Anthropic has signed an agreement to take part of a proposed $32 billion data centre campus on Queensland’s Western Downs, near Dalby, with the company aiming to have its section online in 2027. It is the first Australian data centre deal by a frontier AI developer, and it landed in the same week that OpenAI said it was in talks to train models in Australia and Sharon AI bought a former paper mill site near Albury. For a business owner the question is not whether this is big news for Queensland. It is what, if anything, it changes about how you should buy and build AI. The honest answer is: some things, meaningfully, and later than the headlines suggest.

What has actually been signed

The facts, from the ABC’s report on 16 September 2026 and Converge Digest’s summary of the announcement:

• The campus is the Western Downs Digital Park, proposed by Singapore-based developer Zerra DC on a site of about 725 hectares roughly 37 kilometres from Dalby, in a consortium with Macquarie Capital and Dexus-backed Australian Data Centres.
• The whole project is put at $32 billion, with a planned peak capacity of 2.16 gigawatts. The ABC reports it would draw as much power as about 1.5 million average Australian households and take four to six years to build in full.
• Anthropic has signed lease documentation for the first stage only, and aims to bring its section online in 2027. The lease is subject to Foreign Investment Review Board approval, and the developer still needs council planning approval.
• The capacity is intended for inference, meaning running Anthropic’s Claude models to answer user requests, rather than training new models.
• The ABC reports Anthropic would cover the full cost of connecting to the grid, and that the facility uses a closed-loop, air-cooled design with water use comparable to an office building.

Queensland Premier David Crisafulli called it “a major win that will deliver more jobs”. Federal MP David Littleproud warned that “power reliability and cost for existing homes, farmers and businesses cannot be sacrificed”, and local group Save Our Darling Downs raised concerns about the energy draw and the prospect of gas-fired generation. The Greens criticised the deal. Expect the power debate to continue; it does not change what the facility will do.

What onshore inference changes

Data residency becomes a real option, not a workaround. Today, most Australian businesses using frontier AI models send their prompts and data offshore, usually to the United States, and manage the privacy and contractual implications of that. Inference capacity in Queensland means the possibility of Australian-hosted processing for those models. For industries with data-handling obligations, such as health, finance, legal and government suppliers, that removes a barrier that has kept a lot of AI projects at pilot stage.

Latency drops. A round trip to a US data centre adds a noticeable delay to every AI response. Onshore inference makes real-time uses such as voice agents, live customer chat and in-app assistants feel immediate. If you have tested an AI voice agent and found the pauses awkward, this is part of the fix.

Supply becomes more reliable. Businesses that have hit rate limits or capacity constraints during peak demand will benefit from more compute in the region. That matters most for anyone building AI into a product rather than using it at a desk.

Procurement conversations get easier. “Where is our data processed?” is often the first question a board, an auditor or an enterprise customer asks. “Australia” is a much shorter conversation than a description of overseas transfer safeguards.

What it does not change

The timeline is 2027 at the earliest, for the first stage. Anything you build today runs on the infrastructure that exists today. Do not defer an AI project on the assumption that onshore capacity is imminent; plan for offshore processing now with a path to move later.

Residency is not the same as compliance. Processing data in Australia does not, by itself, satisfy the Privacy Act, a customer contract or an industry code. You still need to know what data you send to a model, why, how long it is retained and who can access it. The location is one line in that assessment.

Model choice still matters more than location. The right model for a task depends on accuracy, cost, speed and how well it handles your data. A model hosted nearby that gives worse answers is not a better model.

The economics of your project do not change. Inference gets cheaper over time regardless of where it runs. The cost that determines whether an AI project pays for itself is usually the integration, the data work and the change management, not the tokens.

What to do now

1. Inventory the AI you already use and record where each tool processes data. Most businesses are surprised by the answer. This is the baseline you will need when onshore options arrive.
2. Write down your data rules. Which categories of data can go to a third-party model, under what contract terms, and which cannot. Do this once, apply it to every tool.
3. Build for portability. If you are integrating an AI model into your systems, keep the model behind an interface you control so that switching provider or region later is a configuration change, not a rebuild.
4. Start the projects that do not depend on residency. Internal knowledge assistants, document drafting, customer-service triage and process automation on non-sensitive data can start today and deliver value before 2027.
5. Queue the projects that do. Anything blocked by a data-residency requirement should be scoped now so it is ready to move when Australian capacity is available.
6. Watch the power story. Energy cost and reliability on the Western Downs will be political for years. It will not stop the facility, but it may shape timing and pricing.

The bigger picture

This week’s announcements, taken together, say that the major AI developers now consider Australia worth building in. That is good for Australian businesses in a way that has nothing to do with sovereignty rhetoric: more local capacity means faster, cheaper, more compliant AI for the companies that use it. The businesses that benefit first will be the ones that already know what they want to build, have their data rules written, and have kept their options open.

How Digital One Agency helps

Digital One Agency builds custom AI solutions, AI voice agents and business automation for Australian businesses, and we design them to be portable from the start, so a change of model provider or hosting region is a configuration decision rather than a rebuild. If you are not sure which of your AI projects can start now and which should wait for onshore capacity, the free AI audit will map that out.

The infrastructure is coming. The advantage goes to the businesses that are ready for it. Book a meeting with Digital One Agency to plan your next AI project.

Related services

Digital One Agency delivers AI development solutions, AI integration services for digital transformation, private AI for data that must stay under your control, and secure AI system development, backed by AI consulting services that produce a costed plan in 30 days.

Frequently asked questions

What has Anthropic actually agreed to in Queensland?

Anthropic has signed lease documentation for the first stage of the Western Downs Digital Park, a proposed $32 billion data centre campus near Dalby developed by Zerra DC with Macquarie Capital and Dexus-backed Australian Data Centres. The lease is subject to Foreign Investment Review Board approval, and Anthropic aims to have its section online in 2027.

Will this let Australian businesses keep their AI data in Australia?

The capacity is intended for inference, meaning running Claude models to answer requests, so it opens the possibility of Australian-hosted processing for those models from 2027. Whether a given product or plan offers Australian processing will depend on what Anthropic makes available, so treat it as an option to plan for rather than something to rely on today.

Does processing data in Australia make an AI tool compliant?

No. Data residency is one factor. You still need to know what data you send to a model, why, how long it is retained, who can access it, and whether that meets the Privacy Act, your customer contracts and any industry code that applies to you.

Should we wait until 2027 to start AI projects?

No. Projects that do not depend on data residency, such as internal knowledge assistants, document drafting and customer-service triage on non-sensitive data, can start now. Scope the residency-dependent projects so they are ready to move when onshore capacity is available, and build with a portable architecture so switching is a configuration change.

How big is the Western Downs data centre?

The ABC reports a $32 billion project on about 725 hectares that would draw as much power as around 1.5 million average Australian households and take four to six years to build in full. The developer's planned peak capacity for the campus is 2.16 gigawatts.