'India can’t be left behind’: Krutrim AI deploys DeepSeek for AI push

Riding on the popularity of the popular Artificial Intelligence (AI) chatbot by Chinese startup DeepSeek, Ola Founder and CEO Bhavish Aggarwal has said his AI startup, Krutrim, will deploy DeepSeek's powerful AI models on domestic servers, ensuring data privacy while slashing training costs.

"India can’t be left behind in AI. @Krutrim has accelerated efforts to develop world-class AI. As a first step, our cloud now has DeepSeek models live, hosted on Indian servers. Pricing lowest in the world," Aggarwal says on X.

He adds that the company will reveal the details of the Krutrim AI lab, state-on-the-art (SOTA) model, research progress, and open source on February 4, 2024.

Krutrim was launched in February 2024 following the announcement of a $50 million financing at a $1 billion valuation, led by Matrix Partners India.

Krutrim AI Studio offers access to a range of pre-trained AI models that can be customised to meet specific business needs. Apart from DeepSeek models, Krutrim-led AI Studio also hosts different AI models from Meta, Google, OpenAI, HuggingFace, Stable Diffusion. Krutrim AI Studio claims it can reduce AI development time by up to 60%.

Among benefits, the Krutrim AI Studio platform shows offerings like faster time-to-market, cost savings of up to 25% compared to in-house development, access to state-of-the-art AI advancements, and reliable performance from thoroughly tested models.

The development at Aggarwal-led Krutrim comes a day after Union Minister Ashwini Vaishnaw's said on the sidelines of the India AI Mission in Delhi, where he said the government plans to host DeepSeek on Indian servers, which will address data privacy concerns and power India's AI capabilities.

He told the media that within a short time frame, India has established the first major pillar of the India AI Mission: the Common Compute Facility. According to the minister, against a target of 10,000 GPUs, India has empanelled 18,693 high-end GPUs. Among them, 12,896 are H100s, 1,480 are H200s, and 742 are MI325 and MI300X GPUs, he said.

“To put this in perspective, DeepSeek was reportedly trained on 2,000 GPUs. This Common Compute Facility is a crucial component of the India AI Mission,” Vaishnaw said. He also said some have raised questions over spending on the India AI mission, but DeepSeek was able to make a really powerful model at just $5.5 million.

Meanwhile, as China's DeepSeek AI model disrupts the global AI landscape, some experts in India argue the country must move beyond "short-termism" and adopt a long-term approach to building a robust AI ecosystem. DeepSeek AI models, V1 and R1, perform on par with several advanced AI models developed by American companies like OpenAI and Meta, but at a fraction of the cost.

While OpenAI and Google’s AI models cost around $100 million to train, DeepSeek’s latest model was developed for just $5.6 million. Venture capitalist and former adviser to US President Donald Trump, Marc Andreessen, described this development as "AI's Sputnik moment."