AI mission crosses 34K GPU capacity, 3 new startups to help develop models

Hindustan Times

a day ago

India has added 15,916 graphics processing units (GPUs) to its national artificial intelligence infrastructure while selecting three new startups to develop indigenous AI models, the government announced on Friday, accelerating efforts to build sovereign AI capabilities that developers can tap into.
According to Union minister for electronics and information technology (Meity) Ashwini Vaishnaw, who made the announcement at an event in Delhi, seven companies will provide the additional GPU capacity — which takes the total in the IndiaAI compute portal to 34,333 units ——at subsidised rates of ₹67 per hour, substantially below the average market rate of ₹115 per GPU hour.
'When we began the India AI mission, we had an initial target of 10,000 GPUs, which we at the time thought would be a lot,' Vaishnaw said. 'But we got a promising response from the industry.'
The government has already empanelled 18,417 GPUs in the first tranche, with 14,000 units currently accessible through the IndiaAI compute portal that became operational in March.
Vaishnaw announced a third round of GPU empanelment will be completed soon, suggesting continued expansion beyond the original targets.
The IndiaAI mission centres on offering AI technology access through government-subsidised infrastructure, including a marketplace of sorts for access to compute (which typically requires expensive hardware) and government databases.
Among these efforts is the development of indigenous large language models, the technology that underpins AI chatbots such as ChatGPT.
The three startups that were chosen from 506 proposals will develop LLMs optimised for India's 22 major languages and hundreds of dialects.
The new startups join Sarvam AI, which was chosen in April to build a 120 billion parameter model, in developing indigenous AI systems trained on Indian datasets.
The government has loaded 367 datasets onto AI Kosha, a platform supporting AI research across India.
Soket AI will develop a 120-billion parameter open-source model optimised for Indian languages, targeting defence, healthcare and education applications. CEO Abhishek Upperwala set a one-year delivery timeline, saying 'India needs a strong base model that others can build on.'
Gnani AI is building a 14-billion parameter voice system for real-time, multilingual speech processing.
Gan AI focuses on text-to-speech technology with a 70-billion parameter multilingual model targeting 'superhuman' capabilities.
The expansion, coupled with the hardware aspect, represents a dramatic scaling from the programme's original ambitions. 'When we began the India AI mission, we had an initial target of 10,000 GPUs, which we at the time thought would be a lot,' Vaishnaw said. 'But we got a promising response from the industry.'
The government has already empanelled 18,417 GPUs in the first tranche, with 14,000 units currently accessible through the IndiaAI compute portal that became operational in March 2025.
'Common compute is a very important part of the principle of democratising technology,' Vaishnaw said. 'Technology should not be left in the hands of a few.'
Rather than building government data centres, officials opted to empanel private providers who invest in the infrastructure while receiving 40% subsidies. The strategy yielded bids already more than 40% below prevailing market rates before the additional subsidy was applied.
'The government thought building data centres was going to take longer, so they decided to empanel service providers and private sector players who will invest in the compute infrastructure,' a senior official told HT.
The subsidised pricing represents a significant departure from international GPU markets, where Nvidia H100 hourly rates range from $2.99 to $9.98 across major cloud providers in 2025, according to estimates.
The seven newly empanelled companies are Cyfuture India, Ishan Infotech, Locuz Enterprise Solutions, Netmagic IT Services, Sify Digital Services, Vensysco Technologies, and Yotta Data Services. Cyfuture emerged as the dominant supplier, offering multiple chip types including Nvidia's H100 and AMD's MI300X processors, while Yotta Data Services provides access to Nvidia's latest B200 chips.