Nvidia Is More Than Doubling Australia's Compute
Eight local partners, up to 2 gigawatts of AI factory capacity by 2027, against a national installed base of 1.6 GW. Nvidia supplies the platform. Someone else supplies the land, the power, and the balance sheet.
Nvidia announced on September 10 that it is working with eight Australian cloud and infrastructure partners to expand land, power, and shell capacity for NVIDIA DSX AI factories — up to a 2-gigawatt buildout by 2027.
The partners: Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC, and AirTrunk.
Australia's current total data centre capacity is roughly 1.6 GW. The announced expansion is larger than everything the country has built to date.
Read the structure, not the gigawatts
The number that gets quoted is 2 GW. The number that matters is who pays for it.
Nvidia's contribution is the DSX platform — accelerated computing, networking, software, and ecosystem support. The eight partners contribute land, power contracts, shells, construction capital, and operations. The partners own and operate the facilities.
This is the shape Nvidia has been repeating all year, and it is worth naming precisely: Nvidia is not building AI infrastructure. Nvidia is underwriting demand certainty for people who build AI infrastructure, in exchange for those facilities being Nvidia-shaped from the foundation up.
A data centre designed around DSX — power density, cooling topology, rack layout, network fabric — is not trivially convertible to someone else's accelerator later. The moat is not the GPU. The moat is a decade of concrete poured to Nvidia's dimensions.
Two gigawatts is an energy story
Australia is an unusual choice on compute-demand grounds and an obvious one on energy grounds.
The country has abundant solar and wind, large amounts of empty land near transmission, and a grid that has spent fifteen years learning to absorb intermittent generation. It also has a national appetite for anchor industrial loads to justify transmission investment.
But 2 GW is not a rounding error on the National Electricity Market. It is a load comparable to a significant fraction of a state's demand, arriving in a compressed window, in a system already managing a coal retirement schedule. The interconnection queue, not the silicon supply, is the binding constraint on this plan — and interconnection is where every ambitious data centre buildout of the last three years has actually slipped.
The presence of IREN in the partner list is a clue about how that gets solved. IREN came out of bitcoin mining, which means it already owns the skill the AI buildout is short of: siting large, flexible, power-hungry compute next to generation and negotiating the offtake. Several of the fastest AI capacity additions globally have come from mining companies converting rather than developers greenfielding, for exactly this reason.
Sovereign AI, restated
Nvidia's framing here is the same one it has used in Korea, Japan, India, and the Gulf: nations need domestic AI compute, and Nvidia will help assemble it. Jensen Huang has spent 2026 arguing for AI infrastructure in every country.
The argument is genuinely sound. Latency, data residency, regulatory jurisdiction, and industrial policy all point toward local capacity. Australian researchers and startups training on Australian infrastructure is better than the alternative on every axis except cost.
It is also, viewed from Santa Clara, a demand-diversification strategy of unusual elegance. Hyperscaler capex is concentrated in perhaps six buyers whose spending is now large enough to be politically and financially scrutinised. Sovereign programs add dozens of buyers whose purchasing logic is national rather than ROI-driven, and who are far less likely to negotiate hard or defect to a competitor mid-cycle.
Both things are true at once. Australia gets real capacity it would not otherwise have. Nvidia gets a customer base that is structurally sticky and politically insulated.
What Australia should actually check
Three questions determine whether this lands as capacity or as a press release.
Is the power contracted or aspirational? "Land, power and shell capacity" covers a wide range of commitment. Signed PPAs and connection agreements are capacity. Options on sites are intent.
Who fills 2 GW? Australia's domestic AI demand does not currently consume this. The realistic fill is a mix of local enterprise, university and research load, and — most of it — international customers renting Australian compute. That is a fine business, but it makes the facilities an export industry subject to global price competition, not a sovereign capability.
What happens at the next platform transition? DSX facilities built for 2027 silicon face the same question every generation: does the next Nvidia rack drop into the same shell, or does the density jump force a retrofit? Nvidia says the capacity is designed to host multiple generations. That claim is the entire depreciation schedule.
The read
Nvidia is systematically converting national energy endowments into GPU demand, and doing it with other people's capital. Australia has sun, land, and transmission ambition; it now has a plan to more than double its compute base with eight partners carrying the construction risk.
If the power contracts are real, this is one of the more sensible sovereign programs announced this year — Australia has genuine structural advantages here, unlike several countries that have signed similar deals on the strength of ambition alone.
The gigawatts are the headline. The offtake is the business.