By Brad Ferris · 19 September 2026
On 16 September the Australian Financial Review reported that Anthropic had signed its first lease for an Australian data centre, a 725 hectare campus at Western Downs Digital Park near Dalby in Queensland, to be built by Macquarie Capital and Singapore's Zerra DC. The AFR put the cost at about $30 billion. Capital Brief reported $32 billion. iTnews, carrying Reuters, reported planned capacity of 2.16 gigawatts, around 1,500 construction workers, first operation from 2027, and FIRB approval still outstanding.
The number is not the story. One word in the announcement is: the campus is for inference, not training.
Training is the model learning. Inference is the model answering. Every time someone in your business asks a model to summarise a supplier contract, draft a reply to a customer, or code an invoice, that is inference, and today it happens on hardware in someone else's jurisdiction. From 2027, if this build lands as described, a meaningful slice of it can happen in Queensland.
That matters because of the sentence I have heard more than any other in Australian mid-market boardrooms over the past two years. Some version of: we looked at it, but our data cannot leave the country. In financial services, health, legal, and anything touching government work, that sentence has ended the conversation before it started.
It is a real constraint. It has also been doing a lot of work it was never entitled to do.
The first is where your inference runs right now. A great many Australian businesses reached the residency objection by reasoning about it rather than by asking. If you run Microsoft 365 or Google Workspace, some of the AI features in front of your staff are already served from configurable regions, and your agreement may already say more than you think. That is a half day of work with your provider, not a strategy project, and it resolves the question for a surprising number of workloads.
The second is which workloads are genuinely blocked. Once you separate the truly constrained data from the data that merely felt sensitive, the blocked list usually gets short. Customer identifiers and clinical records sit on it. Your quoting process, your supplier correspondence, your internal policy library and most of your back office do not.
What is left after that exercise is the honest scope of the problem a Queensland data centre would solve for you. For most operators I speak to, it is smaller than the story they have been telling.
Two days after the lease news, the AFR reported that Anthropic had joined OpenAI in a stand-off with the federal government over copyright, confirming it will not train models in Australia until the law lets it pay for content rights. Read the two announcements together and the shape is clear. Australia is being built out as a place where models run, not a place where models are made.
For an operator, that distinction has a practical consequence. Onshore inference can answer a data residency question. It does not give you any influence over what the model knows, how it behaves, or what it costs. Those remain decisions made elsewhere, and they are the same supply relationship every AI buyer has been managing all along.
There is also a bill attached to the build that few AI roadmaps have accounted for. Capital Brief reported that the managing director of the company that owns the leased land expects the campus to draw grid power from existing coal plants for an extended period. In the same week, the Office of AI opened consultation on mandatory national AI standards covering data centre power supply, grid connection, water efficiency and siting, with submissions closing 9 October and legislation targeted for early 2027. If your business carries both an AI plan and an emissions commitment, those two pages of the board pack have started talking to each other.
None of this is certain. The figures do not reconcile across outlets, FIRB has not cleared the deal, and nothing has been built yet. Australian AI infrastructure announcements have been running well ahead of Australian AI infrastructure for a while now, and a 2027 date should be read as an intention rather than a delivery.
But the direction is firm enough to plan against, and it hands you something more useful than a headline. It hands you about eighteen months in which the objection still holds and the excuse quietly stops.
Spend them on the work that onshore compute will never do for you. Write down how your processes run, so an AI system can be pointed at something more solid than tribal knowledge. Decide who owns which decisions and which of those a machine may touch. Clean up the data you would want a model reading. Pick two workloads that are not blocked by residency at all and get them working properly, so that when capacity lands you are extending something that works rather than starting from nothing.
The businesses that move in that window will meet 2027 with the constraint removed and the groundwork done. The rest will meet it having lost their best reason for standing still, which is a harder position than it sounds.
Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.