Nvidia’s 2GW Australia Build Puts Operators on the Hook

Nvidia is working with eight Australian operators on up to 2 gigawatts of AI factory capacity by 2027, a build equal to 125% of the country’s current computing load. On September 10, the shares fell 2.4% to $218.21. The release sold land, power and shell, not a chip order, and the people who have to deliver those three things are not in Santa Clara.

Eight Australian Partners Are Building Toward 2GW

On September 9, 2026, Nvidia said it is working with Australian cloud partners and infrastructure firms to expand land, power and shell capacity for multiple generations of NVIDIA DSX AI factories, with up to a 2-gigawatt buildout by 2027. The named group is Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC and AirTrunk. Those firms will run the sites. Nvidia is supplying the DSX platform, accelerated computing, networking, software and partner support.

Data Centres Australia, citing DC Byte, puts the country’s current computing capacity at 1.6 GW. Two gigawatts is 125% of that load. If the full target is added rather than substituted, the stack would rise to 3.6 GW, more than double what is in place now. Nvidia did not attach a capital figure, a GPU purchase volume for the whole programme, or a site-by-site switch-on calendar to that 2 GW line.

Raj Mirpuri, vice president of global AI clouds and infrastructure ecosystem at Nvidia, put the pitch in energy terms. He said AI factories turn energy into intelligence, and that DSX is a full-stack platform compatible with CUDA software and updated over the life of the plant. Oliver Curtis, cofounder and co-CEO of Firmus, called Project Southgate a tangible commitment to building Australia’s AI future and said the work is meant to serve demand at home and across Asia-Pacific.

CDC, IREN and Sharon AI Hold the Concrete Commitments

Only some of the eight names came with a number attached. Sharon AI, a Nasdaq-listed Australian cloud firm, is deploying up to 68,000 NVIDIA GPUs with Quantum InfiniBand and Spectrum-X Ethernet, using DSX for startups, enterprises and research groups. James Manning, cofounder and CEO of Sharon AI, said Australia has the energy, connectivity and ambition to lead in AI, and that the potential becomes real only when organisations can reach that compute without giving up security or sovereignty.

IREN is applying the DSX reference design across its development book, including an 800-megawatt Bundey campus in South Australia. Daniel Roberts, cofounder and co-CEO of IREN, said building at this scale means joining data centres, compute and software, and that the DSX design is a repeatable blueprint for bringing capacity online. CDC, the Infratil-owned operator, already runs more than 550 megawatts across Australia and New Zealand, with a further 800 megawatts under construction. Greg Boorer, founder and CEO of CDC, said those halls use 100% renewable electricity and zero-water cooling certified for Nvidia gear.

THE EIGHT FIRMS ON THE 2GW LIST

The rest of the list is thinner on figures. Megaport is widening access to Nvidia compute through Latitude.sh so inference sits closer to users. ResetData is building for customers that need Australian data residency; Marcel Zalloua, joint CEO, pointed to work with ProCan at the Children’s Medical Research Institute. NEXTDC, through chief customer and commercial officer David Dzienciol, said it is rolling liquid-cooled, high-density plant, including DSX halls linked to the large public clouds. AirTrunk, through Damien Spillane, chief customer and innovation officer, is adding powered shells and direct-to-chip cooling. Those are construction and interconnection jobs first.

Australia’s Grid Already Holds 66.9GW of Requests

Oxford Economics Australia, in a June 2026 forecast prepared for the Australian Energy Market Operator, estimated rated data-centre capacity at 2.2 GW in FY26, with about 200 MW added over the prior year. Rated capacity is the maximum draw written into a connection agreement, covering IT load and the building around it. Average utilised draw sat near 580 MW in FY26, because leased halls are not always full, new rooms ramp for years, and operators keep spare plant for uptime. Data centres accounted for 2.8% of Australian electricity use in FY26, up from 1.3% five years earlier.

AUSTRALIA’S CAPACITY STACK

• Computing load now: Data Centres Australia, citing DC Byte, puts current computing capacity at 1.6 GW, the base for Nvidia’s 125% comparison.
• Rated grid contracts: Oxford Economics and AEMO put rated data-centre capacity at 2.2 GW in FY26.
• What the fleet actually draws: Utilised load averaged about 580 MW in FY26, well below both paper figures.
• The request pile: As of June 2026 there were 66.9 GW of requests to connect at mixed stages of maturity.

That 66.9 GW pile is a leading indicator, not a build schedule. Oxford Economics warns it should not be read as the load that will show up on the grid. AEMO’s about 36% of the 2025 project list was later cancelled, including more than 30% of projects that network businesses had classed as committed, and 225 projects sat in the 2026 connection process. Proposed connection capacity had risen from 38 GW the year before. The bottleneck has moved from chips to electrons, and a 2027 date has to live inside that queue, not beside it.

Under AEMO’s central Step Change case, data-centre electricity use on the National Electricity Market rises from 5.1 TWh in FY26 to 15.5 TWh by FY30 and 33.7 TWh by FY36. New South Wales and Victoria account for over 85% of that NEM load through FY36, reaching about 12% of each state’s grid-supplied electricity by FY30 and 19% by FY36. Near-term growth is still mostly existing sites ramping and expanding. Oxford Economics says new centres become the main driver only after FY30, which is after Nvidia’s 2027 target. CommBank, in an August 2026 note, put about 6 GW of potential capacity in the national pipeline and estimated the wider Australian build at around $150 billion by 2030.

What Nvidia’s DSX Platform Puts in Those Halls

DSX is how Nvidia turns eight separate landlords into one kind of plant. The company’s product page describes it as a platform that unifies design, simulation, operations and partner tools to help builders run AI factories optimized for lowest token cost. In Nvidia’s usage an AI factory is a site built to turn power and data into model output, rather than a hall assembled from mixed parts. The Australia note says the same stack will host multiple generations of Nvidia compute, which is the commercial point: the building is specified around one vendor’s boards, networks and software.

Through our partnerships with Australian NCPs, this infrastructure can support multiple generations of NVIDIA compute and expand local access to accelerated computing and NVIDIA Nemotron open models, helping Australian innovators build AI at home.

Raj Mirpuri, vice president of global AI clouds and infrastructure ecosystem at NVIDIA, September 9, 2026 announcement

DSX PIECES THAT REACH THE GRID

• Reference design: Validated layouts from chip and cluster through networking, storage and the building.
• DSX Sim: Tools to find bottlenecks and test tradeoffs before steel goes up, including digital twins of the plant.
• DSX OS: Open, modular software for bringing halls online, keeping fleets healthy and running multi-tenant loads.
• MaxLPS: Power software that Nvidia says can provision up to 40% more GPUs inside the same megawatt budget.
• DSX Flex: Grid-facing controls that take load-shedding, demand-response and price signals and shift work, on-site renewables and storage in reply.

Flex is the piece that admits the hidden party. If Bundey, Project Southgate and the CDC rooms are going to land on an eastern grid already crowded with requests, the software has to talk to the network, not only to the GPUs. MaxLPS is the other tell: Nvidia is selling a way to put more boards in a fixed power cap because the cap, not the board, is what Australia is short of. Heidi, a clinical software firm, and Atlassian are the two Australian product names Nvidia put on the workload side, both using Nemotron open models. They are proof of local software demand. They are not 2 GW of paying training load.

Data Center Sales Already Supply 92.5% of Nvidia Revenue

The Australia note arrived two weeks after Nvidia reported fiscal second-quarter results for the period ended July 26, 2026. Revenue was $96.2 billion, up 106% from a year earlier and 18% from the prior quarter. Data Center revenue was $89.0 billion, up 117% year on year, or 92.5% of the company. GAAP gross margin was 75.0%. GAAP earnings were $2.46 a diluted share. Guidance for the third quarter is $108.0 billion, plus or minus 2%, with no China data-centre compute in that outlook.

Jensen Huang, Nvidia’s founder and chief executive, said AI has reached its inflection point, that its tokens are productive and profitable, and that compute is now revenue. That is the standard the Australia headline has to meet. A country-level DSX programme counts when it turns into racks that bill. Until then it sits in the same bucket as other sovereign build stories Nvidia has grouped around the same platform, including work in Korea and Japan named on the August 26 call. Those programmes advertise the playbook. They do not, on their own, fill a $108.0 billion quarter.

Huang’s line also explains the tape. At $96.2 billion a quarter, 2 GW of Australian shell is a future conversion problem, not a print. The partners still have to pull power, pour halls and then buy the boards. Some of that buying is already visible in Sharon AI’s 68,000-GPU figure. Most of the 2 GW target is still a construction and connection plan.

Shares Fell 2.4% on a Target Without Chip Orders

Nvidia shares traded at $218.21 on September 10, down $5.46, or 2.4%. The move followed a release that said the quiet part in the first bullet: the collaboration is to expand land, power and shell. Investors who wanted a purchase order did not get one. The people with skin in the next 16 months are the eight operators who must finish rooms, and the network businesses and AEMO staff who must connect them, in a country where utilised data-centre draw is still about 580 MW.

WHAT WE KNOW

• The names: Eight Australian firms are on Nvidia’s list, and they will operate the plants.
• The date: The published target is up to 2 GW by 2027, equal to 125% of the 1.6 GW computing load now.
• Hard figures inside the list: Sharon AI has up to 68,000 GPUs; IREN has named an 800 MW campus; CDC has more than 550 MW in service in Australia and New Zealand and 800 MW underway.
• The grid maths: Rated capacity is 2.2 GW, requests are 66.9 GW, and a large share of last year’s list did not survive.

WHAT IS UNCONFIRMED

• GPU volumes: No programme-wide board count, partner purchase commitments or delivery schedule was disclosed.
• Money: No capital figure was put on the 2 GW target as a whole.
• Electrons: Nvidia did not say which generation plants, which substations or which new wires feed the 2027 date.
• Paying load: Heidi and Atlassian are named users of Nemotron tools, not contracted occupiers of 2 GW.

Power is now the scarce input, and the 2027 clock runs inside a connection process that already chewed through about 36% of the previous year’s list. Compute follows that power to places such as Bundey, where IREN already holds an 800 MW campus, and to Firmus halls that still need a live feed. A 2 GW DSX map is a real design win for Nvidia if those rooms switch on. It stays a slide until they do.

Disclaimer: This article is news reporting and analysis of a company announcement, electricity-market figures and a share-price move, and it is for information only. It is not investment advice, a recommendation to buy or sell Nvidia or any partner stock, or a forecast of returns. Readers who are considering a trade should consult a licensed financial adviser who can review their own position, time horizon and risk limits. Revenue, capacity, queue and price figures reflect the company statements, Oxford Economics and AEMO material, CommBank estimates and market prints cited here and can change as later filings, connections and trading sessions arrive.