Flexible AI Data Centers to Power the Grid: Emerald Conductor Demonstrates 25% Load Reduction in Phoenix

Overview

On a scorching Phoenix day, Emerald AI demonstrated how flexible AI computing can support the electricity grid. Their Emerald Conductor software orchestrated AI workloads across multiple data centers to cut peak power draw by 25% for three hours, without compromising training, fine tuning, or large language model tasks run on Nvidia GPUs. The demo highlighted a path to greater grid reliability and the potential to unlock large-scale AI investment by leveraging existing transmission and storage resources.

What this means

The presentation argues AI-enabled demand flexibility could turn data centers into active grid partners, reducing strain during peak hours and enabling faster integration of renewables, while Google and Nvidia share progress in scalable AI computing and grid collaboration.

Executive Summary

The transcript presents Emerald AI's pioneering work on flexible AI infrastructure designed to harmonize AI data center activity with electric power grids. By coupling an on-site software brain, the Emerald conductor, with spatiotemporal flexibility, the team demonstrates how AI workloads can be modulated in time and moved geographically to align with grid capacity. A May 2025 Phoenix, Arizona demonstration used a cluster of 256 GPU servers to reduce power consumption by 25% for three hours during peak demand, while maintaining performance for both flexible and inflexible AI tasks. This was achieved through temporal flexibility for batchable workloads and spatial flexibility via virtual transmission across regional data centers, leveraging fiber networks to relocate tasks without user disruption. The result is a compelling case that flexible AI could unlock up to 100 gigawatts of new data center capacity on existing grids and catalyze trillions of dollars in AI investment, all while supporting a cleaner energy mix by better integrating solar and wind power.

The narrative highlights broader industry momentum, including Google’s progress in scaling AI technologies nationally and internationally, and Nvidia’s role in delivering high-performance GPU hardware for demanding AI workloads. The DC Flex initiative and future demonstrations, including collaborations with National Grid in the United Kingdom, are positioned as pathways to demonstrate and extend the concept. The talk emphasizes that AI data centers can function as “shock absorbers” for grids, reducing energy prices and deferring grid upgrades by using flexible AI infrastructure that can adapt to real-time energy availability. It also outlines a roadmap toward power-friendly AI factories and next-generation data centers capable of dynamically aligning with grid conditions while supporting the rapid deployment and operation of large language models and other AI workloads.

Key Takeaways

Flexible AI computing can decouple heavy energy demand from grid peaks, enabling rapid AI growth without overburdening the grid. Temporal and spatial flexibility are foundational concepts for this model, allowing workloads to pause or relocate to regions with spare capacity. The approach envisions a broader ecosystem with on-site batteries and other energy equipment to maximize grid support and renewable integration. The collaboration with Nvidia and the broader energy-technology community signals a shift toward AI-enabled grid resilience and cleaner energy use as AI scales globally.