"Technological autonomy in AI is a strategic necessity, not a choice." Critically analyse with reference to recent indigenous LLM initiatives.

Technological autonomy in AI means a nation controls its own models, data and compute. The IndiaAI Mission was approved in March 2024 with ₹10,371.92 crore [2]. It has now released EKA-1 (Soket) and Dhīr 2B/5B (ZenteiQ) [1]. These models show that autonomy is necessary, but they also show it is still only partial.

Why autonomy is a strategic necessity

• Security: Dhīr runs offline and on-premise, so data stays on the device. This suits defence and sensitive government work [1].
• Control over weights: A foreign model fine-tuned on Indian data is still a foreign model; its owner controls the licence and the updates. MeitY presents indigenous models as building strategic autonomy [1].
• Linguistic inclusion: EKA-1 covers all 22 Eighth Schedule languages, which foreign models serve poorly [1].
• Economic: EKA-1 is aimed at BFSI and cybersecurity. Subsidised public compute lowers the cost of entry, and 237 projects have been approved for it [3].

Where the claim is overstated

• Hardware dependence: GPUs are made mainly in one country [3]. Indian RISC-V-based processors under the National Supercomputing Mission are only at an early stage [3].
• Rented compute: Most capacity comes from private empanelled providers. The state's own secure cluster is only about 3,000 GPUs [3].
• No independent testing: The India AI Governance Guidelines ask the IndiaAI Safety Institute to build benchmarks and evaluation datasets, and these are still pending [4].
• Cost and scale: EKA-1 has 26.3 billion parameters [1], far fewer than frontier models. The Economic Survey advises building application-driven systems rather than capital-intensive frontier models [5].

Assessment At the model and data layer, autonomy is a necessity. Here India has chosen small, affordable models built first for Indian languages. At the chip layer, autonomy is still an aim, so India must pursue it in stages rather than as total self-sufficiency.

Chips (imported) → Compute (largely rented) → Models & data (indigenous) India's AI stack: autonomy is strongest at the model layer

India's own LLMs secure autonomy where it matters most, even though the layers beneath them still depend on others. The way forward is to plan IndiaAI together with the India Semiconductor Mission and to publish independent benchmarks for Indian-language models. Building resilient compute, as the OECD recommends [6], would turn "Making AI in India and Making AI Work for India" into lasting strategic autonomy.