On 9 September 2026 Nvidia announced that eight Australian operators (Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC and AirTrunk) would deliver up to 2 gigawatts of AI factory capacity by 2027. The pitch is onshore compute for banks, miners, universities and government. It is welcome news for Australian users who today rent GPUs in Singapore or Virginia. It is not, on the evidence published so far, digital sovereignty.
The strongest case for the announcement
Sovereignty advocates have a serious argument. Data that stays on Australian soil is subject to Australian courts, Australian privacy law and Australian security agencies. Latency falls. Regulated buyers such as banks and agencies can meet residency obligations without negotiating offshore exceptions. And a domestic compute market gives local start-ups an alternative to queuing for capacity abroad. Nvidia's release says the participating providers "will operate the AI factories", with Nvidia supplying the platform, accelerated computing, networking and software. Australian companies, on Australian land, would run the facilities.
That is a real improvement over the status quo, and it deserves to be taken seriously rather than dismissed as marketing.
What the announcement does not say
The release is thin where it matters. Trade coverage of the announcement notes that it gives no capital figure for the programme as a whole, no timetable per site and no named paying customer. The only users the release names are Heidi and Atlassian, and those appear as users of Nvidia's Nemotron models, not as buyers of the new capacity. It also says nothing about where the electricity for 2 GW will come from, though CDC separately cites 100% renewable power for its own estate.
The pieces that are specific are partner-level. Sharon AI says it is deploying up to 68,000 Nvidia GPUs. IREN points to its 800-megawatt Bundey campus in South Australia. CDC says it operates more than 550 megawatts across Australia and New Zealand, with a further 800 megawatts under construction. Those figures describe existing or planned assets that the partners would have been building anyway. The headline 2 GW is best read as an aggregate of ambitions, not a binding commitment.
Location is not control
The sharper critique is architectural. A data centre in Western Sydney that runs Nvidia hardware, Nvidia networking and Nvidia's software stack is physically Australian and technologically American. Export-control policy, licensing terms, firmware updates and supply allocation all sit with a US vendor and, ultimately, US law. If Washington restricted advanced-chip shipments, or Nvidia prioritised other customers, an Australian "AI factory" would have no domestic lever to pull.
That does not make the deal a mistake. No country at Australia's scale builds a competitive AI stack from silicon up, and pretending otherwise would waste public money. But it does mean the word "sovereign" is doing marketing work. A more honest framing is resilience: more capacity, onshore, under Australian jurisdiction for data, with a dependency on US technology that is acknowledged rather than hidden.
Canberra already has a yardstick
The government is not starting from zero. Its data centre expectations, published on 23 March 2026 as a deliverable of the National AI Plan, set out five expectations for data centre and AI infrastructure developers: prioritise the national interest, support the energy transition, use water sustainably, invest in Australian skills and jobs, and strengthen research and local capability. The first covers national security and data sovereignty. Commonwealth regulatory assessments are to prioritise proposals most closely aligned with the expectations, and energy-intensive proposals that are not aligned will not be prioritised. The expectations expressly cover large-scale AI compute centres, the "AI factories" Nvidia is describing, and they sit alongside existing law rather than replacing it.
That design is proportionate. It uses approvals leverage rather than a new statute, and it leaves the commercial structure to the market. It also gives the public a way to test Nvidia's announcement: which of the eight partners will show how their projects meet each expectation, and when?
The same plan describes GovAI, a centralised, Australian-based hosting platform for agencies. If government workloads are the sovereignty-critical ones, that is where hard requirements belong, in procurement terms that can be audited, rather than in a blanket demand that all private compute be domestically controlled.
What a proportionate response looks like
First, ask for disclosure rather than prohibition. Operators seeking priority treatment under the expectations should publish site-level timetables, power sourcing and water use. Vague gigawatt totals should not earn favourable assessment.
Second, tie sovereignty claims to specifics. If a provider markets "sovereign" capacity to banks or agencies, it should state who holds encryption keys, who can administer the platform, and what happens if vendor support is withdrawn. Consumer-protection law already polices misleading claims, and the ACCC can enforce it.
Third, avoid compelling a particular vendor mix. Mandating domestic hardware that does not exist would raise costs, slow deployment and deter the very investment the country wants. Encouraging a multi-vendor environment, and keeping open the option of other accelerators, is cheaper insurance than local-content rules.
Fourth, treat energy as the binding constraint. A 2 GW target is only as credible as the grid connections and generation behind it, and that is where the government's "support the energy transition" expectation will bite.
The bottom line
Nvidia's announcement is a useful signal that global capital sees Australia as a place to host AI. But a press release with no capital figure, no per-site schedule and no named customer is a statement of intent. Australia should welcome the investment, hold the partners to the published expectations, and be candid that onshore compute running on a US stack buys resilience and jurisdiction over data, not independence.