AI’s next bottleneck runs through water infrastructure

On 22 July, Reuters reported that a proposed 612-megawatt data centre near Sydney had abandoned plans to use recycled water for cooling.

NEXTDC, which is developing the S7 facility in collaboration with OpenAI, had planned to bring treated sewage water to the site. The design would have avoided using drinking water and reduced part of the facility’s electricity demand. It required a pipeline.

That pipeline could not obtain planning permission on the timetable needed by the project. NEXTDC ended discussions with utilities and moved to closed-loop liquid cooling. The replacement uses no ongoing water at the site and, according to the DCByte analyst quoted by Reuters, will consume more electricity.

The episode is useful precisely because it is so ordinary. Nobody discovered a flaw in artificial intelligence. Nobody ran out of GPUs. A large proposed technology project met a local infrastructure dependency with its own owner, approval process and calendar.

The cloud had found its postcode.

Computing demand is only the beginning

Much of the AI infrastructure discussion revolves around three inputs: chips, electricity and capital. Each matters. The IEA’s Electricity 2026 report expects data centres to account for around half of US electricity-demand growth through 2030.

Figures at that scale naturally attract attention. They can also make the infrastructure underneath them look deceptively simple.

A data centre is a physical industrial facility. It connects to a particular electricity network, occupies land in a particular planning jurisdiction and may rely on a particular water or wastewater system. Its cooling architecture influences both resource demand and operating cost. Local communities and public authorities also have a say.

These systems were not designed to expand at the speed of AI demand.

The official NSW Planning Portal currently lists S7 at the “Prepare EIS” stage. The proposed development has a long way to travel through the approval process. Reuters reported that the recycled-water pipeline could not be delivered on the project’s timetable, despite NEXTDC seeing merit in that approach wherever enabling infrastructure exists.

This is where the word ready becomes slippery. A project may have a credible customer, suitable technology and strong strategic interest. Delivery still relies on organisations outside the developer’s control. They must approve, finance, build and operate the supporting system.

For a project model, those parties are more than stakeholders in a colourful diagram. They control dependencies.

Water saved, pressure transferred

Closed-loop cooling addresses the immediate water-infrastructure problem. It also changes the electricity profile.

Reuters quoted DCByte analyst Sarvaesh Mohan saying that the new approach would be more energy-intensive if the reported configuration proceeds. This observation applies to the specific substitution at S7. Other low-water cooling systems may have different energy characteristics.

The distinction matters because Sydney’s electricity network is already under pressure. Transgrid has said it has limited capacity to connect new large data centres without augmenting the bulk transmission network. Developers are expected to fund the expansion. A Transgrid executive also told a New South Wales parliamentary inquiry in May that existing regulation was not designed for such large, clustered growth in electricity demand.

The cooling redesign therefore shifts part of the project’s exposure from water infrastructure to power infrastructure. It may still be the best available engineering decision. The cost, timetable and risk allocation need to move with it.

This is a recurring feature of complex clean-technology and infrastructure projects. A design team optimises one component, and the invoice appears somewhere else in the system. The spreadsheet usually calls that row “other.”

System boundaries are useful here. A design can look efficient when the analysis ends at the facility fence. Extend the boundary to include grid reinforcement, pipelines, treatment capacity and permitting, and the project may tell a different economic story.

What a complete project assessment needs

Six questions provide a practical starting point.

1. Are the resources locally available?

National or regional averages are weak evidence for a specific site. The relevant issue is whether sufficient water and power are available under realistic operating conditions, including peak periods and future demand.

2. Can the networks deliver them on schedule?

Resource availability and utility capacity are separate questions. A region may have water while lacking the pipe needed to carry it. A grid may produce enough electricity overall while the local connection remains constrained.

3. Who owns the missing infrastructure?

Pipelines, substations and treatment capacity need an owner, a budget and a delivery programme. If the project depends on another institution’s capital plan, that timetable belongs in the investment case.

4. Which permissions remain unresolved?

Each supporting asset follows its own approval path. At S7, the official project itself is still at the “Prepare EIS” stage, and the recycled-water proposal depended on separate pipeline permission.

5. What happens to operating economics?

Changing cooling technology affects capital cost, electricity use, maintenance and reliability. The most buildable option may carry a different margin than the preferred design.

6. Who carries delay and cost risk?

When a grid upgrade or permit arrives late, someone absorbs the cost. The contract structure should reveal who. Silence merely postpones the discovery.

None of these questions is especially novel. Their importance comes from asking them together. One unresolved dependency can alter construction timing, capital requirements and returns across the entire project.

The policy direction is changing

Australia is beginning to make these dependencies explicit.

In a speech on 15 July, Prime Minister Anthony Albanese announced plans for national rules covering the location, energy and water demands of large data centres. The proposed framework would require future operators to underwrite new power supply, pay their full grid-connection costs, minimise water use and fund any additional water infrastructure.

The framework has been announced, although it is not yet law. National Cabinet discussion and legislation are still to come. It nevertheless sends a clear commercial signal: supporting infrastructure may increasingly become a direct project cost.

That changes more than sustainability reporting. It can affect site selection, financing needs, delivery schedules and projected returns. A proposal with a smaller headline capacity and a complete utility pathway may prove more valuable than a larger scheme waiting for public infrastructure to catch up.

For capital providers, the most material question may sit outside the conventional technology assessment. Who controls each dependency, and can those organisations deliver on the same schedule?

The pace of the slowest system

AI demand will continue to grow quickly. Pipes, grids and approval processes will continue to behave like infrastructure.

The S7 case does not prove that water will constrain every data centre, or that closed-loop cooling is a poor choice. It shows how a project can be redesigned when one local dependency fails to arrive, and how the replacement can strengthen another constraint.

Before announced megawatts become forecast revenue, the physical and institutional delivery system deserves the same scrutiny as the chips and customers.

AI may operate at machine speed. Its infrastructure will scale at the pace of the slowest organisation, permit or network it depends on.

---

No composite case is used in this field note. It discusses a named project that is publicly reported and still in the approval process — not a completed facility, and not a client engagement. Statements about NEXTDC, OpenAI, Transgrid and the announced Australian framework describe the position as reported at the time of writing in July 2026 and may since have changed.