Everyone hopes to build an AI startup. After leaving OpenAI, Ilya Sutskever has started his own venture called Safe Superintelligence. The startup is bound to raise billions of dollars in pursuing its goal of building ASI, but what about Indian startups that are aiming to do fundamental AI research?
Speaking with AIM, Soket AI Labs founder & CEO Abhishek Upperwal revealed the numbers required to build a research-based AI startup. So far, the company has already built Pragna-1B, which is a foundational model specifically for Indic languages.
Upperwal said that, currently, funding is just enough to make-do for AI research within a startup. “Yes, there are fewer funds that are available as compared to any foreign markets, but I also believe that we can maybe make-do with that particular fund and then ultimately grow in scale after the seed stage,” said Upperwal.
He explained that for the seed stage in India, a funding of $5 million or $10 million is still a decent amount. This roughly translates to around INR 40-50 crores. This is still very little when compared to the 100s of millions raised by companies in the West.
“If VCs can trust these companies in the generative AI space, we can definitely do wonderful stuff for sure,” added Upperwal.
Where Does the Money Go?
The gap in funding is because of the market. India is a smaller market, therefore the ticket sizes are still way smaller compared to the West. But Upperwal said that the ticket sizes being small pose a problem when compared to the work that startups have to do at the foundational layer.
Upperwal gives the example of building an estimated 7 billion parameter foundation model out of India. He said that the cost for the compute alone would be close to $2 million. For reference, one NVIDIA H100 costs around INR 30 lakhs, or $36k.
To build a 7 billion parameter model, considering a six month time frame, a startup would require at least a dozen NVIDIA GPUs for the training period, taking into account time for other factors like inefficiencies.
This is while considering that the model is built in one shot. “It takes a lot of experiments, and checkpoints, or the path that you are taking fails, so you need to rebuild from the previous checkpoint,” explained Upperwal.
Earlier, Upperwal had told AIM that it took the company six months to train the Pragna model, which involved many experiments with different models and a total of 150 billion tokens. It took close to 8000 GPU hours on NVIDIA A100s to train the model.
Accounting for all of these, an ideal amount to do a lot of foundational work in AI at the seed stage is anywhere around USD $7-15 million, which is close to around INR 125 crore.
All of this is including the cost of running the business such as hiring talent and paying bills, and does not include the cost of making the models ready for production or inference. That would increase the funding requirement to, at least, more than double.
In the same conversation, Speciale Invest founder and partner Arjun Rao said that Indian VCs are interested in investing in the development phase, more than the research phase of AI. It would take a lot of time to research, build a model, compare it with others, and then figure out how it can be commoditised, which is something VCs are still figuring out.
Assuming they’re working with a team of eight people and a funding of approximately $10 million (or INR 82 crore), an AI startup in India can survive for about 2.33 years at the current estimated monthly expenditure of INR 2.92 crore, which does not include inference costs.
Assuming that the $2 million as Upperwal mentioned above goes towards compute just for training a 7 billion model for around six months, it still would account for 80.2% of the expenditure for 2 years, with the rest going towards salaries and other expenditure.
This calculation assumes that the monthly costs remain constant and does not account for potential increases in expenses due to scaling, inflation, or other operational changes.
How Much Are Indian Startups Getting?
While these numbers seem reasonable, they are comparatively lower when looking at the global standard set by OpenAI, Anthropic, or Mistral, who have raised billions of dollars.
For reference, Sarvam AI, which has announced its intention of building foundational AI models, has raised a total of $41 million. Pranav Mistry-led and Reliance-backed TWO.AI has raised $20 million for building their SUTRA line of models, and Krutrim raised $50 million becoming India’s first generative AI unicorn.
This is just for initial research. SML CEO and founder Vishnu Vardhan emphasised the huge investment required to build and scale complex AI models. In an exclusive interview with AIM, Vardhan disclosed his plans of raising $200-300 million for the same. The company only recently launched Hanooman, its own foundational model.
“That’s the kind of money we need to launch this kind of a product. We’ve already spent tens of millions of dollars, but that won’t work,” he said, about building a GPT-5 level model in India.
Even if you add the total funds raised by companies in India, it’s still nothing when compared to OpenAI raising billions single-handedly. It definitely would cost a lot more to build an AI startup in India to compete with the West.
