Flexible data centers could ease grid bottlenecks

Startups, utilities and chipmakers are testing ways for computing facilities to reduce electricity use during grid stress. The approach could speed connections, but critics warn it cannot replace new generation and transmission.

Emerald AI is preparing to deploy its Conductor software at a new facility in Virginia’s Data Center Alley, with partners including Nvidia and Digital Realty. The system is designed to reduce a data center’s power draw when grid demand spikes while preserving the most urgent computing work, a model its backers describe as a “power-flexible AI factory.”

The push reflects a broader effort to fit data centers into existing electricity systems rather than waiting years for new power plants and transmission. Studies cited in the sector suggest major gains from limited flexibility: a 2025 Duke University report found the US grid could support an additional 76 gigawatts if facilities reduced usage just 0.25% of the time, or about 22 hours a year. In Oregon, Aligned Data Centers plans to install a 31-megawatt battery in May 2027, helping Portland General Electric offer 80 megawatts of added capacity without new power plants.

Conductor has moved from simulations to increasingly demanding tests, including control of server racks with 256 Nvidia A100 GPUs and a 25% power reduction for three hours. Its largest trial is planned in Manassas, where it will manage a 96-megawatt hyperscale AI factory on a live grid. Supporters say flexibility can improve reliability and speed interconnection, while skeptics argue it remains an optimization tool, not a substitute for expanding generation, transmission and distribution.

72

Impact Score

AMD and Rackspace plan dedicated AI compute rollout

AMD and Rackspace have finalized a phased deployment for dedicated AMD-based compute across Rackspace data centers. The capacity is aimed at regulated enterprise workloads, including clinical AI and large-scale inference.

Lexar tests SSD offloading for local AI models

Lexar is developing an AI-focused SSD approach designed to cut DRAM demand when running large language models on consumer PCs. Internal tests show the company’s storage offloading can load models that traditional local frameworks struggle to run with limited memory.

NVIDIA Blackwell leads MLPerf Training 6.0

NVIDIA’s latest MLPerf Training 6.0 results put Blackwell across every benchmark in the suite, including new MoE workloads. Partner systems from Microsoft Azure and CoreWeave highlighted large-cluster runs on Llama 3.1 405B and DeepSeek-V3 671B.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.