Public AI infrastructure: Daron Acemoglu and Pascual Restrepo have argued that even when a powerful technology displaces labour, it can create new tasks over time. The internet did this. So did social media, which produced occupations that barely existed a generation ago: influencers, video editors, digital marketers, platform managers.
Artificial intelligence may follow the same path. Stock markets are already rewarding AI firms with rich valuations. Global corporations are pouring money into foundation models. The best-known models have come from American firms. China is catching up quickly with its own systems. India, meanwhile, is already a large user of AI models built elsewhere.
READ | UPI MDR debate: Why free payments need a cost model
A few Indian firms are building niche models. But India’s effort to create a frontier-scale model remains thin. Without a stronger push, India could repeat an old pattern: a large consumer market for mobile phones, metro systems and digital technologies, but not a significant creator of core technology.
India has not been inactive. The IndiaAI Mission, approved with an outlay of ₹10,371.92 crore, seeks to provide common compute capacity, datasets, startup support and indigenous foundation models. The government says more than 38,000 GPUs have been onboarded and 12 teams have been shortlisted for foundation models. These are useful beginnings. They are not yet proof of capability. The test is whether India can move beyond subsidised access to compute and scattered model-building projects to durable public AI infrastructure.
Two Indian experiences are relevant. One comes from telecom in the 1980s. The other from digital payments in the past decade.
C-DOT and India’s telecom lesson
In the 1980s, India’s telephone network was constrained by weak domestic capability and imported technology designed for richer markets. Foreign systems were expensive. They were not built for India’s scale, climate or rural conditions.
The Centre for Development of Telematics was set up by the Department of Telecommunications to solve this problem. C-DOT was given a small budget and a hard deadline. Its task was to build an indigenous switching system.
READ | RBI Payments Vision 2025: Inclusion gaps remain despite UPI success
It delivered a low-cost, modular and upgradable switch that could work in harsh rural conditions. The technology was then transferred to several Indian manufacturers. C-DOT also helped develop local component suppliers.
This mattered. Competition among domestic manufacturers brought down prices. Lower prices helped telephone networks spread. The lesson was not autarky. It was public investment in a core capability, followed by private production and competition.
UPI and the payment infrastructure model
The Unified Payments Interface offers a second template. Its success did not come from a single app. It came from years of institutional design by the Reserve Bank of India and the National Payments Corporation of India.
NPCI was created in 2008 to provide affordable, high-quality payment infrastructure. UPI built on two already diffused technologies: the mobile phone for consumers and the Immediate Payment Service for banks.
The crucial design choice was to let non-banks work with banks and build consumer-facing apps. These firms marketed aggressively. Banks and non-banks competed for transactions. The result was a public payment rail with private innovation at the edge.
UPI exceeded most expectations. It also challenged payment technologies built by dominant multinational networks. Once again, the lesson was not state ownership of the entire system. It was public infrastructure, open access and competition.
READ | UPI global expansion faces the real cross-border test
Public AI infrastructure for India
If AI is indeed a general-purpose technology, India cannot rely only on private foreign models. The question is not technological prestige. It is access. India’s labour force will need AI tools to train for new tasks, raise productivity and participate in future markets.
This requires public investment. Building a serious AI model will be expensive. It also needs leadership that combines frontier technical competence with an understanding of frugal innovation. India’s constraints are not incidental. They are central to the design problem.
Compute remains the binding constraint. A public AI model is not a one-time software project. It needs GPUs, data centres, power, engineers, datasets and repeated training runs. India can lower the price of compute access, but it remains dependent on global semiconductor supply chains. That makes institutional design more important. Scarce compute must be allocated to models, datasets and applications that meet Indian public needs, rather than spread thinly across subsidy-seeking projects.
Such an effort will take time. The institutional structure must recognise this. Once a capable Indian model is ready, private firms can be brought into the ecosystem, as in C-DOT and UPI. They can adapt it to sectors, languages and geographies.
A publicly funded Indian AI model would democratise access, reduce usage costs and blunt the market power of incumbent platforms. Incentive policies can help adoption. The long-run externalities from domestic capability are likely to exceed the fiscal cost.
India should not make dependence on private AI models its default strategy.
Chidambaran G Iyer is Associate Professor, Centre for Development Studies, Thiruvananthapuram.
READ |