How AI Factories Are Redefining Data Centers and How Cadence Is Helping

Data centers are becoming AI factories—production lines for intelligence — where power, cooling, shell, and compute must be designed and operated as a single system to achieve sustainability, performance, and profitability objectives.

At CadenceLIVE India 2026, Sherman Ikemoto, sales group director for Cadence data center tools and software, presented how simulation-driven digital twins can help turn every watt into measurable AI value.

From Traditional Data Centers to AI Factories

For years, data centers served as facilities for storing, processing, and moving information. They hosted applications, supported cloud services, and kept digital operations running reliably in the background. In business terms, they were often treated as overhead: essential infrastructure to be operated as efficiently and cost-effectively as possible while supporting the parts of the business that generated revenue.

AI factories are changing that model. Their output training runs, inference capacity, and model tokens can be sold or monetized directly, turning infrastructure from a cost center into a profit center. That shift changes every design and operating decision: additional investment in throughput can be justified by higher output, downtime becomes lost revenue priced by the hour, and capital planning is increasingly tied to expected future production rather than treated solely as an expense line.

This economic shift sits atop a major technological transition. Large-scale AI training and inference workloads require dense GPU clusters, high-bandwidth networking, coordinated power delivery, and advanced cooling to turn electrical power into accelerated compute at scale. As a result, data centers are evolving from capacity-driven facilities into performance-driven AI production environments, where GPU utilization, latency, cooling efficiency, uptime, and tokens per watt increasingly determine the value delivered by the infrastructure.

Preparing for Zettabyte-Scale AI Systems

The infrastructure build-out for AI is being described as one of the largest in history, with projections suggesting tens of trillions of dollars in investment over the next two decades. At the same time, the amount of data being created, captured, replicated, and consumed worldwide continues to grow dramatically and is projected to surpass 700 zettabytes (ZB) by 2030, driven heavily by the rise of generative AI and machine-generated workloads. International Data Corporation’s (IDC) Global DataSphere research tracks this expansion, highlighting why the infrastructure supporting AI has to evolve beyond traditional models.

Even the language we use to measure data is changing. In 2022, the General Conference on Weights and Measures expanded the International System of Units with new prefixes such as ronna and quetta, to express quantities beyond yotta-scale values. That matters because digital systems are moving toward levels of scale that were once unimaginable. Zettabytes are already part of that conversation.

Why AI Infrastructure Needs a New Design Model

In traditional data centers, facilities and IT systems are designed separately, manufactured independently, and integrated only after the data center becomes operational. AI factories require a unified design approach from the start because performance depends on how power delivery, thermal management, compute placement, network topology, and workload orchestration work together.

This shift also changes how AI infrastructure must be engineered and operated. A workload deployment decision can affect power draw, thermal behavior, coolant flow, network performance, and token output, so operators need a shared, physics-aware view of the entire environment—from chip to chiller.

Digital twins provide that foundation. In line with the Uptime Institute’s data center management maturity model, they help move infrastructure operations from basic monitoring toward automation, analytics, and AI-assisted optimization by continuously modeling, monitoring, and optimizing systems against physical constraints.

Scaling Compute, Cooling, and Coordination

The evolution toward AI factories has been shaped by both advances in computing and physical limits. The industry first focused on scaling up compute output at the chip level, driven in part by Moore’s Law. The rise of GPUs added massive parallel processing capability, but large AI workloads soon required scaling out across racks, pods, clusters, and entire facilities.

Cooling followed the same trajectory. Air cooling served conventional server environments well, while direct-to-chip liquid cooling and hybrid cooling are increasingly important for high-density AI infrastructure. Immersion cooling may become relevant for future architectures as rack power densities continue to rise.

Tokens per Watt: A New Measure of Value

AI factories also change the economics of data center management. In a traditional one-megawatt data center, a 25% reduction in PUE could yield approximately $200,000 to $300,000 in annual energy savings, depending on local electricity costs. That is a meaningful operational gain. In a one-megawatt AI factory, however, the same efficiency improvement can be applied to additional token generation, potentially creating an estimated $7 million to $12 million in incremental annual revenue. The management question shifts from “How do we save energy?” to “How much more intelligence can we produce from every watt?”

That is why tokens per watt is becoming a more useful measure of AI infrastructure value. It connects electrical, thermal, and compute performance directly to business output.

Power Planning and the Sustainability Challenge

Electricity supply has become a strategic design variable for AI factories. The International Energy Agency (IEA) projects that electricity generation for data centers will grow from 460 TWh in 2024 to more than 1,000 TWh by 2030 and to 1,300 TWh by 2035. Renewables are expected to meet nearly half of the added demand through 2030, while natural gas and coal together could account for more than 40%. Nuclear power, including small modular reactors, is expected to play a larger role later in the decade.

For AI factory operators, these trends make power planning, grid readiness, cooling efficiency, and carbon emissions core design considerations. Every watt now represents both a capacity constraint and a potential source of AI output, making accurate modeling essential from design through operations.

AI as Part of the Sustainability Solution

The sustainability picture is also more nuanced than demand growth alone. IEA analysis indicates that CO2 emissions from electricity generation for data centers are expected to peak around 2030 before gradually declining as the electricity mix becomes cleaner. At the same time, AI could help improve energy systems through better grid optimization, demand forecasting, materials discovery, industrial efficiency, and climate modeling.

The real challenge is designing and operating AI factories so that every unit of energy delivers efficient, useful output.

How Cadence Helps Enable the AI Factory Buildout

Cadence supports this shift by combining accelerated computing, physics-based simulation, digital twin technology, and AI-driven optimization. The Cadence Reality Digital Twin Platform enables teams to model data center infrastructure from design and commissioning through operations, helping them evaluate decisions on power, cooling, and performance before physical implementation.

Cadence Reality DC Design, within the Cadence Reality Digital Twin (DT) Platform, and Cadence Celsius EC Solver support the design of data center components and systems. Cadence Reality DC Insight in the DT platform extends these capabilities into operations by automatically synchronizing the digital twin with telemetry and providing an optimization capability to manage performance, cooling, and capacity.

Connecting design and operations is the Cadence DC Elements database, which provides behavioral models for data center components that can be selected and placed into a digital twin, like building blocks. If a required DC element is not available, Cadence will provide it as part of the software support agreement.

Cadence tools can also create AI surrogates of servers and complete data center environments, enabling physics-accurate results in near real time. In computational fluid dynamics (CFD), the physics engine underlying the digital twin, solving large-scale data center models, can be computationally intensive and time-consuming. AI surrogates accelerate this process, allowing teams to evaluate scenarios and obtain comparable results almost instantaneously.

Cadence is working with NVIDIA to advance AI factory digital twins by integrating the  Cadence Reality Digital Twin workflows with the NVIDIA Omniverse DSX() Blueprint to design, simulate, and optimize large-scale NVIDIA Grace Blackwell and Vera Rubin AI factories for training and inference before deployment. As AI factories shift their focus to tokens per watt—the number of model tokens processed per unit of power—Cadence system analysis and data center simulation tools, accelerated by NVIDIA libraries, enable teams to evaluate GPU power settings, system configurations, cooling architectures, and control strategies in a high-fidelity digital twin.

In a joint 10-megawatt (MW) AI factory use case, modeling reduced-power GPU operation through MaxQ showed up to 17% improvement in tokens per watt and the potential for billions of dollars in incremental annual revenue per gigawatt at scale. The result highlights how simulation-driven design can help optimize AI factory performance, energy efficiency, and business output before deployment.

Building AI Factories Responsibly

AI factories represent a new infrastructure model for producing intelligence at scale, in which power, cooling, compute, networking, operations, and economics must be designed together. The buildout will demand more energy, more sophisticated infrastructure, and more coordination across disciplines. It will also create new opportunities to optimize engineering decisions before they are implemented physically.

What we used to know as the data center is evolving into a tightly integrated, high-performance engine for intelligence. Cadence solutions help make that evolution more practical by connecting chip-to-chiller physics, design, and operational digital twins, validated infrastructure elements, accelerated simulation, and AI-driven optimization. If AI factories are built and managed intelligently, every watt, every chip, and every system can contribute to meaningful output.

Watch the on-demand webinar: To learn how simulation-driven digital twins can help maximize AI factory profitability and connect the broader data center ecosystem, watch Cadence Reality Digital Twin: Maximizing AI Factory Profitability and Connecting the Ecosystem Through Simulation.