Towards flexible data centres

Can data centres be transformed from static loads into dynamic, grid-aware assets?

This question is being put to the test in an Australian pilot involving local energy technology company FlexSysAI, web-hosting company ResetData, CSIRO and the University of Queensland (UQ). The pilot comes at a time when governments around the world are grappling with how to regulate data centre energy use while managing an increase in AI infrastructure demand.

“Data centres are forecast to become some of the largest single points of electricity demand, and unlocking demand flexibility could help support the grid during periods of stress,” said Dr Sean Lawrence, an Energy Systems Research Scientist at CSIRO. “This pilot will provide valuable real-world evidence about whether flexible AI workloads can help support electricity system reliability as Australia’s future energy system evolves.”

FlexSysAI is supplying its platform to see if data centre demand can be made more flexible by shifting GPU-based AI workloads across locations and time without impacting service. ResetData is providing the sovereign AI infrastructure and live AI factory environment to test the technology in practice, while CSIRO and UQ are supporting validation of the pilot.

FlexSysAI is hopeful that its platform will facilitate the rollout of data centres by optimising grid capacity and allowing more data centres to connect to current infrastructure. The company’s early modelling estimates its platform has the potential to optimise and dynamically adjust workloads on NVIDIA H200 GPUs at ResetData by 20–50% within seconds of a grid signal.

The FlexSysAI platform is now live, managing the power and workloads of an H200 cluster at ResetData’s AI-F1, a publicly available Australian sovereign AI factory. CSIRO and UQ will independently analyse and evaluate data from the pilot to better understand the role flexible AI workloads could play in supporting electricity system reliability.

“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 Victor Feoktistov, co-founder of FlexSyAI.

“This reduces strain on the energy system, lowers operators’ energy costs and unlocks additional capacity in existing infrastructure, meaning quicker connections to the grid.”

In theory, electricity networks would be able to serve additional data centre connections if those data centre loads could be curtailed during peak stress periods (typically occurring for 0.25–5% of the year). However, network operators currently lack operational data on whether AI data centres can reliably provide this flexibility.

The pilot uses FlexSysAI to categorise AI workloads into distinct ‘Flex Tiers’, in order to ensure critical tasks are protected while elastic training jobs provide reliable load-shifting.

When deployed at scale, the platform is designed to address a major bottleneck to the rollout of digital infrastructure, by connecting live electricity market and grid conditions with AI workload balancing. It may help data centre operators to connect to the grid sooner, pay less for power, access renewable electricity when it is abundant, and get paid to support the grid, while allowing operators to maintain control over when and where workloads are shifted.

Professor Frederik Geth, UQ-Springfield Chair in Energy, said UQ’s role was to take the operational data the pilot generates and use it to examine how data centre flexibility could be represented in future network planning and regulatory frameworks.

“That’s a gap that’s been missing from the conversation around AI data centre growth, and it’s the piece of the puzzle that determines whether this kind of flexibility actually gets recognised and rewarded in how networks are planned,” Geth said.

Image credit: iStock.com/gorodenkoff