The 2 GW AI Factory: How AI Infrastructure Is Reshaping Power, Cooling and Grid Planning
A 2 GW figure describes infrastructure intended to exist by 2027, not power being consumed today. It is a statement about connections, transformers and cooling plant. Almost all the electricity delivered leaves as heat, so cooling decides the practical limit. Part of this load is now being made interruptible.
What was announced
On 9 September 2026 NVIDIA announced a collaboration with Australian data centre operators to expand AI infrastructure capacity, describing a goal of up to a 2-gigawatt buildout by 2027. On 10 September CSIRO announced something quieter and, for an energy engineer, more interesting: a pilot that treats AI computing load as something the electricity network can ask to move.
Company and pilot statements, not measured outcomes
The infrastructure announcement names several operators and describes facilities using direct liquid-to-chip cooling and liquid-cooled high-density infrastructure. One operator is described as running more than 550 megawatts across Australia and New Zealand with a further 800 megawatts under construction, and one participant as deploying up to 68,000 GPUs. The 2 GW figure is explicitly a buildout target for 2027, not capacity in service.
The grid pilot, run with CSIRO and university validation, sorts AI work into tiers so that critical jobs are protected while elastic training work can be shifted. Its participants state that workloads could be optimised by 20 to 50 per cent within seconds of a grid signal, and that networks could connect additional data centre load if that load can be curtailed during the hours of peak stress, which they put at roughly 0.25 to 5 per cent of the year. These are the pilot's own figures, published at its launch, and no measured results have been released.
What a gigawatt figure states
Gigawatt announcements are reported as though they were consumption. They are not. A capacity figure describes what a site could draw once built and energised; consumption depends on how much gets built, when it is energised, and how hard it is worked. The distinction is the same one every plant engineer already makes between installed capacity and metered demand.
| Stage | What it describes | Evidence behind it | What it still does not tell you |
|---|---|---|---|
| Announced | An intention, with a date attached | A press release | Nothing is energised; plans change |
| Under construction | Steel, transformers and cooling being installed | Construction and equipment orders | Commissioning dates commonly move |
| Operational capacity | Connected and able to draw power | An energised connection agreement | Says nothing about how loaded it is |
| Actual demand | What is drawn, hour by hour | Metered energy over a period | The only figure comparable to generation output |
This also settles the comparison people reach for first. A 2 GW data centre programme is not "two nuclear reactors" or "a city". A generator's rating describes output it can supply; a data centre's rating describes load it can take. Comparing them is only meaningful once both sides state utilisation over the same period, and no such figure has been published here.
Where the energy goes
The physics is familiar to anyone who has sized a chiller. Electrical energy delivered to computing equipment does no lasting mechanical work. It becomes heat, essentially all of it, and that heat must be removed continuously or the equipment throttles and then fails. The cooling plant that removes it draws power from the same connection, which is why total site demand always exceeds the computing load.
For an industrial reader this reframes the whole subject. A large AI facility is a heat-rejection plant with computing equipment attached. Its constraints are the ones you already know: connection capacity, transformer rating, pump and fan power, approach temperatures, water use, and what to do with a continuous stream of low-grade heat.
Cooling method and heat grade
The cooling method named in these facilities matters for one reason above all: the temperature at which heat leaves the building. Moving air through a room rejects heat at close to room temperature, and heat at that grade is almost worthless. Bringing liquid directly to the chip removes heat at a higher temperature and at much greater density.
| Method | How heat is captured | Grade of heat leaving | Consequence for reuse |
|---|---|---|---|
| Air cooling | Heat leaves as warm air at room temperature | Lowest grade; usually rejected to atmosphere | Reuse needs a heat pump to be worth anything |
| Rear-door or in-row liquid | Heat captured closer to the rack | Moderate grade | Reuse possible for space heating in some climates |
| Direct liquid-to-chip | Coolant meets the hottest component | Highest grade available in this class of plant | Best chance of a usable reuse temperature |
Industrial energy practice and AI infrastructure meet at this point. A facility rejecting hundreds of megawatts of heat at a usable temperature could supply a district heating network, or process pre-heat where the temperature level matches the demand. Whether that is realised depends on a heat customer within economic pipe distance, on the return temperature that customer needs and on who pays for the connection. Those are siting questions decided years before any of this equipment is ordered.
Load that can move
The assumption that data centre demand is flat and non-negotiable has shaped grid planning for a decade. The pilot announced on 10 September tests the opposite: that AI work can be sorted into what must run now and what can wait or move, so the elastic part can be reduced when the network is under stress.
Training work is the obvious candidate because a job interrupted and resumed loses time rather than validity. Serving a live model is not: latency is the product. The reason this matters beyond data centres is that it establishes the precedent an industrial energy manager has argued for years, that a large load willing to be interrupted for a few hours a year should get a connection sooner than one that is not.
What this means for manufacturers
Expect competition for connections, not for electrons
Large new loads compete for grid connection capacity in the same regions. Where a site expects to electrify process heat or add compressors, start the connection conversation earlier than the equipment decision.
Read gigawatt claims as capacity
When an announcement affects your regional planning, ask which of the four stages in the table it describes. Announced and operational are years apart.
Treat flexibility as an asset you may already have
If a data centre can be paid to curtail, so can a plant with thermal storage, a flexible batch schedule or interruptible refrigeration. Establish which loads can move before a network operator asks.
Look at heat availability locally
A liquid-cooled facility nearby rejects heat continuously. Whether any of it is worth capturing depends on the temperature level, the distance, the cost of the connection and whether a heat pump is needed to reach a usable temperature.
Do not import data centre efficiency metrics
Facility-level ratios designed for computing loads do not describe a factory. Keep measuring energy per unit of production.
Limits and open questions
- Announced capacity is not built capacity. No public data shows how much of the 2 GW will be energised by 2027, and construction timetables commonly move.
- No utilisation figure exists. Without hours at load, no annual energy figure can be derived, and any such figure circulating is an assumption rather than a measurement.
- The flexibility figures are pre-result claims. The 20 to 50 per cent range and the 0.25 to 5 per cent of the year come from the pilot's own launch material. The pilot exists precisely because these have not been demonstrated at scale.
- Heat reuse is site-specific. Grade, distance and a willing customer decide it. Nothing in an announcement establishes any of the three.
- No cost or price data. Nothing published supports a statement about what this capacity costs to build, or its effect on regional electricity prices.
None of this capacity is in service today. The heat it will reject is the one certainty in the announcement, and which share of the load can be interrupted decides how much of it gets connected.
Questions
Does 2 GW mean two gigawatts are being consumed?
No. The figure describes infrastructure capacity intended to be built by 2027. Capacity is the size of the connection and the equipment. Consumption depends on how much of it is built, energised and loaded, and for how many hours. A capacity figure and an energy figure are different quantities and cannot be compared directly.
Can a data centre be compared to a power station of the same gigawatt rating?
Only if the comparison states utilisation. A power station's rating describes what it can generate; a data centre's rating describes what it can draw. Two plants with identical ratings deliver very different annual energy depending on how many hours they run at what fraction of capacity.
Where does all that electricity go?
Almost all of it becomes heat. Electrical energy delivered to servers is converted to low-grade heat that must be removed continuously, and the cooling equipment that removes it draws further power. This is why cooling design, not chip supply, often sets the practical limit on a site.
Is direct liquid-to-chip cooling different from ordinary data centre cooling?
Yes. Bringing liquid to the chip removes heat at a higher temperature and at far greater density than moving air through a room. That raises the temperature at which heat leaves the building. Temperature is one condition for reuse. Distance to a heat customer, the load profile, the return temperature required and the cost of any heat pump decide the rest.
Why would a data centre reduce its load on request?
Because grid capacity is scarce and connection queues are long. If a load can be reduced during the small number of hours when the network is under most stress, the network can connect more load overall. The pilot described here tests that premise.
Sources
- NVIDIA: AI infrastructure capacity expansion with Australia's data centre ecosystem. Company announcement. Published 9 September 2026. Checked 12 September 2026.
- CSIRO: Australia-first pilot to turn data centres from static loads into grid-aware assets. Research organisation announcement of a pilot. Published 10 September 2026. Checked 12 September 2026.
Suggested citation: Inzonex Research (2026), The 2 GW AI Factory: How AI Infrastructure Is Reshaping Power, Cooling and Grid Planning, published 12 September 2026. Capacity, cooling and flexibility figures are attributed to the organisations that published them. The stage table, the cooling-grade table and both diagrams are Inzonex analysis.