India Should Stop Counting Its AI Chips, Focus On Results, Pg8

Key Highlights:

• The IndiaAI Mission has sanctioned 193 lakh hours of chip time for 237 projects.
• This sanctioned chip time is significantly more than the 2.8 million hours used by Chinese lab DeepSeek to build its V3 model.
• Bengaluru-based startup Sarvam developed indigenous AI models using Nvidia H100 chips, comparable to models from OpenAI and Alibaba.
• The article suggests India should prioritize AI results tailored for Indian contexts over merely counting chip allocations.
• A strategic shift towards packaging and assembly layers of the semiconductor supply chain, like the Micron plant at Sanand, is recommended over leading-edge fabs.

Detailed Insights:

• Despite significant chip time allocation, Sarvam's models, while commendable, still lag behind top global AI models.
• Access to advanced chips is constrained by potential US export curbs, as reported by Bloomberg in 2025 regarding Malaysia and Thailand.
• Building AI models is only the first step; the critical next stage involves making them useful through reinforcement learning environments.
• India possesses a strong human capital base in various sectors (legal, medical, agricultural) suitable for creating and annotating data for AI training.
• The article advocates for investing in experts to develop tasks and marking schemes for Indian-language and subject-specific AI tests.
• The IndiaAI Mission has a budget of Rs 10,372 crore, suggesting ample funds for such data annotation and expert engagement.
• India's hardware strategy should focus on less controlled segments like packaging and assembly, which are less susceptible to export controls.
• International collaboration with countries that have capital and engineers but lack advanced fabs could provide a hedge against compute limitations.

Scientific/Technical Concepts Involved:

• Reinforcement Learning Environment: A setup where an AI model learns by performing tasks and receiving feedback (rewards or penalties).
• AI Accelerators: Specialized hardware components designed to speed up artificial intelligence computations, such as GPUs.
• Lithography: A critical process in semiconductor manufacturing used to create patterns on silicon wafers for integrated circuits.
• Open-model weights: The parameters of a pre-trained AI model that are publicly available, allowing others to use and fine-tune the model.