• FlexSysAI and ResetData are testing flexible AI workloads on a live Nvidia H200 cluster with CSIRO and the University of Queensland
  • The pilot will examine whether less urgent computing work can move around periods of grid stress without affecting important workloads

The fact

FlexSysAI has begun an Australian pilot with ResetData, CSIRO and the University of Queensland to test whether AI computing demand can respond to changing electricity-grid conditions.

The platform is running on an Nvidia H200 cluster at ResetData's AI-F1 facility. It groups workloads by how flexible they are, protecting critical tasks while allowing less urgent training jobs to be reduced or shifted when the grid is under pressure. CSIRO and the University of Queensland will independently analyse data from the trial.

FlexSysAI's early modelling suggests workloads on the ResetData GPUs could be adjusted by 20–50% within seconds of a grid signal. That figure describes changes to computing workloads, not a measured reduction of the same size in the facility's total electricity use.

The partners say electricity networks still lack real operating data showing how reliably AI facilities can provide this kind of flexibility. The pilot is intended to help fill that gap.

The assessment

The idea behind the trial is simple: not every AI job has to run at exactly the same time. If less urgent work can be delayed or moved when the grid is under pressure, a data centre could reduce some of its demand while keeping important jobs running.

The harder part is what happens afterwards. Work that is paused still has to be completed, so the trial needs to show that shifting it does not create delays for customers or simply move a large block of electricity demand to another time. That is why the researchers are looking at real operating data rather than relying only on modelling.

For BTW readers, the useful result will be evidence showing how much demand can actually move, how quickly the facility can respond and whether delayed workloads still finish as expected. Only then can network operators judge whether flexible AI computing is useful when planning the grid.

What to watch

Watch for CSIRO and the University of Queensland to publish results showing how much workload was shifted during grid-response events and what happened afterwards. Evidence that delayed jobs were completed without affecting service would provide a stronger test than the current modelling alone.