Nvidia Says Up to 2 Gigawatts for Australia by 2027. 44% Is Dated 2028.

Fourteen terawatt-hours. That is what 2 gigawatts of AI factories consume in a year at 80 percent load, enough for about 2.5 million Australian homes. No coverage we could read of Nvidia's September 9 announcement mentions it: eight partners in Australia expanding "land, power and shell capacity", meaning empty buildings, for up to 2 gigawatts of Nvidia-based AI factories by 2027. The National Electricity Market delivered 178 terawatt-hours in 2024-25. This load would be 8 percent of it, on a 15-month horizon.

Coverage called it a doubling; on the Australian Data Centre Forecast Report's 1.5 gigawatts operational at the end of 2025, it is a 2.3-fold jump. Set the announcement beside the grid operator's August forecast and the partners' own published dates. Both put the load after 2027.

Half the Decade's Growth in Fifteen Months

In August the Australian Energy Market Operator published its 2026 Electricity Statement of Opportunities, whose central forecast has data centers growing from about 5 terawatt-hours a year now to 34 by 2035-36, a rise of 29 terawatt-hours over a decade. Nvidia's 2 gigawatts, at 80 percent utilization, would add 14 of those 29, nearly half the ten-year growth, inside fifteen months. Three readings fit. The forecast is low; or this is not energized load; or AEMO may have counted these projects and spread them across the decade, in which case "by 2027" is on the envelope, not the electricity.

Each row is one division: the announcement, or the queue it must pass through, against a published Australian figure.
Quantity	Number	Compared with	Result
2 GW, one year (our estimate)	14.0 TWh at 80% load	178 TWh NEM consumption, 2024-25 (AEMO)	7.9% of the grid
Same load vs forecast growth	14.0 TWh at 80% load	29 TWh AEMO data-center growth to 2035-36	48% of a decade of growth
Existing fleet plus announcement	1.5 GW + 2 GW	1.5 GW operational, end 2025 (ADC Forecast Report)	2.3x the existing fleet
Requests AEMO expects to be built	8.0 GW	66.9 GW of connection requests, June 2026	12% queue throughput

AEMO's queue is the second half of the story. In June it held 66.9 gigawatts of data-center connection requests across 225 projects, and its central case lets about 8 gigawatts through, roughly 12 percent; more than 40 percent of the projects tracked since 2025 have dropped out or gone backwards, and mature fleets draw only 45 to 51 percent of the capacity they contracted. Apply the queue-survival rate to 2 gigawatts and you get 240 megawatts, a queue-wide base rate, not a forecast for these eight; apply the mature-draw ratio instead and you get about 1 gigawatt, reached over most of a decade.

What the Partners' Own Dates Add Up To

Nvidia published no timetable, so we built one from the companies' own statements, and its largest piece is IREN's Bundey campus in South Australia, 800 megawatts and 40 percent of the total; IREN signed its transmission connection agreement in June and expects power from 2028.

Firmus, whose Project Southgate targets 1.6 gigawatts by 2028, planned up to 150 megawatts across Tasmania and Melbourne by mid-2026, a target now past with delivery unconfirmed, and its St Leonards site in Tasmania has a supply agreement with Aurora Energy for up to 104 megawatts and a start date in early 2027; counting the two as separate gives 254 megawatts; as overlapping stages, 150. Sharon AI has announced up to 68,000 GPUs and no date, which at 600 GPUs per megawatt is 113 megawatts, or closer to 160 at the 424 per megawatt its earlier financing implied. NEXTDC's S6 Sydney is 13.5 megawatts of IT load and its S5, in development, is 80 megawatts-plus. ResetData's 72 megawatts of dedicated generation, secured by Centuria, is expected from early 2028.

Each row is one partner's published capacity and date, where published; 872 MW, 44 percent of the 2 GW, is dated 2028.
Partner	Publicly stated piece	Public date	Counted (MW)	Excluded, with reason
IREN	Bundey, SA: 800 MW	Power from 2028	0	800 MW, dated 2028
Firmus	Southgate 150 MW; St Leonards 104 MW; 1.6 GW by 2028	Mid-2026 target; early 2027; 2028	150 to 254	1.6 GW target, subsumes counted stages
Sharon AI	Up to 68,000 GPUs (113 to 160 MW)	None	0	113 to 160 MW, no date
NEXTDC	S6 13.5 MW IT; S5 80 MW+	S6 operating; S5 in development	13.5	80 MW, no date
ResetData	72 MW dedicated generation (Centuria)	Early 2028	0	72 MW, dated 2028
CDC	>550 MW operating (Australia and NZ); ~800 MW in development	None for AI factories	0	~800 MW in development, no AI-factory date
AirTrunk	>1.2 GW across five campuses; DGX-Ready	None for AI factories	0	>1.2 GW operating, no AI-factory date
Megaport	Network provider	n/a	0	No capacity

Counted: megawatts with a company-published start year of 2027 or earlier. Excluded does not sum: 872 MW is dated 2028, the rest undated. Firmus's 1.6 GW target subsumes the counted stages and is in no total. Sharon AI's megawatts and the Firmus range are our arithmetic.

Add the pieces with public dates before the end of 2027, Firmus and S6, and you get roughly 165 to 270 megawatts, 8 to 13 percent of the target; add Sharon AI and S5, sized but undated, and the ceiling is about 360 to 510 megawatts, a quarter. The pieces dated 2028, IREN and ResetData, are 872 megawatts, 44 percent. Everything else is real capacity with no AI-factory date on it.

The South Australia Problem

Bundey matters because of where it is. South Australia's operational demand peaked at 3,274 megawatts in February 2025, by AEMO's 2025 South Australian Electricity Report, so an 800-megawatt campus is 24 percent of that peak. Rooftop solar there met 99.9 percent of demand at 1:30 pm on August 31, a winter record. The energy exists at noon and an AI factory runs at midnight. IREN's connection agreement provides four 330-kilovolt feeders without network upgrades, not 800 megawatts of firm overnight supply, and no statement we could read names a supplier.

What It Costs, and Who Gets Paid

If all 2 gigawatts are built, this announcement is worth $50 to $80 billion of chips to Nvidia, which told analysts in August that a gigawatt of Blackwell is about $25 billion of revenue and one of Vera Rubin about $40 billion; on the dated-by-2027 share, that is $4 to $11 billion. Building it costs $70 to $120 billion on Bernstein's 2025 and Jensen Huang's per-gigawatt estimates.

What You Can Do

Readers of data-center announcements: convert gigawatts to terawatt-hours and divide by the grid, then ask for dates. "Up to" plus a year is a marketing envelope; a transmission agreement with an energization year is a fact. This deal contains exactly one of the latter. It says 2028.

Grid planners could track the eight partners by energization date against AEMO's two base rates. IREN holds the largest number and this deal's only transmission agreement, so its risk is the calendar year, while Firmus awaits a Hydro Tasmania decision on 450 megawatts; none of this is investment advice.

Limitations

All figures come from search summaries of the companies' releases, AEMO's 2026 forecast and press coverage, not from the linked documents, and no partner was asked for comment. Nvidia has not said whether its 2 gigawatts is facility power or IT load; if IT load, the terawatt-hour figures rise by roughly 10 percent. Eighty percent utilization is a modeling choice; at 60 percent the figure is 10.5 terawatt-hours. The 8 percent grid share counts ResetData's behind-the-meter generation as grid load; homes are counted at 5.5 megawatt-hours a year; the 1.5-gigawatt baseline is a forecaster's; and AEMO's ratios are queue-wide averages that eight well-capitalized partners may beat.

The Strongest Case Against This Analysis

Nvidia never promised 2 gigawatts of energized load in 2027; it described partners expanding land, power and shell capacity for up to that much, which is a pipeline claim, and pipelines are meant to exceed forecasts. AirTrunk and CDC already operate halls that need only retrofitting, Firmus has contracted 600 megawatts of South Australian supply, and AEMO's 12 percent is dragged down by speculative applicants, while these are the eight best-capitalized partners in the country and the forecast this article uses as its denominator is the thing such announcements are designed to move. That argument is right about what was promised and wrong that the arithmetic is unfair: the partners' own statements put 44 percent of the total in 2028, the retrofit path still needs the same megawatts from the same grid, and the forecast is the operator's central case, built from a June queue thirty times the size of this announcement. The behind-the-meter point partly lands, though the ledger already sets ResetData's 72 megawatts aside.

The Bottom Line

Two gigawatts is a real number. Its date is a ceiling, not a commitment. That load would be half of the grid operator's decade of data-center growth, and 44 percent of the total is dated 2028 by the partners themselves. Australia may well get its 2 gigawatts, on the schedule the transmission agreements set.