Nvidia's '2GW' Australia Buildout Is a Chip-Sales Channel, Not Capex: The Eight Builders Own the Debt, Nvidia Books the Margin

On September 9, Nvidia said it would support "up to a 2-gigawatt" buildout of AI data centers in Australia by 2027, in partnership with eight local operators. The number reads like a national event, and on this continent it is: the country's largest data center operator runs just over half a gigawatt in service. But the announcement frames the deal in a way that invites you to picture Nvidia writing enormous checks. The details say the opposite.

The eight firms named — Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC, and AirTrunk — "operate the AI factories," supplying the land, power, and the physical shell. Nvidia delivers the DSX platform, the accelerated computing, the networking, the software, and ecosystem support. That split of labor is the whole story, and it is worth reading twice, because it determines who bears the cost and who books the revenue.

In this structure, the AI factories sit on the operators' balance sheets, financed with their own debt and equity. Nvidia books a sale — chips, network gear, software — each time an operator racks a GPU cluster, and it books that revenue up front, before a single watt flows, at gross margins near 74 percent. Nvidia's trailing capital expenditure, about $7.4 billion a year against roughly $127 billion in free cash flow, is the tell: the company sells the means of AI production rather than owning production. The builders carry the construction, refinancing, and power risk; Nvidia carries none of it.

So flip the headline. The "2GW" is forward revenue visibility for Nvidia's chips, not a Nvidia balance-sheet commitment — which is why both qualifiers, "up to" and "by 2027," are doing such heavy work. The scale is nonetheless real in context. Sharon AI alone says it is deploying up to 68,000 Nvidia GPUs at gigawatt-scale for enterprises, research, and government. IREN is building its 800-megawatt Bundey campus in South Australia. CDC already operates more than 550 megawatts across Australia and New Zealand, with another 800 under construction. Layering 2GW of new fabric on top of that installed base would roughly double the country's compute capacity on a short timeline — an "AI at home" wave built on locally hosted, sovereign compute.

The demand side is what makes Nvidia frame it as a partnership rather than a purchase. Nvidia calls its AI factories a "new investable asset class", a machine that turns "energy into intelligence," and it cites local builders like Heidi and Atlassian already using its open Nemotron models to build regional healthcare and search tools. The pitch to Australia is that its startups, universities, and enterprises no longer have to rent compute from U.S. clouds; they can build AI onshore. Nvidia profits either way, because it wins the sale whether the operator is an American hyperscaler or a Melbourne startup.

That leaves the risk exactly where the announcement places it: on the builders. Power is the binding constraint almost everywhere in this buildout, and several partners are pairing the data centers with new power-generation projects of their own. "Up to 2GW by 2027" is an ambition, not a signed and funded contract, and the operators are the ones who must finance it, obtain the grid connection, and deliver on time.

For an Nvidia holder, none of that moves the needle by itself — this is one datapoint in a global pattern of capital flowing into Nvidia-based AI factories, and Nvidia captures its margin regardless of which operator wins a given site. The structure is the durable lesson, and it cuts against the surface reading of the headline: Nvidia's own spending stays small while the debt that funds its growth piles up on its customers' books. That asymmetry, not the round 2GW figure, is what a reader should take from the announcement — and the number to watch going forward is not the headline capacity but whether the builders can actually fund and power it.