India's Armed Forces are formalizing a sovereign military AI stack built on domestic startups Sarvam AI and CoRover.ai (developer of BharatGPT), with a proposed ₹300 crore Centre of Excellence and existing operational systems including Operation Sindoor and the Battlefield Surveillance System. The strategic framing is explicit: India is transitioning from experimentation toward an indigenous military AI ecosystem under the Atmanirbhar Bharat self-reliance doctrine, rather than fine-tuning foreign models. For anyone watching the multipolar AI landscape, this is one of the first non-Western military stacks built around named domestic startups rather than US-frontier-lab integrations.
Sarvam AI is providing predictive analysis, generative AI, and military-specific LLMs through its Chanakya vertical, plus agentic systems planned for the Centre of Excellence; the focus is on offline deployment where systems operate without internet access. CoRover.ai's BharatGPT is pitched as a multilingual, multimodal, multichannel indigenous model supporting voice, video, and text on-device, with no internet dependency — which matters for field conditions and data security. Operation Sindoor (May 2025 conflict) used AI for targeting enemy assets with what the Armed Forces report as "about 94% accuracy" using 26 years of historical data. The harness, baseline, and target taxonomy aren't disclosed in the reporting, so that number is an institutional claim rather than a comparable benchmark — worth flagging because military procurement metrics rarely get the disclosure that frontier labs are now starting to publish. Other existing systems: Battlefield Surveillance System, CIDSS (Command Information and Decision Support System), ECAS (Electronic Intelligence Collation and Analysis), TRINETRA integrated with Project SANJAY. The proposed ₹300 crore Centre of Excellence is under MoD discussions; Sarvam hadn't formally confirmed the partnership at time of reporting. Named officials backing the direction: General Upendra Dwivedi (Indian Army) and Lieutenant General Rajiv Kumar Sahni (Director General, Electronics & Mechanical Engineers).
The Indian procurement framework iDEX plus DRDO partnerships plus the Atmanirbhar Bharat framing add up to a deliberate parallel-stack strategy — same shape as Helsing in Germany, Anduril in the US, Rebellion AI in the UK, but explicitly oriented toward domestic models rather than fine-tunes of OpenAI, Anthropic, or Google. The BharatGPT-on-device piece is operationally the most interesting: military field use is the hardest deployment case (no connectivity, data-security gating, latency floors, ruggedized hardware), and any model that actually ships in that environment will have learned things cloud-only stacks haven't tested. The sovereign-AI thesis — that nations don't want to depend on US frontier labs for their military and government AI — has been a talking point for two years. India is the largest non-Western, non-Chinese case study actually building it. Watch for similar moves from Brazil, Indonesia, UAE, and Saudi Arabia over the next 18 months as each tries to assemble its own combination of domestic startups + state procurement + indigenous-model framing.
Most of the procurement details here are "under discussion" or "recently selected" — not confirmed deployment. The 94% targeting accuracy from Operation Sindoor wants independent corroboration; military procurements rarely publish the harness disclosure that's becoming table stakes for frontier lab benchmark claims. The faster procurement pathways the source article flags as a bottleneck are the practical constraint — Indian AI capability is being built faster than the contracting infrastructure that buys it. For anyone tracking the multipolar AI buildout: India's choices on Sarvam vs. fine-tuning Llama vs. licensing Mistral set a precedent for the next dozen sovereignty-focused governments. If BharatGPT and Sarvam Chanakya ship in production military contexts and work, the math changes on whether non-frontier-lab models can be trusted with hard real-world stakes — and if they don't, the sovereign-AI thesis takes a credible hit.