Tech and AI
Nvidia and eight Australian operators plan up to 2 gigawatts of AI factory capacity by 2027, as the grid operator forecasts data centre demand rising nearly sevenfold by 2036
By Staff Writer | 11 September 2026
The chipmaker announced on 9 September that it is working with Firmus, Sharon AI, IREN, ResetData, Megaport, CDC, NEXTDC and AirTrunk to expand land, power and building shell capacity in Australia for what it calls AI factories, with a buildout of up to 2 gigawatts by 2027. The operators will own and run the sites; Nvidia supplies the compute, networking and software. The national grid operator's August outlook puts data centre consumption on the eastern market at about 5 terawatt hours this year and 34 terawatt hours by 2035 to 2036, and says 5.4 gigawatts of large projects were already in the transmission connection queue at the end of March.
The announcement of Wednesday 9 September 2026 is not a single contract. It groups eight Australian cloud and infrastructure companies under one programme and one headline number: up to 2 gigawatts of capacity designed to host what Nvidia calls its DSX platform, a full-stack design covering facilities, computing, networking, software and reference designs. The participating providers will operate the sites. Nvidia delivers the platform, the accelerated computing, the networking, the software and what it describes as ecosystem support.
Who is building what
The release sets out what each operator brings. Firmus is expanding a programme it calls Project Southgate. Sharon AI is deploying up to 68,000 Nvidia graphics processors. Megaport is widening access through its subsidiary Latitude.sh, and ResetData is offering Nvidia computing on a cloud platform for customers that need data residency in the country. NEXTDC says it is developing liquid-cooled, high-density facilities, and AirTrunk is expanding what it calls AI-ready powered shells, including direct-to-chip liquid cooling.
For the engineering readership the two largest physical commitments are in the quotations. IREN's co-founder and co-chief executive, Daniel Roberts, placed the design at a single campus of 800 megawatts.
Building AI infrastructure at scale requires integrating every layer: data centers, compute and software. Combining IREN's vertically integrated platform with the NVIDIA DSX AI factory reference architecture provides a repeatable blueprint to bring capacity online at scale. We are applying that blueprint across our development portfolio, bringing IREN's proven international delivery and operating experience to our 800-megawatt Bundey campus in South Australia.
Daniel Roberts, co-founder and co-chief executive, IREN, 9 September 2026
Greg Boorer, founder and chief executive of CDC, said his company "operates more than 550 megawatts of capacity across Australia and New Zealand, with a further 800 megawatts under construction and additional growth planned to support world-leading, high-density workloads for critical infrastructure". Damien Spillane, chief customer and innovation officer at AirTrunk, said that "AI is reshaping hyperscale infrastructure in Australia, driving new approaches to location, capacity, cooling, power and connectivity".
What the grid operator says
The number that matters to anyone pricing or programming these sites is not the 2 gigawatts but the connection queue. The Australian Energy Market Operator's 2026 Electricity Statement of Opportunities, published in August, records that at the end of the March 2026 quarter eleven large data centre projects of more than 5 megawatts each, representing 5.4 gigawatts of maximum demand, were progressing through the transmission connection process on the National Electricity Market.
The same report forecasts data centre electricity consumption on that market rising from approximately 5 terawatt hours in 2025 to 2026 to 34 terawatt hours by 2035 to 2036, taking the sector from around 3 per cent to approximately 13 per cent of electricity supplied through the grid. The operator says it has changed its forecasting method for the sector to reflect how it behaves in practice: project attrition during the connection process, the gap between the capacity a project asks to connect and the electricity it actually draws, and the gradual ramp-up of demand over time. It also notes that around 15 gigawatts of coal and gas generation is scheduled to retire over the decade while consumption grows by more than 40 per cent.
Read together, the two documents describe a programme whose delivery risk sits in the substation rather than the server hall. Nvidia's release lists land, power and shell as the three things being expanded, in that order. Of the three, only power depends on a third party with a queue, and the operator's own figures show that queue already carries more than twice the capacity this programme promises.
Nvidia gave no total investment figure and named no date for any individual site beyond the 2027 horizon for the programme as a whole. Whether the eight operators can energise 2 gigawatts inside fifteen months is a question the release does not answer, and the grid operator's next quarterly connection data will answer it better than any further announcement.