Not All Should Go South: The Pragmatic AI Strategy for Secondary Powers

The main shortcoming of this work is that it missed exploring the space of options for AI strategy in development, jumping too quickly to the other stages of the AI cycle.

It seems very plausible to me that Brazil's most valuable bet to remain relevant is to build more compute and lease it to the top AI labs worldwide. Even if 'developing' models would be unprofitable for them, renting compute definitely wouldn't if they can ensure some reasonable rule of law provisions around it. See arguments here: https://newsletter.forethought.org/p/how-can-the-middle-powers-avoid-getting

Beyond that, I did not find the other arguments in the text particularly tight in driving their conclusions. They chose a really important question, but could have done better in working towards their conclusions

This is a clearly written, well-organized, and timely paper, and the disaggregation of AI strategy into development, deployment, and usage is a potentially useful lens. A few directions would meaningfully strengthen it.

First, the contribution would land harder with a sharper target. The programs cited as cautionary examples, Brazil and India, are in practice already pursuing largely non-frontier, deployment-and-fine-tuning strategies, so it would help to clarify whether the argument is against frontier pretraining from scratch (which few policymakers actually fund) or against non-frontier "sovereign-sufficient" model training (which can be a defensible, auditable, locally-aligned capability investment at relatively modest cost). Drawing that frontier-versus-usable distinction explicitly would prevent the recommendation from reading as broader than intended, and it is arguably the distinction that decides whether domestic training is a misallocation or a reasonable bet. A quantitative comparison of economic effects across the three layers, even a rough one, would also give the central claim empirical grounding it currently relies on argument to supply.

Second, the paper's strongest move, disaggregation, could be applied one level deeper to the two layers it recommends. Deployment spans fine-tuning, post-training, hosting, and the open-versus-closed choice; usage spans building proprietary applications versus adopting commercial ones, and differs substantially between public and private sector contexts. Specifying which of these components suit which kinds of country would turn a directional orientation into actionable guidance. This connects to the acknowledged lack of country-level calibration: because the paper's output is a prescriptive sequence, tying it even loosely to national compute, talent, and institutional capacity would do a lot to make it policy-relevant, and would be a natural next step alongside the deferred analysis of real national plans.

Third, the argument would be more persuasive if it engaged the strongest opposing evidence. The macroeconomic case currently rests on Acemoglu's conservative estimate; bringing in the more transformative projections (e.g. Brynjolfsson) and addressing them directly would strengthen rather than weaken the paper, since the transformative-gains scenario is the main challenge to its thesis. Similarly, the load-bearing premise that development is a poor investment, and several supporting points sourced to industry and venture commentary, would benefit from firmer analytical or empirical backing. The component on building a domestic AI-safety research community is promising but would be stronger with an explicit account of the mechanism by which that expertise translates into influence over how frontier models are built.

On presentation, the writing is clear and the structure logical; dialing back the confident register in a few places would help the strong underlying argument read as the careful analysis it is rather than as an opinion piece.

The economic core is sound and lands a real point, that for secondary powers pretraining at the frontier is a poor use of scarce capital. Brazil's showcase supercomputer goal, an unreachable top-five global ranking, is a clear real-world instance of exactly that failure. The paper overreaches, though, by treating every domestic model as a frontier bid. As far as I can tell, India's state-funded models are deliberately small and multilingual, built because open frontier models handle low-resource languages poorly, a genuine comparative-advantage case for some local development that the paper excludes by assumption.

The development-deployment-usage split and the pivot from sovereignty to resilience are already standard in current policy writing, including the BCG report the authors cite. It would have been far more useful to test the framework against real national plans and derive concrete, plan-specific guidance, which is exactly the step they defer. The safety dimension, which is what places the paper in this track, is also the thinnest, since component D's main lever, influencing frontier labs through the bargaining power of the buyer, assumes a market mass that no single secondary power has.