Nvidia lines up eight Australian data centre partners for one AI factory architecture

The gigawatt figure is easier to audit than it first looks. IREN's Bundey campus in South Australia is an 800MW site ^[3], and CDC says it already operates more than 550MW across Australia and New Zealand with a further 800MW under construction ^[4]. That is 2,150MW of built or building capacity from two of the eight names, against a 2,000MW programme target for 2027 ^[1]. Those are whole-facility figures rather than AI-specific ones, and no capital figure is attached to the programme ^[5], so what the two gigawatts describes is a ceiling drawn around a buildout that was already running, with Nvidia's architecture named as the thing inside it.

The word factory carries the commercial weight. In Nvidia's usage an AI factory covers facilities, compute, networking, software and reference designs built around the DSX platform and the CUDA ecosystem ^[6]. Daniel Roberts of IREN said that building AI infrastructure at scale requires integrating every layer ^[14], which is the honest case for buying a stack from one supplier and also the entire risk in doing it. Eight landlords converging on one reference architecture makes a workload easy to move between the eight and awkward to move anywhere else.

Sovereign AI is the pitch, and for Australia there is something behind it: data residency rules, locally trained models, industries that would rather not send their data offshore ^[10]. What sovereignty means here is that the racks sit on Australian soil while the accelerators, the interconnect, the software stack and the reference architecture are American ^[11]. TNW's read is that this improves on shipping the data to Virginia without amounting to independence ^[12], and that the distinction lands differently in Canberra with Washington already leaning on Australia over its platform regulation ^[13].

The demand side of the list is thin. Six of the eight partners are landlords or connectivity providers rather than model builders, and the only customer-side voice in the announcement is a healthcare startup describing what it can do with an Nvidia model ^[15]. TNW reads that as capacity committed well in advance of the demand that would justify it, on the assumption the demand arrives ^[16]. The balance sheets underneath are already visible: NextDC has been issuing convertible notes to fund this class of capital expenditure, and AirTrunk sits at the centre of American money going into Australian data centres ^[18].

For anyone who has to sign for capacity, two axes decide it. First, does the data have to stay onshore as a matter of law or client contract. Second, can the workload leave CUDA without a rewrite. Onshore and portable is the quadrant where a list of eight is worth something, because they can be played against each other on rack date and price while an exit stays open. Onshore and locked means the premium is the product, and the two things to press for are a per-site delivery date and contracted firm power, neither of which the announcement provides ^[9]. Offshore and portable can wait for the price to come down. Offshore and locked was settled before this week.

Two gigawatts lands in a grid whose operator has spent several years warning about firm capacity in the eastern states ^[7]. The commitment comes from a chipmaker that builds silicon and software, not power stations, and the announcement said nothing about where the electricity comes from ^[8]. The eight names give Australian buyers a choice of floor space and a single choice of stack.