FlexSysAI tests AI workload orchestration on live H200 cluster at ResetData's AI-F1 facility

Australian AI workload orchestration firm FlexSysAI has launched a live pilot with ResetData, CSIRO, and the University of Queensland to test whether AI compute loads can be shifted in response to electricity market signals without degrading service^[1]. The platform is already running, managing workloads on an Nvidia H200 cluster at ResetData's AI-F1 facility in Melbourne^[1].

What the pilot is testing

Early modelling by FlexSysAI estimates the platform can adjust workloads on the H200 cluster by 20-50% within seconds of receiving a grid signal^[1]. The system uses a tiered classification scheme - "Flex Tiers" - that separates workloads by their tolerance for disruption:

• Critical inference jobs are ring-fenced and receive no performance reduction
• Elastic training jobs are designated as shiftable and absorb the load response
• The platform connects live electricity market prices and grid conditions to the workload scheduler in real time^[1]

CSIRO and the University of Queensland will independently analyse data from the pilot to evaluate whether flexible AI loads can reliably support electricity system reliability^[1].

Why network operators need the data

Electricity networks could accommodate additional data centre connections if those loads can be curtailed during peak stress periods, which typically account for just 0.25%-5% of the year^[1]. The barrier is not technical in principle - it is evidentiary. Network operators currently lack operational data on whether AI data centres can reliably deliver that flexibility^[1].

That is the gap the pilot is designed to close. ResetData's AI-F1 - described as Australia's first publicly available sovereign AI factory - provides a live production environment rather than a controlled laboratory, which is what makes the validation credible to network planners^[1].

The NEM as a proving ground

FlexSysAI's CTO Angelo Perera has argued that Australia's National Electricity Market is a deliberately demanding test environment. Its price volatility and five-minute settlement structure reward demand that can respond quickly to supply conditions, making it a harder and more informative context than more heavily regulated markets.

Victor Feoktistov, co-founder of FlexSysAI, said early results show the platform can "rapidly respond to electricity market signals and shift AI workloads to when and where power is more readily available," adding that the approach "unlocks additional capacity in existing infrastructure, meaning quicker connections to the grid"^[1].

FlexSysAI is not alone in this space. Emerald AI has developed a platform called the Emerald Conductor that mediates between the grid and data centres, and has completed several demonstration projects in real-world environments^[1]. The FlexSysAI pilot is notable for its independent academic validation layer and its use of a sovereign AI facility as the test site.

What to watch

The pilot's output will be an independently validated dataset on how much load flexibility a GPU cluster can reliably offer and under what conditions. If CSIRO and the University of Queensland confirm the modelled 20-50% adjustment range, that evidence could be used by Australian network operators to justify higher connection allowances for data centres that commit to curtailment obligations - a dynamic already emerging in other jurisdictions, including Ireland and parts of the US.