Australia should focus on hosting frontier AI training compute rather than developing its own sovereign AI capabilities from scratch, according to Janet Egan, director of artificial intelligence policy at the Institute for Progress. Egan believes the immense cost and compute requirements make developing frontier models infeasible for Australia, but hosting training compute could allow the country to shape the technology's trajectory and capture economic gains.
Australia's strategic approach to AI development and regulation could significantly impact its future economic competitiveness and its ability to manage the societal implications of advanced AI technologies.
Janet Egan, director of artificial intelligence policy at the Institute for Progress, has cautioned against Australia attempting to develop its own sovereign artificial intelligence capabilities from scratch, citing prohibitive costs and technical barriers. Speaking in Canberra, Egan suggested that a more feasible approach for Australia would be to position itself as a key player in the AI supply chain by hosting frontier training compute.
Egan estimated that developing a frontier model independently could cost around $500 billion, a sum she believes is beyond Australia's reach without risking significant financial loss. She noted that some existing "sovereign AI" initiatives rely on refining open-source models from other countries, rather than true indigenous development. Instead, she advocates for Australia to leverage its potential to host training compute, which would allow it to influence the technology's direction and benefit from its economic gains.
Addressing concerns about labor force disruption due to AI, Egan emphasized the importance of a strong tax base to support citizens through transitions. Hosting AI training compute could provide this opportunity. She also touched upon copyright issues, suggesting a central fund managed by the Australian government could remunerate local creatives whose work is used to train AI models, noting that AI companies appear more sensitive to licensing precedents than to price.
Egan pointed to the UK AI Security Institute as a successful model, praising its substantial funding, strong political backing, and ability to attract researchers. However, she noted a potential decline in its global policy influence due to restrictions on access to frontier models by countries like the U.S. Egan believes a robust AI safety institute is crucial for Australia, especially if it succeeds in attracting significant AI training compute.
Pick the topics you care about. Get only what matters, on your cadence.