Type: Explained

Why in the News?

India has made its first major push into foundational AI model training by releasing domestically developed 35B and 105B parameter LLMs using subsidised Graphics Processing Unit (GPU) infrastructure under the IndiaAI Mission. With over 36,000 GPUs commissioned and 4,096 allocated to select firms, the move marks a strategic shift from dependence on foreign frontier models to state-supported indigenous AI capability.

Why Is Training Large Language Models on Indian Soil Financially and Logistically Challenging?

1. GPU Dependence: Requires high-end Graphics Processing Units for model training and inference; combined hardware and electricity costs run into millions of dollars.
2. Electricity Intensity: Compute-heavy training increases power consumption and operational expenses.
3. Capital Requirements: Large upfront investment limits private-sector experimentation in foundational AI.
4. Data Constraints: Internet training corpora disproportionately represent English and European languages.
5. Token Inefficiency: Indian language tasks require more tokens due to translation layers, increasing inference cost.

How Has the IndiaAI Mission Lowered Entry Barriers for Domestic AI Firms?

1. Public Compute Infrastructure: Commissioned 36,000+ GPUs in domestic data centres operated by firms such as Yotta.
2. Cluster Allocation: Provided 4,096 GPUs through a shared government compute facility.
3. Subsidised Access: Enabled startups and researchers to train and deploy models at relatively nominal fees.
4. Institutional Facilitation: Ministry of Electronics and Information Technology supports long-term indigenous AI capacity.
5. Ecosystem Development: Encourages domestic research, experimentation, and AI entrepreneurship.

How Does the Mixture of Experts (MoE) Architecture Improve Cost Efficiency in Model Deployment?

1. Selective Activation: Activates only a fraction of parameters during inference rather than the full network.
2. Compute Reduction: Lowers electricity consumption compared to dense models.
3. Inference Efficiency: Enables large models such as 105B parameters to run at lower operational cost.
4. Scalable Design: Allows domestic firms to optimise performance without matching trillion-parameter scale.
5. Cost Competitiveness: Enhances feasibility of AI deployment in education, healthcare, and governance contexts.

Does Parameter Size Alone Determine Strategic AI Capability?

1. Model Scale: Domestic models at 35B and 105B parameters remain smaller than global frontier systems.
2. Contextual Alignment: Designed for Indian languages and domestic sectoral use.
3. Sector-Specific Model: A 17B multilingual model developed for education and healthcare applications.
4. Incremental Scaling Strategy: Prioritises contextual performance before expanding model size.
5. Capability Gap: Comparative benchmarking with frontier systems remains limited.

How Does Linguistic Data Imbalance Affect Digital Inclusion?

1. Language Dominance: English and European languages dominate global internet datasets.
2. Indian Language Underrepresentation: Limits model accuracy in vernacular contexts.
3. Translation Dependence: Machine translation remains inferior to native-language modelling.
4. Governance Impact: Weak vernacular performance may affect citizen-facing digital services.
5. Inclusion Objective: Indigenous LLMs aim to strengthen equitable AI access.

What Transparency and Accountability Concerns Arise from Publicly Funded AI Infrastructure?

1. Open-Source Ambiguity: Models described as open but not fully accessible on major global platforms.
2. Limited Independent Scrutiny: Restricted external evaluation affects benchmarking.
3. Public Investment Oversight: Large-scale GPU subsidies require measurable performance assessment.
4. Benchmark Transparency: Absence of publicly standardised comparison metrics.
5. Energy Governance: Limited disclosure of sustainability audits for compute-intensive infrastructure.

Way Forward: Strengthening Indigenous AI Capacity

1. Transparent Benchmarking: Establishes clear performance metrics for publicly funded LLMs against global standards to ensure accountability.
2. Green Compute Standards: Mandates energy-efficiency norms and renewable integration for GPU-intensive data centres.
3. Vernacular Data Expansion: Builds high-quality Indian language datasets through public–private collaboration.
4. Outcome-Linked Subsidy: Links GPU allocation and funding to measurable innovation and adoption outcomes.
5. Regulatory Framework: Defines standards for data governance, algorithmic transparency, and institutional accountability.

Conclusion

India’s entry into foundational LLM training marks a shift from AI consumption to domestic capability creation. Public compute subsidies under the IndiaAI Mission reduce entry barriers but require transparent benchmarking, fiscal oversight, and sustainability safeguards. Long-term competitiveness will depend on strengthening vernacular data ecosystems, improving cost-efficient architectures, and institutionalising regulatory accountability.

PYQ Relevance

[UPSC 2023] Introduce the concept of Artificial Intelligence (AI). How does AI help clinical diagnosis? Do you perceive any threat to privacy of the individual in the use of AI in healthcare?

Linkage: Indigenous LLM development strengthens AI capability for governance and sectoral applications such as healthcare diagnostics. It simultaneously raises concerns of data protection, algorithmic transparency, and privacy, core issues highlighted in the 2023 AI question.

Why in the News?

GPT-4 introduced a new design that activates only selected parts of its system for specific tasks, similar to how the human brain works. At the same time, AI models are now approaching the brain in scale but consume far more energy. This contrast between similar size and very different efficiency has made the AI-brain comparison a major policy and technological issue.

How does the scale convergence between AI models and the human brain raise governance and infrastructure challenges?

1. Parameter Expansion: GPT-3 contains 175 billion parameters; newer models approach trillions, nearing the brain’s ~100 trillion synapses. Scale increases computational dependency and infrastructure concentration.
2. Data Centre Energy Demand: Training and operating large AI models require megawatts of electricity. Ensures rising carbon footprint and grid stress.
3. Hardware Dependence: AI training relies on high-performance GPUs originally developed for video gaming. Strengthens semiconductor concentration risks.
4. Digital Infrastructure Concentration: Massive parallel computation requires clustered data centres. Facilitates market dominance by few global technology firms.
5. Strategic Autonomy Concern: Nations lacking advanced chip fabrication capacity face technological dependence. Impacts India’s semiconductor mission and AI self-reliance goals.

In what ways does mixture-of-experts architecture influence regulatory and accountability frameworks?

Mixture-of-Experts (MoE) is a type of Artificial Intelligence model design where: instead of using the entire neural network for every task and the system activates only a few specialised parts (“experts”) for each input.

1. Selective Activation: GPT-4 activates specialised network portions for specific tasks. Enhances computational efficiency but complicates traceability
2. Modular Processing: Resembles the brain’s region-specific activation (language, vision, movement). Raises issues of explainability in AI outputs.
3. Sparse Routing Mechanism: Routes input through selected pathways rather than full network. Challenges transparency audits.
4. Task-Based Resource Allocation: Adjusts computational effort based on difficulty. Requires regulatory standards for algorithmic accountability.
5. Governance Implication: Fragmented internal processing complicates liability assignment in AI-generated harms.

Why does energy efficiency disparity between AI and the human brain matter for sustainability policy?

1. Metabolic Efficiency: Human brain operates at ~20 watts of power. Demonstrates biological optimisation.
2. Event-Driven Signalling: Biological neurons activate selectively and sparsely. Conserves energy.
3. Digital Arithmetic Dependence: AI systems perform continuous high-precision computation. Increases electricity consumption.
4. Carbon Footprint Risk: Large-scale AI training elevates emissions through energy-intensive data centres.
5. Green AI Imperative: Necessitates energy-efficient chip design, including neuromorphic hardware and spike-like operations.

How do differences in feedback mechanisms and learning processes impact ethical and institutional oversight?

1. Deep Feedback Loops: Brain processes signals forward, backward, and laterally. Enables contextual interpretation.
2. Contextual Meaning Formation: Human cognition integrates prior knowledge. Reduces rigid output behaviour.
3. Feed-Forward Architecture: Most LLMs rely on stacked layers without true recurrence. Limits adaptive contextual reasoning.
4. Statistical Learning Model: AI identifies probabilistic patterns from text corpora. Does not “understand” meaning intrinsically.
5. Regulatory Concern: Absence of embodied cognition raises risks of hallucinations, misinformation, and biased outputs.

What are the implications of AI’s divergence from biological intelligence for public policy and strategic planning?

1. Non-Biological Scaling: Machines are not constrained by evolutionary limits. Enables rapid parameter expansion.
2. Super-Computational Potential: AI may surpass humans in speed and pattern recognition.
3. Efficiency Trade-off: AI sacrifices energy efficiency for computational speed.
4. Neuromorphic Research: Attempts to mimic spike-based operations to reduce power usage.
5. Policy Imperative: Requires anticipatory regulation balancing innovation and risk mitigation.

How does AI’s hardware dependency influence economic concentration and digital sovereignty?

1. GPU Dominance: AI training dependent on limited global chip manufacturers.
2. Capital Intensity: High infrastructure cost restricts entry to large corporations.
3. Data Concentration: Models trained on massive datasets inaccessible to smaller players.
4. Regulatory Challenge: Ensures competition law scrutiny in AI markets.
5. National Security Dimension: AI capability linked to defence, cyber security, and economic competitiveness.

Conclusion 

AI is approaching the human brain in scale but remains fundamentally different in design and efficiency. While the brain operates with minimal energy and deep contextual feedback, AI depends on massive computation and data infrastructure.

The key policy challenge lies in balancing innovation with sustainability, accountability, and digital sovereignty. Future AI development must focus not just on scale, but on efficiency, transparency, and alignment with human values.

PYQ Relevance

[UPSC 2023] Introduce the concept of Artificial Intelligence (AI). How does AI help clinical diagnosis? Do you perceive any threat to privacy of the individual in the use of AI in healthcare?

Linkage: Directly linked to GS-3 (Science & Technology) under AI applications and data governance, and GS-4 (Ethics) regarding privacy, accountability, and algorithmic decision-making. The AI-brain debate strengthens this theme by highlighting efficiency, bias, and regulatory concerns in healthcare systems.