The Scarce Megawatt Is the Peak Megawatt
Nvidia is working with Australian infrastructure partners to deliver up to 2 gigawatts of AI-factory capacity by 2027.¹ On September 10, a FlexSysAI-led pilot, with CSIRO and the University of Queensland independently evaluating the results, began testing something that may matter just as much: whether AI workloads can move when the electricity grid needs them to. Early modelling from FlexSysAI suggests 20% to 50% of NVIDIA H200 workloads in the ResetData pilot environment could be dynamically adjusted within seconds of a grid signal.²
Those announcements look like two different stories. They are actually the same infrastructure problem viewed from opposite sides. The first asks how much additional power AI needs. The second asks how much of that demand must be served at exactly the same time.
Australia already shows why the distinction matters. In the first quarter of 2026, average data-center demand in the National Electricity Market was nearly 600 MW. Yet 11 data-center projects representing 5.4 GW of ultimate load were progressing through transmission connection processes.³ That is roughly 9 times the existing average load. AEMO explicitly warns that ultimate connection capacity is not required on day one, projects ramp over time, and some developers pursue multiple connection options.
So: connection capacity ≠ realized load
That is the first important distinction.
The second is between energy and capacity.
AEMO’s latest Electricity Statement of Opportunities forecasts data-center electricity consumption in the NEM rising from approximately 5 TWh in 2025-26 to 34 TWh in 2035-36.⁴ That is about a 6.8× increase, equivalent to roughly 21% compound annual growth over ten years. Data centers would move from about 3% to roughly 13% of grid-supplied electricity.
That energy growth is enormous.
But electricity networks are not built only around how many megawatt-hours customers consume over a year. They must also be capable of surviving the hours when many large loads demand electricity simultaneously.
This is where the pilot becomes interesting.
The FlexSysAI pilot release says electricity networks could potentially serve additional data-center connections if loads can be curtailed during periods of peak grid stress, which it describes as typically only 0.25% to 5% of the year.²
There are 8,760 hours in a year.
8,760 × 0.25% = 21.9 hours
8,760 × 5% = 438 hours
The implication is not that Australia has a power constraint for only 22 to 438 hours. Electricity supply, transmission, reliability and local network constraints are far more complicated than that.
The implication is more specific. Part of the infrastructure may need to be sized around a relatively small number of difficult hours.
That changes the economic question.
If some training jobs or other delay-tolerant workloads can move outside those periods, AI infrastructure begins to resemble an industrial demand-response asset. Compute can move across time, and potentially across location, while critical workloads remain protected.
The architecture becomes:
AI demand → connection capacity → peak MW → workload flexibility → GPU utilization → compute revenue
The scarce denominator is no longer simply megawatts. It is firm megawatts at the constrained node and hour.
The International Energy Agency provides the economic anchor. Demand response can reduce peak-capacity requirements, defer grid investment and improve system resilience.⁵ But the IEA also identifies the counterargument: AI-focused data centers are extremely capital-intensive, roughly ten times more capital-intensive than aluminum smelters in its analysis, which makes simply curtailing expensive compute economically costly.⁶
So flexible compute is not free power.
The real trade is:
value of uninterrupted GPU utilization versus value of avoiding scarce peak-grid capacity
That is where a new infrastructure layer may emerge.
A version of this distinction already exists in grid design through non-firm connection agreements, which can allow faster grid access in exchange for limits on consumption at constrained times. The IEA estimates such arrangements could enable 750–900 GW of advanced-stage projects globally⁷. If AI workloads can reliably fit that structure, compute flexibility becomes an interconnection asset, not merely an operating feature.
That could affect connection queues, electricity contracts, data-center location strategy and ultimately project economics.
Software that orchestrates compute around the power system would then do more than reduce an electricity bill. It could influence how quickly physical AI infrastructure can be energized and how efficiently existing grid capacity is allocated.
The question is no longer simply how many gigawatts AI will need. It is how many of those gigawatts must be guaranteed during the grid’s hardest hours.
References
1. NVIDIA, NVIDIA Expands AI Infrastructure Capacity in Partnership With Australia’s Data Center Ecosystem, September 9, 2026. NVIDIA announcement
2. CSIRO / FlexSysAI, Australia-first pilot to transform data centres from static loads into dynamic grid-aware assets, September 10, 2026. CSIRO pilot release
3. Australian Energy Market Operator, CEO speech at Australian Energy Week 2026, June 10, 2026. AEMO data-centre connection data
4. Australian Energy Market Operator, 2026 Electricity Statement of Opportunities, August 25, 2026. AEMO 2026 ESOO release
5. International Energy Agency, Electricity 2026: Flexibility, 2026. IEA demand-flexibility analysis
6. International Energy Agency, Energy and AI, 2025. IEA Energy and AI
7. International Energy Agency, Electricity 2026: Grids, 2026. IEA, Electricity 2026: Grids
Comments