Beyond the Summit: Why the Government Must Put Money Where Its Mouth Is

The Artificial Intelligence Impact summit is on in Delhi. There is natural interest in the subject and artificial hype over the presumed leadership role India is playing in modulating use, access and governance of the technology on behalf of the Global South. The interest is welcome, and can be leveraged to educate the public to distinguish AI from merely smart software, ridding them of the confusion of all use of extensive technology with AI.

India’s digital public infrastructure is most useful and welcome, but it is not AI; that goes for India’s startup ecosystem and research. This might seem obvious but when high-powered tweets from the government cite these as India’s strides in AI, it points to generalised confusion about what constitutes AI.

When computers can take decisions in the absence of human intelligence, assessing, on their own, data and what is required to be done, that is AI at work. Once sufficiently trained, say by being fed thousands of images of a cat, machines can learn to recognise a cat, when it sees it in whole or part. In some special cases, using multiple layers of neural networks, machine learning can result in deep learning, which underlies contemporary generative AI, capable of creating pictures and writing.

Also Read: Time to Launch Strategic AI Limitation Talks?

Right now, the AI we have is narrow AI, capable of solving problems in defined fields. The expectation is that better and broader narrow intelligence would lead to artificial general intelligence, or AGI, capable of thinking like humans. Many worry that AGI would evolve into artificial super intelligence, which is not just better than human intelligence but also better to a degree that could pose an existential threat to human survival — the stuff of dystopian science fiction.

Creating AI calls for people trained in advanced linear algebra, probability and calculus, and huge data processing capability, available from large numbers of ultra-fast parallel processing units originally designed to generate graphics for videogames and so, still called graphic processing units or GPUs. When large arrays of GPUs crunch data, they generate heat, which must be removed using cooling, consuming gigawatt-hours of power for the purpose. Data centres house large arrays of GPUs and high bandwidth memory chips, which store data and process them as required.

For the kind of business applications, to perform which the American AI firm, Anthropic’s Claude model has developed software tools, called agents, posing a threat to ranges of work performed by India’s information technology service companies and IT-enabled services companies, small models are sufficient, as the Economic Survey points out. Small models that employ up to 10 billion parameters can perform work in narrow areas. India’s Sarvam AI right now works on 2 billion parameters. Google’s Gemini 3, Open AI’s, as well as Anthropic’s larger models work on one trillion-plus parameters.

The more the number of parameters, the more capable the AI model — up to a point. Clever programming is able to do better training and better retrieval of the relevant information, using smaller numbers of parameters. China’s DeepSeek model, which caused a major upset in US stock markets early in 2025, uses software to bring into play only the relevant parameters, instead of making all the parameters work at the same time.

For business applications, India needs to develop or adapt available open-source models from American and Chinese companies. Companies like Sarvam train their AI on Indian data, including Indian languages. Odisha and Tamil Nadu have tied up with Sarvam to create services and public utilities, data centres with tens of megawatts of capacity open to startups, researchers and the like.

More state governments would do well to set up such facilities, to supplement the effort by the Centre to procure ultra-fast processors and make their processing capacity available to Indian researchers and businesses.

However, as the Economic Survey notes, the biggest constraint for India would be the availability of high-end GPUs, which are, at present, designed by a handful of companies like NVIDIA and Google, and manufactured by Taiwan’s TSMC at its plants in Taiwan and the US (Biden’s industrial policy, complete with the Chips Act and huge subsidies, succeeded in getting TSMC to set up its fabs in the US, and now Trump is taking credit for this success). High Bandwidth Memory (HBM) chips are a constraint for even those with access to advanced chips from NVIDIA, and will be for India as well.

India is not just another country which faces the threat of AI converting traditionally labour-intensive industries into robot-intensive ones and having to retrain workers and create new industries to absorb the displaced workers. India cherishes strategic autonomy. It cannot allow its national security to be rendered vulnerable by not being able to deploy the kind of AI its potential adversaries can bring to bear.

While Indian businesses must develop narrow AI models, a government-led initiative must target developing an AI model with hundreds of billions of parameters, apart from narrow models for different strategic capabilities. India must, likewise, develop the capability to design and fabricate its own advanced processing and memory chips. India has the raw talent needed, it must be educated to acquire the needed skills for AI and chipmaking.

China has succeeded in developing its own chip manufacturing capability, even if it still trails the NVIDIA-TSMC duo’s cutting edge. So can India, provided the sufficient political will and funds for risk-taking research can be summoned.

India’s retirement savings should be able to invest in venture funds that invest in hard-technology startups working to produce every single element of the chip-making ecosystem, the assemblies, the sub-assemblies and components, instead of preferring to invest the bulk of their funds in government bonds. The Employees’ Provident Fund and the National Pension System command investible funds of trillions of rupees.

Indian talent slaves away in global capability centres creating intellectual property for foreign companies. Such talent must be given a chance to work on the components and kit designed to solve India’s strategic challenges. This will call for public funds, retirement savings and the huge amounts of the public’s savings that today go into the stock market only to inflate the prices of a few thousand companies to unsustainable levels. Indians would love to invest in large pools of risk capital that finance research and development in AI and associated chips, if only they are given an opportunity to.

This is where policy must focus. Summits are useful to generate enthusiasm, but to convert that enthusiasm into national power, focused policy and action are required.

Also Read: 
Not Just Make in India, Chips Need to be Made by India
We Don’t Need No Data-Centre Coddling
Those Who Most Need to Understand AI Don't Get It
Beyond the Hype: What AI Means for Your Career
AI and the Global South's Next Leap Forward