Dispatches from Scottsdale: Thoughts On Data Center Power Needs From DTECH

Research evidence excerpt

Dispatches from Scottsdale: Thoughts On Data Center Power Needs From DTECH

ute resources up or down based

on local grid capacity, or even seamlessly shift workloads geographically (e.g., from Virginia

to Illinois) via fiber-optic connections to avoid local grid congestion. Research indicates

that managing just 0.5% of the maximum uptime of these facilities could unlock the ability

to interconnect 100 gigawatts of large electricity load. Investors should look for operators

utilizing flexible compute, as utilities are beginning to offer "provisional" or "connect and

manage" agreements that allow flexible data centers to connect to the grid years earlier than

rigid ones.

"DC Flex", an initiative spearheaded by the Electric Power Research Institute (EPRI) in

partnership with companies like Nvidia, transforms AI data centers into grid assets by enabling

them to operate as extremely large, highly responsive demand-response units.

Through integrated software and operational frameworks, DC Flex allows AI facilities to

dynamically manage their power draw in several key ways that were highlighted at the

conference:

■Software-Driven Workload Modulation: DC Flex utilizes reference designs and software

stacks, such as "DSX Flex," which create a digital "handshake" between the data center and

the grid. This allows AI factories to automatically scale their compute workloads up or down

based on real-time local power availability.

■Geospatial Load Shifting: The initiative has successfully demonstrated the ability to

seamlessly shift AI compute workloads across geographic locations in real time. For

example, operators can shift active workloads from a facility in Virginia to one in Illinois via

fiber-optic networks, immediately reducing the power burden on a congested local grid