India to build its own foundational AI model; to host DeepSeek on Indian servers soon: Ashwini Vaishnaw

India is set to join the global AI race by launching foundational artificial intelligence (AI) models on the lines of OpenAI’s ChatGPT, Google’s Gemini, and China’s DeepSeek R1 over the next few months, according to the country’s IT minister.

“The foundational models made in India will be able to compete with the best of the best in the world,” electronics and information technology minister Ashwini Vaishnaw said at a press briefing on Thursday. “With algorithmic efficiency, we can create these models in a much shorter time frame. We will have a world-class foundational AI model in just a few months.”

The development comes in the backdrop of the debut of DeepSeek R1 on 20 January, which rattled global AI pioneers with its inexpensive model built at a fraction of the cost taken to create ChatGPT and other models.

Foundational AI models can perform a wide range of tasks and form the bedrock for creating AI applications. Large-scale datasets and high-end GPUs (graphics processing units) are required for training foundational AI models. Currently, GPUs made by US-based Nvidia dominate the foundational AI ecosystem, followed by other chipmakers such as AMD and Intel.

To support the training required to build the foundational AI model, India’s central government will launch a common compute facility with 18,693 GPUs, of which the first 10,000 GPUs will come online shortly, Vaishnaw said.

The GPUs will be supplied by 10 private sector companies including Yotta Data Services, Tata Communications, CMS Computers, E2E Networks, Jio Platforms and NxtGen Datacenter. These companies have between themselves secured 12,896 Nvidia H100 GPUs, 1,480 Nvidia H200 GPUs, and 742 MI325 and MI325X GPUs from AMD.

The compute facility will be available for all entities including startups, researchers and developers through an online portal that is expected to go live next week. The Centre will create a committee of experts and government officials to review requests for access to these GPUs, and approve them based on purpose and time-period of the access sought.

The common compute facility will provide average rate for AI compute ₹115.85 per GPU hour versus the global benchmark of $2-3 (about ₹170-260), making it highly attractive for entities to train the AI models.

To achieve this cost, the government will provide a subsidy for a period of four years, Vaishnaw said. “For high-end compute, it’s ₹150 per GPU hour. We will give 40% subsidy, reducing the cost to less than ₹100 per hour. Ours is the most affordable compute facility, at this point of time,” the minister said.

Vaishnaw added that the government was in touch with six major start-ups that will be able to build the models in six to 10 months.

Separately, proposals have been sought to build applications. “Eighteen applications have been selected for the first round of funding. These are in three themes—agriculture, learning disability, and climate change,” he said.

Should India create its own AI model?

Chinese company DeepSeek created ripples when its AI model R1 overtook ChatGPT as the top-ranked free app on Apple’s appstore. Built with $5.6 million, a fraction of the cost of other models like OpenAI, the model’s popularity has challenged the AI dominance of US-based companies and even led to billions being shaved off NVidia’s market cap.

OpenAI’s chief Sam Altman said that while R1 was an impressive model at the price it was offered, his own company delivered much better models. He had earlier dismissed India’s attempt to create competing AI models with a $10 million budget.

“It’s totally hopeless to compete with us on training foundation models,” he had said at an Economic Times Conversations event in June 2023. For perspective, a report by The Independent said that OpenAI spends $700,000 per day for keeping its public chatbot, ChatGPT, simply running—that’s over $250 million per year just on the running cost of a free app.

According to Vaishnaw, India is not late in creating its own AI model. He said the decision was not in response to potential impact from DeepSeek, but rather the availability of compute power at super low costs that India will be able to offer.

“We’ve been closely coordinating with startups building foundational models since the AI Mission was approved in March 2024,” the minister said. “The algorithmic efficiency and quality of data sets used for training has been captured. A common compute facility was required to train the models, which gives us a huge advantage versus other countries. That’s why we’re calling for proposals now.”

He expressed confidence that the country would have a ‘world class’ foundational model. “DeepSeek was trained with more than 2,000 GPUs, while the ChatGPT version was trained with about 25,000 GPUs. The technical partners who wanted to participate in this (AI) mission, they have started working and investing. Our focus will be on utilising the power of AI for solving population-scale problems,” he added.

The minister further said that DeepSeek would be hosted on Indian servers after security protocol checks so that users, coders, developers can benefit from its open-source code.

Right move

AI experts said India’s move was in the right direction geopolitically as well as economically as it was aiming to capture the tailwinds from AI for its overall economic growth.

“India seeks to grow (its GDP) to $10 trillion, and eventually to a $30+ trillion economy on its journey to become a developed nation,” said Ankush Wadhera, managing director and partner, Boston Consulting Group. “The largest force behind this increase is going to come from AI and frontier technologies, and only the US, China and India have the overall wherewithal to capture a disproportionate share of the AI rush.”

Wadhera said the Centre’s current move is a concrete way forward to build GPU capacity, which is over and above the processing capacity already in India through the Airawat HPC, Reliance’s facilities, C-DAC and more.

“Alongside, the Centre is also looking to build indigenous LLMs, all of which will contribute to accelerating India’s work to build its own foundational AI IP. This is a fundamental requirement, and it is a good place for India to start,” Wadhera said, adding that the government’s decisions will encourage global chip firms to focus on India.

Rutuja Pol, lead, government affairs at Ikigai Law, said India’s push to develop foundational models will be an important part of its geopolitical play, since its model will help cater to its own geographical and linguistic diversity.

“China, on this note, has showed that developing foundational capacities is possible,” Pol said. “Reducing the reliance on global models to build its own will be key for India in the long run.”

India trusted partners amid chip ban

In response to questions on the impact of GPU and AI chip export curbs by the US on India’s AI mission, Vaishnaw said India was seen as a “trusted partner”, indicating that it was unlikely to be impacted by the sanctions. “We respect IP rights and considerations in technology and that is why we have signed MoUs on very important technologies,” Vaishnaw said. “We believe that the trust that has developed will be important in every consideration and regulation that any country brings.”

On 14 January, former US president Joe Biden signed an executive order restricting the total number of GPUs that any company can order. India, while not being on the US blacklist, was not placed on the whitelist either—causing concerns around India’s ability to access chips. To be sure, all mainstream AI chips are owned by the US today.

The Indian government launched the IndiaAI Mission with a ₹10,372 crore outlay to strengthen the country’s AI ecosystem. The mission had seven pillars, including enabling 10,000 GPUs for AI compute infrastructure, creating local data sets, developing applications besides creating AI skill sets, financing for start-ups, and safe and trusted AI. The government will also provide data sets for development of the AI models.

On Thursday, Vaishnaw also announced AI safety institutions that will be set up in a hub-and-spoke model, where multiple institutions can partner to provide and develop tools, frameworks and processes needed for AI safety.

The projects approved for AI safety include areas of machine unlearning (IIT Jodhpur), synthetic data generation (IIT Roorkee), AI bias mitigation strategy, explainable AI framework (Defence Institute of Advanced Technology, Pune), privacy enhancing strategy (IIT Delhi, IIIT Delhi, TEC), AI ethical certification framework, AI algorithm auditing tool, and AI governance testing framework.