AI superpark in Bengaluru to generate data for world models

Bharat1.ai, a Bengaluru-based AI firm, is launching an AI research and innovation park in a 70-acre campus that will house AI researchers, startups, enterprises and academic institutions in partnership with Nvidia.

Umakant Soni, co-founder, said the idea of the park is to create AI models grounded in the real world and the park is equipped with robots, GPUs and researchers, which will create data for the real world.

“The biggest challenge with large language models (LLMs) is that they are probabilistic in nature and do not have a sense of the real world. It is important to create AI models that are grounded in real world situations,” he said.


The total cost of the project is estimated to be $650 million, of which $250 million will be in equity and $400 million in debt, said Soni and added that they are in talks with institutions for funding.

This is coming at a time when companies are looking at an alternative to LLMs that are based on transformer architecture. Industry experts such as Yann LeCun are advocating World Models that can take sensory inputs from the real world given the limitations LLMs present.

While LLMs are good with text and reasoning tasks such as coding, LeCun, in his keynote at the AI Summit’s research symposium in New Delhi, said that they do not work for high-dimensional continuous data such as image, video or sensor data. He pointed out that you need systems that can learn from sensor inputs.

Soni concurred that at this juncture building agents on top of them has drawbacks. This is the idea behind creating AI Superpark.

In the first phase, it will start at 500,000 square feet and create 1 exabyte of data every year, and will be powered by a combination of solar power and nuclear power, through small modular reactors. It will eventually house 25,000 AI researchers and engineers with 4 million square feet of commercial R&D space for labs, product companies, startups and applied research.

Nvidia will co-create an ecosystem with their computing software stack, compute infrastructure and domain expertise.