Anthropic Plans to Spend $15 Billion to Secure 1.4GW of Computing Power in Australia
Alina Collins
Anthropic plans to deploy 1.4 GW of data-center capacity in Australia at a cost of roughly $15 billion, with the first 1 GW live by end of next year — the first time an AI company has committed to power-plant-scale infrastructure offshore.
How big is 1.4 GW, really?
1.4 GW equals the output of several large nuclear reactor units — enough to run hundreds of thousands of top-tier AI accelerator chips (specialized processors that handle AI workloads) at full load.
This means → Anthropic is not renting a few server halls; it is locking in a small-grid-scale energy footprint.
A confidential tender document sets a deadline: at least 1 GW must be operational by the end of next year, leaving less than two years to build.
Why not build in the United States?
U.S. data-center construction is hitting bottlenecks. QTS, owned by Blackstone, just canceled the 850-hectare DigitalGateway project in Virginia after sustained local opposition and lawsuits.
In plain terms = land, power, and community resistance are a triple chokepoint for large U.S. builds.
Australia offers ample land, stable energy, and a favorable climate. Anthropic signed an MOU with the Australian government in March covering AI-safety research and the national AI agenda — the policy runway is already cleared.
Where is the money coming from?
Annualized revenue has surged from $9 billion at the start of the year to over $44 billion. Inference-infrastructure gross margin has jumped above 70%, delivering the company's first profitable quarter.
Last month Anthropic closed a $65 billion funding round at a $965 billion valuation and formally filed for an IPO.
This means → revenue up nearly 5×, first-ever profit, and a near-trillion-dollar valuation — all three had to land simultaneously to underwrite a $15 billion physical buildout.
Has the compute-supply playbook changed?
Anthropic is shifting from pure cloud leasing to a hybrid model: own-build + mega-partnerships + cloud rental.
Beyond Australia, it recently signed a roughly $45 billion deal with SpaceX, an $1.8 billion agreement with Akamai, and continues to buy chips and cloud capacity from Google.
This reflects a broader shift: top AI companies no longer want to be tenants of cloud providers — they are starting to control the physical layer the way telecom operators do.
Where is the risk in this bet?
Committing $15 billion and bringing 1 GW online by late next year — if inference-demand growth slows or chip efficiency leaps forward, capacity could overshoot.
In plain terms = building power-plant-scale infrastructure is a bet that AI compute demand will only grow, not shrink, over the next two years.
Whether this hybrid model can deliver enough capacity before the IPO will be the key test of whether the company's near-trillion-dollar valuation holds up.
Content is for reference only, not financial advice.