The Window is Closed

Good evening Reader,

For a few days, a model called Fable made something blindingly obvious that policy papers still talk around. It showed how it feels when an AI system doesn’t just autocomplete, but actually thinks with you – tracking your intent, iterating on ideas, and staying “alive” across long arcs of work.

Then it vanished.

Anthropic, the company behind Fable, turned it off. For people who had woven it into their workflows, it felt like losing their wings. The details of that product decision matter less than the lesson it contains for countries like Australia: we are building our future on cognitive infrastructure that can be withdrawn at any moment, by actors we do not control.

A three‑year window we treated as a seminar series

In 2022, Anthropic’s leaked Series C pitch deck predicted that whoever trains the best 2025/26 models will be “too far ahead for anyone to catch up in subsequent cycles.” Their argument was simple: once you reach a certain capability level, your own models start helping to design and optimise the next generation. The frontier becomes self‑accelerating.

Andrew Curran’s recent essay argues that we have just lived through the window that deck described: roughly February 2023 to February 2026. If a state wanted to stand up a genuinely sovereign, frontier‑grade AI stack – models, data pipelines, and compute – that was the moment to start.

Some people treated it like a race. Elon Musk’s xAI is the example Curran reaches for: a high‑capital, high‑focus push to get near the frontier in a little over two years.

Most governments treated it like a seminar series.

The European Union spent much of the window drafting and re‑drafting rules that, whatever their merits, made training sovereign frontier models harder rather than easier. Australia commissioned discussion papers, launched consultations, and announced pilots. We did not do the one thing a window like this demands: give someone a mandate, a budget, and a deadline to build a sovereign stack.

We had the money. We had access to talent. What we lacked was an institution willing to say “this is critical infrastructure” and act accordingly.

Innovation commons vs institutionalised dependence

Innovation commons are governance arrangements that let many actors build on shared infrastructure without being expropriated. They are about pooling data, compute, and knowledge under rules that give participants confidence to invest.

Australia’s AI response has been almost the opposite.

Instead of constructing a shared, sovereign layer, we are drifting towards a future where:

“National” AI services are thin wrappers over US or Chinese frontier models.
Key sectors – finance, health, education, transport, defence – bolt those models directly into their operations via vendor contracts.
Responsibility for the system as a whole is fragmented across agencies, regulators, and procurement processes, none of which are mandated to ask the uncomfortable question: what happens if the upstream tap is turned off?

Fable’s disappearance is a small‑scale rehearsal for that scenario. It is easy to imagine a future in which a model embedded across Australian government services and industry is embargoed overnight – for commercial, political, or security reasons. The effect would not look like a simple IT outage. It would look like a sudden cognitive blackout across systems that had come to rely on that model to plan, triage, draft, and decide.

This is what institutionalised dependence looks like in an AI‑saturated world.

Compute as uranium: the harder phase is starting

Curran’s second claim should ring alarm bells in Canberra: compute is becoming the new uranium.

Over the next few years, the chips and large‑scale compute needed to train frontier models are likely to be:

Monitored, licensed, and rationed as strategic assets.
Locked behind export controls, even between allies.
Concentrated in a small number of countries that already sit at the frontier.

For Australia, that removes an option we quietly assumed we had: the ability to buy our way into the race later. The “we’ll see how it plays out, then invest” strategy fails in a world where access to the raw materials is itself a foreign‑policy tool.

The likely trajectory looks like this:

Frontier labs and their home states offer attractive access deals, with enough local branding that governments can claim they have “national” models.
As domestic systems entangle themselves with these models, dependence deepens.
If and when strategic interests diverge, access terms change – through price, latency, functionality, or outright embargo.

From that point on, you are not a partner. You are a client.

Open‑source models will help in some domains, but they cannot solve structural shortages in compute, talent, and capital at national scale. And as open models approach frontier capabilities, they will themselves become targets for regulation and control.

“Too late” doesn’t mean “no agency”

Accepting that Australia missed the easiest sovereign AI window is uncomfortable but necessary. It is probably unrealistic, in mid‑2026, to imagine we can spin up a fully sovereign frontier lab on our own and keep pace with Anthropic, OpenAI, Google, or the major Chinese players on their timelines.

But the conclusion should not be “we are spectators now.”

The game has shifted from “can we be a frontier lab?” to “what roles can a country like Australia still claim in a concentrated landscape?” For a tech‑policy community, there are at least four concrete agendas:

Regime designer, not pure rule‑taker
Australia can move early on how frontier models are integrated into domestic systems: assurance regimes, safety and accountability standards, procurement rules that insist on auditability and fallback paths, and real parliamentary and public oversight.
Commons architect for data, evaluation, and adapters
We can build shared infrastructure that frontier models must plug into: Australian‑controlled datasets in key sectors; evaluation suites for local context and harms; and “adapters” that sit between models and critical systems, so that providers can be swapped without ripping out everything above them.
Intermediary sovereign layer
Instead of each department or company signing a separate deal with foreign labs, we can create national institutions that sit between providers and users – negotiating contracts, managing integration, enforcing standards, and making sure no single external vendor has one‑hop control over essential services.
Resilience and “graceful degradation”
We can deliberately design for the possibility that upstream access is throttled or removed: investing in regional compute with trusted partners, maintaining non‑AI and human capabilities in critical roles, and planning for how systems fail in slow, manageable ways rather than all at once.

None of this is as glamorous as training a frontier model. All of it is squarely in the domain of tech policy and institutional design.

What an Australian AI commons could be

If we take the Fable moment and Curran’s warning seriously, the next step is not another consultation. It is building institutions.

An Australian AI commons would look less like a new strategy document and more like a piece of infrastructure with a charter:

A publicly chartered, independently governed “AI Commons Corporation” responsible for shared datasets, adapters, and evaluation tools.
A national AI compute “grid,” treated as critical infrastructure, with reserved capacity for public‑interest and safety‑critical use.
Procurement rules that make “commons‑compatibility” non‑negotiable: if you want to sell models into Australian systems, you integrate with our adapters, submit to our evaluation, and support our resilience plans.

We cannot reopen the 2023–2026 window. But we can choose whether the remaining windows – around open infrastructure, shared governance, and institutional resilience – are wasted in the same way.

For a tech‑policy community that cares about sovereignty, this is the uncomfortable question Fable leaves behind: if an American lab can make us feel winged for a week and ground us at a keystroke, what exactly do we mean when we say “national AI capability”?

Share this on LinkedIn and Twitter to spur discussion

— The Editor