At a glance
A new report by Mandala, commissioned by Microsoft, argues AI is both a driver of electricity demand and one of the fastest levers to get more out of Australia’s existing grid.
Australia needs about 6,000km of new transmission, at a cost the report says has tripled since AEMO’s 2024 plan, with lead times of eight to twelve years.
Mandala found no hard regulatory barrier to AI in the grid. The blockers are industry risk aversion, data locked in silos, and a funding model that puts roughly 70% of network revenue on the capital side of the ledger.
The report leans on global IEA figures for the size of the prize, but those are worldwide estimates, not Australian outcomes.
Microsoft, which says it was the first hyperscaler to sign the Australian Government’s data centre expectations, frames hyperscale load as a steadying presence on the grid, a claim the Australian industry backs with its own renewables and grid spending.
The two-way case: AI as grid infrastructure
Australia is preparing to spend heavily on transmission it does not yet have. AEMO’s 2026 Integrated System Plan puts the country on a path to roughly 6,000km of new transmission, and the Mandala report puts the cost of that build at three times AEMO’s 2024 estimate, with lead times of eight to twelve years and, by the report’s account, an added burden on taxpayers of about A$600 a year. That build is necessary. It is also slow, and it arrives while demand from renewables integration, electrification and data centres is landing all at once.
The report, prepared by economics advisory firm Mandala and commissioned by Microsoft, makes a narrower argument than the usual data centre energy debate. Its case is that AI is not only a source of new load but one of the few tools that can release capacity and efficiency from the network Australia has already built, without waiting on the next decade of construction. The relationship runs in both directions: data centres depend on reliable power, and the grid increasingly depends on cloud and AI to run.
The report’s sharpest Australian finding has little to do with technology, Mandala identified no law prohibiting AI in the power system. What it found instead was a funding model that pays networks to build. The Australian Energy Regulator sets network revenue over five-year cycles, and capital costs, the return on capital plus depreciation, account for roughly 70% of the allowance. AI is typically an operating cost, so a system that is cheaper and faster than a physical upgrade can still be the harder investment to justify.
Neara estimates that up to 10GW of latent capacity already sits in NSW's distribution network. Data centre operators have separately offered to run as flexible, grid-supporting load. Both point at capacity the existing network could release without new build. What the Mandala report adds is AI inside the grid's own operations, rather than AI as a customer queuing to connect.
What AI already does in the grid
Adoption in Australia is real but shallow. Generators use AI to predict equipment failures and sharpen wind and solar forecasts. Networks fuse drone, LiDAR and satellite imagery to find faults and manage vegetation before it takes a line down. Retailers automate customer service and tune tariffs. These are useful gains, but the report classes them as incremental, tools that speed up existing human workflows rather than change how the system is run.
The bigger opportunity is whole-system optimisation in real time: matching millions of distributed solar panels, batteries, electric vehicles and flexible loads against supply as conditions change. Mandala draws on the International Energy Agency to size that opportunity, citing around 175GW of transmission capacity that AI-enabled tools could unlock, a further 450GW to 700GW of large loads that grid-enhancing technologies could connect without new infrastructure, and roughly US$110 billion a year in savings from wider use of AI in operations. Those figures are global, not Australian, and the report presents them as evidence of what is possible rather than a forecast for the NEM. Read that way, they set the ceiling; the barriers set how much of it Australia reaches.
Three barriers slowing AI adoption
The sharpest finding is an absence: Mandala identified no hard regulatory barrier prohibiting AI in the power system. Australia’s technology-neutral approach, applying existing law instead of writing AI-specific rules, is treated as workable. The constraints are cultural, institutional and financial instead, and the report groups them into three.
Barrier | What it looks like in Australia | International contrast |
No shared strategy or direction | Utilities view AI through a risk lens tied to their critical-infrastructure status; market bodies have issued little guidance on where AI fits the rules | The European Commission is publishing a roadmap for AI in grid optimisation and demand-side flexibility |
Weak investment incentives | Network revenue is set through a five-year determination where capital costs make up roughly 70% of the allowance, so software and analytics face a higher bar than poles and wires | The UK uses a total-expenditure (TotEx) model and an innovation fund to de-risk AI investment; California runs its EPIC fund |
Siloed data | Operational data sits with individual participants, with metering data constrained under the National Electricity Rules and the Privacy Act, and no obligation on distributors to share granular network data | The EU is building a Common European Energy Data Space; the UK is reforming smart-meter data access |
Source: Certified Strategic Editorial, Mandala “Unlocking a Virtuous Cycle” report for Microsoft, July 2026.
Risk aversion and weak signals
Utilities are critical-infrastructure providers operating under frameworks such as the Security of Critical Infrastructure Act, and the report describes an industry that weighs AI’s risks more heavily than its opportunities. The consequence of a failure is visible and political, so operators default to proven methods. Market bodies have not set out where AI belongs within the existing rules, which leaves leaders who are conceptually willing without a clear path to deployment. Clearer guidance from governments and market bodies, the report argues, would give the sector confidence to move without waiting for anyone to go first.
A funding model built for physical assets
The second barrier carries the clearest Australian policy hook. Networks are regulated monopolies whose revenue the Australian Energy Regulator sets over five-year cycles, and that framework was built to compensate large capital assets. Capital costs, the return on capital plus depreciation, make up about 70% of the allowed revenue, with operating expenditure and tax accounting for the remaining 30%. AI typically lands on the operating side. A software solution that is cheaper and faster can still face weaker incentives than a physical one that grows the asset base. Mandala points to the UK’s shift to a TotEx model and its Strategic Innovation Fund, and to California’s EPIC fund, as ways other markets have started to value software alongside physical assets. The AEMC reviewed expenditure bias in 2018 and 2019, so the question is already live in Australia.
Data locked in silos
The third barrier is data. AI needs large volumes of real-time, high-quality data, and in the NEM that data is fragmented across generators, networks, retailers and metering coordinators, much of it sensitive and subject to privacy rules. Kerry Schott, the former Energy Security Board chair quoted in the report, puts it plainly: “We now have more data than ever, but it isn’t being fully utilised.” Work is already under way to change that. ARENA has pitched an Open NEM model to the AEMC, the AEMC’s Integrated Distribution System Planning process is looking at low-voltage network reporting, and a late-2024 determination set out to give distributors mandated access to basic smart-meter data. None of these are settled, and how far they go will decide how much of the grid becomes legible to AI.
Data centre load and the generation behind it
For the operators Certified Strategic covers, the report’s most relevant claim is about load. Mandala frames hyperscale data centres as a source of consistent, predictable load that can support the grid rather than only strain it, provided that growth is managed. Microsoft, which says it was the first hyperscaler to sign the Australian Government’s data centre expectations, uses the report to align itself with that position. That claim is Microsoft’s own, made alongside its A$25 billion Australian infrastructure commitment announced in April 2026, and is worth reading as positioning rather than an independent finding.
Whether the claim holds turns on whether new load arrives with new generation behind it. The operators listed in our directory of data centres in Australia point to their own spending. Data Centres Australia members offset 70% of their power use with renewables they fund themselves, have underwritten 1.5 TWh of new renewable generation through long-term contracts, and are committed to A$10.3 billion in grid and energy infrastructure investment by 2030, figures set out in our analysis of the 1 July bill cuts. Operators also meet their connection and network augmentation costs upfront.
The counter-case is that load arrives faster than generation. Greenpeace Australia Pacific argued in a May 2026 report that data centres are adding large loads without underwriting enough new generation to match, drawing instead on clean energy built for other users. The Climate Council set seven conditions on how the build-out should proceed, four of which Australia already meets. Those critiques are aimed at the load, while Mandala is arguing about the operations, so both positions can hold at once.
The commercial logic for operators is straightforward. A grid that runs closer to its real limits, with AI releasing latent capacity and smoothing distributed supply, connects new load faster and passes lower network costs to consumers, the thread running through South Australia’s bid to pair new supply with new demand. For an industry still contesting its social licence, a credible account of AI data centres helping the grid work harder is a commercial asset in its own right.
Mandala is candid that the argument matters commercially, writing that demonstrating AI’s benefits to affordability and reliability will be crucial to growing social licence for further data centre build-out. The commissioning party’s own energy commitments are also in the frame. Bloomberg reported on 6 May 2026 that Microsoft is weighing whether to delay or abandon its 2030 target of matching 100% of its hourly electricity use with renewable energy purchases, a standard more demanding than the annual matching most of its peers use. Microsoft has made no public announcement, declined to comment on the internal debate, and told TechCrunch it continues to look for opportunities to maintain its annual matching goal. The company said in February 2026 that it had met its annual carbon-negative milestone, and in April 2026 it said it was working with Chevron and Engine No. 1 on a West Texas gas plant that could eventually generate up to 5GW.
What to watch
The report’s practical value is a short list of Australian levers, each attached to a live process. The AEMC’s Integrated Distribution System Planning rule change is the venue for low-voltage data reporting. ARENA’s Open NEM pitch tests whether system data opens up. Any move toward a TotEx-style revenue framework, or an AI-in-energy fund modelled on the UK and California, would signal that the investment barrier is moving from debate to action. And the NSW Legislative Council’s data centre inquiry, reporting 30 September, is the nearest forum where data transparency could be picked up as a lever in its own right. The report’s own conclusion is that none of these barriers are fixed, and that industry, government and the technology sector have to move together for any of them to shift.