Anthropic has signed an agreement for its first Australian data-centre site, according to Reuters, with the planned facility focused on AI inference rather than model training. The deal is useful to study because the location, power design and regulatory setting show how frontier AI infrastructure is becoming a regional deployment problem, not only a model-training problem.
What Anthropic agreed to build in Australia
Reuters reported on September 16, 2026 that Anthropic signed its first Australian data-centre lease agreement. The planned campus is about 2.16 gigawatts and is located roughly 250 kilometres from Brisbane. The site is being developed by Singapore-based Zerra DC and is expected to begin operations in 2027. The project still requires approval from Australia's Foreign Investment Review Board, so the reported agreement is not the same as a completed operating facility.
The Guardian separately reported a value of about $31.9 billion for the wider development and described the Anthropic portion as covering about 725.5 hectares. These figures describe the development context; they should not be treated as proof that Anthropic has already spent that amount.
Why inference changes the infrastructure question
The most important technical detail is the intended workload. Reuters says the site will be used for inference, not model training. Training and inference have different infrastructure patterns: training needs large coordinated compute runs, while inference needs capacity that can serve user requests reliably and with predictable latency.
Anthropic has already described major compute partnerships elsewhere. In April 2026, the company said its Amazon agreement could provide up to 5 GW of new capacity for training and deploying Claude, including Trainium2 and Trainium3 capacity. That makes the Australian project part of a wider infrastructure strategy rather than an isolated experiment.
The power and cooling design matters too
Reuters reported that the Australian site is planned around renewable energy and a closed-loop, air-cooled system intended to reduce water use. The Guardian reported that the project is connected to the Braemar power station and that the first stage is planned for next year. The sources differ on some policy details around the site's electricity supply, so those details should remain attributed rather than presented as a settled national rule.
For an inference campus, cooling is not a side issue. AI accelerators turn electrical power into both computation and heat. A design that reduces water demand can matter in regions where data-centre growth is constrained by water availability even when electrical capacity exists.
Who is affected by the project
Cloud and AI customers are the indirect users of this capacity because more regional inference infrastructure can add serving capacity. Australian governments and local communities are directly affected through land use, electricity demand, construction and jobs. Infrastructure providers such as Zerra DC are also central because the project depends on a data-centre campus rather than a standalone server installation.
What remains unconfirmed
The site has not yet become a production Anthropic facility. Regulatory approval remains pending, and the final operating timetable can change. The public reports also do not establish which Claude models, hardware configuration, network topology or customer workloads will run at the Australian site. Those details should not be inferred from the 2.16-GW headline number.
What this means for AI infrastructure teams
The practical lesson is narrower than “AI needs more data centres.” Anthropic is adding geographically distributed inference capacity while choosing a site where power, cooling, regulation and land can support a large deployment. Teams planning AI services should therefore treat model serving as a capacity and location problem: where requests are served, how much power is available, how heat is removed, and which regulatory approvals are required can affect the product as much as model capability.
Three questions readers should ask next
First, when will the facility actually become operational? Second, which Claude workloads will be served there? Third, how will the project balance electricity demand, cooling, local infrastructure and regulatory requirements? The current public reporting does not answer all three, which is important context for interpreting the announcement.