## Inside Project Olympus 2.0: Azure Preps Next-Gen Liquid-Cooled Racks for Blackwell and Rubin Deployments 🧊⚡🏗️ As artificial intelligence workloads transition rapidly toward multi-trillion-parameter frontier models, autonomous agentic systems, and complex reasoning pipelines, traditional air-cooled data center infrastructure has officially hit a thermal brick wall. To future-proof its global fleet for NVIDIA’s extreme-density **Blackwell (GB200/GB300 NVL72)** and upcoming **Vera Rubin (Rubin NVL72)** architectures, Microsoft Azure is rolling out **Project Olympus 2.0**—a complete reimagining of its open-source data center rack specifications tailored exclusively for closed-loop liquid cooling and high-density power delivery. --- ### 1. The Thermal Breaking Point: Why Project Olympus Needed a Rewrite 🚧🔥 First launched a decade ago as Microsoft’s open contribution to the Open Compute Project (OCP), the original Project Olympus established the modular standard for air-cooled hyperscale server chassis and power distribution units. However, those designs were engineered for a bygone era of modest 10kW to 20kW server racks. * **The Power Density Crisis:** Modern rack-scale architectures like the NVIDIA GB300 NVL72 pull upwards of **120kW to over 140kW per rack**, packing dozens of GPUs and high-frequency CPUs into tight physical footprints. * **The Air Cooling Ceiling:** Moving air can no longer efficiently siphon heat away from silicon running at sustained multi-volt loads. Without advanced liquid management, high-performance nodes throttle instantly or risk catastrophic thermal shutdown. * **The Olympus 2.0 Pivot:** Project Olympus 2.0 abandons legacy airflow assumptions, treating the server rack not merely as a shelf for hardware, but as a tightly integrated, fluid-managed hydraulic and electrical unit. --- ### 2. Under the Hood of Project Olympus 2.0 Racks 💧⚙️ Designed hand-in-hand with hardware partners to support Azure's massive multi-node deployments for OpenAI and internal frontier models, the new rack specification introduces several foundational breakthroughs: * **Closed-Loop Precision Cold Plates:** Custom-engineered micro-channel copper cold plates are sealed directly over GPUs, CPUs, and high-speed switch ASICs. A specialized dielectric fluid mixture (circulating via internal manifolds) siphons heat away right at the die. * **Two-Thirds Cooling, One-Third Compute Topology:** The physical layout of the rack has been dramatically altered. In high-end configurations, up to two-thirds of the internal enclosure space is dedicated to advanced plumbing, manifolds, power shelf integration, and closed-loop heat exchangers, leaving the remaining third for ultra-dense compute trays. * **Standalone Heat Exchanger Units (HEUs):** To prevent facility-wide water retrofitting headaches, Olympus 2.0 incorporates modular, closed-loop liquid-to-air or liquid-to-coolant distribution units that manage thermal rejection efficiently while keeping blue water consumption down. --- ### 3. Scaling for Blackwell, GB300, and the Horizon of Rubin 🚀🗺️ Project Olympus 2.0 serves as the underlying physical backbone for Azure’s most ambitious infrastructure rollouts: * **Blackwell and GB300 NVL72 Dominance:** Azure relies on these custom liquid-cooled architectures to deploy massive clusters of NVIDIA GB200 and GB300 NVL72 virtual machines. By utilizing ultra-fast intra-rack NVLink interconnects (delivering up to 130 TB/s of bandwidth inside the rack), Olympus 2.0 keeps multi-node memory pools tightly synchronized without thermal throttling. * **Paving the Way for Rubin:** Looking toward the future, these liquid-cooled rack blueprints are being forward-architected to ingest NVIDIA’s upcoming **Vera Rubin NVL72 platforms**, ensuring Azure can seamlessly scale next-generation multi-node clusters as they hit production. --- ### The Bottom Line 🌟📈 Hardware innovation is no longer just about raw silicon performance—it is entirely dependent on thermal and power engineering. Through **Project Olympus 2.0**, Microsoft Azure has successfully converted a massive engineering challenge into a competitive advantage, building the liquid-cooled foundation required to power the next generation of autonomous AI at global scale!