100 MW Hyperscale AI Reference Architecture Data Center
A Purpose-Built Blueprint for Next-Generation AI Compute
As AI workloads demand unprecedented power density and operational efficiency, data center infrastructure must evolve to keep pace. The 100 MW Hyperscale AI Blueprint is a reference architecture developed by Siemens in cooperation with nVent and NVIDIA, purpose-built for next-generation AI compute environments. Designed around NVIDIA DGX™ GB200 SuperPOD-class racks operating at 127 kW per rack, this Tier III fault-tolerant solution integrates Siemens industrial-grade electrical systems with nVent direct-to-chip liquid cooling technology to deliver rapid deployment, high compute density, and operational continuity. Whether you're planning new capacity or scaling existing infrastructure, this blueprint provides a clear path to minimizing time-to-compute and maximizing tokens-per-watt.
Why nVent?
At nVent, we believe safer systems ensure a more secure world. We connect and protect our customers with inventive electrical solutions. We design, manufacture, market, install and service high performance products and solutions for mission-critical equipment, buildings and essential processes. Our solutions are helping build a more sustainable and electrified world.
Watch: Siemens & nVent Discuss the Future of AI Data Center Infrastructure
At Supercomputing 2025, Siemens Head of Data Center Vertical John de Boer joined nVent to discuss how the industry is navigating the challenges of next-generation AI infrastructure.
From the 10X disruption of Blackwell chips to combining liquid cooling loops with power automation systems, this conversation explores how operators can intelligently integrate key technologies—and why collaboration across the ecosystem is essential for building infrastructure that's scalable, efficient, and ready for what's next.
Watch: 10X Future of Data Centers – Scaling in the AI Era
In this webinar hosted by S&P Global Market Intelligence, experts from Siemens and nVent explore the infrastructure challenges reshaping the data center industry. As AI workloads drive rack densities from 15 kW to well over 100 kW—with roadmaps pointing toward 600 kW to 2 MW per rack—the conversation covers what it takes to design, power, and cool the next generation of AI factories. Topics include the shift to liquid cooling, the integration of OT and IT systems, reference architectures that accelerate deployment, and strategies for balancing performance with sustainability goals.
Designing for the Unknown: Cooling Strategies When Your 2028 Roadmap Is Already Obsolete
Building a data center for the AI era has become a high-stakes game of planning a cross-country road trip using a map that changes daily. You know your destination is unprecedented compute density, but the route—defined by explosive GPU power trajectories—is constantly shifting. The 2-megawatt rack, once a distant forecast, is now looming on a horizon that is arriving years faster than construction cycles.
Reference Architecture FAQs
What is direct-to-chip liquid cooling in AI data centers?
Direct-to-chip liquid cooling delivers coolant directly to the heat source, the CPU or GPU, via cold plates mounted on the processor. Unlike traditional air cooling, which struggles at high power densities, direct-to-chip cooling can handle rack loads of 100 kW and above, making it essential for modern AI data centers running high-density GPU workloads.
What power density does the NVIDIA DGX GB200 SuperPOD require per rack?
The NVIDIA DGX GB200 SuperPOD operates at approximately 127 kW per rack. This level of power density far exceeds the capacity of conventional air-cooled infrastructure and requires purpose-built liquid cooling and electrical systems to maintain performance and operational continuity.
What is a Tier III fault-tolerant data center?
A Tier III data center is designed with redundant components and multiple distribution paths, allowing maintenance and upgrades to take place without taking systems offline. With an expected uptime of 99.982%, Tier III facilities are the standard for enterprise and hyperscale AI deployments where downtime is not an option.
How do you cool a 100 MW hyperscale AI data center?
Cooling a 100 MW AI data center requires a combination of direct-to-chip liquid cooling and high-capacity thermal management infrastructure. At rack densities of 127 kW and above, air cooling alone is insufficient. Liquid cooling loops must be tightly integrated with power distribution systems to maintain efficiency, reduce PUE, and support continuous operation at scale.
What does tokens-per-watt mean in AI data center design?
Tokens-per-watt is a measure of AI compute efficiency, specifically the number of tokens a model can process per watt of energy consumed. It is a key metric for hyperscale AI operators looking to balance performance with energy costs. Optimising tokens-per-watt requires an infrastructure approach that maximises compute density while minimising cooling and power overhead.
How do you scale data center infrastructure for AI workloads?
Scaling for AI workloads requires moving beyond traditional data center design. Key considerations include adopting liquid cooling for high-density GPU racks, deploying reference architectures that accelerate build times, integrating industrial-grade electrical systems, and planning for rack power roadmaps that could reach 600 kW to 2 MW per rack in the near future.