Inside India’s scramble for AI independence

In Bengaluru, India, Adithya Kolavi felt a mix of excitement and validation as he watched DeepSeek unleash its disruptive language model on the world earlier this year. The Chinese technology rivaled the best of the West in terms of benchmarks, but it had been built with far less capital in far less time. 

“I thought: ‘This is how we disrupt with less,’” says Kolavi, the 20-year-old founder of the Indian AI startup CognitiveLab. “If DeepSeek could do it, why not us?” 

But for Abhishek Upperwal, founder of Soket AI Labs and architect of one of India’s earliest efforts to develop a foundation model, the moment felt more bittersweet. 

Upperwal’s model, called Pragna-1B, had struggled to stay afloat with tiny grants while he watched global peers raise millions. The multilingual model had a relatively modest 1.25 billion parameters and was designed to reduce the “language tax,” the extra costs that arise because India—unlike the US or even China—has a multitude of languages to support. His team had trained it, but limited resources meant it couldn’t scale. As a result, he says, the project became a proof of concept rather than a product. 

“If we had been funded two years ago, there’s a good chance we’d be the ones building what DeepSeek just released,” he says.

Kolavi’s enthusiasm and Upperwal’s dismay reflect the spectrum of emotions among India’s AI builders. Despite its status as a global tech hub, the country lags far behind the likes of the US and China when it comes to homegrown AI. That gap has opened largely because India has chronically underinvested in R&D, institutions, and invention. Meanwhile, since no one native language is spoken by the majority of the population, training language models is far more complicated than it is elsewhere. 

Historically known as the global back office for the software industry, India has a tech ecosystem that evolved with a services-first mindset. Giants like Infosys and TCS built their success on efficient software delivery, but invention was neither prioritized nor rewarded. Meanwhile, India’s R&D spending hovered at just 0.65% of GDP ($25.4 billion) in 2024, far behind China’s 2.68% ($476.2 billion) and the US’s 3.5% ($962.3 billion). The muscle to invent and commercialize deep tech, from algorithms to chips, was just never built.

Isolated pockets of world-class research do exist within government agencies like the DRDO (Defense Research & Development Organization) and ISRO (Indian Space Research Organization), but their breakthroughs rarely spill into civilian or commercial use. India lacks the bridges to connect risk-taking research to commercial pathways, the way DARPA does in the US. Meanwhile, much of India’s top talent migrates abroad, drawn to ecosystems that better understand and, crucially, fund deep tech.

So when the open-source foundation model DeepSeek-R1 suddenly outperformed many global peers, it struck a nerve. This launch by a Chinese startup prompted Indian policymakers to confront just how far behind the country was in AI infrastructure, and how urgently it needed to respond.

India responds

In January 2025, 10 days after DeepSeek-R1’s launch, the Ministry of Electronics and Information Technology (MeitY) solicited proposals for India’s own foundation models, which are large AI models that can be adapted to a wide range of tasks. Its public tender invited private-sector cloud and data‑center companies to reserve GPU compute capacity for government‑led AI research. 

Providers including Jio, Yotta, E2E Networks, Tata, AWS partners, and CDAC responded. Through this arrangement, MeitY suddenly had access to nearly 19,000 GPUs at subsidized rates, repurposed from private infrastructure and allocated specifically to foundational AI projects. This triggered a surge of proposals from companies wanting to build their own models. 

Within two weeks, it had 67 proposals in hand. That number tripled by mid-March. 

In April, the government announced plans to develop six large-scale models by the end of 2025, plus 18 additional AI applications targeting sectors like agriculture, education, and climate action. Most notably, it tapped Sarvam AI to build a 70-billion-parameter model optimized for Indian languages and needs. 

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 Gautam Shroff of IIIT-Delhi, referencing the country’s cost-effective, and successful, Mars orbiter mission. 

Jaspreet Bindra, cofounder of AI&Beyond, an organization focused on teaching AI literacy, captured the urgency: “DeepSeek is probably the best thing that happened to India. It gave us a kick in the backside to stop talking and start doing something.”

The language problem

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

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

Global tokenizers, which break text into units a model can process, also perform poorly on many Indian scripts, misinterpreting characters or skipping some altogether. As a result, even when Indian languages are included in multilingual models, they’re often poorly understood and inaccurately generated.

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 the early steps of training foundation models far more complex.

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

For example, Sarvam AI has created OpenHathi-Hi-v0.1, an open-source Hindi language model that shows the Indian AI field’s 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.

Pragna-1B, the multilingual model from Upperwal, is more evidence that India could solve for 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 use complex scripts and agglutinative grammar, where words are formed by stringing together many smaller units of meaning 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 struggle with such 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.

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

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.

“If the path to AGI is a hundred-step process, training a language model is just step one,” he says. 

At the other end of the spectrum are startups with more audacious aims. Krutrim-2, for instance, is a 12-billion-parameter multilingual language model optimized for English and 22 Indian languages. 

Krutrim-2 is attempting to solve India’s specific problems of linguistic diversity, low-quality data, and cost constraints. The team built a custom Indic tokenizer, optimized training infrastructure, and designed models for multimodal and voice-first use cases from the start, crucial in a country where text interfaces can be a problem.

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.

Besides public funding and compute infrastructure, India also needs the institutional support of talent, the research depth, and the long-horizon capital that produce globally competitive science.

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

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

Government-backed bet 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 compete in the global AI arms race but benefit the wider population as well. “If it becomes part of the India stack, you can educate hundreds of millions through conversational interfaces,” says Bindra. 

Sarvam was given access to 4,096 Nvidia H100 GPUs for training a 70-billion-parameter Indian language model over six months. (The company previously released a 2-billion-parameter model trained in 10 Indian languages, called Sarvam-1.)

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. Led by MeitY, the mission is focused on supporting AI startups, particularly those developing foundation models in Indian languages and applying AI to key sectors such as health care, education, and agriculture.

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 AI, and Gan AI. 

The mission also includes plans to launch a national multilingual data set repository, establish AI labs in smaller cities, and fund deep-tech R&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.

According to Abhishek Singh, CEO of IndiaAI and an officer with MeitY, India’s broader push into deep tech is expected to raise around $12 billion in research and development investment over the next five years. 

This includes approximately $162 million through the IndiaAI Mission, with about $32 million earmarked for direct startup funding. The National Quantum Mission is contributing another $730 million to support India’s ambitions in quantum research. In addition to this, the national budget document for 2025-26 announced a $1.2 billion Deep Tech Fund of Funds aimed at catalyzing early-stage innovation in the private sector.

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

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

“We’ve already announced support for Sarvam, and 10 to 12 more startups will be funded solely for foundational models,” says Singh. Selection criteria include access to training data, talent depth, sector fit, and scalability.

Open or closed?

The IndiaAI program, however, is not without controversy. 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 private enterprise and the public good. 

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

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.

“Releasing an open-weight model with efficient inference can democratize AI,” says Hancheng Cao, an assistant professor of information systems and operations management at Emory University. “It makes it usable by developers who don’t have massive infrastructure.”

IndiaAI, however, has taken a neutral stance on whether publicly funded models should be open-source. 

“We didn’t want to dictate business models,” says Singh. “India has always supported open standards and open source, but it’s up to the teams. The goal is strong Indian models, whatever the route.”

There are other challenges as well. In late May, Sarvam AI unveiled Sarvam‑M, a 24-billion-parameter multilingual LLM fine-tuned for 10 Indian languages and built on top of Mistral Small, an efficient model developed by the French company Mistral AI. Sarvam’s cofounder Vivek Raghavan called the model “an important stepping stone on our journey to build sovereign AI for India.” But its download numbers were underwhelming, with only 300 in the first two days. The venture capitalist Deedy Das called the launch “embarrassing.”

And the issues go beyond the lukewarm early reception. Many developers in India still lack easy access to GPUs and the broader ecosystem for Indian-language AI applications is still nascent. 

The compute question

Compute scarcity is emerging as one of the most significant bottlenecks in generative AI, not just in India but across the globe. For countries still heavily reliant on imported GPUs and lacking domestic fabrication capacity, the cost of building and running large models is often prohibitive. 

India still imports most of its chips rather than producing them domestically, and training large models remains expensive. That’s why startups and researchers alike are focusing on software-level efficiencies that involve smaller models, better inference, and fine-tuning frameworks that optimize for performance on fewer GPUs.

“The absence of infrastructure doesn’t mean the absence of innovation,” says Cao. “Supporting optimization science is a smart way to work 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. “I believe that within the next three months, we will no longer face the kind of compute bottlenecks we saw last year,” he says.

India also has a cost advantage.

According to Gupta, building a hyperscale data center in India costs about $5 million, roughly half what it would cost in markets like the US, Europe, or Singapore. That’s thanks to affordable land, lower construction and labor costs, and a large pool of skilled engineers. 

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

While some experts argue that the government’s action, or reaction (to DeepSeek), is performative and aligned with its nationalistic agenda, many startup founders are energized. 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.

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

“Let the big boys in the Valley do it,” he said of building LLMs. “We will use it to create synthetic data, build small language models quickly, and train them using appropriate data.” 

His view that India should prioritize strength over spectacle had a divided reception. But it reflects a broader growing consensus on whether India should play a different game altogether.

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

Correction: We amended Reddy’s name

How generative AI could help make construction sites safer

Last winter, during the construction of an affordable housing project on Martha’s Vineyard, Massachusetts, a 32-year-old worker named Jose Luis Collaguazo Crespo slipped off a ladder on the second floor and plunged to his death in the basement. He was one of more than 1,000 construction workers who die on the job each year in the US, making it the most dangerous industry for fatal slips, trips, and falls.

“Everyone talks about [how] ‘safety is the number-one priority,’” entrepreneur and executive Philip Lorenzo said during a presentation at Construction Innovation Day 2025, a conference at the University of California, Berkeley, in April. “But then maybe internally, it’s not that high priority. People take shortcuts on job sites. And so there’s this whole tug-of-war between … safety and productivity.”

To combat the shortcuts and risk-taking, Lorenzo is working on a tool for the San Francisco–based company DroneDeploy, which sells software that creates daily digital models of work progress from videos and images, known in the trade as “reality capture.”  The tool, called Safety AI, analyzes each day’s reality capture imagery and flags conditions that violate Occupational Safety and Health Administration (OSHA) rules, with what he claims is 95% accuracy.

That means that for any safety risk the software flags, there is 95% certainty that the flag is accurate and relates to a specific OSHA regulation. Launched in October 2024, it’s now being deployed on hundreds of construction sites in the US, Lorenzo says, and versions specific to the building regulations in countries including Canada, the UK, South Korea, and Australia have also been deployed.

Safety AI is one of multiple AI construction safety tools that have emerged in recent years, from Silicon Valley to Hong Kong to Jerusalem. Many of these rely on teams of human “clickers,” often in low-wage countries, to manually draw bounding boxes around images of key objects like ladders, in order to label large volumes of data to train an algorithm.

Lorenzo says Safety AI is the first one to use generative AI to flag safety violations, which means an algorithm that can do more than recognize objects such as ladders or hard hats. The software can “reason” about what is going on in an image of a site and draw a conclusion about whether there is an OSHA violation. This is a more advanced form of analysis than the object detection that is the current industry standard, Lorenzo claims. But as the 95% success rate suggests, Safety AI is not a flawless and all-knowing intelligence. It requires an experienced safety inspector as an overseer.  

A visual language model in the real world

Robots and AI tend to thrive in controlled, largely static environments, like factory floors or shipping terminals. But construction sites are, by definition, changing a little bit every day. 

Lorenzo thinks he’s built a better way to monitor sites, using a type of generative AI called a visual language model, or VLM. A VLM is an LLM with a vision encoder, allowing it to “see” images of the world and analyze what is going on in the scene. 

Using years of reality capture imagery gathered from customers, with their explicit permission, Lorenzo’s team has assembled what he calls a “golden data set” encompassing tens of thousands of images of OSHA violations. Having carefully stockpiled this specific data for years, he is not worried that even a billion-dollar tech giant will be able to “copy and crush” him.

To help train the model, Lorenzo has a smaller team of construction safety pros ask strategic questions of the AI. The trainers input test scenes from the golden data set to the VLM and ask questions that guide the model through the process of breaking down the scene and analyzing it step by step the way an experienced human would. If the VLM doesn’t generate the correct response—for example, it misses a violation or registers a false positive—the human trainers go back and tweak the prompts or inputs. Lorenzo says that rather than simply learning to recognize objects, the VLM is taught “how to think in a certain way,” which means it can draw subtle conclusions about what is happening in an image. 

COURTESY DRONEDEPLOY

As an example, Lorenzo says VLMs are much better than older methods at analyzing ladder usage, which is responsible for 24% of the fall deaths in the construction industry. 

“With traditional machine learning, it’s very difficult to answer the question of ‘Is a person using a ladder unsafely?’” says Lorenzo. “You can find the ladders. You can find the people. But to logically reason and say ‘Well, that person is fine’ or ‘Oh no, that person’s standing on the top step’—only the VLM can logically reason and then be like, ‘All right, it’s unsafe. And here’s the OSHA reference that says you can’t be on the top rung.’”

Answers to multiple questions (Does the person on the ladder have three points of contact? Are they using the ladder as stilts to move around?) are combined to determine whether the ladder in the picture is being used safely. “Our system has over a dozen layers of questioning just to get to that answer,” Lorenzo says. DroneDeploy has not publicly released its data for review, but he says he hopes to have his methodology independently audited by safety experts.  

The missing 5%

Using vision language models for construction AI shows promise, but there are “some pretty fundamental issues” to resolve, including hallucinations and the problem of edge cases, those anomalous hazards for which the VLM hasn’t trained, says Chen Feng. He leads New York University’s AI4CE lab, which develops technologies for 3D mapping and scene understanding in construction robotics and other areas. “Ninety-five percent is encouraging—but how do we fix that remaining 5%?” he asks of Safety AI’s success rate.

Feng points to a 2024 paper called “Eyes Wide Shut?”—written by Shengbang Tong, a PhD student at NYU, and coauthored by AI luminary Yann LeCun—that noted “systematic shortcomings” in VLMs.  “For object detection, they can reach human-level performance pretty well,” Feng says. “However, for more complicated things—these capabilities are still to be improved.” He notes that VLMs have struggled to interpret 3D scene structure from 2D images, don’t have good situational awareness in reasoning about spatial relationships, and often lack “common sense” about visual scenes.

Lorenzo concedes that there are “some major flaws” with LLMs and that they struggle with spatial reasoning. So Safety AI also employs some older machine-learning methods to help create spatial models of construction sites. These methods include the segmentation of images into crucial components and photogrammetry, an established technique for creating a 3D digital model from a 2D image. Safety AI has also trained heavily in 10 different problem areas, including ladder usage, to anticipate the most common violations.

Even so, Lorenzo admits there are edge cases that the LLM will fail to recognize. But he notes that for overworked safety managers, who are often responsible for as many as 15 sites at once, having an extra set of digital “eyes” is still an improvement.

Aaron Tan, a concrete project manager based in the San Francisco Bay Area, says that a tool like Safety AI could be helpful for these overextended safety managers, who will save a lot of time if they can get an emailed alert rather than having to make a two-hour drive to visit a site in person. And if the software can demonstrate that it is helping keep people safe, he thinks workers will eventually embrace it.  

However, Tan notes that workers also fear that these types of tools will be “bossware” used to get them in trouble. “At my last company, we implemented cameras [as] a security system. And the guys didn’t like that,” he says. “They were like, ‘Oh, Big Brother. You guys are always watching me—I have no privacy.’”

Older doesn’t mean obsolete

Izhak Paz, CEO of a Jerusalem-based company called Safeguard AI, has considered incorporating VLMs, but he has stuck with the older machine-learning paradigm because he considers it more reliable. The “old computer vision” based on machine learning “is still better, because it’s hybrid between the machine itself and human intervention on dealing with deviation,” he says. To train the algorithm on a new category of danger, his team aggregates a large volume of labeled footage related to the specific hazard and then optimizes the algorithm by trimming false positives and false negatives. The process can take anywhere from weeks to over six months, Paz says.

With training completed, Safeguard AI performs a risk assessment to identify potential hazards on the site. It can “see” the site in real time by accessing footage from any nearby internet-connected camera. Then it uses an AI agent to push instructions on what to do next to the site managers’ mobile devices. Paz declines to give a precise price tag, but he says his product is affordable only for builders at the “mid-market” level and above, specifically those managing multiple sites. The tool is in use at roughly 3,500 sites in Israel, the United States, and Brazil.

Buildots, a company based in Tel Aviv that MIT Technology Review profiled back in 2020, doesn’t do safety analysis but instead creates once- or twice-weekly visual progress reports of sites. Buildots also uses the older method of machine learning with labeled training data. “Our system needs to be 99%—we cannot have any hallucinations,” says CEO Roy Danon. 

He says that gaining labeled training data is actually much easier than it was when he and his cofounders began the project in 2018, since gathering video footage of sites means that each object, such as a socket, might be captured and then labeled in many different frames. But the tool is high-end—about 50 builders, most with revenue over $250 million, are using Buildots in Europe, the Middle East, Africa, Canada, and the US. It’s been used on over 300 projects so far.

Ryan Calo, a specialist in robotics and AI law at the University of Washington, likes the idea of AI for construction safety. Since experienced safety managers are already spread thin in construction, however, Calo worries that builders will be tempted to automate humans out of the safety process entirely. “I think AI and drones for spotting safety problems that would otherwise kill workers is super smart,” he says. “So long as it’s verified by a person.”

Andrew Rosenblum is a freelance tech journalist based in Oakland, CA.