scx.ai logo

Strategy Blog

Australia's New AI Rules Are About Power, Not Processors

National Cabinet just made power the price of admission for AI data centres. SCX.ai CEO David Keane explains why the new rules reward efficiency-first inference, and why that was always the plan.

By David Keane, Founder and CEO, SCX.ai6 min read

Everyone watching AI is watching models. Benchmarks, releases, price wars. I get it. The models are the fun part.

But the most important AI story in Australia this month had nothing to do with models. National Cabinet has signed off on a single national rulebook for AI data centres. Nine governments, one framework. And buried in the detail is a requirement that will decide which AI infrastructure companies thrive over the next decade: if you want Australian power for AI, you have to bring power of your own.

In August, NSW introduced a framework requiring data centre developers to bring power, not take it. Weeks later, the rest of the country signed on and made it national policy.

The world is bidding for megawatts

Step back from the model releases and look at where the money is going.

Saudi Arabia's HUMAIN program is committing around 100 billion US dollars across 11 data centres. The Stargate UAE project is targeting a full gigawatt of AI capacity. Microsoft alone is pouring 25 billion Australian dollars into its Australian AI buildout. Globally, sovereign cloud spending is tracking toward 80 billion US dollars this year, and most of it is going into concrete, cooling and copper, not chips.

Because here is the thing the industry does not like saying out loud. The binding constraint on AI in 2026 is not model architecture. It is not talent. It is megawatts.

Every serious AI infrastructure project on the planet is now an energy project with computers attached. Australia looked at that reality and made a choice. We are not going to out-gigawatt the Americans or the Saudis, and we are not going to let AI growth crowd households off the grid. If AI wants Australian power, AI underwrites Australian power.

What the rules actually say

The national standards, with legislation expected in 2027, apply to new hyperscale and large AI data centres. Three requirements stand out.

Net generators. New large AI data centres will need to underwrite at least as much new electricity generation as they consume, mostly through renewables and battery firming.

Full grid costs. Operators pay their full share of connection costs, so households are not quietly subsidising AI's power bill.

Water and location discipline. Strict water benchmarks and siting rules keep AI infrastructure from competing with housing and farmland for scarce resources.

Nothing is retroactive. The roughly two dozen projects already approved pass under the old rules. Every build after that wears the full cost.

Some commentary says this will chill the Australian AI data centre boom. I read it differently. It prices the externalities properly, and it clears out the operators whose entire business model was cheap power and someone else's grid. That is not a chill. That is a filter.

Efficiency just became the business model

When we launched Australia's first sovereign AI inferencing node at Equinix SY5 in Sydney, running on SambaNova's inference-optimised silicon, we made a choice that looked like a nice-to-have at the time. Air-cooled systems. Up to 99 percent less water and up to 75 percent less electricity than conventional GPU halls. Everything onshore.

Under the old rules, that was a sustainability story. Good for the website.

Under the new rules, it is the business model. When regulation prices power and water into every new build, the operator who needs a fraction of the electricity for the same work is not the green option anymore. It is the cheap option. Efficiency moved from the marketing slide to the P&L, and it is not moving back.

This is why the NSW framework mattered to us less as a burden than as a signal. We read it as confirmation of the architecture we had already picked. The national rules are the same signal, louder.

Inference is where efficiency compounds

There is one more piece, and it is the piece most commentary misses. Not all AI compute is the same.

Training grabs the headlines. Inference pays the bills. Every chatbot answer, every document summarised, every agent action, every generated video: that is inference. It runs around the clock, forever, and it is where the actual AI economy lives.

Efficiency compounds hardest at inference. If your platform delivers the same tokens for a fraction of the power, then every customer you sign is more profitable, every price you quote is more competitive, and every megawatt you bring online under the new rules goes further. That advantage does not shrink as the market grows. It widens.

This is why I keep saying the race for Australia is not about owning the biggest model. Models are becoming commodities. The race is about delivering the most intelligence per megawatt, onshore, under Australian law.

The race we can actually win

Australia will never be the biggest AI infrastructure market on earth. We do not have the land, the grid headroom, or the cultural appetite to burn power like other countries do.

Good. That was never our edge.

Our edge has always been doing more with less. Farming a dry continent. Building world-class systems with small teams. The new national rules simply codify that instinct for the AI age.

The companies that win here will not be the ones with the most racks. They will be the ones producing the most useful tokens per watt, onshore, without borrowing the grid from the rest of the country.

We have spent two years building exactly that. Now the law is catching up. Honestly, it is a good feeling to be on this side of the equation for once.

David Keane is the Founder and CEO of SCX.ai, Australia's sovereign AI infrastructure company.

Sources

Related Topics

data centre standardssovereign AIAI infrastructureenergynet generatorinferenceefficiencyAustraliaSCX.aipower grid
Australia's New AI Rules Are About Power, Not Processors