When the data center becomes a flexible part of the grid

What the development involves

In a 2025 field trial in Phoenix, software from Emerald AI orchestrated the workloads of a 256-GPU cluster to reduce power by 25% for three hours during peak demand, while maintaining quality of service (arXiv, July 2025).

Google has written 1 GW of demand response into long-term agreements with US utilities (Google, March 19, 2026). An analysis from Duke University estimates that nearly 100 GW of new load could be connected in the US with limited impact if it can be curtailed for short periods (Duke Nicholas Institute).

What benefits are documented

The field trial was a pilot of limited scale. Google's figure refers to contracted capacity, not capacity actually used. At the same time, research from Microsoft, OpenAI and NVIDIA shows that large AI training jobs cause rapid power fluctuations that can affect the grid (arXiv, August 2025). NVIDIA reports that energy storage in the power supply units reduced peak power by 30% in a measured rack (NVIDIA, July 2025).

What it could mean in Sweden

The queue of data center projects at Svenska kraftnät amounted to 7,620 MW in May 2026 (Energiföretagen). Svenska kraftnät has studied conditional connection agreements, under which capacity can be allocated earlier, provided that power draw can be curtailed until the grid has been reinforced (Svenska kraftnät, June 2025).

On-site battery storage can make such a connection possible without affecting data center operations. For a developer, flexibility thus becomes part of the value of the site, not just an operational matter.

What needs to be investigated

What the terms of Swedish agreements will look like. How battery storage, backup power and flexible load can be combined. Which customers can accept flexibility. How power fluctuations from AI loads are managed against the grid operator's requirements.