Cerebras and G42 expand AI supercomputing with Condor Galaxy 3

Cerebras Systems and G42 build Condor Galaxy 3 to deliver 8 exaFLOPs of AI performance. This third cluster in the Condor Galaxy network uses 64 Cerebras CS-3 systems. Each system uses the Wafer-Scale Engine 3, which contains 4 trillion transistors and 900,000 AI cores. The WSE-3 chip provides 125 petaflops of peak AI performance per chip. TSMC manufactures the WSE-3 at a 5-nanometer node. This hardware provides twice the performance of the previous generation at the same power and price. The Condor Galaxy network reaches 16 exaFLOPs after adding this cluster to the existing Condor Galaxy 1 and Condor Galaxy 2. The WSE-3 chip delivers inference and training 20 times faster than the competition and measures 56 times larger than the largest GPU. The WSE-3 provides native support for training with long sequence lengths of up to 50,000 tokens without special software libraries. The CS-3 system uses a wafer packaging solution called an engine block. This infrastructure provides high power density and uniform cooling for the wafer component through an internal water loop. It includes redundant and hot-swappable cooling and power systems to maintain performance.

Global expansion and the Condor Galaxy network

The Condor Galaxy network expands across the United States and India. Condor Galaxy 1 operates in Santa Clara, California, with 4 exaFLOPs of AI compute and 54 million cores. The Condor Galaxy 1 system, which operates out of a high-performance colocation facility in Santa Clara, provides 4 exaFLOPs of AI compute through 64 Cerebras CS-2 systems linked to 72,704 AMD EPYC processor cores. This cluster has trained generative AI models such as Jais-30B, Med42, Crystal-Coder-7B, and BTLM-3B-8K. BTLM-3B-8K holds the top position for 3B models on HuggingFace. The Med42 model, developed with M42 and Core42, surpassed MedPaLM in performance and accuracy after training on Condor Galaxy 1 in a single weekend. The Condor Galaxy 1 provides support for models up to 600 billion parameters and configurations that support 100 trillion parameter models.

In May 2026, G42 and the Government of India formalized terms for Condor Galaxy India. This 8-exaflop cluster uses 64 Cerebras CS-3 systems and will serve India’s sovereign AI ambitions. G42 will work with India’s Centre for Development of Advanced Computing to manage the installation, deployment, and maintenance. This system will support research in genomics, energy, and geospatial analytics.

Middle East infrastructure and data center growth

G42 expands AI infrastructure through investments in the Middle East. HUMAIN plans to build AI factories in Saudi Arabia with a capacity of 500 megawatts. This project uses several hundred thousand NVIDIA GPUs over five years. In the UAE, Microsoft and G42 confirmed a 200-megawatt data center expansion. Khazna Data Centers will execute this project. This follows Microsoft’s $1.5 billion investment in G42 in April 2024. Brad Smith joined the G42 board as part of that agreement. You can see the scale of these projects in the Ajman facility, where Khazna is building a 100-megawatt AI-optimized data center. This facility includes 20 data halls of AI GPUs with 5 megawatts of capacity each. The construction cost for the Ajman facility reaches $272 million. G42 launched the Responsible AI Foundation in February 2025 alongside Microsoft. G42 and DataOne also partnered to create an AI data center in France using AMD hardware. The UAE reclassified the country to Country Group A:5 in July 2026, which eased restrictions on advanced AI technology exports. G42 also received designation as an authorized receiving entity for advanced AI chips.

Will the Condor Galaxy network eventually reach 100 exaFLOPs of capacity?