Rethinking AI for India – What policymakers must aim for

After the last Davos meeting of the World Economic Forum, there is a buzz of both speculation and excitement about India’s optimal approach to artificial intelligence (AI) at this pivotal juncture in its journey of becoming an advanced digital economy. While global developments are increasingly focused on building large-scale frontier models, India’s strategic priorities should be calibrated differently.

India’s economic structure, characterised by a large micro, small, and medium enterprises (MSME) base (of over 70 million entities) where the lack of adequate digitalisation is causing poor efficiencies and inadequate global competitiveness.

Significant agricultural dependence, a diverse linguistic and demographic landscape, and a highly cost-sensitive deployment and usage environments necessitates an AI strategy that prioritises utility, scalability, and affordability over sheer model size.

Recommended three layer AI strategy

Accordingly, a three-layered AI strategy is recommended—selective frontier capability development for strategic autonomy, adoption of open-source AI ecosystems to democratise innovation, and large-scale deployment of distilled, domain-specific models for real-world applications.

At the core of this approach lies a shift from “model-centric AI” to “outcome-centric AI”—where success is measured not by parameter counts, but by economic productivity, inclusion, and service delivery improvements.

On the other hand, it is noted that the global AI ecosystem is undergoing rapid transformation driven by exponential growth in model size and complexity, escalating compute requirements, and increasing concentration of capabilities in a few global firms.

Training costs for frontier AI models are now estimated to range between $100-500 million, with projections suggesting costs could exceed $1 billion for next-generation systems (Stanford AI Index, 2024; industry estimates).

While such investments may be justified in certain contexts, their direct applicability to India’s needs is debatable. India’s priorities include enhancing agricultural productivity; improving SME competitiveness; optimising logistics and supply chains; and strengthening public service delivery. These require applied, domain-specific AI solutions rather than
general-purpose frontier systems.

Opimising model sizing for India

A critical and under-explored issue in AI policy is the question of the optimal model sizing for India. For this, key considerations would necessarily include: First, a discussion of returns, as beyond a certain scale, improvements in model performance are marginal relative to cost increases; second, latency and deployment constraints since larger models are harder to deploy in real-time and low-connectivity environments; and, last but not the least, the cost of inference.

Large models impose recurring operational costs. For most practical applications, smaller, specialised models can achieve near-equivalent performance at significantly lower cost. This calls for a paradigm shift toward “right-sized AI” aligned with specific use cases.

At this stage, it might be good to consider the route that China seems to have adopted. They appear to have made a bet on open source models for targeting broad-base innovation.

In general, AI experts hypothesise that open source models will lag frontier models by 6-9 months; so, in six months, one could potentially have an open source model chat which is as good as an advanced model chat of today.

Moreover, the performance gap is narrowing rapidly and most real-world applications do not require cutting-edge performance. Open models have the additional advantage of offering significant advantages in cost, flexibility, and control.

It is reported also that there is a technique called distillation, by which one can take an existing large open source model and create a smaller model that is specialised for specific tasks like for the agricultural domain and different categories of MSMEs.

Therefore, there obviously needs to be an extensive preliminary research on answers for many questions such as the size of a model needed to solve practical problems with accuracy. China’s strategic emphasis on open AI ecosystems has enabled rapid diffusion of capabilities across sectors, fostering innovation at scale.

For India, open-source AI aligns with the principles of digital public infrastructure (DPI)—open, interoperable, and inclusive. India’s DPI framework—exemplified by Aadhaar, United Payments Interface (UPI), and Open Network for Digital Commerce (ONDC)—provides a strong foundation for AI deployment.

Extending DPI principles to AI can enable open datasets and model repositories, shared compute infrastructure, and interoperable AI services via standard application programming interfaces. This would position AI as a public good accessible to startups, enterprises, and government agencies.

In conclusion, an outcome-oriented AI policy or strategy could clearly be more beneficial for India. In effect, this would mean having the focus not on model size but on measurable economic and societal impact. Such an approach would obviously entail developing sector-specific metrics for evaluating AI performance in real-world applications.

The system should also be one that supports open source models and data sets and is capable of creating national repositories accessible to developers. The policy would also feature AI as DPI and envisage shared platforms for compute, models, and deployment as well as enable interoperability across applications. The above would help provide affordable access to compute resources for startups and academia.

As indicated earlier, while pursuing AI missions in different sectors, the priorities should be to drive adoption in agriculture and MSMEs —manufacturing, logistics, public services, and governance.

Appropriate skilling and capacity building missions would also have to be concurrently promoted. The AI Policy would also have to take a balanced approach between reliance on external models, frontier research capabilities, data governance, and privacy concerns and fragmentation of AI ecosystems without standards. The goal should be to harmonise strategic autonomy with global collaborations and standards.

The writer is President, Broadband India Forum

Disclaimer: The views expressed are the author’s own and do not reflect the official policy or position of Financial Express.