How much AI can a power station support?
A 500 MW power station does not translate into 500 MW of AI computing. Some output may serve other customers. Cooling and electrical infrastructure take a share. The supply must also be available when the data centre needs it. A simple worked example shows where those boundaries matter.
The next question is when power is available
On 10 September 2026, CSIRO published a partner release describing an Australian pilot that adjusts data centre power and AI workloads to grid conditions. The release reports that FlexSysAI is operating on an H200 cluster at ResetData's AI-F1. It presents flexible demand as a way to make better use of available grid capacity. Read the CSIRO announcement.
This is a pilot announcement, not evidence that every AI workload can be interrupted or that a particular saving has been achieved. Its practical relevance is the change in question: how much computing can run within the power available at a given time?
The wider pressure is documented in the IEA's 2026 Key Questions on Energy and AI, which examines rising electricity demand, grid constraints and supply chains. A station's headline capacity is only a starting point for that discussion.
Start with a 500 MW station
Assume 500 MW is the electrical output available for allocation, after the station's own use. For simplicity, ignore transmission losses. Allocate 10%, 25% or 50% to the data centre. The remainder stays outside this calculation.
PUE is the ratio of total data centre energy to IT equipment energy over the same period. Here, a constant PUE of 1.25 is also used as a steady-state power ratio. Divide the facility supply by 1.25 to find the IT budget. The difference covers cooling, power conversion and other facility overhead.
| Allocated share | Facility MW | IT MW | Overhead MW | Facility TWh/year |
|---|---|---|---|---|
| 10% | 50 | 40 | 10 | 0.438 |
| 25% | 125 | 100 | 25 | 1.095 |
| 50% | 250 | 200 | 50 | 2.190 |
Annual energy assumes the facility receives that power for all 8,760 hours of a non-leap year. This is an arithmetic scenario, not a claim that the station can provide uninterrupted supply.
Convert the IT budget into racks, servers and GPUs
Start with the middle scenario: 125 MW delivered to the facility and 100 MW left for IT after facility overhead. Choose a hypothetical rack load of 100 kW, including servers and an allocated share of networking and storage. The calculation is 100,000 kW ÷ 100 kW = 1,000 rack equivalents.
For this example only, assign 90% of the IT budget to complete GPU servers and 10% to external networking and storage. Assume each server contains eight GPUs and draws 10 kW at the selected operating point. That 10 kW includes the GPUs, CPUs, memory, internal drives, server fans and server power-supply losses. It is an invented planning input, not a specification for a named product.
| Budget or equipment | Calculation | Result |
|---|---|---|
| Networking and storage | 100 MW × 10% | 10 MW |
| Complete GPU servers | 100 MW − 10 MW | 90 MW |
| Server count | 90,000 kW ÷ 10 kW per server | 9,000 servers |
| GPU count | 9,000 servers × 8 GPUs | 72,000 GPUs |
| Rack equivalents | 100,000 kW ÷ 100 kW per rack | 1,000 racks |
Per rack equivalent, nine servers use 90 kW and the allocated network and storage share uses 10 kW. Real network and storage equipment may occupy separate racks, so this is a power allocation rather than a physical rack layout. Facility cooling and electrical overhead remain in the separate 25 MW PUE allowance; do not subtract them twice.
The result is sensitive to the chosen equipment. Keeping the server budget at 90 MW and eight GPUs per server, an assumed 15 kW per complete server would give 6,000 servers and 48,000 GPUs. Neither configuration establishes training speed, tokens per second or the number of users served. Those require a named GPU, model, precision, workload and benchmark.
These are operating-power equivalents with no spare power margin. A real design must also check peak draw, electrical and cooling limits, redundancy, physical fit and hourly supply availability. Download the scenario CSV for the assumptions and equipment counts at each allocation.
Lower overhead leaves more power for IT
Keep facility supply at 125 MW and change only PUE. At 1.50, IT receives 83.3 MW. At 1.25, it receives 100 MW. At 1.10, it receives 113.6 MW. These values illustrate sensitivity; they are not promises for a particular cooling system or climate.
The gain from PUE 1.50 to 1.25 is 16.7 MW of IT headroom within the same electricity supply. Whether that headroom can be used depends on installed equipment, rack density, cooling capacity and the electrical distribution system.
Do not call this a measured improvement in AI efficiency. PUE does not count completed training runs, useful tokens or model quality. A more efficient facility can still run an inefficient workload. The Green Grid's PUE guidance explains the metric and its measurement boundary.
What if supply falls for four hours?
Take the 125 MW facility scenario. Suppose supply falls to 100 MW for four hours and PUE remains 1.25. The IT budget falls from 100 MW to 80 MW. Over those four hours, the reduction is 80 MWh of IT energy and 20 MWh of overhead: 100 MWh less facility electricity.
This does not tell us how many jobs can be delayed. Some batch workloads may tolerate a later finish. Interactive services may have response-time commitments. Checkpointing, storage traffic and restart costs can also affect the result.
If deferred work runs later, this is a shift in demand rather than a guaranteed energy saving. The later operating window needs spare capacity, and any restart overhead must be counted. Backup generation or a battery would introduce a separate energy balance.
Power capacity is not computing output
Do not turn these MW values into a GPU count by dividing by a chip's rated wattage. IT power includes the rest of the server, networking and storage. A useful compute estimate needs a defined hardware configuration and workload benchmark.
Method and assumptions
Set the supply boundary
Use electrical MW available at the station export boundary. In these examples, supply is assumed to reach the facility without transmission losses.
Allocate a share
Facility power = station power × allocated share. The selected percentages are scenario inputs, not observed demand.
Separate IT from overhead
IT power = facility power ÷ PUE. Overhead = facility power − IT power. Use matching time and measurement boundaries.
Convert power into energy
Annual facility TWh = average facility MW × operating hours ÷ 1,000,000. Use 8,760 hours only for the stated constant-operation example.
The model excludes fuel use, carbon intensity, electricity price, water use, transmission losses, outages, redundancy and reserve margins. It does not establish the reliability or commercial feasibility of a dedicated supply. A nameplate rating or annual capacity factor alone cannot answer those questions.
Use a real station as a starting point
PowerAtlas can provide a starting reference for station identity and reported capacity. Check the record's source, operating status and whether the capacity applies to one unit or the whole complex before substituting it into the example.
Keep the result labelled as a scenario unless a public source establishes the supply relationship. A nearby station does not prove a physical connection, contracted allocation or available grid capacity. The calculation answers what a chosen allocation could support under its assumptions.
Questions
How many GPUs could a 500 MW power station support?
In this hypothetical example, allocate 25% of the station output and use PUE 1.25 to obtain 100 MW of IT power. Reserve 10 MW for networking and storage. At an assumed 10 kW per complete eight-GPU server, the remaining 90 MW supports 9,000 servers containing 72,000 GPUs. These are scenario inputs, not a hardware specification or a prediction of useful AI output.
How much IT power could a 500 MW station support?
In the example, allocating 25% of 500 MW gives a data centre 125 MW at its electricity boundary. At PUE 1.25, that supports 100 MW of IT equipment and 25 MW of facility overhead. This is a hypothetical allocation, not a supply agreement.
Does 100 MW of IT load mean 100 MW of GPUs?
No. IT load includes servers, storage and networking. GPU capacity and useful AI output depend on the equipment, workload, utilisation and service requirements.
Can annual generation prove that a plant can supply a data centre all year?
No. An annual energy balance does not establish hourly availability. Outages, other customers, transmission constraints and reserve requirements need separate assessment.
Does a lower PUE mean an AI workload uses less energy?
It means lower facility overhead per unit of IT energy within the measured boundary. It does not measure useful computation, model quality or the efficiency of the IT workload itself.
Sources and calculation files
- CSIRO: Australia-first pilot to transform data centres from static loads into dynamic grid-aware assets. Partner release, 10 September 2026. Source for the pilot announcement only.
- IEA: Key Questions on Energy and AI. 2026. Background on electricity demand and infrastructure constraints.
- The Green Grid: PUE, a comprehensive examination of the metric. Source for the PUE definition and measurement context.
- Inzonex scenario calculations. Hypothetical inputs and calculated outputs, 10 September 2026.
Suggested citation: Inzonex Research (2026), How much AI can a power station support?. The scenarios and chart are Inzonex calculations. They are not CSIRO or IEA projections. The illustration was generated with AI.