Résumé IA
Lors du match de l'Euro 2020 entre l'Angleterre et l'Allemagne, des millions de téléspectateurs britanniques ont allumé leur bouilloire à la mi-temps simultanément, provoquant un pic de demande d'environ 1 gigawatt sur le réseau électrique national — l'équivalent d'un réacteur nucléaire standard. C'est ce phénomène, surnommé le "TV pickup", qui a inspiré une démonstration inédite menée en décembre 2025 à Londres par Emerald AI, en collaboration avec NVIDIA, EPRI, National Grid et Nebius. L'expérience s'est déroulée dans une "usine IA" construite sur l'infrastructure NVIDIA de Nebius, équipée de 96 GPU NVIDIA Blackwell Ultra connectés via la plateforme InfiniBand NVIDIA Quantum-X800. En simulant ce même pic d'énergie lié au match de football, le cluster IA a automatiquement réduit sa consommation pour absorber le choc — sans interrompre les charges de travail prioritaires. Cette technologie, baptisée Emerald AI Conductor Platform, ouvre une perspective concrète pour la gestion des réseaux électriques sous tension. Les usines IA, habituellement perçues comme de nouveaux fardeau énergétiques, deviennent ici des actifs flexibles capables d'ajuster leur consommation en quelques secondes selon des signaux envoyés par les gestionnaires de réseau. Lors des tests, le système a respecté 100 % des plus de 200 cibles de puissance définies par EPRI et National Grid, couvrant non seulement les GPU mais aussi les CPU et l'ensemble des équipements informatiques. En pratique, cela signifie que le réseau peut gérer les pics de demande avec les capacités existantes, sans avoir à construire d'infrastructures permanentes surdimensionnées pour les scénarios les plus extrêmes — ce qui contribue directement à limiter la hausse des tarifs pour les consommateurs. Pour les opérateurs de centres de données, l'avantage est également majeur : cette flexibilité leur permet d'obtenir des raccordements au réseau bien plus rapidement, sans attendre des années de travaux d'infrastructure. Après des essais probants dans trois États américains — Arizona, Virginie et Illinois —, Emerald AI a transposé son approche au Royaume-Uni, dans un contexte où la croissance explosive des besoins énergétiques liés à l'IA met sous pression les gestionnaires de réseaux du monde entier.
Impact France/UELes gestionnaires de réseaux européens confrontés à la même explosion des besoins énergétiques liés à l'IA pourraient adopter cette approche pour stabiliser leur réseau sans surinvestissement en infrastructures permanentes.
At the half-time whistle of the UEFA EURO 2020 round of 16 football match between England and Germany, millions of viewers stepped away from their screens in the U.K. to do the same thing at the same time — turn on their kettles. National Grid, which provides electricity for England and Wales, saw a demand spike of about 1 gigawatt — an increase equivalent to the average output of a standard nuclear reactor — in a matter of minutes from this countrywide tea break. Grid operators must carefully manage these demand peaks to keep the system stable, and this could become even more difficult as the grid continues to add large new customers. But what if those new customers could actually be flexible and relieve the grid during periods of peak strain? In a recent white paper , Emerald AI — in collaboration with NVIDIA, EPRI, National Grid and Nebius — showcased how “power-flexible” AI factories can autonomously adjust their power usage during peak demand. For AI factories, this could unlock significantly faster grid connections without waiting for massive, years-long infrastructure upgrades. For the public, it helps limit grid build outs by curbing the peak load that the system needs to serve, helping keep electricity rates affordable for everyday bill payers. Boil the Kettle, Balance the Grid After successful proof-of-concept trials at AI factories in Arizona, Virginia and Illinois, Emerald AI took its flexible grid solution across the pond, last December, bringing the Emerald AI Conductor Platform to Nebius’ new AI factory in London, built on NVIDIA infrastructure — among the first of its kind in the U.K. At the AI factory, the research team ran production-grade AI workloads on a cluster of 96 NVIDIA Blackwell Ultra GPUs connected through the NVIDIA Quantum-X800 InfiniBand platform . The NVIDIA System Management Interface is used to retrieve consistent, seconds-level GPU power telemetry. EPRI and National Grid simulated stress scenarios on the power grid — from lightning strikes to long periods of low wind power supply — and sent signals instructing the AI factory, with the help of the Conductor Platform, to temporarily reduce its power use to relieve grid strain. One of these scenarios was the “TV pickup” phenomenon, where that very same Euro 2020 football match’s energy surge was reenacted. As millions of simulated tea kettles were about to be turned on, the AI cluster ramped down its power use — successfully acting as a shock absorber for the abrupt power surge without disrupting the highest-priority AI workloads running on the cluster. https://blogs.nvidia.com/wp-content/uploads/2026/02/Grid-Responsive-AI-Infrastructure-Chart_v4.mp4 In practice, this means the grid can manage sudden demand swings using existing capacity more efficiently, reducing the need to overbuild permanent infrastructure to meet worst-case peaks and helping keep rates affordable for everyday consumers. “With this technology, AI factories become friendly and helpful grid assets,” said Varun Sivaram, founder and CEO of Emerald AI. “Simultaneously, the AI factories get connected much faster to the grid because they can tap into existing power grids.” Stress Relievers, Not Query Crushers In the Nebius AI factory demonstration, despite the quick ramp down of energy to power the national tea break, Emerald AI Conductor ensured that the simulated high-priority AI workloads performed at peak throughput, while more flexible jobs were slowed down temporarily. Emerald AI recorded 100% alignment with over 200 power targets that EPRI and National Grid instructed the AI cluster to follow for this experiment. “We did tests that go beyond the ones that have been done so far in the U.S. because we tested not just the GPUs, but also the CPUs and everything that sits around it — as well as the total power consumption of the IT equipment,” said Steve Smith, group chief strategy officer of National Grid. “We’ve proved the value that this technology brings.” Scaling London’s Grid at Super Speed London’s power grid is constantly working to meet the ever-growing energy needs of its citizens. Its grid operators — including National Grid — face a key bottleneck: constraints in infrastructure upgrades to connect large customers. Plugging flexible AI factories into the grid with solutions like Emerald AI’s Conductor Platform won’t just help to stabilize energy spikes — it can optimize the use of existing grid infrastructure to propel new industry talent and economic opportunities in the U.K. “We have enormous skills and potential in AI,” said Smith. “We’re never going to be on the scale of the U.S. in terms of data centers, but relative to the size of the country, we could be — and we’re certainly seeing that interest from many of the hyperscalers. So, it gives us the opportunity to play our part as National Grid in helping unlock that economic growth for the country.” Four demonstrations in, Emerald AI and NVIDIA are gearing up to put power-flexible AI factories
