FlexSysAI pilot turns AI workloads into grid flexibility

An Australian startup is testing whether a data center can behave like a dispatchable power asset, selling its ability to pause rather than its ability to consume. FlexSysAI, an AI workload orchestration firm, has begun a pilot with ResetData, CSIRO, and the University of Queensland to shift AI compute in response to live electricity market conditions. Co-founder Victor Feoktistov says early testing on an Nvidia H200 deployment at ResetData’s AI-F1 facility shows the platform can move workloads to when and where power is more readily available.

Launched last month, the FlexSysAI platform ties market and grid conditions directly to workload balancing by sorting workloads into Flex Tiers: critical, latency-sensitive tasks stay protected; elastic training jobs provide the load-shifting; and non-critical work can be reduced when the grid tightens. A training run on an H200 is a controllable load in a way a live inference request is not, so an operator can volunteer part of its compute to the grid without degrading customer-facing work. FlexSysAI’s pitch spans faster grid connections, smaller power bills, access to low-carbon power in surplus periods, and demand-response revenue, and ResetData co-CEO Marcel Zalloua frames the pilot as sovereign infrastructure enabling Australian innovation.

FlexSysAI has company in this niche: Emerald AI’s Emerald Conductor sits between the grid and data centers, orchestrating workloads in real time, and has completed demonstration projects in Phoenix, Chicago, and, with the UK’s National Grid, in Britain. Emerald AI has announced its first commercial project, and records show a $150 million deal announced in late August.

The AI buildout’s bottleneck has moved from land and chips to electrons, and operators are moving from grid tenants to power owners. FlexSysAI is the demand-side variant of that thesis: if the grid cannot supply electrons on demand, the data center can make its demand elastic. A handful of H200s at one facility is not yet a resource utilities will pay for, and FlexSysAI’s claims about faster connections and demand-response revenue are unproven at scale. The first commercial contract that puts a price on shifted compute has not yet been signed.