Data centres are usually treated as fixed electricity loads attached to fixed buildings. Lumen Technologies and Emerald AI are exploring a different model: let the network help move suitable AI work to whichever location has the best combination of available power, cost and performance.
The companies announced on 29 October 2026 that they are evaluating an energy-aware routing pilot across Lumen’s edge and network infrastructure. Emerald would contribute software that manages power-flexible computing.
The collaboration is at the exploration stage. No operating pilot results, customer deployment or independently measured grid benefit were published.
Some AI work can move more easily than other work
A request for AI inference may be served from more than one data centre if the model and necessary data are available at each site. Batch jobs can sometimes wait or move without a user noticing. That creates a potential control lever unavailable to a factory or hospital.
Training runs with tightly synchronised accelerators are harder to relocate after they begin. Data residency, security rules and the cost of copying model checkpoints can also limit movement. Energy-aware routing therefore applies to a subset of workloads, not all computing.
The network must balance more than electricity
A routing engine would consider grid availability and power price alongside latency, bandwidth, security and application requirements. Sending a job to a site with cheap electricity is unhelpful if network delay breaks the service or data cannot legally cross a border.
The Emerald AI announcement says the companies are exploring automated workload shifts between locations. Lumen’s role matters because wide-area connectivity is the mechanism that turns several data centres into a selectable pool.
Flexibility could help grids without reducing total demand
If computation follows periods of renewable surplus or avoids a constrained region, it could make better use of existing generation and connections. A fleet might also reduce load during a grid emergency and catch up later.
That does not necessarily reduce total energy use. Moving a job changes where and when electricity is consumed; extra network traffic and duplicated infrastructure can add overhead. The value comes from flexibility, not free computing.
Emerald has said flexible AI data centres could unlock up to 100 GW on the existing US grid. That is a company estimate and should not be treated as demonstrated capacity.
A pilot must prove the control loop
The technical challenge is closed-loop operation: receive trustworthy grid or price signals, identify eligible work, reserve computing and network capacity, move or redirect the job, and verify service quality. Decisions must happen fast enough to matter without making systems unstable.
The Lumen release explicitly uses forward-looking language. A useful pilot would report how much load moved, for how long, at what latency and cost, and whether operators or grid partners could verify the response.
The architecture is more important than the announcement
AI infrastructure has focused on packing accelerators into ever larger campuses. Energy-aware networking points toward a complementary architecture: distributed sites coordinated as a flexible computing fleet.
Commercial incentives will decide whether operators expose that flexibility. Data-centre owners, cloud providers, utilities and customers would need contracts that share the value of shifting demand while assigning responsibility when a workload misses a service target.
That could reduce pressure for every workload to wait for new power at one location. It will only work where applications, data and commercial contracts permit movement. The collaboration is a credible experiment in aligning networks with grids—not yet evidence that AI demand has become a controllable power resource.
Featured photo: Taylor Vick/Unsplash. Contextual data-centre networking; not a Lumen or Emerald AI pilot.
Leave a Reply