Adithya Kolavi, who witnessed DeepSeek’s destructive language model being unveiled to the public earlier this year, in Bengaluru, India, felt a mix of pleasure and confirmation. The Chinese tech rivaled the best of the West in terms of measures, but it had been built with far less money in much less period. &nbsp,

” I thought:’ This is how we undermine with less,'” says Kolavi, the 20-year-old chairman of the Indian AI company CognitiveLab. Why not us if DeepSeek could do it?

But for Abhishek Upperwal, chairman of Soket AI Labs and designer of one of India’s earliest efforts to develop a base model, the instant felt more poignant. &nbsp,

Upperwal’s unit, called Pragna-1B, had struggled to stay afloat with little offers while he watched world peers raise million. The bilingual model had a relatively low 1.25 billion parameters and was intended to lower the “language tax,” the additional costs that come with India having a large number of languages to help, in contrast to the US or even China. His staff had trained it, but limited tools meant it don’t level. He claims that as a result, the project evolved from a strategy to a solution. &nbsp,

There’s a great probability that we would have been the ones building what DeepSeek only released, he says if we had been funded two years ago.

Kolavi’s joy and Upperwal’s despair reflect the range of feelings among India’s AI builders. Despite its reputation as a global technology hub, China and the US both have strong advantages when it comes to developing AI. That space has opened mainly because India has severely underinvested in R&amp, D, organizations, and technology. Teaching language models is much more challenging than it is somewhere because no one local dialect is spoken by the majority of the people. &nbsp,

India’s software habitat evolved with a services-first thinking, and it has long been known as the world back office for the application business. Giants like Infosys and TCS built their achievements on successful technology supply, but introduction was neither prioritized nor rewarded. In contrast, India’s R&amp, D spending remained at 0.65 % of GDP in 2024, far behind China’s 2. 68 % ($ 476.2 billion ) and US$ 3.5 % ($ 962.3 billion ). The muscles to build and market strong software, from techniques to cards, was just never built.

Within federal organizations like the DRDO and ISRO ( Indian Space Research Organization ), which are isolated hands of world-class research, it is rare that their discoveries make it into human or business use. India lacks the bridge to join risk-taking study to industrial processes, the approach DARPA does in the US. In addition, a large portion of India’s top skill moves overseas, becoming drawn to ecosystems that can realize strong technology and, crucially, finance it.

So when the open-source basis model DeepSeek-R1 immediately outperformed some global contemporaries, it struck a nerve. Indian policymakers had to fight the country’s growing AI system gap and the need to immediately address it as a result of this launch by a Chinese startup.

India listens

The Ministry of Electronics and Information Technology ( MeitY ) contacted candidates for India’s own foundation models in January 2025, which are large AI models that can be adapted for a variety of tasks. Its common sensitive invited private-sector fog and data‑center companies to supply GPU determine capacity for government‑led Artificial research. &nbsp,

Companies including Jio, Yotta, E2E Networks, Tata, AWS companions, and CDAC responded. Through this agreement, MeitY was able to acquire almost 19 000 GPUs at reduced prices, which were then repurposed from private infrastructure and given to basic AI projects exclusively. This triggered a wave of ideas from businesses wanting to create their own versions. &nbsp,

Within two weeks, it had 67 ideas in side. By the middle of March, that amount tripled. &nbsp,

By the end of 2025, the government announced plans to build six large-scale designs in April, as well as 18 more AI software aimed at addressing issues like crops, education, and climate change. Most notably, it tapped Sarvam AI to build a 70-billion-parameter model optimized for Indian languages and needs. &nbsp,

For a nation long restricted by limited research infrastructure, things moved at record speed, marking a rare convergence of ambition, talent, and political will.

” India could do a Mangalyaan in AI,” said IIIT-Delhi’s Gautam Shroff, citing the nation’s successful and cost-effective Mars orbiter mission. &nbsp,

Jaspreet Bindra, cofounder of AI&amp, Beyond, a nonprofit that aims to teach AI literacy, captured the urgency:” DeepSeek is probably the best thing that happened in India. It gave us a kick in the backside to stop talking and start doing something”.

The language issue

One of the most fundamental challenges in building foundational AI models for India is the country’s sheer linguistic diversity. India poses a problem that few existing LLMs are able to handle because of its 22 official languages, hundreds of dialects, and millions of people who are multilingual.

Whereas a massive amount of high-quality web data is available in English, Indian languages collectively make up less than 1 % of online content. LLMs that understand how Indians actually speak or search are hampered by the lack of digitized, labeled, and cleaned data in languages like Bhojpuri and Kannada.

Global tokenizers, which break text into units a model can process, also perform poorly on many Indian scripts, misinterpreting characters or skipping some altogether. In consequence, even when Indian languages are included in multilingual models, they are frequently misunderstood and produced incorrectly.

And unlike OpenAI and DeepSeek, which achieved scale using structured English-language data, Indian teams often begin with fragmented and low-quality data sets encompassing dozens of Indian languages. This makes foundation modeling’s initial stages much more complicated.

Nonetheless, a small but determined group of Indian builders is starting to shape the country’s AI future.

For instance, Sarvam AI created OpenHathi-Hi-v0.1, an open-source Hindi language model that demonstrates how the Indian AI industry is developing its growing ability to address the country’s vast linguistic diversity. The model, built on Meta’s Llama 2 architecture, was trained on 40 billion tokens of Hindi and related Indian-language content, making it one of the largest open-source Hindi models available to date.

Upperwal’s multilingual model, Pragna-1B, provides more proof that India has a solution to its own linguistic complexity. Trained on 300 billion tokens for just$ 250, 000, it introduced a technique called “balanced tokenization” to address a unique challenge in Indian AI, enabling a 1.25-billion-parameter model to behave like a much larger one.

The issue is that Indian languages combine complex scripts with agglutinative grammar, where many smaller units of meaning are stringed together using prefixes and suffixes. Unlike English, which separates words with spaces and follows relatively simple structures, Indian languages like Hindi, Tamil, and Kannada often lack clear word boundaries and pack a lot of information into single words. Standard tokenizers have a problem with these inputs. They end up breaking Indian words into too many tokens, which bloats the input and makes it harder for models to understand the meaning efficiently or respond accurately.

However, according to Upperwal,” a billion-parameter model was equivalent to a 7 billion one like Llama 2″ with the new technique. This performance was particularly marked in Hindi and Gujarati, where global models often underperform because of limited multilingual training data. It served as a reminder that small teams can still push boundaries with smart engineering.

Upperwal eventually repurposed his core tech to build speech APIs for 22 Indian languages, a more immediate solution better suited to rural users who are often left out of English-first AI experiences.

Training a language model is just step one, he says, if the path to AGI is a hundred-step process. &nbsp,

Startups with more audacious goals are at the other end of the spectrum. Krutrim-2, for instance, is a 12-billion-parameter multilingual language model optimized for English and 22 Indian languages. &nbsp,

Krutrim-2 is attempting to solve India’s specific problems of linguistic diversity, low-quality data, and cost constraints. The team developed a unique Indic tokenizer, improved the training environment, and created models for multimodal and voice-first use cases right away, which is important in a nation where text interfaces can be challenging.

Krutrim’s bet is that its approach will not only enable Indian AI sovereignty but also offer a model for AI that works across the Global South.

India also requires the institutional support of talent, the depth of research, and the long-range capital to produce internationally competitive science, in addition to public funding and compute infrastructure.

While venture capital still hesitates to bet on research, new experiments are emerging. Lossfunk, a Bell Labs-style AI residency program designed to entice independent researchers with a taste for open-source science, is now personally funded by Paras Chopra, an entrepreneur who previously founded and sold the software-as-a-service company Wingify. &nbsp,

” We don’t have role models in industry or academia,” Chopra says. ” So we’re creating a space where top researchers can learn from each other and have startup-style equity upside”.

Put your money on sovereign AI.

The clearest marker of India’s AI ambitions came when the government selected Sarvam AI to develop a model focused on Indian languages and voice fluency.

The idea is that it would not only help Indian companies defeat the world’s AI adversaries but also benefit the country as a whole. ” If it becomes part of the India stack, you can educate hundreds of millions through conversational interfaces”, says Bindra. &nbsp,

Sarvam was given access to 4, 096 Nvidia H100 GPUs for training a 70-billion-parameter Indian language model over six months. ( Sarvam-1, a two-billion-parameter model trained in 10 Indian languages, was previously released by the company. )

Sarvam’s project and others are part of a larger strategy called the IndiaAI Mission, a$ 1.25 billion national initiative launched in March 2024 to build out India’s core AI infrastructure and make advanced tools more widely accessible. The mission of MeitY, which is led by MeitY, is centered on assisting AI startups, particularly those that are developing foundation models in Indian languages and applying AI to important fields like agriculture, education, and healthcare.

Under its compute program, the government is deploying more than 18, 000 GPUs, including nearly 13, 000 high-end H100 chips, to a select group of Indian startups that currently includes Sarvam, Upperwal’s Soket Labs, Gnani AIși Gan AI. &nbsp,

The mission also includes plans to launch a national multilingual data set repository, establish AI labs in smaller cities, and fund deep-tech R&amp, D. The broader goal is to equip Indian developers with the infrastructure needed to build globally competitive AI and ensure that the results are grounded in the linguistic and cultural realities of India and the Global South.

India’s wider push into deep tech is expected to generate about$ 12 billion in research and development investment over the next five years, according to Abhishek Singh, CEO of IndiaAI and officer with MeitY. &nbsp,

This includes approximately$ 62 million through the IndiaAI Mission, with approximately$ 32 million going toward direct startup funding. The National Quantum Mission is contributing another$ 730 million to support India’s ambitions in quantum research. A$ 1.2 billion Deep Tech Fund of Funds, which aim to spur early-stage innovation in the private sector, was also announced in the national budget document for 2025-26.

The rest, nearly$ 9.9 billion, is expected to come from private and international sources including corporate R&amp, D, venture capital firms, high-net-worth individuals, philanthropists, and global technology leaders such as Microsoft. &nbsp,

IndiaAI has now received more than 500 applications from startups proposing use cases in sectors like health, governance, and agriculture. &nbsp,

” We’ve already announced support for Sarvam, and 10 to 12 more startups will be funded solely for foundational models”, says Singh. Access to training data, talent depth, sector fit, and scalability are among the selection criteria.

Open or closed?

However, there are some issues with the IndiaAI program. Sarvam is being built as a closed model, not open-source, despite its public tech roots. That has sparked debate about the proper balance between the public good and private enterprise. &nbsp,

According to Amlan Mohanty, an AI policy specialist,” true sovereignty should be rooted in openness and transparency.” He points to DeepSeek-R1, which despite its 236-billion parameter size was made freely available for commercial use. &nbsp,

Its release allowed developers around the world to fine-tune it on low-cost GPUs, creating faster variants and extending its capabilities to non-English applications.

Hancheng Cao, an assistant professor of information systems and operations management at Emory University, believes that the release of an open-weight model that allows for efficient inference can revolutionize AI. ” It makes it usable by developers who don’t have massive infrastructure”.

However, IndiaAI has not stated whether publicly funded models should be made open-source. &nbsp,

According to Singh,” We didn’t want to dictate business models.” ” India has always supported open standards and open source, but it’s up to the teams. Whatever the route, the aim is strong Indian models.

There are other challenges as well. Sarvam AI unveiled Sarvam M, a 24-billion-parameter multilingual LLM built on top of Mistral Small, a successful model created by the French company Mistral AI, in late May. Sarvam’s cofounder Vivek Raghavan called the model” an important stepping stone on our journey to build sovereign AI for India”. However, its download numbers were underwhelming, only 300 in the initial two days. The venture capitalist Deedy Das called the launch “embarrassing”.

Beyond the lukewarm early reception, the issues extend. Many developers in India still lack easy access to GPUs and the broader ecosystem for Indian-language AI applications is still nascent. &nbsp,

The compute question

generative AI’s most pressing bottleneck is generative scarcity, not just in India but all over the world. For countries still heavily reliant on imported GPUs and lacking domestic fabrication capacity, the cost of building and running large models is often prohibitive. &nbsp,

India still imports most of its chips rather than producing them domestically, and training large models remains expensive. For this reason, startups and researchers are focusing on software-level efficiency, which involves smaller models, better inference, and fine-tuning frameworks to maximize performance on fewer GPUs.

” The absence of infrastructure doesn’t mean the absence of innovation”, says Cao. Supporting optimization science is a wise way to operate within constraints.

Yet Singh of IndiaAI argues that the tide is turning on the infrastructure challenge thanks to the new government programs and private-public partnerships. He claims that he believes that we will no longer be dealing with the same kind of compute bottlenecks as we did last year.

India also has a cost advantage.

Gupta estimates that building a hyperscale data center in India would cost about$ 5 million, which is roughly half the price in markets like the US, Europe, and Singapore. That’s thanks to affordable land, lower construction and labor costs, and a large pool of skilled engineers. &nbsp,

For now, India’s AI ambitions seem less about leapfrogging OpenAI or DeepSeek and more about strategic self-determination. The nation is betting that it can chart its own course, whether its approach takes the form of smaller sovereign models, open ecosystems, or public-private hybrids. &nbsp,

While some experts contend that the government’s action, or reaction ( to DeepSeek ) is performative and in line with its nationalistic agenda, many startup founders are upbeat. They see the growing collaboration between the state and the private sector as a real opportunity to overcome India’s long-standing structural challenges in tech innovation.

Nandan Nilekani, the chairman of Infosys, urged India to resist pursuing a me-too AI dream at a Meta summit held in Bengaluru last year. &nbsp,

He referred to building LLMs as “let the big boys in the Valley do it.” ” We will use it to create synthetic data, build small language models quickly, and train them using appropriate data” .&nbsp,

His argument that strength should be given precedence over spectacle received mixed reviews. But it reflects a broader growing consensus on whether India should play a different game altogether.

Bharath Reddy, a researcher at the Takshashila Institution, an Indian public policy nonprofit, says,” Trying to dominate every layer of the stack isn’t realistic, even for China.” ” Dominate one layer, like applications, services, or talent, so you remain indispensable” .&nbsp,