ISAIL AI Recommendations on AI Innovation, and Budgetary Considerations in India for 2025

The ISAIL Secretariat under the leadership of Abhivardhan, Chairperson and Akash Manwani, Secretary-General, issues an Official Statement on the key technological developments surrounding NVIDIA, DeepSeek and OpenAI, to the Government of India as a set of Open Recommendations.

The rapid evolution of AI systems like DeepSeek’s R1 (an open-source, high-efficiency model) and Kimi AI (long-context processing) highlights both the opportunities and vulnerabilities in India’s AI trajectory. Emerging insights into DeepSeek’s compute infrastructure (~10,000–50,000 GPUs), its impact on global semiconductor markets (e.g., NVIDIA shorting), and the strategic ambiguities of export controls (e.g., H20 chip availability) demand a recalibrated approach to India’s AI policy. Aligned with the Durgapur AI Principles—strategic autonomy, inclusive innovation, and ethical sovereignty—ISAIL.IN proposes the following recommendations.

Key Insights from Recent Developments

1. Compute Multipolarity:

   
   

   • DeepSeek’s open-source R1 model, despite limitations akin to OpenAI’s o1, offers India a cost-effective foundation for last-mile AI solutions. However, its reliance on massive GPU clusters (~10k–50k) underscores India’s dependency on foreign compute infrastructure.

   • H20 Chip Dynamics: While restricted for training, the H20 remains viable for deployment (e.g., long-context inference, synthetic data generation). Post-2024 high-bandwidth memory (HBM) export bans make its continued availability critical, but reliance on smuggled/loophole-driven imports is unsustainable.

     
     

2. Tech Multipolarity in Action:

   
   

   • China’s DeepSeek—a “side-project turned global player”—exemplifies how open-source AI can democratize access, indirectly aiding India’s AI Mission by lowering entry barriers. However, this does not negate the need for sovereign compute.

     
     

3. Deployment Compute as Strategic Lever:

   
   

   • Models like R1 and o1 are increasingly used to generate high-quality training data, creating a feedback loop where deployment infrastructure directly fuels R&D. India cannot afford to lag in this cycle.

     
     

4. Budgetary Realities:

   
   

   • While DeepSeek’s efficiency (e.g., 2B-parameter R1 matching larger models) reduces costs, India’s ₹10,000 crore AI budget remains insufficient for scaling infrastructure. Strategic reallocation is essential.

Recommendations

Strategic Affordability Through Open-Source Leverage

DeepSeek’s R1 model—trained at a fraction of the cost of comparable Western models—demonstrates that algorithmic innovation can offset hardware limitations. For India, this signals a path to:

• Last-mile AI solutions: Deploy lightweight, domain-specific models (e.g., agricultural diagnostics, vernacular education tools) using R1’s architecture as a template.

• Cost containment: Focus on fine-tuning and deployment rather than full-stack model training, reducing reliance on high-end GPUs.

• Budget optimisation: Allocate saved funds to critical gaps—Indic-language datasets, AI safety research, and compute-sharing pacts with Global South partners.

This approach does not negate the need for increased AI funding but redirects focus from raw compute acquisition to strategic capability-building.

Avoiding the "Wrapper Economy" Trap

While DeepSeek’s open-source model enables cost-efficient innovation, India must resist the temptation to prioritise superficial "wrapper" applications built atop foreign foundation models. Such an approach:

• Perpetuates dependency: Reliance on external model providers risks digital colonialism, as seen in India’s past struggles with proprietary software ecosystems.

• Diverts talent: Encourages engineers to chase low-value API integrations rather than core R&D in areas like MoE architectures or FP8 optimization.

• Wastes public funds: Government-backed wrapper projects would duplicate private-sector efforts.

Instead, the IndiaAI Mission should:

• Incentivize foundational work: Grants for Bharat-specific model families.

• Mandate open weights for public-funded models: Ensure transparency and prevent vendor lock-in, as demonstrated by Aadhaar’s success.

• Leave wrappers to startups: Allow private players to commercialise applications while maintaining basic standards for data sovereignty.

Legal & Regulatory Upgrades

• Classify AI malpractice (e.g., diagnostic hallucinations) as “digital negligence” under Section 43A of the Information Technology Act, 2000, with fines up to ₹50 lakh. Ensure adjudication via existing Cyber Appellate Tribunal.

• Require healthcare/education AI tools to undergo DPDP-compliant audits for consent management and data anonymisation.

Upskill for AI Autonomy

• Integrate AI craftsmanship (model optimization, synthetic data generation) into the National Education Policy, focusing on GPU-agnostic techniques to mitigate supply-chain risks.

Export Control Realism

• Assume semiconductor smuggling will persist; design AI infrastructure with redundancy (e.g., hybrid cloud-edge systems) to mitigate sudden supply shocks. Collaborate with ASEAN/African nations to build “fallback supply chains.”

Deployment-Centric Innovation Zones

• Create AI Deployment Labs focused on memory-efficient inference, reinforcement learning, and test-time compute optimization—areas where India can lead without exascale GPU dependence. Prioritize use cases like Aadhaar-enabled services, weather prediction, and vernacular content.

Data Trusts with Security Safeguards

• Establish AI Data Trusts to pool anonymized data from critical sectors (health, agriculture), ensuring compliance with the Digital Personal Data Protection Act (DPDPA). Use synthetic data techniques to offset privacy risks.

ISAIL Leaders

Abhivardhan, Founder & Chairperson

Akash Manwani, Secretary-General

Mridutpal Bhattacharyya, Chief Policy Advisor

Sanjay Notani, President, Advisory Council of ISAIL

Bogdan Grigorescu, Vice-President, Advisory Council of ISAIL

Kailash Chauhan, Chairperson, AI Development Committee

Ayush Chandra, Chairperson, Policy Innovation Committee

Supratim Bapuli, Co-Chairperson, AI Development Committee

Kapil Naresh, Bhavana J & Sanad Arora, Special Executive Advisors