India's current top AI supercomputer: 200 petaflops.
— Sidharth (@Cloudwatch199) May 16, 2026
Ranks 188 globally.
The new India - UAE system: 8,000 petaflops.
This will be 19x India's existing national AI capacity in a single deployment, and the largest dedicated AI system outside the US–China axis. https://t.co/lxeVW185sv
India and the UAE are building an 8 EXAFLOP AI supercomputer, jointly developed by C-DAC and G42.
— Sidharth (@Cloudwatch199) May 16, 2026
That would rank among the most powerful AI systems on the planet.
Most people still think AI is a chatbot race. It is NOT.
It is a COMPUTE race.
Compute is what decides who… pic.twitter.com/fYK56CiMzW
Condor Galaxy India is an 8-exaflop AI supercomputer project (equivalent to 8,000 petaflops or 8,000,000 teraflops) being developed through a partnership between India and the UAE. It represents a major leap in India's compute infrastructure for AI. Key Details
The project builds on a 2024 India-UAE digital infrastructure MoU and reflects deepening tech cooperation (alongside other investments, like UAE's $5B commitment).
Deployment details (exact location, full timeline to operational status, exact cost) are not fully public in available sources, but it is expected to be a major national asset once commissioned. This is a fast-moving development as of mid-May 2026.
- Scale and Specs: It consists of 64 Cerebras CS-3 systems, powered by Cerebras' Wafer-Scale Engine 3 (WSE-3) technology. This is a radical wafer-scale architecture where an entire silicon wafer forms a single massive processor (with over 4 trillion transistors and nearly 1 million AI-optimized cores), rather than hundreds of small chips cut from a wafer like traditional GPUs.
- Claims include up to 20x faster training/inference than conventional setups for large models, massive on-chip memory (e.g., 44 GB SRAM), extremely high bandwidth, and the ability to handle models with trillions of parameters more efficiently with less networking complexity and a smaller footprint.
- Current Indian Context: India's existing top AI supercomputers (e.g., AIRAWAT and PARAM Siddhi-AI at C-DAC) offer around 200-410 petaflops combined. This new system would be roughly 19-40x more powerful in a single deployment and rank among the world's most powerful dedicated AI systems, positioned as the largest outside the US-China axis.
- Name and Branding: Part of G42 and Cerebras' broader Condor Galaxy network of AI supercomputers (with prior deployments in the US).
- Primary Partners: UAE-based G42 (global AI/tech group) and India's C-DAC (Centre for Development of Advanced Computing). G42 handles installation, deployment, operations, and maintenance in partnership with C-DAC.
- Additional Involvement: Cerebras Systems (US wafer-scale chipmaker), Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi, and broader India-UAE government backing.
- Timeline and Formalization: Plans were announced earlier in 2026 (around the India AI Impact Summit). A term sheet was signed, and the commercial/operational framework was formalized on or around May 15, 2026, during PM Narendra Modi's visit to the UAE. UAE President Sheikh Mohamed bin Zayed Al Nahyan symbolically gifted a Cerebras wafer-scale chip to PM Modi.
- Governance: Data and operations fall under Indian national jurisdiction for sovereign AI control, reducing reliance on foreign cloud providers.
- For India: Boosts sovereign AI capabilities for training large models, R&D in health/genomics, energy, geospatial analytics, drug discovery, disaster management, and more. It supports the India AI Mission, gives researchers/startups access to frontier compute, and advances self-reliance in AI infrastructure.
- For UAE/G42: Expands G42's "Intelligence Grid" globally, leveraging UAE capital and access to advanced Western hardware (e.g., via partnerships avoiding some restrictions faced by others). It strengthens bilateral ties.
- Broader Context: This is part of a compute race. AI progress depends heavily on raw computational power for training models, simulations, etc. The US restricts high-end exports; this project creates an alternative powerhouse outside the main US-China dynamic, combining Indian engineering talent/depth with UAE financing and Cerebras tech.
Deployment details (exact location, full timeline to operational status, exact cost) are not fully public in available sources, but it is expected to be a major national asset once commissioned. This is a fast-moving development as of mid-May 2026.
Condor Galaxy India: An 8-Exaflop Wafer-Scale AI Supercomputer in Global Context
In mid-2026, India formalized plans for Condor Galaxy India, an 8-exaflop AI supercomputer powered by 64 Cerebras CS-3 systems. This partnership between UAE-based G42, Cerebras Systems, and India's C-DAC marks a significant step in the global AI compute race. While not the absolute largest by raw scale, its architecture, efficiency profile, and sovereign control distinguish it from GPU-heavy hyperscale clusters and traditional scientific supercomputers. Technical Specifications and ArchitectureEach CS-3 system features Cerebras' Wafer-Scale Engine 3 (WSE-3), a single silicon wafer with over 4 trillion transistors and approximately 900,000 AI-optimized cores. It delivers around 125 petaflops of AI compute (sparse FP16), 44 GB of on-chip SRAM (with extreme bandwidth of ~21 PB/s), and avoids the networking overhead typical of GPU clusters. A single CS-3 can handle massive models with far less complexity than distributed GPU setups.
The full Condor Galaxy India cluster thus provides 8 exaflops of AI-optimized performance, tens of millions of cores, and a compact footprint. It is designed for linear scaling with software treating the cluster as a single logical device, excelling at training and inference for trillion-parameter models. Power draw per CS-3 is around 20-25 kW, making the cluster far more power-dense and potentially efficient for AI workloads than equivalent GPU systems in terms of interconnect complexity and memory access. Comparison to Traditional Scientific Supercomputers (TOP500)Systems like the U.S. DOE's El Capitan (~1.8 exaflops Rmax on HPL FP64 benchmark), Frontier (~1.35 exaflops), and Aurora (~1 exaflop) lead the TOP500 list. These prioritize high-precision (FP64) calculations for simulations in physics, climate modeling, and materials science.
Condor Galaxy India targets low-to-mixed precision AI workloads (e.g., FP16/FP8/BF16), where effective performance can be many times higher than FP64 figures. Direct FLOPS comparisons are misleading: a GPU or APU-based exascale system might deliver 5-10+ exaflops in AI-relevant precisions, but Condor Galaxy's wafer-scale design offers advantages in on-chip memory and reduced communication overhead. For pure scientific HPC, traditional systems remain superior; for frontier AI training, the Cerebras approach can be faster and simpler for very large models.
India's prior systems (e.g., PARAM series, AIRAWAT) sit in the low hundreds of petaflops. Condor Galaxy India represents a generational leap—potentially 20-40x more powerful in AI terms—positioning the country as a serious player in sovereign AI. Comparison to Other Dedicated AI SupercomputersxAI's Colossus (Memphis) stands as one of the largest GPU-based AI clusters, with hundreds of thousands of NVIDIA GPUs (H100/H200/Blackwell generations) scaling toward a million. Early phases delivered tens of exaflops in AI compute, with massive power draw (hundreds of MW to GW-scale). It excels in raw scale and benefits from the mature NVIDIA ecosystem (CUDA, vast software support). However, it faces challenges in networking latency, power/cooling demands, and model parallelism complexity for the largest models.
Hyperscale clusters from Google (TPU pods like Ironwood, claiming 10s of exaflops per pod), Meta, Microsoft, and others often reach effective hundreds of exaflops across campuses but are distributed and optimized for both training and inference. Google's TPU v5/v6/v7 generations and similar custom silicon prioritize efficiency at scale.
Condor Galaxy India's 64 CS-3 setup is smaller in node count than these GPU behemoths but competes in effective AI performance due to wafer-scale advantages (e.g., claims of 10-20x faster training for certain large models vs. GPU equivalents, with simpler programming). Previous Condor Galaxy systems (CG-1, CG-2 at 4 exaflops each with CS-2; CG-3 at 8 exaflops with CS-3) were already among the largest dedicated AI supercomputers at announcement.
Cerebras systems shine in scenarios requiring massive on-chip memory and minimal data movement. GPU clusters dominate in flexibility, ecosystem maturity, and sheer aggregate scale. A 64-CS-3 cluster is more like a highly optimized "single logical superchip" at cluster scale, while Colossus is a vast distributed mesh.Strategic and Geopolitical ContextMost top AI compute resides in the U.S. and China, with hyperscalers and national labs dominating. Condor Galaxy India diversifies this landscape through India-UAE collaboration, leveraging Cerebras hardware (U.S.-designed, TSMC-fabricated) while maintaining Indian data sovereignty. This supports India's AI Mission, reducing reliance on foreign clouds and enabling local research in healthcare, genomics, climate, and defense.
It contrasts with U.S.-centric builds (Colossus, DOE systems) or China's domestic efforts, highlighting a multipolar compute future. Power efficiency, deployment speed, and accessibility for researchers (vs. hyperscaler internal use) are key differentiators.Strengths, Limitations, and Future OutlookStrengths:
In mid-2026, India formalized plans for Condor Galaxy India, an 8-exaflop AI supercomputer powered by 64 Cerebras CS-3 systems. This partnership between UAE-based G42, Cerebras Systems, and India's C-DAC marks a significant step in the global AI compute race. While not the absolute largest by raw scale, its architecture, efficiency profile, and sovereign control distinguish it from GPU-heavy hyperscale clusters and traditional scientific supercomputers. Technical Specifications and ArchitectureEach CS-3 system features Cerebras' Wafer-Scale Engine 3 (WSE-3), a single silicon wafer with over 4 trillion transistors and approximately 900,000 AI-optimized cores. It delivers around 125 petaflops of AI compute (sparse FP16), 44 GB of on-chip SRAM (with extreme bandwidth of ~21 PB/s), and avoids the networking overhead typical of GPU clusters. A single CS-3 can handle massive models with far less complexity than distributed GPU setups.
The full Condor Galaxy India cluster thus provides 8 exaflops of AI-optimized performance, tens of millions of cores, and a compact footprint. It is designed for linear scaling with software treating the cluster as a single logical device, excelling at training and inference for trillion-parameter models. Power draw per CS-3 is around 20-25 kW, making the cluster far more power-dense and potentially efficient for AI workloads than equivalent GPU systems in terms of interconnect complexity and memory access. Comparison to Traditional Scientific Supercomputers (TOP500)Systems like the U.S. DOE's El Capitan (~1.8 exaflops Rmax on HPL FP64 benchmark), Frontier (~1.35 exaflops), and Aurora (~1 exaflop) lead the TOP500 list. These prioritize high-precision (FP64) calculations for simulations in physics, climate modeling, and materials science.
Condor Galaxy India targets low-to-mixed precision AI workloads (e.g., FP16/FP8/BF16), where effective performance can be many times higher than FP64 figures. Direct FLOPS comparisons are misleading: a GPU or APU-based exascale system might deliver 5-10+ exaflops in AI-relevant precisions, but Condor Galaxy's wafer-scale design offers advantages in on-chip memory and reduced communication overhead. For pure scientific HPC, traditional systems remain superior; for frontier AI training, the Cerebras approach can be faster and simpler for very large models.
India's prior systems (e.g., PARAM series, AIRAWAT) sit in the low hundreds of petaflops. Condor Galaxy India represents a generational leap—potentially 20-40x more powerful in AI terms—positioning the country as a serious player in sovereign AI. Comparison to Other Dedicated AI SupercomputersxAI's Colossus (Memphis) stands as one of the largest GPU-based AI clusters, with hundreds of thousands of NVIDIA GPUs (H100/H200/Blackwell generations) scaling toward a million. Early phases delivered tens of exaflops in AI compute, with massive power draw (hundreds of MW to GW-scale). It excels in raw scale and benefits from the mature NVIDIA ecosystem (CUDA, vast software support). However, it faces challenges in networking latency, power/cooling demands, and model parallelism complexity for the largest models.
Hyperscale clusters from Google (TPU pods like Ironwood, claiming 10s of exaflops per pod), Meta, Microsoft, and others often reach effective hundreds of exaflops across campuses but are distributed and optimized for both training and inference. Google's TPU v5/v6/v7 generations and similar custom silicon prioritize efficiency at scale.
Condor Galaxy India's 64 CS-3 setup is smaller in node count than these GPU behemoths but competes in effective AI performance due to wafer-scale advantages (e.g., claims of 10-20x faster training for certain large models vs. GPU equivalents, with simpler programming). Previous Condor Galaxy systems (CG-1, CG-2 at 4 exaflops each with CS-2; CG-3 at 8 exaflops with CS-3) were already among the largest dedicated AI supercomputers at announcement.
Cerebras systems shine in scenarios requiring massive on-chip memory and minimal data movement. GPU clusters dominate in flexibility, ecosystem maturity, and sheer aggregate scale. A 64-CS-3 cluster is more like a highly optimized "single logical superchip" at cluster scale, while Colossus is a vast distributed mesh.Strategic and Geopolitical ContextMost top AI compute resides in the U.S. and China, with hyperscalers and national labs dominating. Condor Galaxy India diversifies this landscape through India-UAE collaboration, leveraging Cerebras hardware (U.S.-designed, TSMC-fabricated) while maintaining Indian data sovereignty. This supports India's AI Mission, reducing reliance on foreign clouds and enabling local research in healthcare, genomics, climate, and defense.
It contrasts with U.S.-centric builds (Colossus, DOE systems) or China's domestic efforts, highlighting a multipolar compute future. Power efficiency, deployment speed, and accessibility for researchers (vs. hyperscaler internal use) are key differentiators.Strengths, Limitations, and Future OutlookStrengths:
- Exceptional efficiency for large-model training/inference.
- Simplified software model.
- Compact footprint.
- Sovereign control for India.
- Rapid deployment potential via G42/Cerebras partnership.
- Smaller overall scale than leading GPU clusters (hundreds of thousands of accelerators).
- Less mature ecosystem than NVIDIA CUDA.
- Wafer-scale manufacturing complexity and potential higher per-unit costs.
- Limited to AI-optimized workloads.
