India now has a data platform to train its own AI models, work on own GPU begins

New Delhi: India unveiled its library of non-personal data sets, called AI Kosh, to help companies build large language models (LLMs), the latest step by the Centre to help scale up the growth of artificial intelligence (AI) in the country.

AI Kosh will also be used to give startups approvals to get access to the India AI Mission’s 18,693 GPUs, or graphic processing units, which will be offered at a subsidized fee, said the ministry of electronics and information technology (Meity) on Thursday.

At the same time, said Union IT minister Ashwini Vaishnaw, “extensive stakeholder consultations” have begun to build “India’s own GPUs”—a process that he said will take “three to four years”. “Building GPUs is a very, very difficult process, but we’ll get there. We’ll share more details in due time, as we make progress.”

GPUs are regarded as the most crucial part of building AI models—algorithms that help in generating speech, text or videos just like humans. Since the advent of US tech firm OpenAI’s ChatGPT in October 2022, GPUs have become among the world’s most in-demand commodity—exemplified by a surge in Nvidia Corp’s valuation from $300 billion to $3.6 trillion.

Also read | Hiranandani’s AI infra firm Yotta plans a US listing. But at what cost?

However, India currently does not own the patents to manufacture its own GPUs, which makes the country reliant on US tech firms Nvidia, Advanced Micro Devices (AMD), Amazon Inc, Qualcomm Inc, Intel Corp and others for these prized chips.

“In the long run, the idea is to gain our own advanced chip patents and make them in more sophisticated semiconductor fabrication plants that will come up in India, in the years to come,” a senior government official added regarding India’s move to build a circular supply chain for AI. “Until then, we have plenty of GPUs to access to build requisite foundational AI models.”

Mint reported on 5 February that India was targeting to develop its own AI chip in about five years.

Until then, Vaishnaw said that Meity has received applications from 67 startups to develop foundational models—which will be granted access to Meity’s datasets through the AI Kosh platform.

“To begin with, we will select three to five foundational AI startups, and make the GPU infrastructure available to them imminently,” the minister said. “The platform to offer AI compute is now live, and India’s GPU cost will be among the lowest in the world.”

Also read | Nvidia’s Blackwell GPU may reach Indian shores as early as October

Vaishnaw said that 22 applicants have applied to build foundational AI models, while the rest could build models for a wide range of use cases.

The cost per GPU per hour—a key hindrance for upstarts from entering the field of training proprietary AI models—in India will be ₹67 or $0.77. Global data centres sell GPU access to enterprises at an average cost of $2-2.5 per GPU per hour.

“The compute platform also includes public datasets offered by various ministries, which will give startups access to non-personal data in various languages. These datasets will cover sectors such as agriculture, logistics and more. Even Bhashini, Meity’s vernacular language programme, will make its Indian language data available for startups to build AI models for India’s specific needs,” Vaishnaw said.

Bengaluru-based AI startup CoRover.ai, which sells Indic-language AI applications based on its secondary, OpenAI-based language model BharatGPT, is one of the applicants in Meity’s AI Mission to build its own GPUs.

Also read | Chips, Indic LLMs, flying taxis: Inside India’s tryst with tech independence

The company’s foundational AI model will be consumer-facing and a “general purpose model covering all beats, to which sector-wise expertise will be incrementally added depending on demand,” Ankush Sabharwal, chief executive of CoRover.ai, told Mint.

The compute infrastructure that was made available on Thursday includes GPUs from four companies—Nvidia, AMD, Intel and Amazon Web Services, being supplied to Meity by 10 shortlisted data centre and technology infrastructure vendors. In the long run, the IT ministry’s GPU mandate will be a key revenue stream for these vendors, Yotta Data Services’ CEO Sunil Gupta told Mint on 10 February.

Vaishnaw said the Cabinet-allocated fund of over ₹10,000 crore will be used to “offer 40% subsidy” to the cost of accessing GPUs by startups. However, the minister did not disclose if the Centre will hold stake in the AI models developed or will they be entirely privately owned.

And read | Mint Explainer: The mercurial rise of India-focused LLMs