What the development involves
Researchers from Microsoft, OpenAI and NVIDIA have described how large training jobs cause synchronized power fluctuations, and how these can be dampened at the GPU, rack and facility level (arXiv, August 2025).
NVIDIA has built energy storage into the power supply units of its GB300 platform and measured 30% lower peak power drawn from the grid during training, compared with the previous power supplies in the same rack (NVIDIA, July 28, 2025).
What benefits are documented
NVIDIA's figure is a vendor claim and applies to a single rack. The storage smooths the load but does not deliver energy back to the grid. How these solutions interact with grid operators' requirements is not yet standardized.
What it could mean in Sweden
In discussions with Svenska kraftnät, Sweden's transmission system operator, and the regional network operators, the behavior of the load will become as important as the amount of power required. A facility that can demonstrate a stable load will be easier to connect.
What we are investigating
What requirements may appear in connection agreements. How much can be solved in the IT equipment and how much must be solved in the facility, for example with battery storage or filters. How space and interfaces are reserved from the earliest stages.
