India pays Rs 9 a GB; AI inference must be priced in paise: Sarvam’s Raghavan
Sarvam AI’s Vivek Raghavan urges India to build a full AI stack locally, warning of “digital colony” risks and calling for low-cost, paise-priced inference.
Bengaluru: India risks becoming a digital colony if it continues to export its data and import intelligence, Sarvam AI co-founder Vivek Raghavan said, calling for the country to build capabilities across the AI technology stack instead of depending on global models and infrastructure.“Today, we are exporting our data and importing intelligence,” Raghavan said at Bengaluru Tech Summit 2026 on Tuesday. Drawing a parallel with colonial-era trade, when India exported cotton and imported cloth, he cautioned that the country could face similar dependence in the digital economy.When users interact with global AI models, their data leaves the country and helps improve models developed elsewhere. As these models get better, more users gravitate towards them, reinforcing the cycle. “A lot of people are very proud to say that India is the largest user of ChatGPT. But this is not a good thing,” he said.In June, Sarvam raised $234 million in a funding round led by HCLTech and Bessemer Venture Partners, taking its valuation to $1.5 billion and making it one of India’s newest AI unicorns.Raghavan said technological sovereignty requires India to build capabilities across five layers — compute, data and evaluations, models, harnesses and applications. While the country may not yet be close to manufacturing advanced AI chips, it should ensure that GPUs and cloud infrastructure are located in India, with sensitive and regulated workloads hosted and operated by Indian entities.He called for building Indic datasets and benchmarks to improve model performance in Indian languages and on tasks relevant to the country. India should use the best open-weight models while also developing its own, he said. Open weights alone do not guarantee sovereignty if the training data cannot be audited.Raghavan said Sarvam had built a 105-billion-parameter model with a team of about 20 people, most of them under 25, and was working towards developing a trillion-parameter model from scratch.He stressed that AI products for India must be designed around affordability, scale, local languages and voice. The price of 1GB of mobile data has fallen 97%, from Rs 287 in 2014 to Rs 9 in 2025. Over the same period, monthly data consumption per subscriber surged from 61.66MB to 24GB, nearly 399 times, while broadband connections rose 15.7 times to 94.4 crore, according to figures cited from Department of Telecommunications.Raghavan said AI inference would need a similar affordability model. “A Rs 200 recharge and an Rs 8,000 handset set the ceiling for what a conversation can cost. Inference priced per token in dollars cannot serve this market; it has to be priced per conversation, at Indian scale,” he said.With nearly one billion active internet users and about 500 million people still outside the internet, voice will be crucial to expanding access. AI interfaces must work across India’s 22 scheduled languages and thousands of dialects, rather than expect users to adapt to technology.Raghavan said the larger goal was “universal basic intelligence”, giving every citizen access to an AI agent that understands their language, represents their interests and helps them access services and build skills.Money habits form by age 7. Is your child learning the right ones? Join the TOI Money Masterclass.
Topics in this story
Gathered from external sources. Rights to this text belong to whoever originally published it.