An announcement naming eight companies and one round capacity figure invites one question, and it is the wrong one. How much capacity is Nvidia building in Australia? None: the partners build and operate. So the question it avoids is what Nvidia has committed.
On 9 September Nvidia said it is working with eight Australian cloud and data-centre operators "to expand land, power and shell capacity designed to host multiple generations of NVIDIA DSX AI factories … with up to a 2-gigawatt buildout by 2027." Two of the eight named a megawatt figure. Neither belongs inside that window, by its own company's words, and the sum now circulating against the headline, which adds them to 2,150MW and concludes the 2GW target must be a floor, is built on exactly that mistake.
Five of the eight named no capacity figure at all
Count what each partner actually put a number to, as the table above does off the eight quotes, and the release thins fast. IREN and CDC gave megawatts. Sharon AI gave a GPU count. Firmus, Megaport, ResetData, NEXTDC and AirTrunk gave no capacity figure at all. No dollar, no term, no exclusivity and no committed GPU volume appears anywhere, from anyone.
The 2,150MW sum does not survive either of its two terms
IREN's 800MW plus CDC's 550MW plus CDC's further 800MW is 2,150MW, which is 7.5% above a 2,000MW headline, so the headline must be a floor and the programme bigger than advertised. Both terms fail, and the primary documents say why.
CDC's figures are CDC. Greg Boorer's quote reads: "CDC operates more than 550 megawatts of capacity across Australia and New Zealand, with a further 800 megawatts under construction." A founder is describing his firm's estate: two countries, all workloads, operating plus construction. Nothing says how much of that 1,350MW is a DSX AI factory, how much is Australian, or when any of it lands. An estate figure inside a programme sum adds capacity that existed before the programme.
IREN's 800MW is outside the window. Daniel Roberts is quoted applying the DSX blueprint "across our development portfolio, bringing IREN's proven international delivery and operating experience to our 800-megawatt Bundey campus in South Australia." IREN announced Bundey itself on 3 June, with a transmission agreement securing four 330kV feeder exits, and told the market it was on track to commence energisation from 2028. Energisation is the campus first having power, not the moment a GPU earns revenue. A target closing in 2027 cannot contain it.
Strip both and nothing is left. Across eight named partners, not one attaches a megawatt figure to a 2027 date anywhere in the release. Firmus is the fairest test of that, because Firmus has the in-window capacity: its October 2025 Project Southgate announcement put a 150MW first stage in Tasmania and Melbourne, 54,000 GB300s, at mid-2026, inside 1.6GW through 2028. Those are dated megawatts of the kind this target needs, and they are not here. Oliver Curtis's quote names Project Southgate and no number.
For IREN, both halves of this were already on the record
Nvidia and IREN announced their partnership on 7 May, intending to support "up to 5 gigawatts of NVIDIA DSX-aligned AI infrastructure across IREN's global data center pipeline over time", with deployments "expected to focus on IREN's 2-gigawatt Sweetwater campus in Texas" as the flagship, alongside a five-year right for Nvidia to buy 30 million IREN shares at $70 — $2.1 billion if it exercises. That release names North America, Europe and APAC, and no Australian site. Bundey arrived four weeks after it.
So 9 September restates both halves and joins them: the DSX relationship from May, the campus from June, and Bundey named for the first time as somewhere the blueprint goes. That is new packaging around an existing agreement and an existing site — not a new Australian plan, and not IREN undertaking to put 800MW into the ground by 2027. Nvidia's target is what eight operators might collectively reach; IREN's own number is one campus, and its own date is 2028.
Whether 2GW is a ceiling or a floor is the undisclosed part
"Up to a 2-gigawatt buildout" can mean two opposite things. One is a ceiling on DSX-ready shell that eight operators might collectively reach if demand and financing cooperate. The other is a floor Nvidia is setting on its own attach rate in a country it wants to anchor. Nothing in the release settles it, and every phrase pushes the commitment onto someone else's balance sheet: "up to", "designed to host", "the participating providers will operate".
When Nvidia stands behind a gigawatt it has a house format for saying so — a guarantee, a term, a GPU count, a revenue figure, filed with a regulator — and it used that format two weeks earlier.
Ohio from Nvidia's Q2 FY2027 filing; Australia from the 9 September release
| What is stated | Ohio, PORTS-Pike | Australia |
|---|---|---|
| Capacity | ~4.25 GW | Up to 2 GW |
| Nvidia guarantees | $105.0bn | None disclosed |
| Lease term | 20 years | None stated |
| GPUs named by Nvidia | ~1.5m per generation | None |
| Nvidia revenue named | $150-200bn | None |
| Where it is disclosed | Quarterly filing | Press release |
Ohio is the PORTS-Pike campus in Nvidia's quarterly disclosure filed 26 August: $105.0bn of guarantees entered in August 2026 covering approximately 4.25GW, the site expected to host Nvidia infrastructure under 20-year leases to OpenAI, and Nvidia's own statement that each generation deployed there could be about 1.5 million GPUs, or $150-200bn of Nvidia revenue. The same disclosure carries a separate $3.5bn of land, power and shell guarantees for AI-cloud partners across every partner and geography; nothing identifies whether any of it relates to Australia. The Australia column is the absence of a figure, not a figure of zero - the release may sit inside that $3.5bn line indistinguishably, or outside it entirely.
Nvidia's July-quarter disclosure put $105.0bn of guarantees, entered in August 2026, behind roughly 4.25GW at the PORTS-Pike campus in Ohio, under 20-year leases to OpenAI, and said each generation deployed there could be about 1.5 million GPUs, or $150–200bn of Nvidia revenue. We have set that guarantee book against Alphabet's separately.
The same filing carries the line that matters most here: $3.5bn of land, power and shell guarantees for AI-cloud partners. Nvidia already reports a category for precisely the thing the Australian release describes. Australia either sits inside that $3.5bn indistinguishably or outside it, and nothing published lets a reader tell which.
Priced at Nvidia's own rate, the whole ceiling is about one quarter
Nvidia publishes its own gigawatt-to-revenue conversion, so nobody needs to invent one. $150–200bn divided by 4.25GW is $35.3bn to $47.1bn per gigawatt per generation, which is the same basis as Huang's per-gigawatt figure from the August call: $40bn for Vera Rubin, out of a gigawatt he priced at $60bn all in, land, power and shell being the other third, the part explicitly not Nvidia. At those rates 2GW carries a ceiling of $70.6bn to $94.1bn of potential Nvidia revenue through to the end of 2027.
Nvidia's Data Center segment reported $89.0bn in the July quarter alone, up 117% year on year, and has guided $108.0bn of total revenue for the quarter now running. So the whole Australian target, at Nvidia's own rate, with every megawatt filled and every rack Nvidia's, comes to 0.79 to 1.06 of one current quarter, a fifth to a quarter of the annualised run-rate, none of it contracted. Two gigawatts of shell in one country does not move that line.
For IREN it is smaller and differently shaped: no 2026 megawatt and no 2026 dollar attach to any of it. Bundey is no more than 16% of the ">5GW" of power IREN says it has secured, in a business where AI cloud revenue first passed bitcoin mining only last quarter. What the release adds is reputational: worth something, not worth a number.
One filed figure would flip this
A dollar figure, a term or a committed GPU volume attached to the Australian programme in a filing rather than a release would settle the ceiling-or-floor question; moving Bundey's energisation into 2027 would rescue the sum's second term. Neither exists yet: IREN has published nothing since 3 June that restates the date.
Until one appears, the words carry their own reading. Eight operators have agreed to build shell that Nvidia's reference architecture can fill, Nvidia has committed nothing in public to filling it, and the two partners who brought megawatts brought them from somewhere else.
Nvidia's capacity target, partner list, operating language and all eight partner quotes come from its 9 September release, carried by the wires on 10 September; the count of which partners quantified anything is ours. Bundey's 800MW, the 330kV connection and the energisation-from-2028 wording are from IREN's 3 June release; the 5GW DSX intent, the Sweetwater flagship and the 30-million-share right from the two companies' 7 May partnership release. Project Southgate's figures are from Firmus's October 2025 announcement. Nvidia's Ohio guarantees, the land-power-shell line, segment revenue, the current-quarter guide and the per-gigawatt revenue figures are from its Q2 FY2027 disclosure and CFO commentary, IREN's from its June-quarter release. Every division and ratio is R40 arithmetic on those disclosures. Sharon AI is private; its GPU figure is its own plan, not an order.
