Cadence’s Sherman Ikemoto on Digital Twins, Power Reality and Designing the AI Factory

AI data centers are no longer just buildings full of racks. They are tightly coupled systems where power, cooling, IT, and operations all depend on each other, and where bad assumptions get expensive fast.

On the latest episode of The Data Center Frontier Show, Editor-in-Chief Matt Vincent talks with Sherman Ikemoto of Cadence about what it now takes to design an “AI factory” that actually works.

Ikemoto explains that data center design has always been fragmented. Servers, cooling, and power are designed by different suppliers, and only at the end does the operator try to integrate everything into one system. That final integration phase has long relied on basic tools and rules of thumb, which is risky in today’s GPU-dense world.

Cadence is addressing this with what it calls “DC elements”:  digitally validated building blocks that represent real systems, such as NVIDIA’s DGX SuperPOD with GB200 GPUs. These are not just drawings; they model how systems really behave in terms of power, heat, airflow, and liquid cooling. Operators can assemble these elements in a digital twin and see how an AI factory will actually perform before it is built.

A key shift is designing directly to service-level agreements. Traditional uncertainty forced engineers to add large safety margins, driving up cost and wasting power. With more accurate simulation, designers can shrink those margins while still hitting uptime and performance targets, critical as rack densities move from 10–20 kW to 50–100 kW and beyond.

Cadence validates its digital elements using a star system. The highest level, five stars, requires deep validation and supplier sign-off. The GB200 DGX SuperPOD model reached that level through close collaboration with NVIDIA.

Ikemoto says the biggest bottleneck in AI data center buildouts is not just utilities or equipment; it is knowledge. The industry is moving too fast for old design habits. Physical prototyping is slow and expensive, so virtual prototyping through simulation is becoming essential, much like in aerospace and automotive design.

Cadence’s Reality Digital Twin platform uses a custom CFD engine built specifically for data centers, capable of modeling both air and liquid cooling and how they interact. It supports “extreme co-design,” where power, cooling, IT layout, and operations are designed together rather than in silos. Integration with NVIDIA Omniverse is aimed at letting multiple design tools share data and catch conflicts early.

Digital twins also extend beyond commissioning. Many operators now use them in live operations, connected to monitoring systems. They test upgrades, maintenance, and layout changes in the twin before touching the real facility. Over time, the digital twin becomes the operating platform for the data center.

Running real AI and machine-learning workloads through these models reveals surprises. Some applications create short, sharp power spikes in specific areas. To be safe, facilities often over-provision power by 20–30%, leaving valuable capacity unused most of the time. By linking application behavior to hardware and facility power systems, simulation can reduce that waste, crucial in an era where power is the main bottleneck.

The episode also looks at Cadence’s new billion-cycle power analysis tools, which allow massive chip designs to be profiled with near-real accuracy, feeding better system- and facility-level models.

Cadence and NVIDIA have worked together for decades at the chip level. Now that collaboration has expanded to servers, racks, and entire AI factories. As Ikemoto puts it, the data center is the ultimate system—where everything finally comes together—and it now needs to be designed with the same rigor as the silicon inside it.