‘IndiaAI Does Not Need More Policy PDFs or GPU Ribbon Cuttin…

It has been more than a year since the IndiaAI Mission was announced, yet the country awaits its homegrown LLM. While there have been efforts across compute, datasets and research collaborations, the Mission’s progress remains far behind global labs that are already rolling out next-generation models.

Google just launched its latest model, Gemini 3, which it claims outperforms OpenAI’s GPT-5.1 and Anthropic’s Claude Sonnet 4.5.

Speaking at the Bengaluru Tech Summit 2025, IndiaAI Mission CEO Abhishek Singh acknowledged that India continues to lag the US and China. While global players release new LLMs, the Mission remains focused on building its seven foundational pillars, which are compute capacity, IndiaAI Innovation Centre (IAIC), datasets platform AIKosh, application development initiative, startup financing, IndiaAI FutureSkills and safe & trusted AI. 

On foundation models, Singh said the government is currently supporting 12 initiatives. One of them, a homegrown sovereign LLM being developed in Bengaluru by Sarvam AI, is slated for release in December and will be showcased at the IndiaAI Impact Summit in February, 2026. The model is a 120-billion-parameter foundation model trained on more than 17 trillion tokens, including 17–20% Indian data.

Frontier AI cannot be built on short-term funding cycles, according to Jacob Joseph, VP of data science at CleverTap, an all-in-one customer engagement platform. He told AIM that it needs “deep, patient capital,” especially when India’s R&D spending sits at roughly 0.6% of GDP, compared to the 2–3% committed by countries at the forefront of advanced AI. 

Joseph added that training a world-class model is a years-long, multibillion-dollar journey, with substantial investment going into foundational research that may take time to show commercial payoff.

Meanwhile, Singh said the Mission released its AI safety framework two weeks ago and is trying to balance innovation with safeguards. “We are developing tools for detecting deepfakes or limiting AI-generated content,” he said, adding that these tools will be available on the IndiaAI platform. 

With that structure in place, “we will be able to do much more to ensure that whatever AI we develop improves efficiency, productivity and the reach of services,” Singh said.

Sanchit Vir Gogia, CEO of Greyhound Research, was more direct. He  told AIM, “India does not need more policy PDFs or GPU ribbon cuttings.” He argued that infrastructure must perform, not just exist, which means subsidised compute that is fast, reliable and tiered for real workloads. 

According to him, India also needs to back national AI bets in sectors like healthcare, agri-tech, and financial inclusion with funding that runs from data curation all the way to deployment.

IndiaAI is a well-structured starting point, but its impact will hinge on “velocity, usability, and signal clarity,” Gogia said. While the Mission lays out the right pillars across compute, data, fellowships and startup support, the real test will be whether these translate into tangible traction.

IndiaAI is trying its Best 

“Our ecosystem is still too fragmented,” said Ashutosh Singh, co-founder and CEO of RevRag.AI, explaining why India still struggles with frontier AI research. “India has strong talent, but lacks the density, long-horizon R&D funding, and tightly integrated research groups needed for frontier breakthroughs. Our ecosystem is still too fragmented.”

At the Tech Summit, Abhishek Singh said that the government is scaling up compute capacity, foundation model development and skilling initiatives under the national AI Mission, as it prepares for larger investments in the coming months. Singh said the Mission has acquired 38,000 GPUs, bringing down the effective cost to ₹65 per GPU/hour after subsidies. 

He added that teams from BharatGen, IISc Bengaluru and IIIT-Hyderabad are also progressing on their respective models. “For foundation models, we provide 100% of the compute support that is required,” Singh said, adding that the goal is to build models trained on Indian datasets to reduce dependence on foreign models.

Notably, BharatGen has secured 13,640 H100 GPUs and close to ₹1,000 crore in funding, the single-largest allocation in the country. It already has a series of early releases under its belt — Param-1, a bilingual 2.9-billion-parameter model, Shrutam for speech recognition, and Patram, a vision-language model for document understanding. But scaling to a trillion parameters is a different order of challenge.

The Next Moves India Cannot Miss

Joseph said that for India to catch up would require more than hardware and infrastructure. India needs to give researchers the freedom to run ambitious, high-risk experiments “without friction,” and to build hubs where talent, compute, and capital come together with a shared purpose. “That’s how global labs operate,” he said, adding that building that kind of rhythm will take time.

Meanwhile, under the FutureSkills pillar, IndiaAI is offering fellowships to undergraduate, postgraduate and research students from all disciplines who take up AI projects. 

“These fellowships are not limited to only engineering or science students,” Singh said, adding that they also extend to fields such as medicine, law, commerce and liberal arts. The Mission is also exploring partnerships with industry to expand training programmes. 

MeitY, under the IndiaAI Mission, has also launched ‘YUVA AI for ALL’, “a first-of-its-kind free course that introduces the world of AI to all Indians, especially the youth.”

Gogia said that India must become “the centre of gravity for AI researchers, not their backup plan.” He argued that this requires far more than fellowships. It demands world-class labs, stable infrastructure, academic freedom, and career paths that don’t push talent abroad. 

If India can move these pieces together, Gogia said, it can run its own race; if not, it risks becoming merely a customer in someone else’s system.

The post ‘IndiaAI Does Not Need More Policy PDFs or GPU Ribbon Cuttings’ appeared first on Analytics India Magazine.

Generated by RSStT. The copyright belongs to the original author.

Source