What happened
Anthropic has signed its first data center lease agreement in Australia, securing a campus with a planned total capacity of 2.16 gigawatts, according to foreign media reports dated September 16.
The project still requires approval from Australia's Foreign Investment Review Board and is expected to begin operations as early as 2027.
People familiar with the matter said the new data center will mainly be used for AI inference — handling user requests and generating predictions during the operational phase — rather than for model training.
The project is being handled by data center developer Zerra DC, which plans to obtain electricity through renewable energy purchase agreements, bear grid connection costs itself, and use a closed-loop, air-cooling system.
The news comes as Anthropic, which has voiced calls for slowing down AI, again draws industry attention; CEO Dario Amodei previously published a long article urging the industry to slow the pace of advancing frontier model capabilities to buy time for safety research and risk governance, drawing support from OpenAI CEO Sam Altman and SpaceX CEO Elon Musk.
Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg have expressed partial disagreement, with Huang firmly rejecting 'AI doomerism' and saying the AI industry 'does not need any new laws,' while Zuckerberg argued that introducing third-party evaluation bodies is the industry's best practice rather than advocating for slowing down.
Google DeepMind co-founder Shane Legg said in a September 16 Financial Times interview that AI capabilities are developing 'very, very fast' but that 'capability must not get ahead of safety,' calling Amodei's 'slowing down' proposal 'worth considering in direction' while saying implementation needs further discussion.
Also on September 16, Microsoft AI head Mustafa Suleyman published an article titled 'A Warning About Model Welfare,' questioning Anthropic's approach of involving concepts such as 'consciousness,' 'emotion' and 'model welfare' in model training, arguing AI has no consciousness or feelings and warning that a highly autonomous AI acting on the premise that it has rights or that its welfare is threatened could become harder for humans to control.
Microsoft is an important investor and commercial partner of Anthropic, and Suleyman emphasized he recognizes Anthropic's work on AI safety, with his doubts focused specifically on 'model welfare' and AI anthropomorphization.
Why it matters
The contrast between Anthropic's public calls to slow frontier AI development and its continued infrastructure expansion suggests that 'slowing down' and infrastructure growth may not refer to the same thing.
If the new data center is indeed mainly for inference, it indicates that even if frontier model training slows, AI slowdown does not mean computing demand disappears — demand may shift from training to inference.
The exchange between Suleyman and Anthropic shows the AI safety debate is widening from development speed to questions about model training itself and anthropomorphization.
Key facts
Anthropic signed its first Australian data center lease for a campus with a planned total capacity of 2.16 GW.
The project needs approval from Australia's Foreign Investment Review Board and is expected to start operations as early as 2027.
The data center is mainly intended for AI inference, not model training.
Zerra DC is the developer, and the project plans renewable energy purchase agreements, self-borne grid connection costs, and closed-loop air cooling.
Australia's government plans to set limits on data center resource use starting in 2027.
Local economists expect related Australian investment could reach 150 billion Australian dollars by 2030.
Dario Amodei called for slowing frontier model capability advances; Sam Altman and Elon Musk voiced support, while Jensen Huang and Mark Zuckerberg disagreed in part.
Shane Legg said AI capabilities are developing very fast but safety must not lag behind.
Mustafa Suleyman questioned Anthropic's use of 'model welfare' concepts in training.
What to watch next
Whether Australia's Foreign Investment Review Board approves the project, and whether the planned 2027 start of operations holds.
How Anthropic reconciles its safety messaging with continued infrastructure expansion, particularly if the campus remains inference-focused.
Whether the debate over model welfare and anthropomorphization reshapes how AI labs approach training and third-party evaluation.
