What happens to an AI factory when the cooling changes in seconds?
How NVIDIA uses Simcenter Amesim to explore AI factory cooling design, failure response and controls before physical deployment.
An AI workload starts across thousands of GPUs. Power demand rises, heat loads change and the AI factory cooling system must respond. But what happens if a coolant distribution unit fails at the same moment? Can the remaining pumps ramp up quickly enough? Will rack valves open in time? For NVIDIA, these are questions to investigate before an AI factory reaches commissioning—not only after equipment is installed.
In this NVIDIA GTC Berlin discussion about AI factory cooling, NVIDIA’s Kourosh Nemati describes how Simcenter Amesim provides an environment for end-to-end thermal-fluid modeling and for tuning cooling controls before real commissioning. Simcenter Amesim is Siemens’ physics-based system simulation software, used to model how mechanical, hydraulic, thermal and control components interact as one system. The story is not simply about seeing a data center in 3D. It is about understanding how the cooling system behaves—and using that understanding to inform engineering decisions.
Watch the full conversation
Hear Kourosh Nemati of NVIDIA and the Siemens experts discuss how simulation and digital twins can help engineers investigate liquid cooling for AI factories.
Key takeaways
- Simcenter Amesim supports end-to-end thermal-fluid modeling of AI factory cooling, from the chip to the heat rejection units, according to NVIDIA’s Kourosh Nemati.
- Engineers can investigate pipe sizing, fluid velocities, the number of coolant distribution units (CDUs), redundancy strategies and transient response, for example when a CDU fails.
- NVIDIA describes Simcenter Amesim as an environment for tuning AI factory cooling PID controllers before real commissioning and deployment.
- Behavioral models such as the Siemens intelligent rack valve can be exchanged through the FMI standard, and results can be viewed in a 3D scene in Digital Twin Composer.
- The discussion does not quantify commissioning-time savings or return on investment.
Cooling is now a facility-wide engineering challenge
Nemati describes a shift from racks dissipating around 10–20 kilowatts to racks dissipating hundreds of kilowatts within the same footprint. Liquid cooling addresses the need to remove that heat, but introduces an interconnected network of valves, pipes, coolant distribution units (CDUs), heat rejection equipment and controllers. Delivering enough coolant to one rack is only part of the AI factory cooling problem: engineers must understand how the whole network responds as operating conditions change.
The controls are interconnected too. Rack valves regulate local flow, while CDU and facility-level controls influence the conditions available to those racks. Optimizing each component separately does not establish how the complete cooling chain will behave. NVIDIA’s broader ambition is to bring power, thermal behavior, cooling and controls together rather than optimize them in isolation.
From rack valve to system: modeling liquid cooling with Simcenter Amesim
Simcenter Amesim supplies the physics-based system simulation behind that AI factory cooling investigation. Nemati describes thermal-flow network analysis extending from the chip to the heat rejection equipment. Within that network, engineers can examine component sizing, fluid flow and the interaction between cooling equipment and control logic.
“The Simcenter Amesim thermal flow network analysis is going from chip to the heat rejection units.”
The Siemens intelligent rack valve illustrates how physical equipment becomes part of this digital system. In the demonstrated workflow, its behavioral model includes mechanics, hydraulics and control logic. That model becomes part of a rack-level supercomponent, which in turn is integrated into a larger deployment-unit model in Simcenter Amesim. The speakers also describe using the Functional Mock-up Interface (FMI) standard to exchange behavioral models as part of simulation-ready assets.
This distinction matters: geometry shows where equipment sits; a behavioral model helps explain what it will do. The demonstration connects Simcenter Amesim simulation results to Digital Twin Composer, an NVIDIA Omniverse™-based application, so users can view temperatures, flow and valve behavior in a 3D scene. Simcenter Amesim provides the system-level physics; the connected visualization makes those results easier to explore in their spatial context.
Investigating CDU failure and transient cooling response
Nemati identifies several practical questions for the thermal-flow models: how large the pipes should be, what fluid velocities result, how many CDUs a deployment unit needs and which redundancy strategy is appropriate. Crucially, the investigation goes beyond steady-state capacity. When a CDU fails, the remaining units need time to increase pump output. Engineers need to understand whether cooling delivery remains sufficient during that transition—not simply whether the surviving equipment has enough capacity once it has reached its new operating point.
Changing AI workloads raise another transient question: can the controllers recognize the change and adjust valves and cooling delivery quickly enough? By exploring these scenarios in Simcenter Amesim, NVIDIA can investigate cooling-system response during design, before deployment and operation. The value is earlier insight into the interaction between equipment, thermal loads and controls, rather than a claim that a digital model alone guarantees reliable operation.
PID tuning and virtual commissioning before physical deployment
One of the clearest Simcenter Amesim-specific takeaways is the ability to prepare AI factory cooling controls before physical commissioning. Nemati explains that an end-to-end Simcenter Amesim model, built with carefully prepared simulation-ready assets, provides an environment to work through failure scenarios and tune proportional–integral–derivative (PID) controllers before real deployment.
“Simcenter Amesim is able to basically provide that environment for us… We can do the PID tuning before going to the real commissioning and basically the real deployment of the AI factories.”
The Siemens speakers describe a complementary virtual commissioning workflow: connecting the physical system model in Simcenter Amesim to simulated automation so engineers can test how control logic responds to events such as equipment failures or leaks. Reusing the design model makes the engineering work relevant beyond component sizing—it also supports preparation of the controls that will operate the cooling system.
For AI factory cooling projects, the intended benefit is to resolve more questions before arriving on site and reduce the amount of troubleshooting left to physical commissioning. The transcript supports that engineering rationale and NVIDIA’s use of the modeling environment. It does not provide a measured NVIDIA commissioning-time reduction or a quantified Simcenter Amesim return on investment.
From design to an operating digital twin
The conversation also looks beyond design and commissioning. Siemens describes deploying models to controllers or edge devices as executable digital twins, with potential applications including leakage detection and energy-optimized control. Nemati highlights the importance of connecting equipment telemetry to operational models to understand workload changes and protect cooling delivery.

These operational applications should be understood as a direction for further work, not as verified NVIDIA deployment results from this discussion. The same applies to real-time controller optimization and energy savings: the speakers explain their potential, but do not establish achieved savings from an Simcenter Amesim-based operational twin.
Conclusion: understanding AI factory cooling before deployment
The strongest story is specific: NVIDIA describes using Simcenter Amesim for end-to-end AI factory cooling analysis and as an environment for pre-commissioning PID tuning. The connected workflow brings component behavior, rack models and deployment-unit cooling into a system-level investigation, helping engineers explore sizing, redundancy and transient response before hardware deployment.
For NVIDIA’s AI factories, that is the role of Simcenter Amesim: turning an AI factory cooling design into a system whose behavior engineers can investigate before they depend on it in the real world.
Frequently asked questions
How does Simcenter Amesim support AI factory cooling design?
Simcenter Amesim provides physics-based system simulation for the cooling network. In the GTC Berlin discussion, NVIDIA’s Kourosh Nemati describes thermal-flow network analysis from the chip to the heat rejection units. Engineers can use it to examine pipe sizing, fluid velocities, the number of CDUs per deployment unit and redundancy strategies.
What happens to AI factory cooling if a coolant distribution unit (CDU) fails?
The remaining CDUs need time to increase pump output. Engineers therefore need to understand whether cooling delivery stays sufficient during that transition, not only whether the surviving equipment has enough capacity once it reaches its new operating point. Simulating the scenario in Simcenter Amesim lets teams explore this transient response during design.
Can AI factory cooling PID controllers be tuned before commissioning?
According to Nemati, an end-to-end Simcenter Amesim model, built from carefully prepared simulation-ready assets, provides an environment to work through failure scenarios and tune PID controllers before real commissioning and deployment. The discussion does not quantify the outcome of this tuning.
How do Simcenter Amesim results connect to Digital Twin Composer?
In the demonstrated workflow, Simcenter Amesim simulation results are connected to Digital Twin Composer, an NVIDIA Omniverse-based application, so users can view temperatures, flow and valve behavior in a 3D scene. The speakers also describe using the Functional Mock-up Interface (FMI) standard to exchange behavioral models, such as that of the Siemens intelligent rack valve.
Does the discussion show measured time or cost savings?
No. The discussion supports the engineering rationale of resolving more questions before arriving on site, but it does not provide a measured commissioning-time reduction or a quantified return on investment for Simcenter Amesim. but it does not provide a measured commissioning-time reduction or a quantified return on investment for Simcenter Amesim.