Curating Sovereignties: Mapping AI Factories

In April 2026 Anthropic’s chief executive Dario Amodei signed a memorandum of understanding in Canberra, while approximately $100bn in data-centre investment was moving into Australia. The data centre investment said good things about the Australian Government’s National AI Capability Plan. We were building our AI sovereignty.

Were we? Current policy frameworks identify AI factories as sovereignty-enabling even while they deepen dependences on external data stacks.

This is obviously confusing. By tracing the sources of data centre investment, we can reveal a framework through which to also trace the extension of platform sovereignties into new territories. Yes, we are living in what the Digital Statecraft policy specialists describe as the ‘platform dystopia’.

I propose reading national AI strategy as curation.

States select from constrained menus of financial, institutional and political options to assemble an infrastructure position, and that position can be specified. I distinguish four registers: Locational sovereignty puts data and compute onshore while ownership and model governance stay foreign. Financial sovereignty takes public equity under commercial rather than public-benefit mandates. Operational sovereignty runs a state’s own stack for internal use. Productive sovereignty builds and openly publishes models as public goods, like Switzerland’s Apertus does.

Most national strategies conflate these registers, which is what allows corporate actors to draw public subsidy without returning public value. Introducing a curation frame - which area of this stack do we want to negotiate? - changes the question.

I am using this framework to track the curation of Australia’s AI factories. I draw on a coded dataset of Australian critical infrastructure, data centres alongside the rare-earth and processing sites they depend on, each tagged by ownership, public funding, planning pathway, water risk and governance flags, including whether a sovereign-compute claim is being made against foreign ownership.

The dataset is being published as a live map, so the gap between the claim of sovereignty and its ownership is legible at the level of individual sites.

Worth noting here: The platform also tracks its own compute cost. Because this is AI-assisted work, the material cost of this work must be made transparent.

See the live map.

See the database.

And the Git reference.