Christina Wiremu-Brook, writing recently in the Australian Financial Review, made an uncomfortable observation: every major AI nation, except Australia, has already picked a lane for its AI strategy:
US (Frontier innovation), China (Scale, sovereignty, full-stack capability) Canada (Research engine), UK (Governance and assurance), Singapore (Trusted standards), South Korea (Hardware and applied tech), India (Global services and talent).
Her point is simple: Australia hasn’t chosen a strategic pathway. Without a clearly defined mission, we risk becoming a permanent consumer of AI rather than a meaningful contributor.
I agree. And instead of copying the US, UK, or China, I propose a lane that leverages something truly Australian:
Harnessing our low cost sun power for flexible, low-cost, AI training
I propose an AI infrastructure strategy tailored to our geography, our energy profile, and designed for our on shore needs. Here’s why this is our advantage.
1. AI training does not need 100% peak load data centres Most data centres are designed for always-on workloads: banking systems, airline systems, and operational platforms that cannot pause. They require:
- continuous power,
- redundant diesel backup,
- high water consumption for cooling,
- and 24/7 uptime guarantees.
They are accordingly, expensive to run and demanding to cool. But AI training is fundamentally different. Training a large model is:
- long-running,
- batch-based,
- non-interactive, and
- potentially tolerant of pauses.
An AI model training run can potentially slow down when electricity prices spike and accelerate when power is abundant.
That flexibility is not a burden, it’s an opportunity. This long-running training is different to normal data centre compute tasks including answering AI queries (so-called AI inference) which DO need to run as fast as possible.
2. Turning unused power into a Strategic Asset Australia already produces more solar energy at midday than the grid can absorb. The excess:
- depresses wholesale prices,
- destabilises the grid,
- and literally goes to waste.
Now imagine data centres purpose-built to soak up this surplus. These would not be hyperscaler facilities demanding flawless uptime. They would be interruptible AI training centres that: ✔ ramp up when energy is cheap and green ✔ throttle back when the grid is stressed ✔ reduce wasted generation ✔ deliver significantly cheaper compute for local organisations For sure such workloads may take a bit longer. A one-month training run may take five weeks for example. But for many use cases, that trade-off is entirely acceptable, especially when the job is on-shore, secure, and cost-efficient.
3. The sovereign trade-off For government, defence, healthcare, research, and regulated industries, speed is rarely the single most important factor. Security, control, and affordability often matter more. A sovereign AI training environment, powered by surplus Australian energy, would offer:
- data residency guarantees
- reduced exposure to foreign cloud monopolies
- lower emissions and lower cost
- alignment with existing national priorities Australia will never out-spend the US or out-scale China. But we can: build the world’s most efficient environment for sovereign model training turn our energy surplus into a digital asset become a partner, not a passenger, in the AI ecosystem We are one of the few nations with predictable low cost power during peak solar production.
Instead of exporting electrons via speculative underwater cables, we could convert that surplus directly into sovereign, green, training compute capacity.
This is in effect an arbitrage play, and one that helps the power grid, rather than adding additional loads to it.
The questions we need to answer
This proposal requires thought, not blind enthusiasm. Here are some policy design challenges:
- Pricing: What level of discount makes slower training commercially irresistible?
- Technology: Will future models still require long-running training? (Current evidence suggests yes.)
- Structure: Should this be a public-private initiative to coordinate AI workloads with the grid?
- Power: as battery storage, EVs start to consume more of the peak excess during the day, the power curve will flatten out, so the midday period low cost power might reduce somewhat - we need long-term pricing models that factor this in for the long-term.
- exploits our natural energy advantage,
- strengthens sovereign capability,
- reduces emissions, and
- positions us as a constructive, ethical contributor.
The Choice Ahead
Australia can continue admiring other nations' AI strategies from the sidelines. Or we can define our own lane, one that:
Flexible, low-cost, sovereign AI training is economically realistic, strategically distinctive, and deeply aligned with who we are. So I ask:
What do you think? Is Green & Sovereign the right lane for Australia’s AI future?
PS I realise there are some initiatives underway on this like: the Green Data Centre hub in WA for example and another proposed for Tasmania but I'm talking about a strategic scale investment.
Concept and words by Paul Cooper. Grammar tuned by AI. #datacentres #aitraining #long-running compute #green data #flexible power #grid stability
Leave a Comment