Australia's decision to sign the San Francisco Declaration, reported by the ABC on 17 November 2026, is easy to defend on its own terms. It is also a good moment to ask what kind of AI order this club is building. The declaration was signed by ten countries, including Australia, Japan, India and the UK. According to the ABC, signatories will identify vulnerable supply chains, protect sensitive technology from theft or foreign interference, and coordinate efforts to strengthen domestic industries. It sits under the US-led Pax Silica initiative, which Australia joined at its December 2025 launch.
The strongest case for the declaration
The case for acting is real. Frontier AI depends on a short chain of inputs: rare earths and their processing, advanced chips, fabrication capacity and energy. Australian analysts have been blunt about the concentration. Writing for UNSW, Arup George described TSMC as "a single-point-of-failure in the AI ecosystem" and noted that when essential elements come from a narrow set of suppliers, "even minor disruptions can ripple globally." Olivia Shen of the United States Studies Centre told the ABC that "China could very easily weaponise control over rare earths and processing, while the US could strengthen export controls on chips."
A middle power like Australia, with large mineral deposits and limited chip capacity, has good reasons to want predictable access. A coalition that maps chokepoints before a crisis is better than improvising during one. Diversifying suppliers is also a pro-market goal: it reduces the odds that a single government decision can switch off an industry.
What the declaration actually does
Less than the headlines suggest. Pax Silica began with a declaration signed at a Washington summit on 12 December 2025. Digital Policy Alert's tracker records 24 signatories after the June 2026 summit and Italy becoming the 25th on 31 July 2026. The State Department's summary describes commitments to cooperate on enforcing policies that protect sensitive technologies and critical infrastructure, and to strengthen economic and national-security cooperation. The San Francisco Declaration, which the State Department says was signed on 15 November 2026 at the Pax Silica Founders Retreat, moves the initiative into what it calls a new phase. The ABC reports that this phase extends to power and energy systems, and that the extension "has raised questions about how it would work in practice."
That is the first problem. The text is framed as cooperation and intent, and I could not independently read the full declaration, so any claim about binding obligations should be treated cautiously. Terms like "sensitive technology" and "foreign interference" are undefined in the public descriptions. Whether they mean export-controlled chips, model weights, research collaboration or ordinary commercial contracts will be decided later, by agencies, outside open legislative debate.
Where the risks sit
The pro-innovation concern is not that governments secure their supply chains. It is that vague security language can expand without a test. Three risks stand out.
- Scope creep into open research. Protecting technology from "theft" is uncontroversial when it means trade-secret enforcement. It becomes costly if it grows into limits on publishing, open-source releases or the movement of researchers. Open science and open models have been among the largest drivers of AI progress, and a security bloc has little incentive to weigh those costs.
- A de facto regulatory zone. The Atlantic Council's Trisha Ray argued in April 2026 that Pax Silica "can serve as a vehicle for a de facto global AI regulatory zone." Standards written inside a club of ten to twenty-five governments, without most developing countries or civil society at the table, will bind others through market access rather than consent.
- Fragmentation. The ABC reports that China is offering its own alternative: several ASEAN states, including Indonesia, Malaysia, Vietnam and Cambodia, have backed a rival initiative, the World AI Congress Organisation. Singapore and the Philippines are in Pax Silica. A region split between two rule-sets raises compliance costs for startups, who have the least capacity to absorb them.
A proportionate way forward
None of this argues against Australia's decision. Albanese's line at the ASEAN summit in Manila, that countries "must shape AI" rather than let it shape them, is the right instinct. The question is how the commitments are implemented. Four tests would keep the initiative on the right side of proportionality.
- Publish definitions. Signatories should say in public what counts as "sensitive technology," so that firms and universities can plan and so that restrictions can be challenged.
- Target narrowly. Controls should focus on identifiable chokepoints, such as advanced fabrication equipment and leading-edge chips, rather than general-purpose software or published research.
- Build in review. Restrictions should carry sunset clauses and periodic evidence reviews, since a control that made sense in 2026 may only protect incumbents by 2029.
- Stay open to non-members. Supply-chain resilience is about adding suppliers, not subtracting them. Pairing the security track with market-opening measures, such as shared standards for trusted chip and mineral sourcing, gives smaller economies a reason to join without being asked to pick a side.
Australia can do real good inside the club by pushing these points. As a resource-rich democracy with a strong research sector, it is better placed than most members to argue that resilience and openness are complements. The risk for Australia and for the wider AI sector is not that Pax Silica exists, but that its most consequential choices stay informal, undefined and unreviewed. Governments that sign should be willing to say, in plain terms, what they have promised and what they have not.