Training for India's first large-scale foundational artificial intelligence (AI) or Large Language Model (LLM) has been underway since June 1, according to sources aware of the development.
About 1,500 high-end Graphics Processing Units (GPUs) were allocated to Sarvam AI in May, with training formally kicking off at the beginning of June. An additional 2,500 GPUs are set to be added to the cluster by mid-July, bringing the total to 4,000 GPUs.
"Sarvam has been training workloads for over a month now," said a source on the condition of anonymity. "The GPUs are live, interconnected via InfiniBand networking and are actively being trained."
On January 30, 2025, Union IT Minister Ashwini Vaishnaw announced India’s ambition to develop its LLM, designed specifically to meet the country’s needs. Subsequently, the Indian government, on February 11, earmarked Rs 1,500 crore under the IndiaAI Mission to support the development of indigenous foundational AI models.
On June 24, Moneycontrol reported that the government is offering a 100 percent subsidy on compute infrastructure costs for companies building foundational AI models. This subsidy applies exclusively to foundational model development, while the existing 40 percent subsidy will support other GPU-intensive activities such as inference and applications.
In January 2025, the government announced that it would subsidise GPU access through a coupon-based system up to 40 percent. The IT ministry has empanelled over 34,000 GPUs to ease compute access for startups and researchers till now.
Apart from Sarvam, three startups have received approval from the government so far for developing a foundational model: Gnan.AI, GAN.ai, and Socket.
Also, read: Gnani.AI to build sovereign multilingual LLM with emotion-aware voice
Advantage NVIDIA H100 SXMs
Much of India’s sovereign Generative AI efforts depend on access to high-performance hardware, particularly NVIDIA’s H100 SXM GPUs, which are optimised for training large models.
"People talk about 1,000 GPUs, but they’re delivering 200-300 at best, some even include low-end L40s or L4s, which aren’t suitable for LLM training," the source further said. "For foundational model training, you need thousands of SXM GPUs and InfiniBand. Otherwise, it’s just inferencing, not training."
At present, Sarvam AI is the only one in India actively training a foundational LLM using this level of GPU infrastructure.
Also, read: IndiaAI allocates Rs 111 crore in GPU subsidies; Sarvam bags nearly Rs 99 crore
Discover the latest Business News, Sensex, and Nifty updates. Obtain Personal Finance insights, tax queries, and expert opinions on Moneycontrol or download the Moneycontrol App to stay updated!
