At GTC, Emerald AI highlights its deep integration with NVIDIA systems as well as a progression of technical milestones culminating in a newly completed Oregon demonstration on the full NVIDIA DSX software stack in partnership with Portland General Electric and EPRI.

SAN JOSE, Calif., March 16, 2026 — Emerald AI today announced a newly completed demonstration in Hillsboro, Oregon, in collaboration with NVIDIA, Portland General Electric, and EPRI, that shows how AI factories can respond precisely to utility signals, while maximizing the performance of priority AI workloads. At NVIDIA GTC, Emerald AI is highlighting how its software, integrated with the full NVIDIA DSX software stack, helps operationalize DSX Flex in live AI factory deployments to make next-generation AI factories power-flexible by design.
This demonstration is a key milestone toward this year’s launch of the first commercial-scale AI Factory with DSX Flex, NVIDIA’s 96MW Vera Rubin AI Factory Research Center in Virginia. At scale, power-flexible AI factories could unlock up to 100 GW of grid capacity on the existing U.S. power system.
DSX Flex makes this possible, turning AI factories from static loads into controllable grid assets and creates a new pathway to faster, larger interconnections for advanced AI infrastructure. By enabling large AI facilities to flex demand during limited periods of grid stress, this approach can speed AI innovation, make better use of existing electric infrastructure, and protect energy affordability and reliability for local communities.
Over the last year, Emerald AI—an NVentures portfolio company—and NVIDIA have advanced this model through a series of increasingly sophisticated demonstrations. Emerald AI has now completed five live demonstrations at commercial data centers that have each showcased deep integration with NVIDIA systems.
That progression began with a demonstration in Phoenix, Arizona in partnership with Oracle, NVIDIA, EPRI, and the Salt River Project power utility, which was published in Nature Energy as the first peer-reviewed evidence of AI power flexibility. It expanded to a demonstration of spatial flexibility that shifted AI inference workloads from Virginia to Chicago to maintain low latency inference serving, while relieving winter peak grid load in Virginia; that demonstration ran on NVIDIA Hopper systems on Oracle Cloud Infrastructure and integrated NVIDIA Dynamo framework.
This month, Emerald AI, NVIDIA, National Grid, EPRI and Nebius announced the UK’s first live demonstration of grid-responsive AI infrastructure in London, where Emerald AI orchestrated a cluster of NVIDIA Blackwell Ultra GPUs and reduced electricity demand by over a third in under a minute, while high-priority workloads continued to run.
Now, in collaboration with Portland General Electric and EPRI, the Hillsboro, Oregon demonstration represents the largest and most advanced joint technical demonstration of AI power flexibility by Emerald AI and NVIDIA to date. This demonstration showcased Emerald AI software integrated with the full NVIDIA DSX software stack, including DSX Flex and DSX Max-Q, built on NVIDIA Grace Blackwell Ultra clusters running realistic production-grade AI workloads at scale defined by NVIDIA. The system demonstrated it could meet requested power and ramping targets while maximizing workload performance under power constraints. Portland General Electric sent utility-style dispatch commands across a range of grid-relevant scenarios, including:
Taken together, these milestones trace a clear technical progression across NVIDIA systems—from early field validation, to cross-region inference shifting on Hopper-based clusters, to live power flexibility on NVIDIA Blackwell Ultra rack-scale systems in London, and now to full DSX stack integration on Grace Blackwell Ultra clusters in Oregon. Emerald AI is now ready to integrate with NVIDIA DSX Flex at commercial scale, at the 96 MW Aurora facility in Manassas, Virginia, and test power flexibility at large scale in collaboration with Digital Realty, EPRI, and the PJM Interconnection.
“Emerald AI is proud to team with NVIDIA to demonstrate that AI factories do not need to behave as inflexible, always-on loads,” said Varun Sivaram, founder and CEO of Emerald AI. “Alternatively, we’ve demonstrated through our partnership that they can become software-defined, measurable, and dispatchable assets that accelerate time-to-power for advanced AI infrastructure, while utilizing our existing power system and avoiding expensive grid upgrades to ensure grid reliability.”
“The massive power requirements of next-generation AI factories require a shift toward grid-responsive infrastructure that can balance peak demand with continuous computational performance,” said Vladimir Troy, vice president of AI infrastructure at NVIDIA. “NVIDIA’s work with Emerald AI on DSX Flex enables these large-scale facilities to function as controllable grid assets, unlocking new capacity while keeping high-priority AI workloads on track.”
At GTC, Emerald AI is also sharing technical materials on flex-ready AI factories. These include its report with National Grid, Nebius, and EPRI, “Power Flexible AI Factories: A UK-First Demonstration of Grid Responsive AI Infrastructure,” as well as its recent report with InfraPartners, “Enabling Grid-Integrated, Flex-Ready Data Centers,” to operationalize how AI infrastructure built to the NVIDIA Vera Rubin DSX AI Factory reference architecture can utilize DSX Flex from day 1. Emerald AI founder and CEO Varun Sivaram will speak at GTC on March 18 at 9:00 a.m. in the session “Build Flexible and Secure Power Infrastructure for a Sustainable Future at Scale,” alongside leaders from NVIDIA and across the energy and digital infrastructure ecosystem. Visit the Emerald AI booth at the DSX Infrastructure Pavilion for more details on technical milestones, partnership opportunities, and community benefits of power-flexible AI.