
Share
Project Quicksilver will test whether federated, power-aware scheduling can make distributed AI infrastructure more productive, not just bigger, as data centers hit power and cost walls.
Building more data centers doesn't automatically mean getting more useful compute out of them. That's the core problem RAVEL and Sandia National Laboratories are digging into with Project Quicksilver, a new research collaboration announced this week that uses RAVEL's Orchestrate AI platform to study dynamic, power-aware orchestration of AI workloads across geographically distributed computing environments.
The premise is straightforward even if the engineering isn't: power availability, not raw GPU count, is increasingly the bottleneck constraining AI infrastructure. Utilities can't always deliver megawatts where and when operators want them. Costs vary wildly by region and time of day. Regulatory requirements differ across jurisdictions. Project Quicksilver wants to find out whether orchestration software that's aware of all these constraints can squeeze more productive output from the same power and compute footprint, instead of just scaling horizontally and hoping the grid keeps up.
"The next phase of AI infrastructure leadership won't be won by simply building more," said Denise Muyco, CEO and co-founder of RAVEL. "It will be determined by how efficiently and productively we use every megawatt, every GPU, and every dollar invested." That's a notable framing shift for an industry that's spent the last few years treating capacity as the only metric that matters.
RAVEL Orchestrate AI is built around a concept the company calls federation: managing workloads across multiple, heterogeneous locations as one coordinated system rather than optimizing each site in isolation. In practice, that means the scheduler isn't just asking "which node has a free GPU?" It's asking a more complex question that factors in:
That's a meaningfully different architecture from the siloed scheduling most data centers run today, where each facility manages its own resource pool independently and workload placement decisions rarely account for what's happening at a sister site hundreds of miles away.
"AI infrastructure can't reach its full potential if every data center operates as an island," said Philippa Carroll, RAVEL's Chief Product Officer. "As AI infrastructure becomes more distributed, the challenge is no longer just how much computing capacity we have, but how effectively we can use it."

For readers who work in HPC scheduling, this will sound familiar in spirit to workload managers like Slurm or Kubernetes schedulers, but the power-aware, cross-site federation piece is the novel part. Most production schedulers today optimize for availability and queue depth within a single cluster or cloud region. Extending that logic across geographically distributed, heterogeneous environments, while factoring in real-time power and cost signals, is a harder distributed-systems problem, and it's exactly the kind of thing a national lab is well positioned to stress-test at scale.
Sandia is serving as the initial research lead on Project Quicksilver, with plans already in motion to expand the effort to additional U.S. Department of Energy national laboratories. The project is explicitly multi-phase, and RAVEL says it intends to explore additional use cases beyond the initial scope, including further environmental and performance optimization work. No specific hardware configurations, benchmark targets, or timelines for those later phases have been published yet, so there's not much to independently verify on the technical implementation beyond what's described conceptually.
That's worth flagging. This is a research collaboration announcement, not a results paper. There's no benchmark data here, no reported energy savings percentage, no comparison against baseline scheduling approaches. What we have is a stated research question and a plan to investigate it jointly with a DOE lab, which carries some credibility given Sandia's long track record in HPC and computational infrastructure work, but the actual findings are still ahead of us.
If you want more detail before the formal papers land, RAVEL and Sandia are presenting at SC26 in Seattle. The talk, "Project Quicksilver: Dynamic Power-Aware AI Orchestration Across Distributed Computing Environments," is scheduled for Tuesday, November 17 at 2:30 p.m. CST at the DOE booth. The presentation is billed as covering the experimental approach, planned test phases, and anticipated research outputs, which should give a clearer picture of methodology than this announcement does.
Project Quicksilver sits at the intersection of two trends that have been building all year: the industry's growing anxiety about power constraints on AI scaling, and a push toward smarter, cross-site orchestration as a lever for squeezing more value out of existing infrastructure rather than just building more of it. RAVEL's federated approach, treating distributed data centers as one coordinated system instead of isolated islands, is the architectural bet here, and Sandia's involvement gives it a credible research testbed.
The caveats are real, though. This is early-stage, multi-phase research with no published performance data yet. The claims about efficiency gains are aspirational framing from a vendor announcement, not independently verified results. Anyone evaluating whether power-aware federated orchestration is production-ready should wait for the actual experimental outputs Sandia and RAVEL say are coming, and the SC26 presentation is the first checkpoint to watch for specifics on methodology and test phases. If the approach holds up under DOE-scale testing, it could inform how the next generation of distributed AI infrastructure gets designed and operated, particularly as more labs join the effort.
Tags
Original Sources
RAVEL and Sandia to Present Project Quicksilver at SC26 - HPCwire
↗ https://www.hpcwire.com/off-the-wire/ravel-and-sandia-to-present-project-quicksilver-at-sc26
About the author
Kai built ML infrastructure at a Bay Area startup before developing an obsession with transformer architectures and inference optimisation that eventually pulled him out of product work entirely. A stint at a compute research lab sharpened his instinct for what actually matters in a model release versus what is marketing. He writes from the inside — from the perspective of someone who has debugged the systems he is describing at three in the morning. He is allergic to hype and instinctively drawn to the unglamorous plumbing questions that everyone else skips over.
More from The Engineer →This Week's Edition
13 November 2026
14 articles
Related Articles

Tether Evo's New Research Tackles BCI's Calibration Problem with Cross-Subject Models
Models & Research · 5 min

Why Some Engineers Think the Ocean Floor Is the Future of Data Centers
Environment & Climate · 5 min

Europe's Digital ID Push Tests a New Privacy Tool: Zero-Knowledge Proofs
Policy & Regulation · 5 min
Related Articles

Tether Evo's New Research Tackles BCI's Calibration Problem with Cross-Subject Models
Models & Research · 5 min

Why Some Engineers Think the Ocean Floor Is the Future of Data Centers
Environment & Climate · 5 min

Europe's Digital ID Push Tests a New Privacy Tool: Zero-Knowledge Proofs
Policy & Regulation · 5 min
More Stories
© 2026 Cedar & Bloom. All rights reserved.