Cerebras plans UAE-backed AI supercomputer in India • The Register

Cerebras Systems will power a new AI supercomputing cluster in India capable of up to 8 exaFLOPS of "super sparse" AI compute, the company announced in New Delhi during this week's AI Impact Summit. The installation is a collaboration between the United Arab Emirates' Mohamed Bin Zayed University of AI (MBZUAI) and India's Center for Development of Advanced Computing (C-DAC) and will be deployed by UAE technology group G42. The system will be operated under India-defined governance frameworks, with all data remaining within the nation's borders, and will be made available to Indian universities, startups and small and mid-sized businesses.

"Sovereign AI infrastructure is becoming essential for national competitiveness," G42 India CEO Manu Jain said in a canned statement. "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."

Details

Cerebras told The Register the supercomputer will be powered by its WSE-3 wafer-scale accelerators. According to publicly available performance figures and the company's own disclosures, each WSE-3 is capable of delivering 125 petaFLOPS of highly sparse 16-bit floating point performance. Back-of-the-napkin arithmetic suggests a full system targeting 8 exaFLOPS would likely include about 64 WSE-3 accelerators.

The wafer-scale chips eschew traditional high-bandwidth memory (HBM) in favor of large amounts of on-chip SRAM. Each Cerebras CS-3 system is rated at 23 kW and features 44 GB of SRAM with roughly 21 petabytes a second of memory bandwidth — a figure the vendor contrasts with HBM4 on Nvidia's Rubin GPUs. The architecture's memory bandwidth and on-chip SRAM have made the platform particularly effective for memory-bound AI inference workloads: industry benchmarking cited in the announcement suggests Cerebras can serve gpt-oss 120b High at nearly 2,853 tokens per second per user, more than three times faster than the next-fastest GPU-based inference provider.

The UAE's G42 is a major backer of Cerebras and previously financed the chipmaker's Condor Galaxy deployment effort at an estimated cost of $900 million. G42 has also pursued development of sovereign language models: late last year it released NANDA 87B, an 87 billion-parameter open-weights model trained in Hindi and English.

• Partners named in the announcement: MBZUAI, C-DAC and G42.
• Chip platform: Cerebras WSE-3 wafer-scale accelerators (approx. 125 petaFLOPS each, sparse 16-bit).
• Targeted peak capacity: up to 8 exaFLOPS of super-sparse AI compute.
• Operational commitments: India-defined governance, data to remain within India.

Outlook

The Cerebras deployment arrives as several other large compute projects are being announced in India: AMD with Tata Consultancy Services is set to deploy 200 MW of Helios racks powered by MI455X, cloud provider Yotta has planned a cluster of 20,000 Blackwell Ultra GPUs, and Larsen & Toubro has outlined a giga-scale datacenter rollout beginning with 30 MW in Chennai and 40 MW in Mumbai. For India, the Cerebras-led cluster aims to expand sovereign AI capacity and provide a high-bandwidth option tailored to memory-bound inference use cases, while making advanced compute accessible to academic and commercial users under Indian governance rules.