Australian energy technology startup FlexSysAI has begun an Australia-first pilot with sovereign AI infrastructure provider ResetData, CSIRO and The University of Queensland to test whether GPU-based AI workloads can be rapidly reduced or shifted when electricity networks are under stress.
The technology is now live on an NVIDIA H200 cluster at ResetData’s AI-F1 facility, where demand-response testing has begun. CSIRO is providing independent measurement and verification, while UQ is analysing the operational data and examining how flexible data centre demand could potentially be incorporated into future network planning and regulatory frameworks.
The proposition is relatively simple on paper. Not all AI workloads need to run at the precise moment they are scheduled. Training and other batch or deferrable jobs can potentially be paused, slowed or rescheduled, while latency-sensitive and mission-critical workloads continue running.
FlexSysAI classifies workloads into different “Flex Tiers”, allowing operators to decide which jobs can be curtailed. The platform combines a market and grid forecasting layer with Kubernetes-based workload orchestration and an execution layer that manages power at the GPU cluster.
The interesting question is how much electrical demand this can actually release.
FlexSysAI’s claimed its early modelling suggests power demand from the H200 cluster under test could be reduced by between 20 and 50 percent within seconds of a grid signal by shedding flexible and deferrable workloads.
Big numbers, but that figure is not yet an independently verified result. The company says measuring the actual power characteristics of different AI workloads during flexibility events is one of the principal objectives of the pilot, with CSIRO and UQ expected to independently verify the results.
There is, however, already some live performance data. FlexSysAI says committed throttling on the live cluster is currently landing around two to three seconds after a signal, with the company targeting a sub-two-second response as it further parallelises GPU instruction dispatch.
CSIRO and UQ are independently measuring that response using high-resolution telemetry.
That distinction is important. FlexSysAI is not yet claiming that it has demonstrated a 20-50 percent reduction in the electricity consumption of an entire data centre. The figure relates to the cluster’s power usage and remains an early modelling estimate. The actual degree of flexibility will depend on the workload mix at a particular facility, according to the company.
The pilot is therefore testing a more fundamental proposition: whether AI compute can behave enough like a controllable electricity load to become useful to the grid.
“Early testing has been extremely promising, showing our platform can rapidly respond to electricity market signals and shift AI workloads to when and where power is more readily available,” said FlexSysAI co-founder Victor Feoktistov (above). “This reduces strain on the energy system, lowers operator’s energy costs and unlocks additional capacity in existing infrastructure, meaning quicker connections to the grid.”
From cheaper electricity to grid capacity
There are several potential benefits if they get this right. At the most immediate level, a data centre could reduce its electricity costs by moving flexible workloads away from periods of high prices and towards periods when electricity is cheaper or more abundant.
FlexSysAI has also built the energy-market infrastructure needed to turn that flexibility into a commercial transaction. Flipped Energy holds the electricity retail licence and NEM registration, while FlexSysAI provides the technology platform.
The company says it can currently participate in wholesale market activity and demand response, using CSIP-AUS as the signalling protocol. FCAS and network services are on its roadmap.
Under its proposed model, customers pay nothing to participate. Flipped Energy owns the market positions associated with the flexibility, while FlexSysAI earns a share of the savings and grid-services value generated.
The more consequential proposition is potentially getting around one of the biggest barriers facing new data centres: the availability of firm grid capacity. To demonstrate, we can use the example of a hypothetical 100MW data centre that cannot obtain a 100MW firm connection because the local network does not have sufficient headroom during periods of system constraint.
Instead of asking the network for 100MW of firm capacity, FlexSysAI’s model would allow the operator to offer a defined portion of its demand as reliably curtailable. If that flexibility could be measured, controlled and independently verified, the network might ultimately only need to provide firm capacity for the remainder.
That could potentially allow a facility to connect using existing network capacity rather than waiting for additional generation, storage or network infrastructure.
Early days
But this is where the FlexSysAI proposition moves from demonstrated technology into an experiment. No Australian network is currently recognising curtailable data centre load in a live connection agreement, according to FlexSysAI. The company told W.Media the purpose of the ResetData pilot is to generate the operational evidence needed to make that conversation possible.
The pilot’s final deliverable is intended to be a “Data Centre Flexibility Model” for operationalising curtailment-enabled headroom in the NEM.
That is also where UQ’s role becomes particularly important. The university is using the operational data to investigate how data centre flexibility could be represented in future network planning and regulatory frameworks. UQ-Springfield Chair in Energy Professor Frederik Geth says this is an important gap in the current discussion around AI data centre growth: determining whether the flexibility of these loads can actually be recognised and rewarded when networks plan their systems.
CSIRO’s role is also practical. Rather than relying solely on FlexSysAI’s own performance claims, its Data Clearing House will independently measure the response of the system.
The stated technical success criteria include delivering committed load reductions within seconds, minimising the difference between promised and metered reductions, avoiding sustained power oscillations as GPUs move between power states, and protecting mission-critical workloads.
The project also has an economic target. FlexSysAI says it is looking for a double-digit percentage reduction in facility energy costs from demand response, verified through billing analysis, with the resulting flexibility settling through actual energy-market transactions.
A new form of demand response?
FlexSysAI argues that its approach differs from conventional demand response because control operates at the individual workload level rather than treating the data centre as a single block of demand.
Flexible, batch and other deferrable AI jobs can be shifted or paused, while critical and latency-sensitive workloads are protected. The company says operators can see a live price for their flexibility and are free to sit out an event without penalty.
The company has begun testing demand response across different AI workload types, including inference and video-generation jobs across several models. FlexSysAI also argues that flexible compute can complement rather than replace batteries, generation and network investment.
A battery can respond rapidly to a grid signal but requires capital investment and has finite stored energy. Flexible AI workloads use the compute demand itself as the flexible resource, potentially without requiring additional customer-side infrastructure.
The company sees a future where flexible compute and batteries are co-optimised, with batteries potentially providing ride-through and helping manage the delays around demand-response events.
There are nevertheless important questions still to be answered. For example, how much of a large AI facility’s total load will ultimately prove flexible once cooling, networking, storage and other infrastructure are included? How frequently can workloads be curtailed without affecting customers? What happens when deferred workloads subsequently have to run, potentially creating a rebound in demand? And, most importantly, will electricity networks be willing to treat demonstrated flexibility as equivalent to firm capacity when assessing new connections?
Those questions are precisely why the ResetData pilot matters to the company. FlexSysAI says it is already in commercial discussions with multiple Australian data centre operators and large-load projects across several states, covering both existing facilities and new developments. It has not disclosed the size of that pipeline, citing commercial sensitivity.
The company also sees the Australian NEM as a useful proving ground because of its renewable penetration, price volatility and constraints. If the model can be demonstrated in Australia, it believes it can subsequently be applied to other electricity markets.
There is already international interest in grid-responsive AI computing, with companies developing systems that dynamically orchestrate AI workloads in response to grid conditions. FlexSysAI’s proposition is that workload orchestration is only one part of the problem.
Its differentiation is the combination of workload orchestration with participation in Australia’s electricity market. The company’s argument is that software which can decide when a GPU workload should run is only useful as a grid resource if the resulting flexibility is measurable, controllable and capable of being converted into a settled energy-market transaction.
That makes the ResetData pilot more than a demonstration of GPU workload management. The bigger experiment is whether a data centre can stop being viewed by the electricity system as a large, largely inflexible block of new demand and instead become a controllable grid resource.
If the pilot demonstrates that AI workloads can reliably deliver substantial power reductions within seconds, without affecting critical services, the next challenge for the company will be regulatory and commercial as opposed to technical: convincing networks that such flexibility is sufficiently predictable to count when determining how much new data centre capacity the grid can accommodate.