Thought Leaders
Beyond the GPU: Optimizing the AI Factory as a System
Add Unite.AI to your preferred sources on Google
An AI factory is not a bigger data center. It is specifically built to manufacture intelligent tokens, not to serve web pages and email. When a hyperscaler or neocloud puts billions of dollars and years into building an AI factory, the GPUs doing the computational heavy lifting naturally become a priority. But optimizing the GPU alone does not optimize the factory around it. Power and thermal systems have to support that compute, and the decisions made in one system can have significant consequences for the others. AI exists to maximize complex systems, yet the factories building it have often been optimized one component at a time. Virtual twin technology allows engineers to build and streamline the entire system, before construction even begins.
Why Conservative Design Became the Safest Option
With the looming consequences of a system failure or performance bottleneck so high, early facilities were built with excessive safety margins and hardware. Without a roadmap for this new generation of data centers, building with caution could seem like the safest approach. These margins helped protect against issues such as thermal spikes near the chips, which can leave operators with as little as a minute to respond before hardware fails. In some early facilities, as much as 10 percent of available power was stranded as an unused margin.
AI factories are exposed to temperature, weather, equipment failures, inadequate capacity, design deficiencies, and maintenance issues. With such high stakes, at gigawatt scale, each day of delayed capacity can represent roughly $100 million in lost token-generation revenue. Caution felt necessary when there was no way to predict the outcome of those issues.
The challenge becomes even greater when these risks span multiple, independent systems. An AI factory depends on three tightly coupled systems: compute, power and thermal management. The GPU requires an elaborate power chain to feed it and an equally demanding thermal system to remove the heat it generates. When compute, power and thermal management are treated as separate steps, each optimized within its own domain; teams can miss how a decision in one system affects the others. This creates blind spots and imbalances, particularly when power and cooling are treated as infrastructure requirements to address after the compute decisions are made.
Power and cooling cannot be separate from compute. They are part of the system that determines how effectively the compute can operate. The shift, therefore, is from optimizing individual systems to optimizing the entire factory.
The Case for Co-Optimization
Fixing this means going from optimizing each system independently to co-optimizing the factory as a whole. Physical construction forces decisions into sequence: one system is designed, then another is built around it, and another is added to support it. That makes it difficult to optimize the factory as a whole because each decision narrows the options available for the systems that follow. Co-optimization requires moving that process into a virtual environment, where the entire system can be modeled, changed, and evaluated before anything is physically locked in, creating a dynamic blueprint.
Modeling the Factory Before It Exists
This is where system-level virtual twin technology becomes critical.
A virtual twin uses formal systems modeling to represent compute, power and thermal infrastructure together and can model the AI factory before it physically exists. Unlike a traditional digital twin focused on an existing physical asset, a virtual twin can model a system before it physically exists. It allows engineers to test different configurations and understand how changes in one part of the system affect the others before those choices become physical commitments.
For example, increasing GPU density does not simply mean adding more compute. It can increase electrical load, change heat-rejection requirements, affect equipment placement, and introduce new commissioning considerations. Modeling those relationships together allows engineers to evaluate the trade-offs before construction rather than discovering them as the project progresses.
Virtual twin technology allows engineers to make an expensive mistake virtually before it becomes an expensive one in the physical world. It provides a way to identify potential problems and opportunities in the virtual environment, allowing teams to solve problems before they exist in the physical world. Testing configurations before physical commissioning also means unnecessary design margins; stranded power and safety-margin waste can be caught and removed at the design stage. On a large scale, even small reductions can translate into millions in capital savings and earlier access to revenue.
Making the Model Predictive
The model can also use physics-based simulation to predict system behavior. Where the underlying physics or chemistry is well understood, simulation tools can predict system behavior. When abundant operating data offers a better approach, AI and machine learning models can augment those simulations.
The result is an environment for exploring how the coupled system behaves and orchestrating decisions across engineering domains. The model can also evolve as operational data becomes available, allowing lessons from one facility or operating cycle to inform future designs and decisions.
A Blueprint for the AI Factory
The industry is already assembling the foundation for this approach. NVIDIA’s DSX blueprint is codifying a reference design for gigawatt-scale AI factories. NVIDIA has begun by codifying the compute portion of that blueprint using systems modeling and a virtual-twin simulation approach, and is inviting ecosystem partners spanning power, cooling and construction, including companies like Eaton, Vertiv and Jacobs, to contribute their own portions of the design. The interoperability of reference designs is enabled through SysML which supports system-level definition, while OpenUSD provides a foundation for 3D-centric data.
The goal is not for every company to optimize its own component and then expect those pieces to fit together. Compute, power, cooling and facilities expertise need to be brought together so that each best-in-class technology contributes to the performance of the entire system.
Each industry can optimize within its own domain, but creating a powerful final product requires those systems to be optimized together. As this approach matures, it’s also what lets AI infrastructure move beyond isolated, one-off pilot projects toward enterprise-scale systems capable of managing highly complex operational environments.
Optimizing the Factory, Not Just the GPU
The GPU will remain central to AI-factory performance. But buying the best GPUs does not, by itself, optimize the factory around them. The next generation of AI infrastructure will depend on engineering compute, power and thermal systems as one integrated system, modeling those relationships before major physical commitments are made, and continuously using what the system learns to improve what comes next. We can’t predict the future or prevent every mistake, but virtual twin technology can help teams plan for uncertainty and optimize AI factories as a system.
英文原文如下；中文版正在随流水线自动整理，稍后刷新即可。