India and Yotta Place Orders for 40,000+ Blackwell GPUs
Nvidia's Blackwell Ultra platform is seeing massive adoption through two major deals. India is deploying 20,000 Blackwell Ultra GPUs in a $1 billion national AI infrastructure initiative. Separately, Yotta and Nvidia announced a $3 billion partnership to create an AI supercluster in Asia, also deploying over 20,000 Blackwell Ultra GPUs.
- The Blackwell B200 GPU, featuring 208 billion transistors, offers up to a 57% faster training performance compared to the H100. For inference workloads, it is designed to be significantly more efficient, potentially lowering energy consumption by up to 25 times compared to the H100. - Yotta's AI supercluster will be located at its D2 hyperscale data center in Greater Noida, which has the capacity to scale from 60 MW to 250 MW. This facility is part of Yotta's broader plan to potentially scale to over one million GPUs within three to five years. - India's national AI initiative, known as the IndiaAI Mission, is a government program backed by over $1 billion to enhance the country's AI compute capacity and foster the development of sovereign AI models and applications. The initiative aims to establish a robust AI infrastructure with a minimum of 10,000 GPUs through public-private partnerships. - The GB200 "superchip" configuration, which combines two B200 GPUs with a Grace CPU, is estimated to be priced between $60,000 and $70,000 per unit. A complete server rack with 72 of these superchips could cost around $3 million. - The Blackwell architecture introduces a second-generation Transformer Engine and new floating-point formats like FP4 and FP6, which can double the performance and the size of models that can be supported in memory while maintaining high accuracy. - Yotta, a part of the Hiranandani Group, operates large-scale, Tier IV certified data centers in India, including Asia's largest in Navi Mumbai, focusing on sovereign cloud and AI services. - The move towards custom silicon is a growing trend among hyperscalers like Google, AWS, and Microsoft, who are designing their own ASICs (Application-Specific Integrated Circuits) to optimize AI workloads and reduce reliance on third-party chipmakers. This build-versus-buy decision is driven by the need for greater performance-per-watt and control over the hardware-software stack for specific AI tasks. - The infrastructure required for Blackwell GPUs is substantial, demanding liquid cooling, 800-gigabit networking, and power densities that exceed the capabilities of many existing data centers.
