So, Sam Altman Was Right About Indian AI Startups

When OpenAI CEO Sam Altman arrived in India in 2023, the room was full. Founders, engineers, and policymakers sat shoulder to shoulder, driven by the same belief: India could build its own large AI models. The mood was hopeful. Confident. Almost restless.

Then came the line that no one forgot.

“It’s totally hopeless to compete with us on training foundation models,” he said.

Altman was not trying to provoke. He was speaking as the head of OpenAI, the lab training frontier models at a scale few could imagine. Thousands of GPUs, budgets that looked more like national projects than startup bets. But in that room, the comment felt dismissive. Even patronising. 

The pushback was quick. Some Indian founders challenged him openly. Others took it as fuel.

Two years later, the anger has cooled. The reality has not.

Altman later softened his words. Now, he has said that India can be a full-stack AI leader, while speaking to TOI. Training frontier models, he argued, was costly even then. Countries like India would do better by building full-stack sovereign capabilities instead. 

Around the same time, India’s response gathered pace. Tech Mahindra backed Project Indus and delivered a model within two months, spending less than $5 million. The IndiaAI Mission has now lined up 12 firms to work on foundational models, with early results expected at the India AI Impact summit. 

Notably, Sarvam, BharatGen, Gnani and Tech Mahindra, among others, are expected to release models this week. However, none of them focuses on competing with GPT on a global scale, but instead on use cases.

Things were moving on the surface. But underneath, the tone had changed. This year, the Economic Survey of India delivered a rare moment of honesty. 

“Attempting to close this gap would involve prohibitive fiscal costs towards what is increasingly becoming an unsustainable approach to AI development,” read the survey, referring to the frontier models competing with Western firms.

The Hard Truth?

In plain terms, trying to outspend Silicon Valley or Beijing may simply not be practical. That realism is now showing up in how founders talk about their own strategies.

Mansoor Rahimat Khan, co-founder and CEO of Beatoven.ai, does not mince words. Looking back, he says, Altman was largely right. “There was unnecessary sensationalism but no tangible outcomes from India in the foundation AI space due to a lack of research and funding,” he tells AIM.

He argues that the opportunity lies elsewhere. “Building specific applications for India-specific problems is the way forward. It is already too late to even be able to compete with the US and Chinese foundation models. Given the pace at which our bureaucracy and VCs move when it comes to fundamental research, it is increasingly evident to… compete on the application layer.”

His view reflects what many younger startups have discovered the hard way. Training large models is not only expensive. It is slow, risky, and demands deep research talent that takes years to build. Most startups do not have the time or capital to wait that long.

Customers, on the other hand, are not waiting. They want products that work today.

Ankush Sabharwal, co-founder and CEO of CoRover and the mind behind BharatGPT, agrees, but for a different reason. The problem, he says, is not just cost. It is purpose. 

“I always believed that focused, purpose-led models are the way forward for India’s AI journey. We should focus on developing specialised models that address specific domains, problems and needs,” he tells AIM.

He frames it with a simple analogy. “When we need to get a job done, we’d rather hire someone with one specialised degree (a Master’s or PhD) than someone with 10 PhDs.” In the same way, he explains, building mega models to do everything often does very little well. In India’s case, focus matters more than grandeur.

Instead of chasing giant models, Sabharwal suggests stitching together many small ones. He says that India should focus on solving real-world problems and then scaling its models and solutions via the Mixture of Experts (MoE) approach or with a battery of small models—each specialised in a specific domain, operated by an intelligent router model adopter.

His stance is pragmatic rather than ideological. Let others build the giants, he says. Someone should. India’s strength lies in building models that are good enough to solve real problems.

The country no longer seems obsessed with winning the race for the biggest AI model in the market. Instead, it is asking a better question. Does it even need to run that race at all?

Sarvam’s trajectory offers a telling example. Over the past year, the company faced criticism since the launch of Sarvam-M, a model built on top of Mistral with added Indic tokens. Some dismissed it, calling it “embarrassing”. But today, Sarvam sits at the centre of India’s sovereign AI conversation.

Instead of chasing English language benchmarks, the company has gone deep into areas most frontier labs overlook. In recent days, the company has made several announcements, like text-to-speech through Bulbul 3, speech-to-text with Saaras, as well as vision models and ASR models capable of reading and speaking 22 Indian languages. 

Sarvam is building natively for India, something the IndiaAI Mission has been encouraging among the 12 startups selected to build foundational models. Sarvam has finally demonstrated the importance of Indic AI models, and the use cases may now finally come to fruition.

What Not Everyone is Ready to Accept

Still, not everyone is ready to step away from the frontier just yet.

Vishnu Subramanian, founder and CEO of Jarvislabs.ai, now part of E2E Networks, looks at the challenge through a technical lens.

“If the question is whether we can compete with a company training a frontier model with $10 million and with just three engineers—the question that was asked to Sam Altman a few years ago—I think it will still be difficult in 2026,” he tells AIM.

His reasoning is simple. The goalposts keep moving. Frontier models are not fixed milestones. Every year, one or two new variants of the models are released, each displaying sharp jumps in capability. “And it’s not just the models; there is also a deployment aspect to it. Deployment alone is becoming very expensive, and to capture the market is also going to be very expensive.”

This matter’s in India’s context. In the Union Budget, the IndiaAI Mission was allocated just ₹1,000 crore for the year, even as inference costs continue to rise once models are deployed in the real world.

Subramanian explains that even if $10 million were allocated to a research institute, the likely outcome would be a solid open source model, but not one that competes with proprietary models. He points to DeepSeek as an example. While the company claimed its model cost only $6 million to train, a closer look shows it had access to one of the largest A100 clusters in China. 

Yet, he stops short of saying India should not try at all. “Ideally, we should build frontier models. That should be one of the end goals, but we have to understand that it is a journey.”

His advice is to start small. For example, training a one-billion-parameter model doesn’t require the same level of computing as a 100- or 200-billion-parameter model. “Experimenting on a one-billion model is comparatively easier, and we should evaluate any models that we build on a global leaderboard. This would create a path to create frontier models,” he explains.

In other words, learn to walk before trying to sprint.

Prashanth Kaddi, partner for engineering, AI and data at Deloitte, pushes back on the idea that India should simply abandon model development and rely entirely on Western systems.

“We need at least significant fine-tuning of those models, if not building models specific to India, because we have a cultural context and a language context that is quite unique. We cannot just lift and use models trained on data that is predominantly Western in context,” he explains.

His argument sits between ambition and restraint. Instead of pouring billions into frontier bets, he suggests starting with adaptation.

“Maybe we can start small. Several people have taken open-source models and customised them for local languages. Maybe we can start on that path—build some SLMs, for instance.” Frontier models, he suggests, can remain a longer-term goal.

For Kaddi, timing matters. India has often skipped one phase of infrastructure and jumped straight to the next. Landlines never really took off before mobile phones arrived. Perhaps AI will follow a similar path.

“One of the things we tend to do in India is wait for that leapfrog moment, when we can move quickly… So maybe, as a culture, we are waiting to ensure that this is real.”

There is Real Cause for Hope

Nikhil Malhotra, chief innovation officer at Tech Mahindra and global head of Makers Lab—part of the IndiaAI Mission—sees the period after Altman’s moment as a turning point for the ecosystem itself.

“The India AI Summit is a culmination of the significant momentum built across the Indian AI ecosystem. These initiatives have helped catalyse the ecosystem by accelerating skilling, model development, and demonstrating India’s capability to lead in applied AI,” he tells AIM.

What became clear is that while large, general‑purpose models have value, India’s real needs are better served by domain‑specific, efficient, and locally relevant AI systems. “The diffusion of AI across sectors, rather than a singular focus on massive models, will be critical for India’s growth story. The idea that AI will become an invisible digital infrastructure for every citizen is what makes this initiative truly transformative,” he adds.

Therefore, two major shifts stand out to Malhotra. First, a sharper focus on India’s AI strengths and capabilities. Second, greater visibility into where and how these models are being applied across.

He has long argued that AI is less about showpiece models and more about invisible plumbing. Project Indus, which is set to launch another bigger model focusing on education at the summit later this month, follows that thinking. 

“When generic LLMs interact with agents, the experience can often become inconsistent. Smaller, more focused models built for a specific purpose will work far better with agents. In short, India’s approach should prioritise context‑rich, efficient, and purpose‑built AI architectures rather than competing on size,” he says.

Across these conversations, a pattern emerges. Very few founders are still talking about beating OpenAI at its own game. Most are talking about where India can win without trying to replicate Silicon Valley’s spending power.

There is a quiet acceptance that foundation models have become capital-heavy infrastructure plays. They resemble power plants or telecom networks more than scrappy startups. Only a handful of players in the world can afford that race.

India’s strength may lie elsewhere. Distribution, languages, domain knowledge, and millions of real-world use cases in government, agriculture, healthcare, and small business.

Altman’s comment might have once felt like an insult. Today, however, it reads more like a constraint.

And constraints, history shows, often shape better products. In that sense, the question then is no longer whether Altman was right. It is whether India needed to hear it. Judging by the shift in strategy, the answer appears to be yes.