The AI revolution is driving a dramatic surge in computing capacity. Data centres are being built to meet demand for training AI models and inference. Most of this capacity is built by private entities for commercial clients and allocated, in a still supply-constrained market, largely to the highest bidder.
Yet AI is also a tool for public good. It can help address intractable problems and enable innovations that markets alone may not support. This is where compute dedicated to public good comes in.
Pawsey Supercomputing is an Australian National Tier I advanced computing facility that accelerates cutting-edge scientific research for the benefit of the nation. It is an unincorporated joint venture between Australian universities and CSIRO, funded by the Western Australian and Federal Governments. It houses Setonix, the most powerful research supercomputer in the Southern Hemisphere and launched as the fourth greenest in the world.
Mark Stickells, CEO and a member of the APAC AI & Sustainability Council, shared at a recent Council session how compute for the public interest is managed differently from private-sector compute—and how its success is measured.
Public vs Private Computing Infrastructure
The difference between Pawsey and private data centres lies less in technological capability—Pawsey’s infrastructure can arguably surpass that of many commercial facilities—and more in:
- the nature of its clients and the problems they solve;
- how it allocates capacity and manages demand; and
- the social licence under which it operates.
First, its work is almost entirely in the public interest: human health and disease, climate and environmental data, astronomy, energy science, critical minerals, agriculture and water. Its users include hospitals, environmental agencies, climate scientists and industry. More than 1 billion compute hours have been allocated in 2026 to researchers from Australian universities, research institutions and government organisations. Public compute answers questions that have beneficiaries but not necessarily customers: a climate projection 50 years out, a rare disease affecting a few thousand people, or a sky survey whose value may not be apparent for a generation.
Second, capacity is allocated not by ability to pay, but through a national and partner computational merit allocation scheme. Compute is priced according to energy consumption: users do not buy machine time, but account for the energy they consume.
Finally, social licence has been central from Pawsey’s inception. Its facility is on an aquifer, which it uses and recharges, saving millions of litres of water annually. It also reduces energy consumption by:
- saving 20–30% through hardware tuning and matching CPU/GPU speeds to workloads;
- helping researchers optimise code, reducing run times by up to 90%;
- adjusting workloads to grid conditions;
- piloting AI with a local startup to move workloads to where renewable power is available—moving workloads to power, rather than power to workloads; and
- exploring quantum AI to optimise data-centre energy consumption.
Pawsey demonstrates that allocating resources for public-good goals need not follow market principles to be efficient. Public compute can foster innovation, deliver significant social value and operate with a strong social licence.
