RAVEL Plans Power-Aware Scheduling Study for Distributed AI

RAVEL announced on November 12, 2026 a research collaboration with Sandia National Laboratories on Project Quicksilver, a multi-phase investigation into power-aware scheduling of AI workloads across geographically distributed computing environments. Phase One will cover three Sandia locations and compare air-cooled and immersion-cooled systems.

The study will examine whether power availability, operating conditions and workload requirements can guide scheduling and placement decisions to address energy, capacity and cost constraints. Its purpose is to inform data-center design and operations by investigating how to make power and computing capacity more productive, rather than assuming that additional capacity automatically produces more useful infrastructure. The announcement describes planned measurements, not completed findings or demonstrated savings.

RAVEL CEO and co-founder Denise Muyco said the project will investigate how orchestration can make AI infrastructure investments more productive while supporting more responsible energy use.

Three-Site Experimental Design

Phase One will evaluate the two cooling environments across three Sandia locations connected by a wide-area network. Box–Behnken experimental design will be used for the air-cooled environments and Plackett–Burman for the immersion-cooled environments. Network bandwidth and speed will be varied alongside other experimental factors.

The energy-efficiency analysis will compare identical workloads across cooling environments and account for compute, IT infrastructure and data-center operations. Measurements will include energy use, TUE and PUE-adjusted efficiency, and potential heat-reuse benefits. That scope connects workload placement to facility and cooling overhead rather than measuring computing equipment in isolation.

A separate total-cost-of-ownership analysis will assess how scheduling and placement affect computing costs across distributed environments. It covers capital and operating expenditures associated with infrastructure, power, cooling, networking and operations. A third workstream will measure AI workload time-to-completion, putting execution time alongside energy consumption and cost so researchers can examine the trade-offs between those objectives.

Federated Orchestration Across Locations

RAVEL describes Orchestrate AI as a federated orchestration layer for managing AI workloads across heterogeneous, geographically distributed resources. The intended model coordinates locations as one system rather than optimizing each environment independently. Its power-aware scheduling and policy-driven management account for available power, cost, performance requirements, infrastructure conditions, workload needs and operational priorities.

Dave Martinez, engineering program and project lead for Infrastructure Computing Services at Sandia, described the collaboration’s scope: “Under a Cooperative Research and Development Agreement with RAVEL, we’ll explore whether dynamic workload placement can make better use of available power and computing resources while balancing performance and cost.”

That experimental framing matters because moving or rescheduling a workload may change more than its energy consumption. The test design also tracks networking and operating costs, cooling environments and execution time, allowing the researchers to assess placement decisions across several infrastructure constraints.

Laboratory Expansion and SC26

Sandia is the initial research lead. The collaborators plan to expand Project Quicksilver to additional US Department of Energy national laboratories and explore further environmental and performance-optimization use cases.

RAVEL and Sandia are scheduled to discuss the experimental approach, planned test phases and anticipated research outputs at the Department of Energy booth during SC26 on November 17, 2026.

Project Quicksilver’s central question is whether treating separate data centers as a coordinated system can improve energy use and computing economics without undermining workload performance. Its measurement plan makes that a combined infrastructure question: scheduling decisions affect power consumption, but cooling, networking, capital costs and completion time also determine whether those decisions make distributed computing more productive.