India will not build OpenAI.

What the team has actually catalyzed.

Across TE partners.

Four numbers that define the moment

The IndiaAI Mission commits USD 1.24B over five years (~₹10,300 crore). The compute-subsidy mechanism reduces effective GPU cost to ~₹65–115 per hour - 10–20× cheaper than global. Training cost for an Indic fine-tune of a Qwen, Llama, or DeepSeek backbone has dropped from notional USD 5–10M to closer to USD 0.5–1M per fine-tune. This is the single most consequential cost-side shift in Indian AI venture economics.

India has 1.4 billion users speaking 22 official languages plus 700+ dialects. The Digital Public Infrastructure stack - Aadhaar, UPI (~16B monthly transactions), ONDC, Bhashini, DigiLocker - provides transactional and identity rails that AI agents can call into at population scale. No other country offers anything equivalent.

The Indian voice-AI market is forecast at USD 153M in 2024 growing to USD 957M by 2030, a 36% CAGR. The Bharat consumer - the next 500 million users coming online - is voice-first by default.

Robot foundation models reached USD 39B at Figure's Series C and USD 5.6B at Physical Intelligence's Series B. The Indian analogue, where vision-language-action robotics will run on Indian factory floors, warehouses, ports, and agricultural operations, is structurally underpriced by two to three orders of magnitude.

Angle A - AI for India and Bharat

Vertical AI in regulated Indian industries is the largest near-term venture opportunity. Qure.ai is the template - 18 FDA clearances (the most in the US for lung-cancer AI), deployments in 90+ countries, USD 65M Series D, Time 100 Most Influential Companies 2025. We expect to back 8–10 Qure-class companies in health, legal, agriculture, climate, education, and public services.

Voice and low-literacy interfaces are structurally where Indian vertical AI will find its largest commercial breakthroughs. Skit.ai, Slang Labs, Reverie, Yellow.ai are early demonstrations.

Indic foundation and voice models - Sarvam AI, selected by the IndiaAI Mission to build India's sovereign LLM in collaboration with AI4Bharat at IIT Madras, has open-sourced Sarvam-30B and Sarvam-105B. AI infrastructure - Yotta, Jio-NVIDIA, Tata-NVIDIA, E2E Networks - is the third pillar of Angle A. India will have 80,000+ GPUs deployed or in pipeline by end-2026.

Angle B - AI for deeptech

AI applied as an R&D primitive across bio, materials, manufacturing, robotics, and space. Isomorphic Labs raised USD 2.1B in October 2025. Microsoft published MatterGen in Nature in January 2025. Physical Intelligence raised USD 600M and open-sourced pi-0. The AI-meets-physical-world boundary is being redrawn.

AI × Bio. Aganitha is the lead Indian player. Bugworks uses AI-supported target identification in its AMR pipeline. The structural opportunity is the in-silico CRO 2.0 model serving global biopharma at India price points.

AI × Materials and chemistry. MatterGen-class generative materials models are now in Indian academic labs. We expect 4-6 AI×materials seed positions over the deployment period.

AI × Manufacturing and robotics. CynLr is the lead Indian visual-intelligence-for-industrial-robotics company. Vision-language-action robot foundation models for Indian factory floors, with strategic LP relationships at Tata Motors, Bharat Forge, Bajaj Auto, Mahindra, and Reliance providing pilot facilitation.

AI × Space and earth observation. Pixxel and GalaxEye are the lead Indian players. India's launch infrastructure combined with satellite manufacturing creates a vertically integrated AI×space stack few other countries can match.

AI × Defence. IIT Kanpur's defence-drone programme has produced AI-enabled kamikaze drones, anti-drone systems, and stealth ISR drones. The TE Defence-Tech Translation Sprint at IIT Kanpur is the primary funnel.

Physical AI - India's third-pole opportunity

Physical AI - robot foundation models running on real hardware in factories, warehouses, ports, fields, and operating theatres - is now the central frontier of applied AI. Figure raised a USD 39B Series C in late 2025. Physical Intelligence raised USD 5.6B and open-sourced pi-0. AGIBot, three years old, has already shipped roughly 10,000 robots. The bits-and-atoms boundary is being redrawn at a pace that resembles the early years of mobile or cloud, and the centre of gravity is no longer pure-software AI.

Recent field observations from China make the geography of this race vivid. China has built 'vibe manufacturing' - a dense, hyper-responsive electronics supply chain across Foxconn, BYD, LYitech, and thousands of tier-2 and tier-3 firms that lets robotics start-ups iterate hardware at software speed. State-funded data factories teleoperate humanoid robots 24/7 to generate proprietary training data; in some cases the state buys robots, generates the data, and sells it back. Hardware is impressive (kinematics, locomotion, acrobatics); intelligence and manipulation dexterity remain unsolved. The US holds the counter-advantages: developer velocity (full Claude, Cursor, Copilot, agentic-dev tooling - all of which are blocked or thin in China) and chip-architecture dominance through NVIDIA, on which even Chinese robotics still runs.

India sits in a structurally unique third-pole position that neither the US nor China occupies on its own. Indian engineering teams have full, unrestricted access to the entire US software-developer stack - Claude, Cursor, Copilot, and agentic-dev tools that Chinese engineers cannot legally use. Indian engineering teams have full, unrestricted access to NVIDIA silicon - Jetson Orin, Thor, H100, H200 - without the US export controls Chinese firms face. India has a fast-growing electronics manufacturing base across Sricity, Pune, Chennai, Greater Noida, and Hyderabad where PLI-anchored capacity for PCBAs, edge compute, and precision components is doubling every two to three years. And Indian labour cost on the teleoperation and data-factory side is genuinely competitive with China. The result is a position no one else has - US-speed software development on top of NVIDIA silicon, integrated with manufacturing depth that is growing toward China-grade density, at Indian cost.

What Indian entrepreneurs should build. Vision-language-action foundation models for Indian factory floors, warehouses, and ports - CynLr is the early demonstration; we expect a generation of CynLr-class companies out of IIT Madras, IISc, IIT Bombay, and IIT Delhi robotics labs. India-grade data factories - teleoperation centres at Indian BPO scale generating labelled, multi-environment, multi-modal training data for global model developers. Edge compute and PCBA design IP for the embodied-AI stack - the parts of the strategic technology pyramid (high-performance compute, smart actuators, multi-layer PCBAs) that the rest of the world is racing to localise. Robotic manipulation and dexterity systems - the unsolved problem the China visit explicitly confirmed. Agri-robotics priced for Indian smallholder operations. Defence-grade physical AI - drones, ISR, autonomous ground and underwater platforms - where IIT Kanpur and our Defence-Tech Translation Sprint are the primary funnel. And robotics-as-a-service business models for Indian industrial buyers - the operating layer that converts a one-time hardware sale into a decade of recurring revenue.

The China + 1 dimension amplifies all of this. Global brands and OEMs diversifying away from China-only supply chains need a credible alternative for physical-AI components, sub-systems, and finished platforms. India is the only large economy that combines manufacturing depth at scale with full alignment to US-led technology, governance, and export-control regimes - and the window for Indian founders to capture this is the next thirty-six months.

What we do not invest in

Pure-play foundation-model start-ups attempting to compete with OpenAI, Anthropic, or Google at the frontier. Sarvam's sovereign-LLM positioning is the exception, not the template.

Pure-play AI infrastructure scale-ups at Yotta scale - capex required is in strategic-LP and sovereign-capital territory, not venture-capital territory.

Pure-play hype-cycle AI consumer apps without defensible distribution moats.

AI ethics consultancies, AI training data labelling companies, and AI-enabled productivity SaaS for the developed-market enterprise.

The contrarian view we hold

The frontier-model layer is not where most AI venture value will accrue. The application layer - regulated industries and embedded software-hardware combinations - will capture more value over the next decade. Harvey, Hippocratic AI, Sierra, and Glean already validate this at the global level. Indian vertical AI, with Qure.ai as the template, will produce 2–3 Qure-class outcomes per year between 2026 and 2030.

AI×robotics outcomes will not be dominated only by US and Chinese players. Indian-founded robotics start-ups with vision-language-action models on Llama, Qwen, or pi-0 backbones have a structural cost and demand advantage that consensus has not repriced. CynLr is the early demonstration; we expect five to seven CynLr-class companies from IIT Madras, IISc, and IIT Bombay robotics labs over the next five years.

Why full-stack wins in Indian AI

Pure-model AI is an expensive race to a thin pricing layer; vertical and applied AI in India is a recurring-revenue business sitting on top of a workflow. The Indian customer - a hospital, a bank, a farm, a government agency, a factory - does not buy 'AI'; it buys an outcome (a diagnosis, an underwriting decision, a yield forecast, a payout, a finished part). The companies in our book that have compounded best are full-stack operators: model + deployment + data ownership + workflow + recurring service.

The pattern shows up across every vertical. Qure.ai and 5C Network sell radiology AI as a clinical-reporting service. Kaleidofin sells AI credit scoring wrapped in a lending product and a collections operation. DeHaat and CropIn sell AI agronomic intelligence wrapped in input distribution, advisory, and offtake financing. Sarvam sells voice AI as enterprise and government deployment, not as raw inference. Aganitha and Cradle-style companies sell generative-design models bundled with computational-chemistry services and IP co-creation.

The compounding logic is straightforward - every operational engagement produces proprietary data, which improves the model, which improves the next engagement. Indian patient diversity, agricultural variation, language diversity, and financial-informality complexity are unique data sources that pure-model rivals in San Francisco cannot replicate. Founders who build the model and the operating layer together capture the value; founders who only build the model end up renting it to whoever runs the workflow.