Eight names, one blueprint

Nvidia said on 9 September, in a release datelined Melbourne, that it is working with eight Australian data centre and cloud companies on what it calls AI factories, targeting “up to a 2-gigawatt buildout by 2027”. The partners are Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC and AirTrunk — a list that covers both the incumbent colocation operators and the newer AI-specific builders.

The common element is NVIDIA DSX, which the company describes as a full-stack AI factory platform spanning facilities infrastructure, computing, networking, software and reference designs, compatible with the CUDA ecosystem. In practice that means Nvidia is selling a design as much as it is selling silicon: the partners build and own the sites, and DSX standardises what goes inside them.

Tower cranes over a large construction site
Announced capacity and energised capacity are separated by grid connections and planning approvals. Tom Van Dyck · pexels · Pexels License

The two specifics worth having

Most of the announcement is framework. Two numbers are not. Sharon AI is deploying up to 68,000 Nvidia GPUs, connected with Quantum InfiniBand and Spectrum-X Ethernet networking. And IREN is applying the DSX blueprint across its portfolio, including an 800-megawatt campus at Bundey in South Australia — which on its own is close to half the headline figure.

That matters because “up to 2GW by 2027” is a ceiling, not a commitment. Gigawatts of announced capacity and gigawatts of energised capacity are different quantities, separated by grid connections, transformer lead times and planning approvals, and the release does not break the total down by site or by year.

Energy is the constraint being asserted

“AI factories turn energy into intelligence — the essential resource of the AI economy,” said Raj Mirpuri, a vice president at Nvidia, in the release. IREN co-chief executive Daniel Roberts framed the same point operationally: building at scale, he said, requires integrating every layer — data centres, compute and software.

Fibre optic cables plugged into server racks
Sharon AI is deploying up to 68,000 Nvidia GPUs on InfiniBand and Spectrum-X networking. Brett Sayles · pexels · Pexels License

AirTrunk’s Damien Spillane said AI is reshaping hyperscale infrastructure in Australia, “driving new approaches to location, capacity, cooling, power and connectivity”. None of the three quotes names a power source or a grid agreement, which is the gap between this announcement and, for example, Google’s Finnish deal last week, where the electricity contract was the substance.

What to watch

The test is whether any of the eight publishes a site-level timetable with a grid connection date attached. Nvidia’s role here is to standardise the design and sell the hardware; the two-gigawatt number belongs to eight balance sheets and to the operators of Australia’s electricity network, not to Nvidia. Watch the Bundey campus and Sharon AI’s GPU deployment for the first concrete evidence that the ceiling is being approached.