The Grid AI Will Need
On the 24th of June, Australia's system operator stopped treating compute as a line item in business demand and made it a separately modelled driver of the national electricity plan
Good evening, reader.
Editor's note — There are two ways to read the 2026 Integrated System Plan AEMO released on 24 June. The first is as a transmission document, which is what it is. The second is as a constitutional document, which is what it has become. This piece is about the second reading.
In today's Innovation Commons:
AEMO moves data centres from a sub-line of business demand into the CEO preface of the national plan, alongside electrification and population growth (AEMO 2026 ISP).
11 large data centres totalling more than 5 GW are already working through transmission connection at end of March quarter, with energisation running two years and ramp-up running five to ten (Certified Strategic summary).
The transmission benefit doubles from A$30 billion under Step Change to A$61 billion under Higher Demand, almost entirely on data-centre load.
The federation is being stress-tested by a workload, not a crisis. States own siting, the Commonwealth owns the rulebook, AEMO owns the queue, and none of them owns the chips.
The United States arrived at almost the same institutional answer three weeks earlier through FERC's six show-cause orders (White & Case alert).
What AEMO actually did
In the 2024 plan, data-centre load sat inside the broader business-demand category, undifferentiated from cool rooms and steel mills. In the 2026 plan, AEMO names it on the first page of the chief executive's preface and gives it its own forecast track. Total underlying NEM consumption is projected to almost double, from 205 TWh today to about 390 TWh in 2050. Inside that envelope, data centres grow from roughly 2 per cent of underlying demand today to almost 10 per cent by 2050, compounding at about 25 per cent a year (Certified Strategic).
A system operator does not move a workload from a sub-line into the front matter of its plan unless it has decided the workload is structurally distinct, and once a category is created it is very hard to delete. The 25 per cent compounding rate, applied to a base of 4 TWh today, lands data centres at the same order of magnitude as the entire current Tasmanian load by the middle of the next decade. AEMO has effectively conceded that the NEM is no longer a grid that incidentally serves compute; it is, on a multi-decade view, a compute grid that incidentally serves households.
Two numbers do the heavy lifting. First, the connection queue: 11 large data centres totalling more than 5 GW were in transmission connection at end of March 2026, with applications taking about two years to energisation and a further five to ten years to ramp to full demand. Second, the sensitivity test: under a Higher Demand scenario adding 39 TWh on top of Step Change, the net market benefit of the ISP's transmission projects rises from nearly A$30 billion to A$61 billion. The same wires; double the value; almost entirely because of compute.
A constitutional document in disguise
The transmission plan is a constitutional document because it now allocates capacity, cost, and consequence across four jurisdictions that do not normally co-legislate.
The states own siting. Victoria, South Australia and Tasmania have dedicated data-centre policies; New South Wales is drafting one with a Legislative Council inquiry due 30 September; Queensland, Western Australia, the Northern Territory and the ACT are competing on broad digital strategies rather than standalone plans. The state policies vary in ambition and have nothing in common except their object.
The Commonwealth owns the rulebook. In March 2026 the Australian Government published its Expectations of data centres and AI infrastructure developers under the National AI Plan, requiring operators to bring new clean energy, pay full grid-connection and network costs, and run as flexible loads supporting the grid. The Expectations are not a binding instrument, but they sit over every state plan as a uniform overlay and they pre-commit the Commonwealth to a particular bargain: you may scale, if you bring your own firmed supply and absorb your own externalities.
AEMO owns the queue. Connection studies, hosting capacity, and Minimum System Load procurement now route through one agency whose statute predates the workload it is being asked to allocate among. The Minimum System Load tender, which contracts pre-positioned controllable resources to soak up surplus daytime solar, names data centres alongside electric vehicles and industrial loads as the technologies it most wants to recruit. AEMO is no longer just balancing supply and demand; it is paying particular customers to be the daytime sink of last resort.
None of them owns the chips. Sovereignty of the upstream supply (wafers, lithography, the model weights) sits with foreign firms and one or two foreign treasuries. The institutional design Australia is settling on, by drift more than by debate, is to compensate for that upstream dependency by tightening the grip on what we do control: where the load lands, what it pays, and how it behaves on stress.
The American mirror
Three weeks earlier, on 18 June, the United States Federal Energy Regulatory Commission issued six show-cause orders to every regional transmission organisation under Section 206 of the Federal Power Act, requiring them to explain, within 60 days, why their existing tariffs are just and reasonable in respect of large-load interconnection (White & Case alert). Large load is defined as 50 MW or more interconnecting above 69 kV.
The reform menu reads as if Australia and the United States are drafting from the same memo. RTOs must require cost-recovery agreements with minimum financial contributions secured by strict credit support, to address the so-called no-show problem where a data centre fails to materialise and the network upgrade is stranded. They must offer new transmission services tailored to flexible loads that can curtail under system stress. They must restrict the netting of behind-the-meter generation against wholesale demand. They must publish searchable, filterable data on large-load additions and their associated upgrade costs. Chairman Swett's language was direct: free-riding by data centres is not an option.
The AEMO ISP and the FERC show-cause orders are not coordinated. They are convergent. Two federations, with very different politics, arrived independently at the same institutional answer in the same fortnight: condition the connection, price the optionality, push flexibility onto the load. The convergence is the news, not the documents.
Second-order effects nobody is naming yet
Three structural moves follow from this that the headline numbers obscure.
The first is that the grid stops being a passive utility and becomes an active counterparty. When AEMO contracts a data centre to lift load during high-solar daytime periods, the data centre is not a customer in the old sense; it is a contracted balancing resource that happens to do useful work in the foreground. The economics of a hyperscale build now include a non-trivial revenue line from being available to absorb surplus solar — which means the unit economics of compute in Australia depend on the daily solar cycle in a way they do not in, say, Virginia. This is a feature, not a bug, and it is one of the few structural advantages a sovereign-AI strategy could lean on.
The second is that the federation is being reshaped by the workload. State approval power was always real but was rarely exercised at the limits, because no industry asked for that much electricity in that much of a hurry. A 5 GW queue at energisation timescales of two years forces the question that the federation has historically been polite enough to leave unanswered: which jurisdiction's planning regime governs an installation whose only physical footprint is a substation, whose customers are global, and whose revenue line items are denominated in foreign currency. The answer being settled by AEMO's queue, by the National AI Plan Expectations, and by state-by-state policy drift is not the answer that would have been chosen in a designed system. It is the answer a federation lands on when a workload arrives faster than the constitutional machinery.
The third is that the binding constraint on Australian sovereign AI is no longer compute, capital, or talent. It is transmission planning. The chips can be imported; the engineers can be hired; the equity is available. The 500 kV transmission corridors needed to land a gigawatt of new load in the Sydney–Newcastle–Wollongong arc cannot be hurried. The Higher Demand sensitivity loads concentrate in South Australia (+18 TWh), the Sydney–Newcastle–Wollongong corridor (+11 TWh), Melbourne–Geelong (+4 TWh) and Gladstone (+3 TWh) — and the optimal development path under that sensitivity stays within A$60 million of the least-cost candidate. The grid is robust to the demand, it is the build is what is fragile.
What the front matter is really for
Documents have grammar. A line item in business demand is a line item. A separately modelled driver in the CEO preface is a constitutional acknowledgement. AEMO is, in effect, telling the rest of the Commonwealth: this workload is one of the things this country will be organised around. Treat it accordingly.
The follow-through has to come from elsewhere. State planning approvals need a published, predictable hosting-capacity register so siting decisions can be made on grid evidence rather than on lobby access. The National AI Plan Expectations need teeth — at minimum a default tariff schedule for who pays for what, and at maximum a Commonwealth power to refuse connection to a development that refuses to bring its own firmed supply. AEMC and AER will need to ratify, or push back on, the new transmission services AEMO and the connection rule changes implied by the ISP, and they will need to do it in a way that does not leave residential customers carrying the upgrade cost of a workload that is, on the balance of evidence, a paying counterparty.
The frontier here is not what gets built. It is what gets institutionalised around what gets built. The grid is becoming the most important institutional surface in the country, because the workload that pays for it is the workload that decides who gets to do AI on Australian soil.
A separately modelled driver. In the preface. That is the news.
— The Editor
P.S. Share this on LinkedIn/X to spark conversations with your colleagues.