"Public compute infrastructure is the backbone of India's sovereign AI ambitions." Examine this in light of the IndiaAI Mission's foundation-model pillar.

Training a foundation model needs thousands of GPUs, which few Indian startups can pay for. The IndiaAI Mission (₹10,371.92 crore, approved 2024) [4] therefore pairs its Innovation Centre (Foundation Models) pillar with subsidised public compute. The November 2026 release of EKA-1 (Soket) and Dhīr 2B/5B (ZenteiQ) shows this approach working [1]. Compute is the backbone of India's sovereign AI, but it is not the whole skeleton.

How public compute supports sovereign models

• Lower entry barrier: The state pays for compute and research costs, and private firms build the models. This produced EKA-1, a 26.3-billion-parameter model covering all 22 scheduled languages [1].
• Scale: Shared capacity passed 45,000 GPUs by June 2026 [3]. 237 projects have been sanctioned 93.18 lakh GPU hours [2].
• Strategic use: A 3,000-GPU secure national cluster is being built for sovereign and strategic uses [5]. It suits offline, on-premise models like Dhīr, which is aimed at defence [1].
• Autonomy: Owning the model weights ends dependence on foreign licences and updates. MeitY calls this technological and strategic autonomy [1].

Where the backbone is weak

• Imported chips: GPUs are made mainly in one country. Indian RISC-V processors under the National Supercomputing Mission are still being designed [2].
• Rented, not owned: Most capacity comes from 15 empanelled private providers. State-owned capacity is small, such as a system of about 1.1 EFLOPS at the NIC Data Centre [2].
• Sanctioned is not used: GPU hours are reported as allotted, not consumed, so the public cannot see what the subsidy produces [2].
• Compute alone is not enough: Datasets, talent and independent testing matter as much. The India AI Governance Guidelines ask the IndiaAI Safety Institute to build benchmarks that do not yet exist [6].

Way forward

• Plan the AI and semiconductor missions together to secure the hardware layer.
• Publish GPU hours actually used for each funded model.
• Require open Indic benchmarks before models are used in government.

Public compute has turned India's sovereign-AI ambition into working models, but it is necessary rather than sufficient. The OECD judges national compute on capacity, effectiveness and resilience [7]. India has built capacity; it must now secure chips, track usage and verify quality. Only then will "Making AI in India" also mean "Making AI Work for India" [1].