AI firms want India to fast-track compute infra
At present, companies like Yotta, which procures GPUs from Nvidia, are providing chips on a subscription model at an hourly rate of $2.5. Artificial intelligence (AI) infrastructure and software companies like AMD, Nvidia, Ola Krutrim, Reliance Jio, Microsoft, People + AI, and Yotta have urged the government to expedite creating compute infrastructure in the country as well as focus on sovereignty in AI systems.
This assumes significance as the government is looking to create a compute infrastructure through procurement of graphic processing units (GPUs) for startups that are looking for a variety of these chips to train their models. In absence of any funding support from venture capitals, currently a lot of startups are not getting access to the right compute infra.
At present, companies like Yotta, which procures GPUs from Nvidia, are providing chips on a subscription model at an hourly rate of $2.5.
Comments from Aggarwal assume significance as the government is looking to procure 10,000 GPUs to create compute infrastructure in the country in a public-private partnership mode.
The cost or the capex for procuring GPUs for the government is expected to be around Rs 4,000 crore. Industry executives point that moving fast on the procurement of GPUs can help the government to realize the use cases, the need for more GPUs and varieties, and pivot on the need fast. India needs to be the intelligence capital. To achieve that, we know the algorithm and we know how to source compute, said Vishal Dhupar, MD, South Asia, Nvidia.
According to Dhupar, the country is ready to have sovereignty in the AI infrastructure once it is able to bring the elements of skillsets, technology awareness among people, and the country’s digital experience, together with the infrastructure.
Sovereignty is crucial for India to not get affected by any constraints on the usage of AI models or tech from countries like the US, etc, in case of geopolitical conflicts. Through the IndiaAI mission, the government is looking to develop its own fundamental large-language model trained on Indian datasets.
Anil Nanduri, head of AI acceleration office, Intel, said, Today, the whole AI ecosystem is importing models from the West and trying to collaborate it in the domestic world. So, it’s more of an application model.
According to Nanduri, for India AI models, there is a need to have unique datasets. The same will also have the potential of exports if the cost model is set right.
Currently, most LLMs are trained on the western culture and preferences, and that also increases the cost for Indian startups when they try to leverage it for their own services owing to the difference in data sets the models are trained on.
Mohit Sewak, who handles AI research and developer relationship for Nvidia South Asia, said the country needs multimodal AI models that can consider diverse India cultures.
Yotta Data Services, which is a Hiranandani group company, has been providing startups with Nvidia GPUs. The company is currently witnessing 70% demand for GPUs on subscriptions from global companies and 30% from India, said co-founder, MD, and CEO Sunil Gupta.
Gupta expects the demand from Indian startups and enterprises for GPUs will increase going forward, and the government along with the private sector need to create the compute capacities.
Tanuj Bhojwani, head of People + AI said, the government needs to know on how it’s going to spend the first Rs 1,000 crore on the AI mission.
According to Bhojwani, for sovereignty, self-sufficiency, and self-determination are key. We want people of India to have the benefits of using AI, especially it will be applied in education, in healthcare, etc. And if there is somebody else, who can put a stop to that, that is what you should be predicting against first, Bhojwani said.