How should the global south think and act about compute policy?
Adre Novais Xavier Rodrigues, Bruno De Biase Deo
We test whether the compute targets in Brazil's national AI plan (PBIA) are actually buyable with the budget the plan committed, and whether that spending is effective. We built an agent-based model with a 2,000-scenario Monte Carlo layer that turns each country's AI budget into installable frontier compute (FP64 PFLOPs, the TOP500 metric), gated by chip access, energy, and an endogenous chip price set by global demand. Brazil is the subject; the Gulf states, India, Egypt and Nigeria enter as competitors for the same scarce chips.
The plan's two compute targets, funded from one budget, diverge sharply: doubling national compute by 2027 is feasible in 100% of scenarios, while placing Santos Dumont in the TOP500 top five by 2029 is feasible in 0% (about 4% of the bar). The showcase target fails on framing and scale, not chips or energy: those add nothing, and even a tenfold budget reaches only ~7%. The verdict holds across every parameter extreme. Our takeaway for compute policy in the Global South: national compute goals should be written as absolute, budget-anchored targets sized to a country's own needs (running services, research, and evaluating deployed AI), not as positions in a moving global ranking that recedes at the frontier's growth rate.
The research question is interesting and a strong result could have actionable policy implications. Sadly, the methodology chosen was not the best fit to answer their research question.
Monte-carlo simulations are best suited for cases where randomness affects the outcome. In this case, it does not. Their main results are outcomes that happen 0% and 100% of the time, no randomness involved given their input parameters. And instead of explaining what are the key parameters the conclusion depends on and how they can be used to mechanistically deduce the final conclusions, the authors get lost on the details of the model instead of trying to explain in a couple sentences what's going on. I suspect the answer might be something like "if there's a few countries with bigger budgets than you that also want to have the largest supercomputer, it's really hard to beat them". Precise modelling is still really valuable, but only when it clarifies the final explanation of the mechanism, instead of hiding it.
Still, the writing was clear and direct, which made it easy and fun to read.
This is a rigorous and unusually well-presented modeling exercise: the provenance labeling, single-file reproducibility, robustness checks, and the disclosure that a validation fix strengthened rather than weakened the result are all exemplary. The central finding, that absolute, budget-anchored targets are reachable on a fixed budget while ranking targets move away at the frontier's growth rate, is correct and policy-relevant.
Three things would sharpen it. First, the actors that constitute the TOP500 frontier are the same actors Brazil competes with for chips, yet they enter only as an exogenous bar rising at a fixed 2.2×/yr. This makes the showcase verdict partly definitional: the model assumes the frontier runs away and then concludes Brazil cannot catch it.
Modeling at least the hyperscaler/frontier dynamics endogenously would test whether the result is discovered or assumed.
Second, the generalization to "the Global South" overreaches against the paper's own evidence: the Gulf plans clear feasibility thresholds within the plausible growth range, so the operative variable is budget-scale-relative-to-target-framing, not geography. Reframing the lesson around target design (absolute-and-needs-sized vs ranking-pinned) would be both more defensible and more useful, since not every relative target is infeasible for every endowment.
Third, the constructive half of the argument is asserted but not modeled. The doubling target is met 17× over in 100% of scenarios, so it functions as a trivial floor rather than a meaningful test, and the genuinely interesting question the paper raises (what absolute target sized to Brazil's actual evaluation, audit, and public-service needs would cost) is never quantified. The paper diagnoses the bad target with precision and leaves the good target as a principle; modeling even a rough needs-based figure would complete the contribution.
Cite this work
@misc {
title={
(HckPrj) How should the global south think and act about compute policy?
},
author={
Adre Novais Xavier Rodrigues, Bruno De Biase Deo
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
