Photo: Bokeh / Tomedia. Urban Aqueduct and Construction Site.
The proposed Western Downs Digital Park near Dalby, in regional Queensland, has been reported as a $32 billion development. Anthropic has signed an agreement to use part of it, and the ABC reports that the company would use it to run Claude, its AI assistant, rather than train new models.[1]
If you mostly use Claude to ask questions or write something, that sounds like an enormous amount of infrastructure. Surely training the model is the bit that requires all those computers. Once you’ve done that, how much hardware do you need to let people use it? Most of us experience the whole thing as a box we type into, which doesn’t give us much to go on when somebody proposes building a facility this large.
I think Australia needs more serious computing infrastructure. I would like us to be less dependent on accessing it somewhere else, and I don’t think putting “AI” in a development announcement automatically makes the development bad. But a project this large needs a better explanation than being told this is the future. We need to understand what would run there, what it takes to keep it running and what having it here would do for us.
Answering questions takes a lot of computers
Training is the process that produces the model by adjusting its internal parameters. Inference is the work involved in running that trained model. The distinction makes sense, but it can leave you with the impression that most of the work is finished by the time you get access to Claude.
There is still a considerable amount happening every time you use it. The model processes your input and generates a response in tokens, the pieces of text it works with. It needs memory for the model and for the information associated with the requests it is handling. Large models can be spread across multiple graphics processors, or GPUs, and longer conversations and more simultaneous requests put more pressure on that memory. Then everyone wants their answer quickly. You can’t make people wait half an hour every time they send a message and expect them to keep paying for it.[2]
The simultaneous part is where the scale starts making more sense. Imagine 10,000 people each receiving 50 tokens a second. That’s half a million output tokens every second, before processing what those people sent in. Those are illustrative numbers, not a benchmark for Claude or a calculation of what Dalby needs. But they help explain why looking at your own conversation gives you a fairly poor idea of what the service has to handle.
There are ways to make this more efficient. Requests can be batched together, some work can be cached, and software can schedule the hardware so it spends less time sitting idle. You don’t need a completely separate copy of the model permanently assigned to every person. Even with those efficiencies, though, the provider still has to deliver all of those responses at a speed people will tolerate.[2]
Then you have agents. A coding agent can read files, decide what to change, call a tool, inspect the result and make another model call. It might change something, run the tests, find a problem and go through the process again. One instruction from you can result in a whole sequence of inference calls. That is already a different workload from asking for one paragraph and closing the window.[3]
And the GPU servers are only part of what you need to operate the service. Before a request reaches them, something has to check who you are, whether you can access the service and where to send the work. Load balancers distribute requests. Serving software schedules them. Monitoring tells the operators when things are slowing down or failing. There is networking and storage around all of that, plus the power, cooling and spare capacity needed to keep it running when equipment breaks.[4]
Some of those jobs run as software on shared machines; others are spread across many servers. I wouldn’t pretend that this is a blueprint of Claude’s infrastructure. It is the sort of work involved in providing a large service, and it explains why buying GPUs is only one part of the job.
So yes, an inference facility can be enormous without training a new model. Whether this particular development needs to be this particular size still depends on how much work its customers will send through it. I can explain why inference needs substantial infrastructure without knowing whether a developer’s demand forecast will turn out to be right.
Anthropic has signed a lease
The development itself is Western Downs Digital Park. Dexus’s announcement on 16 September describes a consortium involving Australian Data Centres, Zerra DC and Macquarie Capital. That consortium has entered into lease documentation with an Australian subsidiary of Anthropic for the first stage, subject to approvals. The project website describes a wider plan of up to four phases.[5]
Anthropic is obviously a major part of the story. But there are separate decisions involved in developing the campus, leasing capacity and building later stages. Reading about the entire proposed development as though every part of it is already being built and filled with Claude servers makes it much harder to understand what has been agreed to.
I find the way these projects get described frustrating. You see the company name, the dollar figure and “AI data centre”, but a figure for the first building can mean something very different from a figure for the eventual campus. That gets confusing quickly when you’re trying to work out how much water or electricity the development would need.
The ABC’s written report identifies Zerra as the developer and says Anthropic has agreed to use part of the proposed development. It also includes the claim about low operational water use.[1] Those details help explain what is being proposed. A headline about Anthropic and $32 billion gives you very little basis for deciding whether the development makes sense.
Before treating another data centre as unnecessary duplication, I’d want to know what it is offering. A company renting server racks to other businesses, a cloud provider and a facility running large AI models can have different customers and different hardware requirements. Calling them all data centres tells you surprisingly little about whether they are interchangeable.
Australia already has networks of these facilities. Amazon Web Services, or AWS, groups its infrastructure into regions, including Sydney and Melbourne. Each region contains multiple availability zones, which can themselves contain more than one physical data centre.[6] More providers and locations could give us useful alternatives. The combined demand they place on water and electricity still needs to be considered wherever they are built.
The cooling system changes the water demand
GPU servers generate heat, and getting rid of it takes equipment, energy and, with some cooling designs, a continual supply of water. That is a reasonable thing to ask about in regional Queensland. But you need to know which cooling system is being proposed before assuming how much water it consumes.
Western Downs says it will use closed-loop water circulation, with heat released to the outside air rather than through evaporative cooling. So there can still be liquid circulating inside the system. What they’re avoiding is the continual evaporation and replacement of cooling water. “Air-cooled” is a fairly easy term to misunderstand if you take it to mean there is no water anywhere in the machinery.[7]
The developer’s current estimate for the first building is about 21,000 litres of operational water a day. Around 8,000 litres would come from recycling on site, leaving about 13,000 litres of raw-water demand, primarily to be met through rainwater harvesting. Those are modelled figures for the first building. We don’t yet have an operating campus whose consumption we can measure.[7]
Construction is a separate demand. The current project website estimates a peak of about 380,000 litres a day across the site, with a one-off requirement of roughly 27 million litres per building for commissioning and initial filling. It proposes using processed water from nearby coal seam gas operations as the main construction supply, subject to agreements, availability and approvals.[8]
I don’t see anything wrong with making use of an existing industrial by-product like that, provided it is treated appropriately and the supply is available. Recycling water and reusing it seems like a fairly sensible approach. Pointing out that some other industry uses more water wouldn’t settle whether this development is suitable for Dalby, though. The supply and demand in that particular area are what we have to work with.
The website says groundwater is not a planned source, but retains it as a last-resort contingency requiring further investigation, entitlements and approvals. Water delivered from outside the region is another backup. I’d want to understand how those arrangements work during an extended dry period, how much water can be stored, and where the backup supply would come from. A rainwater calculation over a whole year still has to translate into enough water being available when it is needed.[8]
That leaves me fairly comfortable with the principle of the proposed cooling and water arrangements. There is enough detail here to have a more useful conversation than assuming every AI facility has the same water requirements. There is also enough still dependent on agreements and approvals that I wouldn’t call the water question finished.
Electricity is another substantial part of the proposal. A dedicated connection to the high-voltage network at Braemar would supply the site, and the developer says it will fund the dedicated connection and project-specific network infrastructure. The electricity would still come from the Queensland system. The website says the generation and contracting arrangements for each stage will be developed as the campus grows.[9]
That’s where I want more detail. The computers, cooling and supporting equipment all need power. A dedicated connection gets that power to the site; additional generation has to be available to meet the demand. I want to know how those schedules line up as each stage opens, who pays for the work and what happens if the electricity supply is delayed. Batteries can help manage fluctuations and interruptions, but they also need electricity to charge them.
I don’t think the size of the announcement tells us what will happen to household bills. I also wouldn’t accept a promise that bills will be unaffected without seeing the arrangements behind it. Funding a connection is one commitment. Having enough electricity available for the new demand is another thing that needs to be worked through.
Having the servers here helps
Part of my support for more Australian infrastructure is very practical. Accessing a service hosted overseas adds network delay, and Australia is a long way from a lot of the places where these services run. Having suitable capacity closer to the people using it can reduce that delay and give providers more options for distributing their workload.
A facility in Queensland could serve users across Australia and overseas. Providers can route requests between locations according to latency, capacity and availability. The location of a building doesn’t tell us where all its users will be, or guarantee that every Australian request will go there. It also won’t remove the time the model itself takes to generate an answer.[10]
Still, I’d like Australian businesses to have more options here. I’d like us to be able to run more of the systems we rely on without assuming access to overseas infrastructure will always be affordable and available. The case for having domestic capacity makes sense to me even before we get into the much bigger argument about AI.
Sovereignty is where I think we need to be more precise about what we’re gaining. If an American company operates a service from an Australian facility, we have gained local infrastructure. Control of the service can still sit with that company. It can still determine who gets access, what it costs and the terms under which we use it.
Where the data sits, who can access it, who controls the encryption keys and whether we can move to another provider all affect how much control we have. I’d want to know whether Australian organisations can access the capacity and what happens if the provider withdraws the service. A local address alone doesn’t answer those questions.
There are practical reasons for choosing Australian hosting, including requirements that can apply to particular organisations and contracts. But the Privacy Act doesn’t impose a blanket rule that all personal information must stay here. Australian privacy obligations can also apply when covered organisations disclose information overseas. Moving data out of the country doesn’t automatically remove responsibility for what happens to it.[11]
Leasing an overseas company’s service from a local building can improve our options while leaving us dependent on that company to keep providing it. I’d still count that as useful Australian capacity. I would be more careful about calling it independence.
Privacy needs people to enforce it
Keeping data in Australia still leaves the organisation holding it responsible for how it is used. A business needs to understand what it can collect, who should have access and what happens when something goes wrong. The location of its servers is one part of those decisions.
The obligations already vary depending on the business and the information involved. Most businesses with annual turnover of $3 million or less are generally exempt from the Privacy Act, with exceptions including health service providers and businesses that trade in personal information. So a small business can have privacy obligations that another business of a similar size doesn’t.[12]
If we’re going to extend those obligations to more businesses, I want to know who audits them, who enforces them and what compliance requires in practice. Australia already has a federal privacy regulator, the Office of the Australian Information Commissioner, or OAIC. Whatever rules we settle on, somebody has to investigate failures, and businesses have to understand what they are expected to implement. There are a lot of different ways to collect, store and share information. Writing a policy doesn’t tell you whether the system underneath it is doing what the policy says.
Organisations also need to understand what they’re asking people to agree to. Where an organisation relies on consent, the OAIC says that consent needs to be informed, voluntary, current and specific, and given by someone with the capacity to understand it. Telling somebody a practice exists doesn’t necessarily establish that they have consented to it.[13] There is more involved than putting something in the sign-up form and deciding everyone agreed.
I’d like us to be able to build this stuff
I don’t particularly like the idea of Australia continuing to use all these services while assuming the infrastructure behind them should always be somebody else’s problem. We’d still be buying access, depending on the provider and accepting their terms. I’d like Australian businesses to have more options about where their systems run and who they rely on to keep them running. Having more of that capability here gives us something to work with.
That is why I find the argument around projects like Dalby frustrating. There are useful things being proposed here, and there are questions the developers still need to answer. The people living nearby are entitled to ask what happens to their water and electricity. We should also be asking how much control Australian customers would have over the services they buy. I want those questions answered well enough that we can make a decision about the actual development.
If the developers can show that the supply arrangements work, existing users are protected and there is demand for what they’re building, then I’d like us to get on with building it. I want Australia to get better at providing the infrastructure we keep saying we need.
Sources
ABC News, AI giant Anthropic signs agreement for $32b Queensland data centre, 16 September 2026. The reported cost and intended inference workload are attributed to this coverage.
NVIDIA, Mastering LLM Techniques: Inference Optimization and Optimizing Inference for Long Context and Large Batch Sizes with NVFP4 KV Cache. These describe general serving techniques, not Claude’s private implementation.
Anthropic, Building effective agents.
NVIDIA, Triton Architecture. Triton documents scheduling, batching, model execution, health checks and performance metrics; it is an example of serving infrastructure.
Dexus, Australian Data Centres – Western Downs Digital Park, 16 September 2026, issuer announcement distributed through EQS; Western Downs Digital Park, project overview. Checked 17 September 2026.
AWS, AWS Region in Australia and Availability Zones.
Western Downs Digital Park, September 2026 water fact sheet and current water FAQ. Developer estimates and proposed design, checked 17 September 2026.
Western Downs Digital Park, current water strategy and construction FAQ. The live page supplies the 380 kL/day construction estimate, 27 ML commissioning figure and groundwater-contingency wording. Checked 17 September 2026.
Western Downs Digital Park, electricity FAQ. Proposed connection, funding and supply arrangements, checked 17 September 2026.
AWS, Latency-based routing. An example of routing between regions, not a claim about Anthropic’s customer-routing plans for Dalby.
OAIC, Sending personal information overseas and APP 8: Cross-border disclosure of personal information. Specific duties and exceptions depend on the arrangement.
OAIC, Small business. Checked 17 September 2026.
OAIC, APP 3: Collection of solicited personal information, especially paragraphs 3.32–3.34.