Cerebras Systems has embarked on an ambitious project to establish a powerful AI supercomputer in India, aiming to achieve up to 8 exaFLOPS of super sparse AI compute. This initiative was unveiled during the AI Impact Summit in New Delhi, marking a significant collaboration between the Mohamed Bin Zayed University of AI (MBZUAI) in the UAE and India’s Center for Development of Advanced Computing (C-DAC).
The deployment of this advanced computing system will be managed by G42, a prominent UAE technology firm, which seeks to enhance the nation’s sovereign compute capabilities. G42 is notably one of Cerebras’ major investors, having previously supported the chip startup’s Condor Galaxy deployment, estimated to cost around $900 million.
G42 has carved a niche in developing sovereign AI models tailored to various languages, exemplified by the recent release of NANDA 87B, an 87 billion parameter model trained in both Hindi and English. Manu Jain, CEO of G42 India, emphasized the importance of sovereign AI infrastructure for national competitiveness, stating, “This project brings that capability to India at a national scale, enabling local researchers, innovators, and enterprises to become AI-native while maintaining full data sovereignty and security.”
While G42 will oversee the system’s deployment, it will operate under governance frameworks defined by India, ensuring that all data remains within the country’s borders. Once operational, the supercomputer will be accessible to Indian universities, startups, and small to mid-sized businesses.
Cerebras has confirmed that the supercomputer will utilize its WSE-3 wafer-scale accelerators. Preliminary calculations suggest that the system will likely include 64 of these accelerators, each capable of delivering 125 petaFLOPS of highly sparse 16-bit floating point performance. These accelerators are distinctive as they do not depend on expensive high-bandwidth memory (HBM) typically found in GPUs from competitors like Nvidia and AMD; instead, they leverage fast on-chip SRAM.
The CS-3 systems from Cerebras feature 44 GB of SRAM, providing an impressive 21 petabytes per second of memory bandwidth—approximately 1,000 times faster than the HBM4 on Nvidia’s latest GPUs. Originally designed for AI training, this rapid SRAM performance has also proven advantageous for memory-bound AI inference tasks. Reports indicate that Cerebras can serve the gpt-oss 120b High model at nearly 2,853 tokens per second per user, significantly outpacing the next fastest GPU-based inference provider.
This announcement arrives shortly after AMD and Nvidia revealed their own large-scale deployments in India, further highlighting the growing competition in the AI supercomputing landscape.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
