Australia could spread some of its growing AI computing needs across hundreds of old telephone exchanges rather than concentrating them in mega data centres, a Senate inquiry has heard.
Rob Nicholls, a senior research associate at the University of Sydney’s Centre for AI, Trust, and Governance, put the idea to the Senate Environment and Communications References Committee’s inquiry into artificial intelligence and data centres.
“I wouldn’t be at all surprised seeing lots of Telstra telephone exchange buildings hosting small inference-level data centres,” he said.
Telephone exchanges are buildings traditionally used to house equipment that connects local phone and telecommunications networks, with sites spread throughout Australian suburbs and towns.
Nicholls estimated each site could hold one or two megawatts of computing capacity.
Telstra has 483 local exchange buildings, Nicholls told the committee, located within four kilometres of more than 85 percent of Australia’s population.
Related Stories
Used this way, he said, the combined network could offer gigawatts of computing capacity, “but it doesn’t look like a gigawatt data centre.”
Could Run AI, But Not Train It
Nicholls said the model would only suit “inference,” or running an AI model that has already been trained, rather than training new models from scratch.
Training runs can draw power “in the gigawatt range” while underway, he said, but demand falls to “pretty close to zero” once a run ends, leaving large, purpose-built sites idle in between.
The proposal comes amid debate over whether Australia needs large data centres specifically to train AI models, rather than relying on computing services already available overseas.
Independent Senator David Pocock told the panel he had questioned government departments about the rationale for building AI training capacity in Australia.
“They say because we need access to the frontier models, which we already have,” he said.
“They tell me, ‘but we need more access to it, and this is the only way to ensure that we have that access.’ Does that argument stack up?”
Frontier models are the most advanced and powerful AI models available, generally developed by major technology companies.
Sue Keay, director of the UNSW AI Institute, said the government’s argument did not stack up.
“It sounds pretty circular,” she said. “So to protect ourselves from U.S. frontier models, we need access to U.S. frontier models.”
Pocock replied: “Essentially, yeah.”
Questions Over ‘Sovereign AI’
The distinction between hosting AI infrastructure and developing sovereign AI capability has become part of the Australian policy debate.
Liberal frontbencher James Paterson took a different view, arguing Australia risked missing the investment window for “sovereign AI training” if it did not attract the facilities.
“Just because they train frontier models on our soil doesn’t uplift our capability,” she said.
A frontier model remains the intellectual property of the company that developed it, regardless of where the computing infrastructure used to train it is located.
Nicholls said the more important question was not simply whether AI infrastructure was physically located in Australia.
