Verdict: India's sovereign AI strategy should not try to out-spend OpenAI and Google on trillion-parameter frontier models. Instead, it should build purpose-driven, right-sized models for specific government, enterprise, and linguistic use cases — leveraging the IndiaAI Mission's 38,000+ subsidised GPUs and a composite-AI approach that mixes small models, retrieval, and frontier APIs where needed. The evidence from early movers like CoRover (BharatGPT) and Sarvam AI shows that problem-first AI deployed at population scale creates more value than chasing a leaderboard ranking that shifts every quarter.
Last verified: 2026-08-01 · Primary keyword: India sovereign AI strategy · Best for sovereign AI: Sarvam AI (foundation models) · Best for enterprise deployment: CoRover/BharatGPT (composite AI) · Best for Indian-language voice: Bhashini
What Is India's Sovereign AI Strategy?
India's sovereign AI strategy is the national effort — anchored in the Rs 10,372 crore IndiaAI Mission approved by the Cabinet in March 2024 — to build domestic AI infrastructure, foundation models, and applications that keep sensitive data and compute inside India's borders while reducing dependence on US and Chinese AI platforms. Rather than attempting to clone GPT-5 or Claude from scratch, the strategy increasingly favours purpose-driven models trained on Indian-language data for governance, financial services, healthcare, and cybersecurity use cases.
The IndiaAI Mission's compute pillar has deployed more than 38,000 GPUs — nearly four times the original 10,000-GPU target — accessible to registered startups, researchers, and government agencies at a subsidised rate of approximately Rs 65 per GPU-hour, roughly 60% below commercial cloud rates. The mission also funds foundation model development: 12 organisations including Sarvam AI, BharatGen (IIT Bombay), Gnani AI, and Tech Mahindra Maker's Lab were selected to build large and small language models trained on Indian datasets.
Should India Try to Build a GPT-Class Frontier Model?
India should not attempt to build a frontier-class model from scratch in the near term, because the economics don't close. Training a single GPT-5-class model requires hundreds of millions of dollars in GPU compute, a team of elite researchers, and a sustained multi-year inference cost that can exceed $1 million per day. Even DeepSeek's widely cited $5.576 million training figure for DeepSeek-V3 (671B parameters, 37B active per token) covers only the final training run — excluding prior research, failed experiments, salaries, and infrastructure — and used 2,048 NVIDIA H800 GPUs that India does not currently possess in equivalent concentration.
That said, India has already proven it can build competent models for its own use cases. Sarvam AI's Sarvam-105B, trained from scratch with 4,096 government-allocated NVIDIA H100 GPUs, scores competitively on Indian-language tasks — the domain that actually matters for a country with 22 constitutionally recognised languages and over 700 spoken dialects. The key principle: build for the problem, not the parameter count.
The Composite-AI Approach (CoRover's Four-Layer Model)
CoRover, the company behind BharatGPT (India's first indigenous LLM), uses a composite-AI architecture that avoids using a large model for every query. This approach splits work across four layers:
| Layer | What it does | When it's used | Example |
|---|---|---|---|
| 1. Rule-based NLP | Intent classification, entity extraction | When deterministic answers suffice | "Check my account balance" -> API call, no model |
| 2. BharatGPT (RAG) | Grounded answers from enterprise data | When domain-specific knowledge is needed | "Can I carry a pet on the train?" -> retrieves from IRCTC docs |
| 3. Indian-language LLM | BharatGPT for Indic-language generation | When natural-language generation is needed in Indian languages | Citizen services in Hindi, Tamil, Telugu, etc. |
| 4. Frontier APIs (GPT, Gemini) | Complex reasoning, general knowledge | When the task exceeds small-model capability | "Where should I travel in August?" -> uses Gemini for reasoning + web search |
The insight: more than 90% of queries in enterprise and government deployments don't need a frontier model. By routing only the hard 10% to GPT or Gemini and handling the rest with small models or rule-based systems, you cut compute costs dramatically, keep data on-premise for sensitive queries, and avoid vendor lock-in.
How Much Does the IndiaAI Mission Actually Provide?
The IndiaAI Mission provides subsidised GPU compute, foundation model funding, dataset access, and application development grants. Here's what's been deployed as of mid-2026:
| Resource | Amount | Cost to access | Source |
|---|---|---|---|
| Total budget | Rs 10,372 crore ($1.25 billion) over 5 years | Government-funded | PIB India |
| GPUs deployed | 38,000+ (target: 100,000 by Dec 2026) | ~Rs 65/GPU-hour (subsidised) | ET Government |
| Foundation model builders | 12 selected (Sarvam, BharatGen, Gnani, etc.) | Full compute subsidy for model training | Mint |
| Sarvam allocation | 4,096 NVIDIA H100 GPUs + Rs 99 crore subsidy | Part of IndiaAI compute grant | ET Government |
| FY25-26 allocation spent | Rs 973 crore (committed: Rs 2,194 crore to 12 startups) | — | Mint |
For comparison, OpenAI is estimated to burn approximately $1 million per day on inference alone. India's entire FY25-26 AI Mission allocation of Rs 973 crore (~$117 million) would cover roughly 117 days of OpenAI's inference costs — illustrating why head-to-head frontier competition is not a viable strategy at current resource levels.
What Did Sam Altman Actually Say About India in 2023?
In June 2023, during an event in Delhi, OpenAI CEO Sam Altman was asked whether a small team of Indian engineers with $10 million could build something comparable to OpenAI. His response: "It's totally hopeless to compete with us on training foundation models."
The remark sparked outrage in India's tech community. By February 2026, Altman reversed course at the India AI Impact Summit, stating India was "well positioned to lead and shape AI," noting that 100 million Indians use ChatGPT weekly and calling India "the fastest growing market for Codex." The shift reflects both diplomatic pragmatism and the reality that India's sovereign AI stack has matured faster than expected — but it also validates the thesis that India's opportunity was never in head-to-head foundation model competition.
Can Indian Startups Compete Against Free Open-Weight Models?
Indian AI startups can compete against free open-weight models like Meta's Llama and China's DeepSeek by building vertical, problem-specific solutions that commodity models cannot match. The question isn't "can we build a better LLM?" but "can we build a better solution for this specific use case?"
CoRover's approach is instructive: when a client (Perfetti, via WPP) needed a gamified tongue-twister chatbot in Hindi, BharatGPT handled the Q&A but Gemini was used for the gamification logic because BharatGPT couldn't do it well. Rather than seeing this as a failure, CoRover treats it as evidence that composite AI — mixing your own models with frontier APIs — actually sells solutions better than either approach alone. The client pays for the solution, not for which model ran.
For practical guidance on cutting AI token costs while building these composite systems, see our guide on AI token cost optimization for enterprises.
What Happened to Drift (Cautionary Tale)
Drift, a billion-dollar-plus conversational AI company that CoRover's founder admired as inspiration, shut down in 2025 despite its valuation and technology. The lesson: building general-purpose conversational AI that competes head-on with the best in the world is extraordinarily difficult to sustain, regardless of funding. Purpose-first companies that solve specific problems for specific populations have better survival odds.
How Is India Using AI in Government at Scale?
India is already running some of the world's largest AI deployments across government systems. CoRover's BharatGPT-powered virtual assistants serve over 1 billion users across 14+ Indian languages on platforms including IRCTC (Indian Railways), LIC (Life Insurance Corporation), NPCI (National Payments Corporation of India), and the MyGov portal.
| Deployment | Platform | Scale | What it does |
|---|---|---|---|
| IRCTC Assistant | Indian Railways booking | Millions of queries/day | Ticket booking, PNR status, refunds in 14 languages |
| LIC Assistant | Life Insurance Corp | 290M+ policyholders | Policy queries, premium info, claim guidance |
| NPCI Assistant | National Payments Corp | UPI/BHIM ecosystem | Transaction help, dispute resolution |
| Port GPT | Major Ports | First in the world | Port operations, cargo tracking, customs queries |
| AskSarkar | Government services | Citizen-facing | Multi-department navigation in local languages |
These deployments don't use frontier models. They use small, purpose-tuned models with RAG over enterprise knowledge bases — proving that the "right-sized model" thesis isn't theoretical. It's already in production, serving real citizens. For a deeper look at how this fits India's broader enterprise AI strategy, see our analysis of India's enterprise AI skills gold rush.
Should India Build Cybersecurity AI Comparable to Anthropic's Models?
In July 2026, the Indian government directed Sarvam AI and the IIT Bombay-led BharatGen consortium to develop indigenous cybersecurity AI models, reportedly comparable to Anthropic's most advanced models, for critical infrastructure protection across defence, banking, telecommunications, and power sectors. This is an ambitious directive, but the framing matters: it's not about building a general-purpose "Claude clone" — it's about building specialised models for cyber threat detection, malware analysis, vulnerability assessment, and anomaly detection.
India's existing IT Act of 2000 already provides for up to 10 years' imprisonment for cybercrimes, and the government has argued that since most cyberattacks on Indian infrastructure originate from outside India, strict data localisation and sovereign AI deployment can reduce exposure. The real need isn't a Claude-class general model — it's a specialised model trained on Indian attack patterns, network topologies, and threat intelligence data that global frontier models don't have.
For more on India's broader critical-infrastructure defence posture, see How India is defending critical infrastructure from cyberattacks.
What Is the Right Investment Priority for India's AI Budget?
If allocating the next Rs 10,000 crore in sovereign AI spending, the highest-leverage move is funding AI adoption — not just model creation. The government has already begun this shift: under the National e-Governance Division (NeGD), six companies including CoRover have been selected to drive AI adoption across government departments, moving beyond infrastructure building to actual deployment.
The reasoning:
- Infrastructure is necessary but not sufficient. 38,000 GPUs and 12 foundation models don't create value until they're deployed to solve real problems.
- Adoption creates the data flywheel. Every deployment generates domain-specific interaction data that improves future models — data that no frontier model trained on the open internet has.
- Skills compound. Training engineers to build AI applications — not just models — creates the workforce that sustains the ecosystem. See how AICTE is already partnering with vendors on this front.
- Solutions attract repeat spend. Government and enterprise budgets flow toward tools that demonstrably cut costs or improve service — not toward parameter counts.
For a contrasting perspective on how resource-efficient training can still produce strong models, see why data quality is the compute multiplier AI builders overlook.
Is Intelligence Getting Cheaper — or Just the Models?
Intelligence is not getting cheaper; what's getting cheaper is the cost of running a model trained on publicly available internet data. The distinction matters because the next frontier of AI value — collective wisdom that lives in people's heads, in undocumented workflows, in cultural context — is not in any training corpus and cannot be scraped from the web.
CoRover's founder makes this point: if all intelligence were already on the internet, why do we interview people? Every generation contributes new intelligence. The questions that matter to India's 1.4 billion citizens — how to get a train ticket in a regional language, how to access a government scheme, how to resolve a banking dispute — are lived problems, not text problems. Models trained on internet text will always be one step removed from this reality.
This is why purpose-first AI, composite architectures, and deployment-at-population-scale matter more than the size of any single model. For more on emerging architectures that move beyond massive monolithic models, see our guide on Macaron-V1 and mixture-of-LoRA multi-specialist agents.
What This Means for You
If you're building AI products in or for India:
- Don't try to build a frontier model. Use the IndiaAI Mission's subsidised compute (Rs 65/GPU-hour) to fine-tune, not to chase the parameter race. Apply through the IndiaAI portal under MeitY — access is gated but the cost advantage over commercial cloud is real.
- Build composite-AI systems. Mix small models, RAG, rule-based NLP, and frontier APIs. Route each query to the cheapest layer that can handle it. This is how CoRover serves a billion users without a trillion-parameter model. For cost-cutting techniques, see 4 free tools to cut Claude Code tokens by 80%.
- Focus on Indian-language coverage. 22 official languages and 700+ dialects means models like Sarvam-105B (optimised for Indic languages) have a moat that GPT-5 or Claude can't match. Consider Sarvam's open-source models on Hugging Face before defaulting to US-based APIs.
- Adopt, don't just build. The government is funding AI adoption. If you're a startup, position yourself as an adoption partner, not a model lab. The Rs 10,000 crore is increasingly flowing to companies that solve problems, not those that publish benchmark scores.
FAQ
Q: Can India build a model comparable to GPT-5 or Claude?
A: Not in the near term with current resources. India's FY25-26 AI Mission budget is Rs 973 crore (~$117M), while a single GPT-5 training run costs hundreds of millions. India's advantage is in purpose-built models like Sarvam-105B that excel at Indian-language tasks — the domain that matters for 1.4 billion users — rather than competing on English-centric frontier benchmarks.
Q: How many GPUs does India have for AI?
A: The IndiaAI Mission has deployed over 38,000 GPUs as of mid-2026, up from an initial target of 10,000, accessible at ~Rs 65/GPU-hour to registered startups, researchers, and government agencies. The government is targeting 100,000 GPUs by December 2026.
Q: What is BharatGPT and who built it?
A: BharatGPT is India's first indigenous large language model, built by CoRover.ai. It powers virtual assistants across 14+ Indian languages on platforms including IRCTC, LIC, and NPCI, serving over 1 billion users. A 3-billion-parameter variant, BharatGPT-3B-Indic, is available open-source on Hugging Face.
Q: Should Indian startups use DeepSeek or Llama instead of building their own models?
A: Use them where they fit. The composite-AI approach mixes open-weight models (DeepSeek, Llama) for general reasoning with purpose-built small models for domain-specific tasks. The competitive advantage comes from domain data and deployment context, not from the model weights.
Q: What is the IndiaAI Mission budget?
A: The IndiaAI Mission has a total outlay of Rs 10,371.92 crore ($1.25 billion) over five years, approved by the Cabinet in March 2024. Of this, the compute pillar was allocated Rs 4,563.36 crore. For FY25-26, Rs 973 crore has been spent, with Rs 2,194 crore committed to 12 startups through GPU subsidies and grants.
Q: Is India's open AI ecosystem a disadvantage for domestic startups?
A: It's a double-edged sword. Indian startups compete against OpenAI, Google, Meta, and free open-weight models from day one — unlike China, where domestic companies had a protected market. However, the openness accelerates adoption and forces Indian companies to build solutions (not just models) that are genuinely better for specific use cases, which tends to produce more resilient companies.
Discussion
0 comments