Anthropic is running a tender for at least 1.4 gigawatts of Australian data centre capacity, a build-out that could cost up to US$15 billion (about A$21.6 billion) and reach at least 1 GW online by end-2027. 1.4 GW roughly equals Australia's entire operational data centre fleet today, so a single procurement would nearly double national capacity. This is less a real-estate transaction than a bid to lock in training and inference compute for the next generation of Claude models.
TL;DR
- Anthropic issued an RFP in March 2026 for 1.4 GW+ of Australian capacity, met shortlisted developers in Canberra in April, and expects a preferred partner within roughly six weeks.
- Total capex could reach US$15B; target is at least 1 GW operational by end-2027.
- Shortlisted operators: CDC Data Centres, AirTrunk, NEXTDC (ASX:NXT), Iren, and Stack. The work may split across four to five providers.
- Anthropic opened a Sydney office in early 2026 and signed an MoU with the Australian government on AI safety and infrastructure guidelines.
- AEMO has separately disclosed 5.4 GW of large data-centre projects in the grid connection queue — this tender lands in an already crowded pipeline.
What exactly is the deal?
According to reporting on confidential tender documents (AFR, July 5, 2026), Anthropic issued an RFP in March 2026 asking Australian developers to offer at least 1.4 GW of new capacity. It prefers a single 1.4 GW+ site but is flexible about spreading the load across multiple partners. Shortlisted bidders met in Canberra in April 2026, and a preferred-partner decision is expected within roughly six weeks.
The full envelope, up to US$15 billion, covers land, power, buildings, cooling and GPU-cluster fit-out. Anthropic is not publicly listed — it has confidentially filed for an IPO but set no listing date — so funding will likely mix balance-sheet cash, cloud-partner commitments, and developer-financed leases where operators like NEXTDC or CDC front the capital against long-term take-or-pay contracts.
Why is Anthropic building this in Australia?
Three constraints push frontier AI labs toward Australia over adding capacity in the US or Europe.
- Power availability. Training clusters at 100 MW+ scale need firm, expandable grid connections. Australia has land, high-voltage corridors, and a renewable generation pipeline. AEMO's connection queue already includes 5.4 GW of large data-centre projects, signaling grid planners are adapting to this load class.
- Renewable mix. Anthropic frames compute expansion around low-carbon energy. Australia's solar, wind and firming mix makes it easier to align builds with climate commitments than coal-heavy US grids.
- Regulatory posture. A signed MoU with the Australian government covers AI safety research and alignment with domestic data-centre guidelines — a predictable policy environment matters when payback runs a decade or more.
Geographic diversification also reduces exposure to any single jurisdiction's export controls or grid failures. If a US region caps new data-centre interconnections, an Australian campus keeps training on schedule.
Who are the shortlisted developers?
Five operators are named in the tender materials:
- CDC Data Centres — privately held, dominant in Canberra and government workloads.
- AirTrunk — Asia-Pacific hyperscale specialist, acquired by Blackstone in 2024.
- NEXTDC (ASX:NXT) — the largest ASX-listed operator, national footprint.
- Iren (formerly Iris Energy) — Nasdaq-listed, pivoted from Bitcoin to AI compute, hydro-backed sites; its share price moved on the tender reports.
- Stack Infrastructure — IPI Partners-owned, expanding in APAC.
Splitting 1.4 GW across four or five of these spreads execution risk — no single Australian developer has 1.4 GW of shovel-ready capacity, so the realistic outcome is a portfolio of 200–400 MW builds across multiple states.
What does 1.4 GW of AI compute actually mean?
- 1.4 GW roughly equals the electricity consumption of about 1 million Australian homes.
- It matches the entire currently operational Australian data-centre fleet (~1.4 GW across all workloads).
- A single modern training cluster of ~100,000 GPUs draws around 150 MW, so 1.4 GW supports 8–10 frontier-scale training clusters, or a much larger inference footprint.
The practical read: enough compute to train several successive Claude generations and serve Asia-Pacific inference without round-tripping to US regions.
How does this fit the wider AI capex picture?
Anthropic's tender lands inside a broader wave of AI infrastructure spending straining both grids and balance sheets. Hyperscalers run annual capex north of US$50 billion each. The scale-versus-revenue question is real — training costs are rising faster than API revenue, the tension covered in our AI capex-revenue gap analysis. The parallel realignment among hardware suppliers is visible in Foxconn's AI infrastructure pivot, and the chip economics in our AI chip architectural innovation piece.
For Claude buyers, the deal signals two shifts. First, Anthropic is moving from a pure cloud-tenant model toward controlling more of its own stack, as explored in the enterprise AI deployment strategy guide. Second, APAC inference capacity should reduce latency for regional customers using the Claude 5 model line.
What are the risks and open questions?
- Grid connection timing. A 1 GW-by-2027 target depends on network providers granting connections on schedule. AEMO's queue is long; precedent projects slipped 12–24 months.
- Water and cooling. Liquid-cooled halls at this density need reliable water or closed-loop refrigerant. Site selection hinges on local water rights as much as power.
- Planning approvals. Councils in NSW and Victoria are scrutinising data-centre approvals more closely, especially where projects compete with residential grid upgrades.
- Financing. Until Anthropic completes an IPO or raises a fresh round, the $15B figure is an upper bound, not a committed budget. Developer-led financing shifts risk to operators.
- Model-demand risk. If frontier training plateaus or inference shifts to on-device silicon faster than expected, part of 1.4 GW could go under-utilised.
FAQ
Q: When will Anthropic's Australian data centres come online? A: The target is at least 1 GW operational by end-2027, with the full 1.4 GW+ following after. A preferred-partner decision is expected within roughly six weeks of the April 2026 Canberra briefings, though grid-connection timelines are the main schedule risk.
Q: How much is Anthropic actually spending? A: Tender documents point to up to US$15B (about A$21.6 billion) in total build cost — an upper bound, not a signed commitment.
Q: Which companies are bidding? A: Five operators: CDC Data Centres, AirTrunk, NEXTDC, Iren, and Stack. Anthropic prefers a single 1.4 GW campus but is open to splitting across four to five providers.
Q: Why Australia rather than the US or Southeast Asia? A: Australia offers land, a renewable-heavy generation pipeline, AEMO's existing connection queue, and a signed MoU on AI safety and infrastructure guidelines — a combination harder to assemble in more constrained markets.
Q: Does this affect Claude pricing or availability? A: Not directly in 2026. From 2027, more APAC inference headroom should improve availability and reduce latency. Pricing effects depend on how the capacity is financed.
Q: Is Anthropic publicly traded? A: No. Anthropic has confidentially filed for an IPO but set no listing date. Public exposure to the tender flows mainly through NEXTDC and Iren.
Discussion
0 comments