Power, dilution and the AI chokepoint

On 12 June, global access to Anthropic’s new Mythos-class models – the most capable (Opens in new window) ever released – was abruptly disabled. This was in response (Opens in new window) to a White House directive to restrict access to all foreign nationals. Without a way to verify user nationality in real time, Anthropic took the models offline for everyone (Opens in new window).

It was the first time (Opens in new window) export controls have been used to restrict access to AI models (as opposed to the chips or weights underpinning them). And it came only days after Trump signed an executive order requesting (Opens in new window) early government access to every frontier model before public release. An administration explicit (Opens in new window) about wanting to win the AI race has just shown how readily state-level power concentration can be exercised.

This is bad news regardless of whether the United States is friend or foe. Australia’s access disappeared alongside everyone else’s, and neither the alliance nor membership of Five Eyes made any difference. As recursive self-improvement looms over the horizon (Opens in new window), and with it the prospect of artificial general intelligence (AGI), it is plausible that, by early next decade, a single data centre could be the equivalent of a country of geniuses (Opens in new window). This would translate an already decisive lead into an insurmountable one (Opens in new window), with efficiency and productivity gains multiplying across every dimension of state power.

An AGI hegemon would be far more powerful than any country in history, locking in a first mover advantage, hoovering up the majority of global GDP, and relegating everyone else to bystanders. In the most extreme example, concentration of power could permanently entrench (Opens in new window) a global totalitarian regime. Not ideal.

So what does that mean for the vast majority of countries that cannot realistically hope to compete in the race? The key is in ensuring power dilution so that the outcome is not winner takes all.

Pursuing sovereign AI capability (Opens in new window) is an idea that has gained traction recently. Where it can be built, it should be, but competing across the board with superpowers already at the AI frontier is not realistic (Opens in new window).

The handful of labs at the frontier of AI development already wield immense influence. But they are subject to the domestic legislation and regulations of the countries in which they exist. Whether and how these governance levers are pulled is a critical question, but there is nevertheless a power-limiting mechanism (at least for now – superintelligence would probably change that (Opens in new window)).

By contrast, the laws, norms and institutions that shape behaviour between states are voluntary, with no overarching authority to enforce them. The abrupt Mythos shutoff demonstrated starkly how access to cutting-edge models can be withdrawn on a foreign government’s whim, without recourse. For as long as the AI industry’s centre of gravity remains in the US, that power will continue to compound.

The Tech Policy Design Institute’s concept of AI agency (Opens in new window) offers a helpful reframing, focusing on whether a country possesses the ability to influence outcomes – both in defending its interests and capitalising on opportunities – rather than sovereign control. On this metric Australia’s agency is already high in areas like critical minerals and domain specific datasets, though gaps remain in compute infrastructure, data lifecycle management and a handful of other capabilities. Closing these won’t be without friction, as the current copyright reform (Opens in new window) debate illustrates, but a business-as-usual approach will only see agency erode.

How this agency is exercised is key. AI progress has been driven by relentless scaling, but like any digital technology it depends on the physical world – things like supply chains, chip manufacturing, data centre build-outs, and energy supply. Each represents a bottleneck (Opens in new window) on frontier development, and amid a contest between great powers, the strategic dimension is obvious. Translating agency into influence means exploiting these chokepoints to secure a foothold in the AI value chain, and Australia’s mineral resources and vast renewable energy potential are clear assets in this regard.

None of this should be done in isolation: a country the size of Australia must work with like-minded countries to shape global AI rules and norms. An ambitious step would be to push for a comprehensive global AI treaty (Opens in new window) that includes provisions on distributing the windfalls of AI. Canberra can lead regionally (Opens in new window) by being active in AI governance, serious about capacity building, and ready to put its tech diplomacy and research networks to work. The logic of dilution is collective – the more countries that hold a stake, the harder it becomes for any single one to dominate.

Even a benevolent AGI hegemon would leave everyone else’s sovereignty conditional: that is the novel danger in how power concentrates through AI. And if absolute power corrupts absolutely, the prospect is darker still. If Australia wants to avoid such an outcome, it should act while the chokepoints, minerals, and governance forums remain within reach.