Nvidia's 2GW Australia target has no buyer attached

Nvidia has signed eight Australian companies to a target that reads like a national program: 2GW of compute online by 2027. Firmus, Sharon AI, Iren, ResetData, Megaport, CDC, NextDC and AirTrunk will expand land, power and shell capacity to host Nvidia infrastructure, operating the resulting AI data centers on the chip designer's DSX platform.

What the coalition does not carry is a customer: the coverage names no offtaker for any of the eight sites, no anchor tenant behind the 2GW, and no generation contract underwriting it, as land, power and shells arrive as categories of capacity to be expanded rather than assets the partners hold today. Read strictly, that is a supply-side alliance assembled ahead of contracted demand, which makes this an ecosystem announcement with a platform, rather than a campus, as the thing actually being sold. Firmus had a $300 million deal announced on September 4 and ResetData one on September 9, so the coalition is being put together in the middle of an active stretch of Australian AI dealmaking.

Several partners brought specifics: Sharon AI is deploying up to 68,000 Nvidia GPUs on the DSX platform, and Firmus is expanding its Project Southgate initiative. Iren is pairing the DSX AI factory architecture with its own power, land, data centers, GPU deployment and infrastructure operations, work that co-founder and co-CEO Daniel Roberts ties to the company's 800MW Bundey campus in South Australia — the only site-level nameplate the announcement attaches to a partner. Megaport's Latitude.sh subsidiary will widen access to Nvidia compute, while CDC and NextDC appear on the partner roster without the coverage describing their part in it.

Nvidia's framing is aimed past the operators, and its own executive describes the blueprints as the product: "AI factories turn energy into intelligence — the essential resource of the AI economy," said Raj Mirpuri, the company's vice president of global AI clouds and infrastructure ecosystem, describing DSX as a full-stack platform compatible with the CUDA ecosystem, enhanced by software over the life of the infrastructure, and so "more productive, fungible and durable — and a new investable asset class." The pitch to its Australian partners is that capacity built to this blueprint can carry several generations of Nvidia compute and widen local access to accelerated computing and the Nemotron open models.

Investable asset classes need cash flows, and that is the claim to test. The US version of this story became financeable at the point a credit anchor appeared: Nvidia's guarantee, more than its equity, is what would make the $500 billion Ohio campus behind OpenAI bankable, and Nebius has already converted an Nvidia stake into $5.75bn of debt at a four-point spread. Australia's roster has no comparable tenant, and hyperscaler anchors set the infrastructure price while everything without contracted cash flow trades as a merchant shell until proven otherwise — a logic that bites harder where a single number is the only thing resembling a commitment.

Power is the other unpriced line: the partners commit to expanding power capacity, but the coverage attaches no interconnection position, no generation asset and no energy contract to the 2GW figure, and grid access is now the underwriting variable for this asset class — the queue position and the permit carry more value than the rack, and neither is disclosed here. The coalition has promised the number without showing the supply.

Roberts is selling the blueprint itself: "Combining Iren's vertically integrated platform with the Nvidia DSX AI factory reference architecture provides a repeatable blueprint to bring capacity online at scale," he said, adding that the company is applying that blueprint across its development portfolio. Repeatability is what makes a neocloud financeable rather than speculative, and it is the piece of this announcement that does not depend on the 2GW target being hit.

The physical build is already the harder half: Nvidia says the data center operators among its partners have made their facilities suitable for high-density deployments, and AirTrunk is expanding AI-ready powered shells across the region with direct-to-chip liquid cooling built for those workloads. Chief customer and innovation officer Damien Spillane lists location, capacity, cooling, power and connectivity as the things AI is reshaping in Australian hyperscale — the line items that decide whether a shell can host a high-density deployment at all. Nvidia's platform promise is software that improves the asset over its life; the coolant loop and the powered shell have to exist before the first GPU is racked, and their cost lands on the operators.

The announcement to watch from here is the first named counterparty attached to one of these eight sites — a tenant, an offtake contract, a generation deal — because that is the point at which Australian AI capacity starts pricing as infrastructure rather than as an option on Nvidia's roadmap. Until then the operators financing land, shells and coolant loops are carrying the construction risk for an ecosystem whose demand has yet to be contracted, and the partner that publishes a contract first will be the one whose capacity the market can actually underwrite.