Physics Underpinning Decisions: Simulation‑Trained AI Optimizes Tokens per Watt

The rapid rise of AI factories is pushing data center infrastructure beyond the limits of traditional planning. Rack densities have increased by an order of magnitude, liquid cooling is becoming standard, and AI workloads no longer behave predictably. In this environment, capacity decisions-where to deploy new AI systems, how far to push existing infrastructure, and when to invest-must move beyond static, worst-case design assumptions that obscure true headroom, drive over-provisioning, and still fail to quantify operational risk.

This is where the Cadence Reality Digital Twin Platform introduces a new approach: AI Capacity Agents, powered by physics-trained AI surrogate models and tightly integrated with high-fidelity simulation. Instead of sequential planning and then validation, Cadence Reality combines both concurrently, allowing teams to explore options rapidly and verify the best candidates with computational fluid dynamics (CFD) simulation before committing to the changes for deployment. This shifts the paradigm from static planning of individual components to a continuous predict-and-optimize workflow of entire systems.

Why Traditional Capacity Planning Breaks Down for AI Infrastructure

AI workloads place simultaneous stress on power, cooling, data, and space. High-density GPU systems, like NVL72 rack scale systems, operate close to thermal and electrical limits, leaving little margin for error. Full CFD simulations provide the required accuracy, but they are computationally heavy and impractical for evaluating dozens-or hundreds-of workload placement options during day-to-day operations.

As a result, teams are often forced into an uncomfortable tradeoff: speed without confidence, or accuracy without agility.

Cadence Reality software resolves this tension by introducing AI surrogate models directly into the digital twin workflow.

Training AI Surrogates on Physics, Not Conjectures

AI surrogate models in Cadence Reality software are trained using synthetic data generated by detailed, physics-based solvers, such as those used in Cadence Celsius EC and Cadence Reality DC Design. These high-fidelity simulations capture complex thermal and airflow behavior, combined with emulated control system interactions across a wide range of operating conditions.

From this training data, neural networks learn to mimic the thermal response of servers, cabinets, and the performance response of the AI factory-producing models that can infer results in seconds rather than hours. Crucially, these surrogates inherit their accuracy from the underlying physics. They are grounded in principled simulation, not statistical averages or simplified heuristics.

This approach aligns with broader best practices in digital twins, where physics-trained AI surrogates are increasingly used to accelerate prediction while preserving physical realism.

Real-Time Insight at Operational Scale

Once trained, AI surrogates of the IT servers are embedded directly into the Cadence Reality Digital Twin, enabling near-real-time thermal inference. This allows teams to answer questions that were previously too slow to answer:

• Can this new AI rack be deployed somewhere in the existing hall without exceeding thermal limits?
• Which locations have sufficient power and cooling headroom?
• How does workload placement affect risk across the rest of the facility?

Instead of running a full CFD simulation for every scenario, the surrogate provides fast, physics-informed predictions that guide decision-making-dramatically expanding the number of options that can be explored, quickly.

Speed Matters: From Minutes to Seconds

Even a well-configured CFD CPU-solver simulation can take a couple of hours or more to evaluate a single placement scenario. At the AI-factory scale, this quickly becomes a bottleneck.

By contrast, AI surrogate inference inside the Cadence Reality Digital Twin Platform can deliver predictions in less than 10 seconds. This orders-of-magnitude improvement transforms capacity planning from a batch process into a dynamic and interactive one. Operators can explore "what-if" scenarios on demand, and AI Capacity Agents can continuously scan the facility to identify feasible deployment options across power, cooling, and space constraints.

Full CFD remains essential-but its role changes. Instead of being used everywhere, it is applied selectively, where accuracy and granularity matter most.

GPU-Accelerated CFD for Fast Validation

When high-fidelity validation is required, the Cadence Reality Digital Twin platform, which integrates NVIDIA Omniverse libraries and OpenUSD, leverages GPU-accelerated CFD to dramatically reduce simulation time. With GPU acceleration on NVIDIA Blackwell, full CFD runs can be completed up to ~32X faster than traditional CPU-only approaches.

This acceleration enables faster validation cycles and broader exploration of alternatives within the same engineering window. AI surrogates narrow the decision space in seconds; GPU-accelerated CFD confirms the best options with physics-grade accuracy. Together, they deliver a workflow that is both agile and rigorous-something neither AI nor CFD can achieve alone.

Figure 1: Cadence Reality AI Capacity Agent offers the best options for quick deployment.

The AI Capacity Agent: From Insight to Action

Building on surrogate models and accelerated simulation, Cadence Reality introduces the AI Capacity Agent-a decision engine that continuously evaluates available capacity across the digital twin.

The agent assesses power, cooling, and space constraints using AI-accelerated inference; ranks feasible deployment options based on performance, efficiency, and risk; escalates selected candidates to full CFD for validation; and generates comparison reports for side-by-side review.

This ensures that speed never comes at the expense of accuracy. AI narrows the field; physics confirms the final decision.

Continuous Learning Through the Digital Twin

As facilities evolve, the digital twin evolves with them. New telemetry, updated configurations, and continuously generated CFD results can be fed back into the training pipeline, allowing surrogate models to be refined over time.

This creates a continuous feedback loop between design intent and operational reality-a core principle of the Cadence Reality Digital Twin Platform. The outcome is not just faster decisions, but better ones, grounded in physics and validated before risk becomes reality.

Extending Agentic AI to On-Prem GPU Data Halls

The same agentic AI principles transforming semiconductor design are now being applied to physical AI infrastructure. Semiconductor companies already using Cadence Agentic AI solutions, such as ChipStack AI Super Agent, can now extend these workflows to their on-prem high-performance GPU data halls through the Cadence Reality Digital Twin Platform.

In this context, the Cadence Reality Digital Twin Platform becomes the physical counterpart to digital design, extending the optimization logic that drives PPA improvements in an AI chip to AI factory design and operations. This allows teams to evaluate GPU cluster deployment feasibility, such as for NVIDIA GB200 and GB300 NVL72 systems, predict thermal behavior under real workloads, and validate final configurations before capital is invested. A physical counterpart bridges the gap between chip design, system validation, and facility readiness, ensuring that the physical environments powering AI are as optimized and reliable as the chips themselves.

From Planning to Predict-and-Optimize

AI factories demand decisions at machine speed-but they still require engineering-grade accuracy. By combining AI Capacity Agents, physics-trained surrogate models, and GPU-accelerated CFD in a single platform, Cadence Reality Digital Twin Platform enables teams to move seamlessly from deployment dilemma to simulation to decision at unprecedented speeds.

The result is a new operational model-one that reduces time to first revenue, unlocks new revenue streams, reduces deployment risk, and enables confident scaling of AI infrastructure. In an era where every kilowatt and every degree matters, speed and accuracy are no longer tradeoffs. They are partners.

Learn more about the Cadence Reality Digital Twin Platform.