Key takeaways

  • An Emerald AI-led test used a 130 kW cluster with 96 NVIDIA Blackwell Ultra GPUs at a Nebius facility in London.
  • The cluster cut power by about 30% within 40 seconds in one test and sustained requested reductions of 10% to 40% for two to ten hours in others.
  • A separate Oracle demonstration shifted inference between 80-H100 clusters in Virginia and Illinois when the Virginia cluster was power-constrained.
  • These are meaningful technical results, but they do not prove that a hyperscale AI campus can run without firm power or that 100 GW of capacity is already available.
  • NVIDIA says its Vera Rubin DSX reference design includes DSX Flex for grid services; commercial-scale deployment statements remain forward-looking.

What Emerald Conductor controls

Power-flexible computing starts with a simple idea: not every AI job needs identical treatment when the grid is stressed. Interactive inference may need stable latency, while some training or batch jobs can tolerate a temporary slowdown.

In a June 2026 industry-authored preprint, Emerald AI and collaborators describe Conductor as the software layer between a grid request and a GPU cluster. The architecture has three main parts:

  1. A grid interface receives a target, start time, duration and ramp requirement from a utility or system operator.
  2. A workload orchestrator chooses control actions such as GPU power caps or delaying lower-priority jobs.
  3. Telemetry and power models estimate the cluster's draw and help the controller stay near the requested target.

The experimental implementation integrated with SLURM job priorities. Critical jobs retained stronger performance protection, while lower-priority jobs supplied more of the power flexibility. That distinction matters: “flexible” does not mean every workload can be interrupted without consequence.

The 130 kW Blackwell Ultra experiment

The main test ran for five days at a Nebius AI facility in London. It used 96 NVIDIA Blackwell Ultra GPUs in a cluster rated at 130 kW and exercised a mix of fine-tuning, multimodal training and batch-inference workloads.

National Grid and EPRI supplied 22 dispatch events derived from historical grid conditions. Across the campaign, those events contained more than 200 individual power targets and ramp constraints. The paper reports that the system met all requested targets in the tested scenarios.

The most concrete results were:

  • About 30% load reduction within 40 seconds during a surprise emergency-dispatch scenario.
  • 40% reduction in roughly one minute in additional tests.
  • Requested reductions of 10% to 40% lasting two to ten hours, with near-baseline throughput maintained for the highest-priority jobs.
  • Repeated power adjustments while jobs continued to arrive and complete.

These measurements show that software can make a working GPU cluster respond to grid-style instructions on several timescales. They do not establish the same performance at a 100 MW or 1 GW campus. The tested cluster was orders of magnitude smaller, and larger facilities add cooling, networking and electrical-system constraints that the paper itself identifies as scaling challenges.

What the Virginia-to-Illinois shift showed

The researchers also tested geographic load shifting between two Oracle data centers: an Ashburn, Virginia, cluster and a Chicago, Illinois, cluster. Each site used 80 NVIDIA H100 GPUs and was rated at 60 kW.

During the test, the team placed a 375 W power limit on the Ashburn GPUs, ramped the site down over 15 minutes and held the limit for three hours. A load balancer redirected inference requests to Chicago, where power draw rose by 3.1 kW as that cluster absorbed work.

There was a user-visible tradeoff. Average time to first token at the constrained Ashburn cluster increased by about 30 milliseconds. Chicago saw a brief latency spike until its autoscaler added capacity.

This is evidence that inference demand can move between prepared regions while one site reduces power. It is not the same as eliminating the electricity demand: some compute and power use moved to another data center.

Where NVIDIA DSX Flex fits

NVIDIA's announcement with Emerald AI and energy partners says the Vera Rubin DSX AI Factory reference design includes a DSX Flex software library for connecting AI factories to grid services. The company also says Emerald Conductor can coordinate compute flexibility with on-site generation, batteries and other behind-the-meter resources.

NVIDIA reports that the companies trialed flexibility at five commercial data centers and expected DSX Flex to reach commercial-scale deployment later in 2026 at the NVIDIA AI Factory Research Center in Virginia. Those statements describe a roadmap and company-reported trial history, not an independently verified commercial fleet.

The same announcement says power-flexible AI factories could help unlock up to 100 GW on the U.S. power system. That is a potential estimate, not a measured result from the 130 kW experiment. It depends on adoption, workload mix, grid rules and whether operators can reliably commit curtailment when the system needs it.

What grid planners can take from the tests

The strongest conclusion is narrower than the marketing headline: a production-style GPU cluster can follow power targets by combining scheduler priorities, hardware power caps and live telemetry, while protecting its most important jobs better than its flexible ones.

For utilities, that creates a possible alternative to treating every proposed data center as a fully firm load in every hour. For operators, it creates a new engineering constraint: they must define which workloads can slow down, measure the service impact and prove that promised reductions are dependable.

The next evidence to watch is commercial-scale performance. Useful disclosures would include facility-level power rather than GPU-only measurements, cooling response, event failure rates, contractual availability and results across hundreds of megawatts. Until then, Emerald AI's tests are a credible demonstration of the control method—not proof that flexible computing has solved the grid-capacity problem.

Sources

This article was researched and fact-checked against the following sources: