NVIDIA and Emerald AI are redefining the relationship between artificial intelligence and power grids, unveiling a new architecture that treats AI factories as dynamic, responsive assets rather than static energy drains. The collaboration, presented at CERAWeek, integrates accelerated computing with real-time energy orchestration to accelerate AI deployment, improve efficiency, and bolster grid reliability. Built upon the NVIDIA Vera Rubin DSX AI Factory reference design and Emerald AI’s Conductor platform, this approach allows AI factories to generate valuable AI tokens while simultaneously adjusting to grid demands, reducing the need for excess infrastructure. “Power is a concern, but it’s not the only concern,” NVIDIA founder and CEO Jensen Huang recently stated, “That’s the reason why we’re pushing so hard on extreme codesign, so that we can improve the tokens per second per watt by many times each year.” This advancement signifies a critical step toward a more resilient and efficient energy future, supported by a growing ecosystem of partners including AES, Constellation, and NextEra Energy.
NVIDIA & Emerald AI Enable Power-Flexible AI Factories
A significant increase in computational efficiency is driving a redesign of AI factory design, positioning them as active participants in grid stabilization. NVIDIA and Emerald AI are leading this shift, unveiled at CERAWeek, by treating large AI deployments not as static power drains but as dynamically adjustable grid assets. This collaboration merges accelerated computing with real-time energy orchestration, promising faster grid connections and improved system reliability for substantial AI operations. Several energy companies, AES, Constellation, Invenergy, NextEra Energy, Nscale Energy & Power, and Vistra, are actively working to expand generation capacity to meet the escalating power requirements of these advanced facilities and plan to collaborate on optimized strategies. NVIDIA founder and CEO Jensen Huang describes this new computing paradigm as a layered system, with energy forming the foundational layer, highlighting the critical importance of efficient power management. Achieving this progress requires collaborative effort across the entire AI stack, from energy provision to application development, to fully realize its potential.
NVIDIA Vera Rubin: Tokens Per Second Per Watt Gains
Power efficiency has become the primary concern in modern AI infrastructure, shifting the focus from raw computational power to metrics like tokens per second per watt as the key determinant of performance. NVIDIA has demonstrated substantial gains in this area, reporting that the number of tokens generated within the same power budget has increased over one million times since the introduction of the NVIDIA Kepler GPU in 2012, culminating in the capabilities of the recently released NVIDIA Vera Rubin platform. This progression isn’t simply about hardware improvements; it necessitates a holistic approach to AI system design. Improving efficiency requires collaboration across the entire AI stack, encompassing energy providers, chip manufacturers, infrastructure developers, and application specialists, to create a resilient digital infrastructure capable of supporting both businesses and consumers globally and deliver a more sustainable future for artificial intelligence.
“That’s the reason why we’re pushing so hard on extreme codesign, so that we can improve the tokens per second per watt orders of magnitude every single year.”
AI-Driven Robotics & Digital Twins Accelerate Energy Projects
Maximo, a solar robotics company originating from AES, demonstrated significant progress in automated solar installation at CERAWeek, completing a 100-megawatt robotic solar installation at the AES Bellefield site. This achievement highlights a shift toward AI-driven construction methods capable of operating at utility scale. Utilizing NVIDIA accelerated computing, Omniverse libraries, and the Isaac Sim framework, Maximo’s approach improves installation speed, safety, and consistency, directly addressing the growing disparity between electricity demand and construction timelines. The company’s success underscores a broader trend of leveraging robotics to compress project schedules and enhance efficiency within the energy sector. TerraPower is also employing advanced simulation techniques to expedite the development of next-generation nuclear power plants.
Working with SoftServe, the company previewed an NVIDIA Omniverse-powered digital twin platform intended to drastically reduce the time required for advanced nuclear plant siting and design; this platform aims to compress design cycles from years to months, accelerating the deployment of TerraPower’s Natrium energy plants while simultaneously improving design quality and grid integration. Representatives at the conference explained that applying AI and simulation to early-stage engineering reduces design cycles from years to months. Beyond hardware and software, workforce development is receiving focused attention, with Adaptive Construction Solutions launching a national registered apprenticeship initiative in collaboration with NVIDIA. This program seeks to cultivate a skilled workforce capable of building and maintaining the infrastructure required for AI factories and advanced energy systems, expanding access to high-demand careers while supporting the rapid expansion of AI-driven power systems; the convergence of AI, digital twins, and workforce innovation is demonstrably delivering faster, more resilient energy infrastructure.
GE Vernova, Schneider Electric Scale AI with Simulation
The convergence of digital twins, validated designs, and integrated infrastructure is rapidly becoming essential for scaling artificial intelligence factories into reliable participants within the electrical grid, according to announcements made at CERAWeek. Companies like GE Vernova, Schneider Electric, and Vertiv are addressing the critical “power-to-rack” challenge by designing AI infrastructure as a unified energy and compute system from the outset, rather than as disparate components. GE Vernova detailed how high-fidelity digital twins, aligned with the NVIDIA Omniverse DSX Blueprint, enable utilities and developers to simulate grid behavior, substations, and AI factory loads together prior to deployment. “Such system-level modeling helps validate interconnection strategies, reduce risk and accelerate time to power in constrained grid environments,” the company stated. This pre-deployment simulation capability is crucial for managing the increasing demands placed on existing grid infrastructure as AI deployments expand.
Schneider Electric unveiled new validated NVIDIA Vera Rubin reference designs and lifecycle digital twin architectures developed with AVEVA, allowing operators to optimize performance per watt through simulation. By modeling power, cooling, and controls within Omniverse, Schneider Electric aims to enable more efficient and predictable operation of AI factories at scale and validate designs before construction even begins. Vertiv is contributing by offering converged, simulation-ready physical infrastructure built on repeatable power and cooling blocks, integrated with the Vera Rubin DSX reference design. This approach reduces both design and deployment complexity, facilitating faster and more confident scaling of AI factories. These combined efforts represent a digital pathway forward, providing the validated architectures and physical infrastructure necessary to transform AI factories into flexible, grid-aware assets capable of efficiently powering the world.
