Why in the News
A case is being made to judge India’s artificial intelligence (AI) push by what its models can do, not by how much chip time it sanctions. The government has told the Lok Sabha that the IndiaAI Mission has sanctioned 93 lakh hours of chip time for 237 projects.
What is the IndiaAI Mission, and why is compute the wrong yardstick?
- What it is: The IndiaAI Mission (2024), under the Ministry of Electronics and Information Technology (MeitY), rents subsidised AI chips to start-ups and researchers, like a shared public gym for computing power.
- Why it was introduced: Training a model needs thousands of costly graphics processing units (GPUs), the chips AI runs on. A separate mission pillar backs home-grown foundation models (general-purpose AI models), including Sarvam‘s.
- DeepSeek comparison: Chinese lab DeepSeek needed about 2.8 million hours of similar chip time for its V3 model. India has sanctioned about three times that.
- Sarvam’s allocation: The Bengaluru start-up got 4,096 Nvidia H100 chips for six months. DeepSeek trained V3 on half as many in under two months.
- The takeaway: India already buys more compute than a world-class lab used, so its gap is in use, not supply.
Why will more chips not close the gap?
- Sarvam’s results: Its home-built models match 2025 open models from OpenAI and Alibaba, the firm says. They still trail the best, and chips do not explain the gap.
- Dependence on US policy: Chips rented abroad depend on United States (US) decisions. Eg. In 2025 the US was reportedly drafting curbs on chip shipments to Malaysia and Thailand.
- Short-lived curbs: Takshashila Institution research finds new technology spreads within a few years, so curbs fade and strategies built on them date fast.
- Free models close behind: The best free models trail paid leaders by about four months (Epoch AI). India can use them and fund AI that understands Indian law, medicine, farming and languages.
What is a reinforcement learning environment, and why should India fund it?
- What it is: A reinforcement learning environment is a set of expert-written tasks with a marking scheme. The model attempts each task, is scored and improves. Eg. A bail application graded against senior advocates’ checklist.
- Why it matters: Making a model useful, such as for advising a farmer, needs fewer chips but more runs and experts. India has them in courts, hospitals, farm universities and language institutions.
- Market value: Meta paid $14.3 billion for 49% of Scale AI, which supplies this kind of data.
- Outside export rules: Chip export controls generally do not reach this work, so India controls this input.
Where should India compete on chip hardware?
- Chokepoints: Accelerators, AI chips such as GPUs, and lithography, machines that print circuits on silicon, can be cut off by export controls imposed arbitrarily.
- Packaging and assembly: This stage, encasing and connecting finished chips, is fragmented, faces no targeted controls and has a few Indian plants.
- Memory packaging at Sanand: Deepening memory packaging with Micron at Sanand is more winnable than a leading-edge fab, a factory making the most advanced chips.
Challenges
- Scarce expert time: Senior advocates and doctors are costly and busy, so marking schemes are slow to write at scale.
- Benchmark gaming: Models tuned to published tests can score well without performing in real use.
- Thin language data: Many Indian languages have little digital text to train and test on.
- Fab dependence: Packaging plants still rely on chips made in foreign fabs.
Way Forward
- Compute for results: IndiaAI should reserve a large chip block for month-long runs by teams scoring best on Indian-language and subject tests.
- Public task banks: Pay experts to write and publish tasks for any Indian lab, a small cost beside the mission’s Rs 10,372 crore outlay.
- Compute buyers’ coalition: Convene countries with capital and engineers but no advanced fab, the group the scrapped US AI diffusion rule had capped.
- Open-weight mirrors: Keep national copies of major open-weight models, whose trained parameters are public, across these countries.
Conclusion
India’s AI effort is judged by the compute it buys, not the capability it produces. Whether IndiaAI ties chip time to measured performance on Indian tasks will show if that changes.
Government Initiatives on AI and Semiconductors
- Digital India Bhashini: Speech and translation tools across 22 Indian languages.
- BharatGen: A government-funded multimodal AI initiative (text, speech and images) for public services.
- India Semiconductor Mission (ISM): ISM 2.0 focuses on chip-making equipment, materials and Indian-owned chip design.
Matching Previous Year Question
“[2026, GS3, 15 marks] What is agentic Artificial Intelligence (AI)? Explain its working. Describe its applications with suitable examples. Discuss the advantages, risks and challenges associated with agentic AI systems.”
