Trusted Trade and the AI Stack

Hi, my name is William Smith. I work at the East Asian Bureau of Economic Research at the ANU, and I'm going to be talking today about trusted trade and the AI stack. One of the recurring questions in AI governance is where control over frontier AI actually sits. Some of it sits at the model layer in alignment work and decisions about which capabilities are developed and released. Some of it sits with labs and cloud providers and in the rules they impose on access. But a great deal of it also sits in the economic infrastructure beneath the model — the various advanced technologies and raw materials that are traded between countries and make frontier training possible.

And, as will be the focus of this talk, in the rules and norms that govern their trade more generally. To see why these rules and norms are relevant to the broader AI safety conversation, imagine that in 2026, the world slid back into the kind of protectionist trade agenda that dominated the interwar period, about a century ago. Advanced AI would take far longer to emerge, simply because no single country could easily assemble the full AI stack on its own. It would be, in the first instance, more difficult, but also more costly and take longer.

Frontier AI depends on cross-border value chains, and so the speed of progress is regulated by the performance of those value chains. The shift in trade I want to focus on in this talk is the move from general openness in trade — that is, no-holds-barred free trade — to a form of selective openness, sometimes called trusted trade. By trusted trade, as you see on the slide, I mean the preferential organisation of international trade between states judged to be reliable or trustworthy partners. A key example would be the China-Japan rare earth minerals dispute that occurred in 2010, which caused China's share of Japanese rare earth minerals imports to reduce from 90% in 2010 to now 60% in 2026. This form of trade — trusted trade — has become increasingly prevalent over the last decade, and, as I'm going to talk about, is increasingly relevant to AI governance.

To get a sense of this shift, my colleagues and I built a corpus of around 26,000 speeches, interviews and press releases from G7 and Australian trade and finance ministers ranging from 2013 to 2025. We first filtered the documents for trade policy relevance. Then we coded each relevant document for whether it promoted trusted trade according to a more precise definition of the concept than the one I just presented. This coding was automated through OpenAI's API to facilitate the scale of the project, and checked against human coding of a subset of the corpus.

On that subset, the automated coding matched the human coding in more than 95% of cases. The resulting chart tells a useful story. For most of the 2010s, as you can see here, the rhetoric is almost absent from trade policy discussions. This was a period in which the World Trade Organisation's dispute settlement mechanisms still worked well enough to discipline discriminatory trade measures, and the United States had not yet abandoned its role as the principal defender of the rules-based economic order. The upward climb begins around 2019, when the United States' refusal to appoint judges to the WTO Appellate Body practically paralysed the enforcement of global free trade.

It continues as the US-China trade war deepens beyond 2018, and the COVID-19 pandemic brought about the direct considerations of national interest in trade. I think there are three relevant lessons that AI governance discussions can draw from the trend that I just identified. First, AI governance discourse is highly concerned with export controls. That is, government restrictions on the export of particular goods. And in recent months, especially US controls on advanced semiconductors and controls aimed at frontier AI models like Fable. While this discussion is important, it captures only part of the broader story about trends in international trade, which will, going forward, undergird the development of frontier AI models.

Second, and relatedly, trusted trade makes AI safety partly a question of economic statecraft. Trade decisions, especially those outside the free trade paradigm, will shape which countries can build advanced systems and on what terms. This issue is particularly important if middle powers such as South Korea and perhaps Australia try to build sovereign AI capabilities of their own. Third, and finally, trust matters unevenly across the AI stack. Raw materials may continue to move through relatively open markets, but advanced technologies are much more likely to be channelled through trusted partners.

This nuance is needed for a satisfactory and comprehensive account of the AI stack. Thank you.