What happened
Emerald AI and its partners tested an AI-factory control system that responds to grid stress by temporarily reducing electricity use. The demonstrations included a London cluster of 96 NVIDIA Blackwell Ultra GPUs, with NVIDIA reporting that priority workloads maintained throughput while more flexible jobs slowed.
NVIDIA’s blog, updated Aug. 20, 2026 and originally published in March 2026, describes a white paper and follow-on arXiv technical paper from Emerald AI, developed with NVIDIA, the Electric Power Research Institute, National Grid and Nebius. The proposal is to make AI factories responsive electricity loads: software would receive signals about grid stress and adjust power consumption while protecting workloads designated as highest priority. NVIDIA presents this as a way for large computing sites to reduce their withdrawals temporarily instead of requiring utilities to plan around their maximum possible demand at all times. These are claims from the NVIDIA source and associated work; this report does not independently verify them.
At Nebius’ new AI factory in London, the teams ran production-grade AI workloads on a cluster of 96 NVIDIA Blackwell Ultra GPUs connected through NVIDIA’s Quantum-X800 InfiniBand platform. NVIDIA says its System Management Interface supplied consistent, seconds-level GPU power telemetry. EPRI and National Grid simulated several grid-stress situations, including lightning strikes, extended periods of low wind generation and a sudden demand increase modeled on the so-called TV pickup effect. In that scenario, the teams reenacted the rapid electricity surge associated with millions of households turning on kettles during a football match. Signals from the simulated grid instructed the AI facility to lower its power use.
NVIDIA reports that Emerald AI’s Conductor Platform reduced the cluster’s power use without disrupting the highest-priority workloads, while more flexible jobs were temporarily slowed. The company says the system achieved 100% alignment with more than 200 power targets issued by EPRI and National Grid during the experiment. The follow-on paper reportedly found that the cluster could reduce power by approximately 30% within 40 seconds during emergency load-reduction testing. The source also says the testing measured CPUs and other equipment around the GPUs, rather than focusing only on accelerator power. It does not provide enough information here to assess the test duration, workload mix, baseline, or independent replication.
The source describes four demonstrations in total, including proof-of-concept trials at AI factories in Arizona, Virginia and Illinois, followed by the London work. NVIDIA and Emerald AI say they are preparing for real-world deployment at the Aurora AI Factory in Virginia, which the source says is scheduled to open this year. NVIDIA also describes two related products: DSX Flex, which translates grid signals into actions across GPUs, racks and jobs, and DSX MaxLPS, which manages power across GPUs, racks and workloads within a fixed budget. The blog does not establish that either product is broadly deployed or that the planned Virginia operation has begun.
Read the primary source: blogs.nvidia.com ↗
Why it matters
Large AI facilities are becoming important electricity customers. If the approach works outside controlled demonstrations, flexible computing demand could help utilities manage peaks and connect data centers using existing infrastructure more efficiently. The source does not establish lower electricity bills, emissions reductions or grid-wide reliability gains.
AI factories can concentrate substantial computing demand in a single location, creating challenges for utilities that must balance generation, transmission and local distribution capacity. The source argues that a facility able to reduce demand during short periods of strain could be treated as a controllable grid asset rather than a fixed load. That flexibility could make interconnection discussions easier because planners would not need to size every part of the system around the facility’s highest theoretical withdrawal. This is a plausible infrastructure benefit, but the source does not quantify how much capacity could be freed or how quickly a utility could connect a flexible site.
The approach also differs from simply shutting down an AI facility. NVIDIA says the London demonstration preserved peak throughput for selected high-priority workloads while slowing jobs considered more flexible. That distinction could matter for services with strict latency or availability requirements, as well as training tasks that can tolerate delay. However, the source does not explain who defines priority, how those decisions are enforced across customers, or what service-level commitments apply when flexible work is curtailed. It also does not state how much computing output was delayed, canceled or shifted to another time.
The public-interest argument is that lowering short-lived peaks could reduce pressure to build permanent generation and grid infrastructure for rare events, potentially helping keep electricity rates affordable. The source uses the example of a sudden kettle-driven demand surge to illustrate why fast response might be valuable. Yet electricity prices depend on many factors beyond one data center’s load profile, including generation costs, transmission investment, regulation and local constraints. There is no bill-impact analysis, emissions accounting or system-level reliability result in the supplied material, so those benefits should be treated as prospective rather than demonstrated.
The work is significant because it connects AI infrastructure with power-system operations, bringing a major technology load into discussions about demand response and interconnection. The collaboration includes a grid research organization and National Grid, and the London test reportedly covered total IT equipment power rather than GPUs alone. Even so, a controlled demonstration cannot by itself show that the approach is economical, reliable during real emergencies or suitable for every type of AI workload. The source also does not discuss cybersecurity, failure recovery, contractual responsibility or the consequences of a mistaken grid signal.
What to watch next
The next significant test is a planned deployment at the Aurora AI Factory in Virginia. Key questions include whether the system performs over sustained commercial operations, how utilities value and compensate flexibility, and whether interruptions, rebound demand or control failures create new costs or risks.
The Aurora AI Factory deployment is the clearest near-term test of whether the concept moves beyond demonstrations. Reporting should establish whether the system is operating with a live utility connection, how often power reductions are requested, how long reductions last, and whether the facility meets its promised computing performance while responding. It will also be important to distinguish a deployment announcement from measured operational results. The source says the Virginia site is set to open this year but does not provide a start date, utility agreement, operating data or independent assessment.
The follow-on arXiv paper warrants close examination for methodology. Important details include the definition of a power target, the response-time measurement, the equipment included in the 30% reduction, the duration and frequency of emergency tests, and the baseline against which performance was judged. Researchers and utilities should also examine whether the reported 100% alignment with more than 200 targets was measured under simulated conditions only and whether it held while workloads changed. Independent replication would help determine whether the results depend on the specific hardware, software configuration or carefully selected scenarios used in the London test.
Scaling is another unresolved issue. The demonstration involved 96 GPUs, while commercial AI factories may contain substantially larger and more varied systems with different networking, cooling and storage requirements. Cutting GPU power may not reduce total facility demand proportionally if other equipment remains active. Repeated throttling could also affect training schedules, inference latency, customer service agreements or demand immediately after a curtailment. The source does not say whether reduced work is rescheduled, abandoned or made up later, nor whether later rebound demand could offset some grid benefit.
Finally, utilities and data-center operators will need rules for participation. Questions include how flexible capacity is measured, whether sites receive compensation, which grid conditions trigger a reduction, how priority workloads are selected, and who bears the cost if automated controls fail. Public reporting should also address safeguards around control signals and recovery procedures. Until those operational, economic and governance questions are answered in real deployments, the evidence supports describing power-flexible AI factories as a promising tested concept rather than an established solution for global grid stability.
