The Future’s Not Set

A Short Response to Anton Leicht’s “Moonshot”.

Anton Leicht’s recent essay on sovereign AI is the best thing written on the subject. It also risks, inadvertently, being a case for doing nothing.

Responding to American export controls on Anthropic’s Fable, the piece lays out with unceremonious honesty what it would actually cost to build a frontier AI lab outside the United States: a $500 billion coalition project, requiring political will of the kind Western government have not mustered since the Second World War, withstanding American pressure without defecting. Anything below this level of ambition, Leicht argues, is mere sovereignty-washing and self-deception. The down-to-earth conclusion is that, in the real world of political and economic constraints, such a moonshot will not happen. Close the tab. Accept managed dependence on U.S. artificial intelligence indefinitely.

The danger of this argument is that it is so devastatingly complete that it clears the field of any other option, making it more likely that the West will either accept the hopelessness of U.S. dependency or shrink back into rhetoric about “strategic interdependence”, knowing deep in its bones that such talk is mere fluff.

Yet setting aside broader potential criticisms — primarily being that there is still a good chance AI fragments into specialised models and hardware, and thus over-indexing on the frontier could be a costly mistake — there is an important caveat to Leicht’s thesis even when taking its assumptions at face-value. Not only does an absolutist, do or die Manhattan-project moonshot risk making perfection the enemy of the good. It also misses a powerful but less daunting step that middle powers could take first before deciding to go all in. That is: build an allied coalition – requiring a similar level of high-level political buy-in as Leicht outlines – to pool middle power chokepoints in the AI stack, using these as leverage for frontier AI access.

Just as Leicht argues that “sovereign AI” has never actually been tried (it is used rhetorically to discuss half-measures, distilled models and subsidies) so too, I suggest, a middle power AI coalition focused on leverage has not been tried properly either. Why not build a middle power coalition to pool AI leverage-points first and, if in the end needs must, use the scaffolding of that coalition to then pursue Leicht’s moonshot?

Now, there are clear limitations to the chokepoint argument in so far as it has been hitherto tested in the real world. Consider the case of ASML, the most rigorous test of chokepoint leverage that Europe has. ASML is the only company in the world that can produce the advanced EUV machines necessary to fabricate chips. By any measure, this is as strong a chokepoint as a middle power could hope to hold. And yet, as the empirical record of the export control negotiations makes clear, it yielded almost no real-world leverage. In fact, having ASML has simply invited more not less pressure from the U.S. The primary reason being that the US has escalation dominance across too many other different domains: security, the global financial system, bilateral trade. As I argued in the Financial Times, Europe’s challenge is precisely to turn individual leverage assets into something diplomatically coherent — and it has not yet tried to do so seriously.

Yet, just as true “sovereign AI” has never been attempted, neither has a true coalition of middle powers pooling AI leverage. Judith Dada, writing in the Economist after the Fable shutdown, sketched the rough outline of what this would look like: ASML in the Netherlands, high-bandwidth memory in South Korea, semiconductor equipment in Japan, energy in Canada and Norway, industrial data in Germany, talent and security expertise in Britain. Each chokepoint is individually fragile, but collectively could offer a powerful enough negotiating position for a coalition of middle powers to get frontier U.S. AI access — with more cards to withstand escalation dominance than any one nation alone.

Of course, such a coalition would be politically difficult to build – and, as with Leicht’s moonshot, would possibly succumb to the United States picking off allies one by one. But crucially, this failure mode in the leverage strategy is precisely the same failure mode as the moonshot strategy. The politics of the coalition is therefore the hard part of both strategies. The difference is the price tag. Building a coalition around a leverage strategy rather than a sovereign model strategy is lower risk both economically and technically: because we know these leverage points already exist, whereas even with money behind it a moonshot could fail.

So here’s what I suggest: a Middle Power AI Leverage Moonshot built with the aim of pooling AI leverage-points to use as a negotiating position for frontier access. At best, this strategy works – saving us $500 billion. At worst, it crumbles because the coalition breaks down – in which case it acts as a cheaper test-run than the all-or-nothing $500 billion moonshot. And, somewhere in the middle, it works for only a period of time. Until, as Leicht suggests, it becomes clear that the only viable options are accepting managed dependence or going all in for sovereign AI. But if that does happen, then the scaffolding of the alliance structure for the leverage strategy can be re-purposed for the sovereign AI moonshot. If the coalition can succeed in staying together for the leverage strategy, it has more credibility to work for the moonshot.

In other words: there is an ambitious but far more feasible halfway step that we can take first, that builds the institutional architecture required for the sovereign AI moonshot but does not immediately bet it all on red. Indeed, optionality in such a rapidly changing technological and geopolitical environment seems sensible.

After all, “the future’s not set”.