Renting Intelligence

Indian enterprises are investing vast sums on AI, but whether this is generating assets or just adding to their subscriptions bills is a question that most organisations cannot answer. IMA India's ‘AI Reality Check’ series examines how this technology is changing the operational reality of Indian enterprise, and what it means for the leaders who run them. Earlier papers in this series looked at how data architecture determines whether AI deployment succeeds, and how organisations are drawing the line between human and machine judgement. This paper steps back to ask a fundamental question: Can India build a viable sovereign AI model of its own, and what would that entail?

India is ChatGPT’s second largest market, with over 100 mn weekly active users. Indian developers are among the biggest contributors to GitHub AI projects and Indian paid users run coding queries at around three times the global median, according to OpenAI Signals. But when Indian enterprises route their operational data through foreign AI models, the inference outputs, the fine-tuning signals and the IPgenerated by those interactions are largely leaving the border and enriching the models of OpenAI, Google, Anthropic and others. Vishal Dhupar, Nvidia's Managing Director for South Asia, put the choice plainly when we spoke to him, 'Right now we are at a junction, whether to rent services or build services.'

The Rent-or-Build Calculus and Why It Has Changed

Two years ago, the rent-or-build choice looked much starker than it does today. Training a frontier model from scratch was the preserve of a handful of US labs, and open-source alternatives were capable but not competitive at the frontier. That has changed. Various open-source models – DeepSeek, Llama, Qwen, Mistral and Nemotron – are now competitive on the benchmarks that matter for enterprises.

This opens up a third path for Indian firms. Rather than training a frontier model from scratch or renting capability indefinitely from a hyperscale, they can take an open-frontier model and post-train it on proprietary data, in Indian languages, deployed on Indian compute. Several labs are already on this path: Sarvam does so across 22 Indic languages; BHARATGen is building India specific foundational models; and AI4Bharat is producing the open research and public datasets the ecosystem depends on.

The Value Transfer that Doesn’t Appear on Any Invoice

When an enterprise routes its supplier contracts, customer call transcripts or manufacturing process data through a foreign LLM, the visible transaction is an API cost against an operational output. What does not appear on any invoice is the signal the provider receives about how its model performs on that enterprise's specific domain. Across thousands of queries over months, that signal trains a model that understands Indian enterprise operations better than it did before. 

Mr Dhupar calls this paying rent on intelligence. ‘Indian enterprises help foreign providers make their models better’, he says, ‘while borrowing back access to those improved models.’ He illustrates the point with a specific example. When Google Maps learned to navigate Indian roads, to say 'culvert' rather than 'speed bump', and to understand unlabelled lanes and informal landmarks, it learned from Indian users. That learning accrued to Google but the data belonged to the original users. The same thing happens each time a business routes operational knowledge through an external model without any mechanism to retain or build on the result.

Numbers make the imbalance plain. According to Observer Research Foundation (ORF), India accounts for ~17% of the global IT services market but captures just ~1% of high-value global technology value pools, including AI, semiconductors and hyper-scale infrastructure. Without domestic capabilities in compute, foundational models and data governance, India risks becoming what ORF calls a ‘net importer of intelligence’, capturing limited downstream gains while the economic rent flows out.

What is Already in Place

Businesses that are building AI applications for Indian customers have access to a data infrastructure that most comparable markets have not yet assembled. India’s UPI system processed over 228 bn transactions in 2025 and its ONDC generates structured transaction data across millions of small merchants previously invisible to formal systems. As of March 2026, DigiLocker held over 9.5 bn documents for more than 676 mn users. Taken together, this is a massive, machine-readable record of citizen and economic life, available for application development at a low cost. Most countries that are trying to build an AI infrastructure are simultaneously seeking to build the sort of data layer that India already possesses. But India’s advantage will not last indefinitely, and the window for building AI applications calibrated to local economic behaviour, languages and institutions is not open-ended.

India's talent position reinforces the case. The Stanford HAI AI Index 2026 ranked India second globally in absolute AI researcher count, at 50,460 and first in AI skill penetration on LinkedIn at three times the global average. The productivity multiplier available to India's software engineers through AI is of the same order as the one China is applying to its manufacturing workforce through robotics. Whether that multiplier compounds inside India or inside the firms whose models local engineers are training is, ultimately, a capital allocation question.

What Sustained AI Investment Produces Over Time

China has demonstrated, at scale and at pace, what it looks like when a country treats AI infrastructure as a national industrial priority. Its four largest internet companies, Alibaba, Tencent, ByteDance and Baidu, have committed a combined $84 bn in AI infrastructure investment by 2027, according to Goldman Sachs. In April 2026, China's Ministry of Industry and Information Technology reported domestic AI computing capacity of 1,882 exaflops, which translates to 1,882 quintillion (i.e. billionbillion) calculations per second, about 6000 times higher than what is reflected in the Top500 list, the benchmark used to compare computing capacity across countries.  

DeepSeek, released in early 2025, performed at near-GPT-4 level and was open-sourced at a fraction of US training costs. The assumption that frontier AI capability is a US prerogative has been broken. What this has produced is an open-source ecosystem that organisations can build upon. The cost of adapting them to proprietary data is no longer the preserve of large technology companies.

What India Has Built, and What Remains

Rather than attempting to compete across the full AI stack, India has focused on specific channels where the case for domestic capability is strongest. Attempting to match frontier foundation model investment would be economically inefficient at the pace the frontier is moving. Under the India AI Mission, Rs 104 bn has been allocated over 5 years across 7 pillars, including a skilling initiative targeting 5 mn people by 2027. Over 38,231 GPUs have been allocated to businesses and researchers with subsidised compute rates that sharply reduce the cost of AI development for startups and researchers.

Last year, the Department of Science and Technology launched BHARATGen, its first government-funded multimodal LLM initiative. Sarvam AI released its 105-billion-parameter open-source model thisFebruary, covering all 22 official Indian languages. The training costs for such models – Sarvam-105Bcost just ~Rs 4.2 bn – sit well below that of comparable US frontier models. 

India can drive widespread AI adoption through smaller, purpose-built models. But, as ORF points out, a disproportionate share of economic rent from AI will continue to flow to the firms controlling foundational LLMs, compute infrastructure and global platforms, so long as India stays primarily on the consumer side of those layers. Generative AI is estimated to contribute $2.6-$4.4 tr annually to the global economy, according to McKinsey & Co, and the returns will likely concentrate in the foundational layers, not the application layer. Full-stack sovereignty would cost hundreds of billions India does not have. The most viable path is therefore to control the layers that matter most for India's economic and strategic interests.

The chips layer is where selectivity will get tested the most severely. The India Semiconductor Mission 2.0, the Micron ATMP facility in Gujarat and the Tata-PSMC fabrication plant at mature nodes are the first steps away from pure import dependence. They are modest relative to the $52.7 bn the US has appropriated under the CHIPS and Science Acts and the ~$47 bn the EU has mobilised under the European Chips Act . The cost of training frontier AI models has been rising at roughly 2.4x a year since 2016, which means that each year a country or enterprise defers investment in sovereign compute, the cost of the infrastructure required to stay competitive increases by more than double.

What this Means for Your Organisation

Most enterprises are somewhere on the AI adoption curve. Very few have answered the more consequential question: Of all the AI work currently running, what will the organisation actually own at the end of it? Renting access to a foreign model and routing proprietary data through it is a legitimate decision for many workloads. Problems arise when this is the only decision, and one made by default rather than by design. The companies that will be in a different position by the end of the decade are those that have separated the workloads where renting makes sense from the ones where their own data is worth more than that.

India's DPI stack, its talent base and the new accessibility of open-source frontier models mean the opportunity to build rather than simply consume is more concrete than ever. Whether enterprises act on it will be determined by the quality of the capital allocation conversation happening, or not happening, in the boardroom over the next few years.