres use about 0.5% of India's electricity, which could approach 3% by 2030; and sectoral adoption, which remains uneven in agriculture, education, healthcare and public services.¹

Much of the paper describes the current landscape. On data centres, India hosts nearly 20% of the world's data but only about 3% of global data centre capacity, with about 960 MW installed and 9.2 GW expected by 2030. Mumbai and Navi Mumbai hold over 25% of live capacity.¹ On compute, the paper says the IndiaAI Mission is building a secure cluster of 3,000 next-generation GPUs for sovereign and strategic applications, and that the IndiaAI Compute Portal operates over 38,000 GPUs and 1,050 TPUs at subsidised rates below Rs. 100 per hour, compared with global rates above Rs. 200. It also credits the National Supercomputing Mission and the India Semiconductor Mission (Rs. 76,000 crore outlay, with 10 approved chip projects).¹

On data, it describes AIKosh as a national repository of AI datasets, models and tools with permission-based access: as of December 2025 it had onboarded 5,722 datasets and 251 models from 54 entities across 20 sectors. It also covers Bhashini, the Geospatial Data Sharing Interface and the Telangana Data Exchange (TGDeX), which aims to create 2,000 AI-ready datasets in five years. On models, it notes that eight startups are being supported to develop sovereign models, with credits and funding covering up to 25% of compute costs through a mix of grants (40%) and equity (60%).¹

The paper also treats legal enablers, citing the Government Open Data Licence - India under the 2012 National Data Sharing and Accessibility Policy, the DPDP Act, 2023 for personal data, and state data centre policies in Maharashtra, Tamil Nadu and Karnataka.¹

As an explanatory brief, the paper does not create programmes or deadlines. It describes measures already under way, mainly the IndiaAI Compute Portal, AIKosh and TGDeX, and suggests a direction for further work.¹ The follow-up the paper itself identifies is technical: examining pathways to an integrated architecture that unifies compute and data across providers, in modular phases.¹ The sources reviewed for this entry did not show any government decision adopting the DPI-for-AI proposal.

As of 15 October 2026 the paper remains published on the PSA website as Version 3.0 dated 29 December 2025.² It was followed by the PSA's white papers on techno-legal AI governance (January 2026) and on indigenous foundation models (March 2026), and a September 2026 discussion paper on cross-border data interoperability for AI.² This entry could not verify whether a later version exists or whether any ministry has acted on the DPI-for-AI proposal.