How are Indian firms training LLMs?

UPSC Prelims + Mains Study Note | GS-III: Science & Technology / Economy

1. At a Glance

• Large Language Models (LLMs) are neural networks trained on vast text corpora to power generative AI services; training requires massive GPU clusters, electricity, and curated data — all scarce in India relative to the US/China. [1]
• The IndiaAI Mission (approved March 2024, ₹10,372 crore) is the primary state instrument enabling Indian firms to train LLMs domestically by subsidising compute access. [2]
• Sarvam AI (Bengaluru) is the flagship case study — it won a Government of India tender and released India's first sovereign foundational LLMs trained entirely on IndiaAI Mission compute in February 2026. [3]
• UPSC relevance: cuts across GS-III (Science & Tech, Economy), GS-II (Government Policy), and India's digital sovereignty / strategic autonomy discourse.

2. Why in the News

• AI Impact Summit, New Delhi (19 February 2026): Prime Minister Modi met Sarvam AI co-founder Pratyush Kumar at Bharat Mandapam. Sarvam released two open-sourced foundation LLMs — Sarvam 30B and Sarvam 105B — both trained exclusively on IndiaAI Mission infrastructure. [4]
• IndiaAI Mission's common compute capacity crossed 34,000 GPUs (later expanding to 38,000+), a milestone underscoring India's domestic AI infrastructure build-out. [1]
• April 2025: Government formally selected Sarvam AI to build India's sovereign LLM — capable of reasoning, designed for voice, fluent in Indian languages, intended for population-scale secure deployment. [3]

3. Background & Evolution

4. Core Static Facts

Implementing Body

• Ministry of Electronics and Information Technology (MeitY) — nodal ministry for IndiaAI Mission [2]
• IndiaAI (dedicated implementation entity under MeitY) manages compute portal, model development calls, and startup support [2]

Budget

• ₹10,372 crore (≈ $1.1 billion) approved by Union Cabinet, March 2024 [2]

Compute Infrastructure

• 38,000+ GPUs empanelled across 14 AI service providers [1]
• 1,050 TPUs also available [1]
• Subsidised rate: ₹65–₹100/hour vs global market rate >₹200/hour [1]
• Data centre locations: Mumbai, Navi Mumbai, Hyderabad, Bengaluru, Noida, Jamnagar [1]
• Eligible users: startups, researchers, academic institutions, government organisations [1]

Sarvam AI Models (Feb 2026)

• Sarvam 30B — 32-billion-parameter Mixture of Experts (MoE), 65K context window; speed-optimised (comparable to Gemini Flash-Lite / GPT-5 mini tier) [3][4]
• Sarvam 105B — 106-billion-parameter MoE, 128K context window; complex reasoning tasks (comparable to Gemini Pro / GPT-5 tier); wins ~90% of Indian-language benchmark comparisons vs GPT-4 [3][4]
• Both models: open-sourced, trained 100% on IndiaAI Mission infrastructure [3]

Key Terminology

• LLM: Large Language Model — neural network trained on billions of text tokens, foundation for generative AI
• Parameter: numerical weight in a neural network; more parameters ≈ greater model capacity
• Token: smallest unit of text processed by an LLM; Indian-language text requires more tokens than equivalent English, raising inference cost [4]
• MoE (Mixture of Experts): architecture where only a subset of model parameters ("experts") are activated per inference pass — lowers compute cost without proportional loss in capability [4]
• GPU (Graphics Processing Unit): primary hardware for LLM training/inference
• TPU (Tensor Processing Unit): Google-designed chip for ML workloads; available on IndiaAI portal

5. Multi-Dimensional Analysis

Economic

• India's domestic LLM training market was entirely dependent on foreign cloud providers (AWS, Google, Azure) pre-IndiaAI Mission; subsidised compute reduces this dependency. [1]
• ₹10,372 crore public investment aimed at catalysing private AI ecosystem; expected to reduce per-GPU-hour costs by 50%+ for Indian startups. [1][2]
• MoE architecture reduces training and inference costs substantially — enabling Indian firms with smaller capital bases to compete with global labs. [4]

Scientific / Technological

• Dual bottleneck for Indian LLMs: (a) scarcity of high-quality Indian-language training data (English/East Asian languages dominate internet corpora); (b) capital-intensive GPU clusters costing millions of dollars per training run. [4]
• Indian-language tokens are structurally less efficient: sentences must often be translated to English for inference, consuming extra tokens and raising costs. [4]
• MoE innovation addresses compute bottleneck; curated Indian-language datasets (being built under IndiaAI's Data Management Unit) address the data bottleneck. [2]
• Sarvam 105B achieving ~90% win-rate over GPT-4 on Indian-language benchmarks signals narrowing of the capability gap with frontier global models. [3]

Geopolitical / Strategic

• Digital sovereignty: training LLMs on Indian soil with Indian data reduces dependence on US/China AI stacks for critical government applications. [3]
• Sovereign LLM mandate explicitly targets "secure population-scale deployment" — implying use for government services (Aadhaar, health, welfare) without data leaving India. [3]
• Aligns with India's G20 AI principles (2023) and NITI Aayog's AI strategy emphasising "AI for all." [2]

Administrative / Governance

• IndiaAI Mission uses a competitive tender model — transparent procurement for sovereign LLM rather than directed allocation; Sarvam selected after open call. [3]
• 14 empanelled private AI service providers supply GPU capacity — public-private partnership model for infrastructure without full state ownership. [1]
• Eligible institutions include academic researchers and government bodies, not just startups — broad access design. [1]

Ethical / Governance

• Open-sourcing Sarvam models (30B, 105B) enables public scrutiny, reduces proprietary lock-in, and allows downstream adaptation by Indian researchers. [3]
• Risk: concentrated state selection of a single "sovereign LLM" vendor could create governance/monopoly concerns if procurement transparency lapses.

Social

• Indian-language capability of LLMs has direct equity implications: 90%+ of India's population is not English-dominant; English-centric LLMs exclude them from AI benefits. [4]
• Voice-first design of Sarvam's sovereign mandate targets low-literacy users — aligns with Digital India's last-mile inclusion goals. [3]

6. Recent Developments (Last 12–18 Months)

• April 2025: Sarvam AI formally selected by Government of India under IndiaAI Mission to build India's sovereign LLM. [3]
• September 2025: Government announced 500 AI Data Labs with ₹988 crore investment to boost AI capabilities. [5]
• February 19, 2026: PM Modi meets Sarvam co-founder Pratyush Kumar at AI Impact Summit, Bharat Mandapam, New Delhi. [4]
• February 26, 2026: Sarvam AI officially releases Sarvam 30B and Sarvam 105B — both open-sourced, trained on IndiaAI Mission compute. [4]
• 2026: IndiaAI common compute capacity crosses 38,000 GPUs and 1,050 TPUs; 20,000 additional GPUs under active procurement. [1]
• IndiaAI Mission expanded with additional startup support mechanisms and affordable compute extensions beyond initial mandate. [1]

7. Prelims Hooks

1. The IndiaAI Mission was approved by Union Cabinet in March 2024 with an outlay of ₹10,372 crore. [2]
2. Nodal ministry for IndiaAI Mission: Ministry of Electronics and Information Technology (MeitY). [2]
3. IndiaAI common compute portal provides GPUs at ₹65–₹100/hour vs global rates exceeding ₹200/hour. [1]
4. As of 2026, IndiaAI has empanelled 14 AI service providers offering 38,000+ GPUs and 1,050 TPUs. [1]
5. Sarvam AI is the Bengaluru-based startup selected by Government of India as the sovereign LLM developer under IndiaAI Mission. [3]
6. Sarvam's Sarvam 30B has 32 billion parameters with a 65K context window; Sarvam 105B has 106 billion parameters with a 128K context window. [3]
7. Both Sarvam models use Mixture of Experts (MoE) architecture — key reason they are less compute-intensive than comparable dense models. [4]
8. In MoE architecture, only a subset of parameters (experts) is activated per inference, reducing computational cost. [4]
9. Sarvam 105B wins approximately 90% of comparisons against GPT-4 on Indian-language benchmarks. [3]
10. Both Sarvam LLMs were trained entirely on IndiaAI Mission infrastructure and are open-sourced. [3]
11. IndiaAI data centres are located in: Mumbai, Navi Mumbai, Hyderabad, Bengaluru, Noida, and Jamnagar. [1]
12. Indian-language text costs more in LLM inference because it requires more tokens — including translation overhead to English. [4]
13. 500 AI Data Labs announced in September 2025 with ₹988 crore investment to expand India's AI data infrastructure. [5]
14. The AI Impact Summit where Sarvam models were unveiled was held at Bharat Mandapam, New Delhi, on 19 February 2026. [4]

8. Mains Relevance

GS Papers:

• GS-III: Science & Technology (AI/ML); Indian Economy (Start-up ecosystem, Digital Infrastructure)
• GS-II: Government Policies and Interventions; e-Governance

Syllabus Headings:

• GS-III: "Awareness in IT, Space, Computers, Robotics, Nano-technology, Bio-technology and issues relating to Intellectual Property Rights"
• GS-III: "Indian Economy and issues relating to growth, development and employment" (Digital Economy)
• GS-II: "Government policies and interventions for development in various sectors and issues arising out of their design and implementation"

Plausible Mains Questions:

1. "Critically examine the challenges faced by Indian firms in training Large Language Models domestically, and evaluate the role of the IndiaAI Mission in addressing them." (GS-III, 15 marks)
2. "The development of a sovereign LLM is not merely a technological achievement but a matter of strategic autonomy. Discuss in the context of India's IndiaAI Mission." (GS-III/GS-II, 15 marks)
3. "What is the Mixture of Experts (MoE) architecture, and why is it particularly significant for AI development in resource-constrained economies like India?" (GS-III, 10 marks)

9. Related Topics to Study Next

10. Common Errors / Trap Areas

1. Wrong ministry: IndiaAI Mission is under MeitY, NOT NITI Aayog (NITI Aayog authored earlier AI strategy papers but does not implement the Mission).
2. Parameter confusion: Sarvam 30B is described as "35 billion parameters" in some reports and "32 billion" in others — the MoE model has 32B active parameters but 35B total; exam questions will likely use the model name (30B/105B) rather than exact count.
3. MoE ≠ smaller model: Sarvam 105B has 106B total parameters — it is NOT a small model; MoE makes it cheaper to run because only a fraction of parameters activate per token, not because it has fewer parameters overall.
4. Confusing Sarvam AI with other Indian AI initiatives: Krutrim (Ola), Hanooman (SML India), BharatGPT are separate LLM projects — Sarvam is the one with the government sovereign LLM mandate.
5. Year of IndiaAI Mission approval: March 2024, not 2023 (the National AI Strategy/NITI Aayog paper was 2018; these are different documents separated by 6 years).