An Activate survey of 244 Indian AI builders shows most deployments depend on Western proprietary APIs, while open source and local model training remain marginal, raising concerns over sovereignty, cost, and long-term control.
India’s AI startup ecosystem may speak the language of open-source, but production reality tells a different story.
A survey by venture firm Activate covering 244 CTOs, developers, and founders finds 74% of deployments rely on proprietary Western models accessed through APIs, while only 13% customise or train their own models and just 13% use open-source models. As the report states, “The ‘open source India’ narrative doesn’t match what’s actually running in production.”
Infrastructure use reflects this dependence. 65% of GPU capacity goes to inference, compared with 21% for training and 14% for fine-tuning. “The majority of this ecosystem is calling APIs, not training models,” the report notes.
Cost pressures reinforce the trend. 60% spend under $2,400 per month, making local training uneconomical. “Investing in training infrastructure makes less sense every month when you can ride the capability curve by calling an API.”
Open-source preference remains weak, with only 4% choosing open weights and 2% selecting Meta’s Llama.
Meanwhile, 62% plan to adopt AI-agentic workflows, despite warnings that “Agentic workflows are also the most expensive and failure-prone pattern”, with cost already the top barrier.
Structural gaps persist. 16% cite lack of sovereign datasets, while experts argue India needs multilingual local data more than GPUs. Talent shortages are operational rather than technical: “The talent bottleneck isn’t in ML engineering. It’s in AI ops.”
Western platforms dominate across the stack, with OpenAI and Anthropic leading text and code tools, underscoring a clear reality: India is consuming AI infrastructure, not building it.
