Steelman analysis

Compute is the substrate. Frontier performance scales with it, training budgets are growing 2.4x a year, and the civilizational gains already on the board --- life expectancy doubling, child mortality halving, extreme poverty collapsing from 44% to 8.5% --- came from exactly this kind of substrate buildout compounding. Every quarter of unbuilt capacity is averted capability that does not arrive. The burden of proof sits on the brake. Build the chips, build the datacenters, build the networking. Physics is the only legitimate stop.

The US lead is 6-18 months and compute is what preserves it. If responsible labs do not secure compute capacity, the marginal frontier run happens somewhere with worse safety culture and worse geopolitical alignment. Leading-edge fab is already a single-point-of-failure in Taiwan --- expanding compute capacity in jurisdictions under safety-conscious governance is the dominant move on expected outcomes. Ceding the substrate doesn't stop the race; it just picks the winner.

Alignment research is compute-bound. Interpretability at frontier scale requires running, probing, and retraining frontier-scale models --- you cannot align what you cannot afford to instrument. Algorithmic efficiency halving every 8 months means unaligned actors will reach dangerous capability regardless; the relevant question is whether the labs doing alignment work have the compute headroom to stay ahead of the capability curve they are studying. Expanding frontier-lab compute is alignment infrastructure.

The capability set of the worst-off is bounded by what medicine, education, and institutional capacity can deliver to them. Mental and neurological disorders are 15-16% of global YLD. NCDs are 74% of deaths. DALY rates vary 3x across regions. None of these bend without tools that don't exist yet, and those tools are compute-gated. Expanding frontier compute is what makes protein-folding, diagnostic, and tutoring capability cheap enough to reach Sub-Saharan Africa rather than stopping at the OECD.

Stewardship of creation includes reducing the suffering of persons made in the image. A child dying of a preventable condition in a high-burden region is a failure of stewardship that more capability could address. Tool-use in service of human persons --- the Catholic and Islamic framing of technology as instrument --- is not only permitted but obligatory when the tool could reach the suffering. Compute that builds instruments serving persons is stewardship, not transgression.

A maxim of systematically under-provisioning the actors doing interpretability work, while algorithmic progress makes dangerous capability cheaper every 8 months for everyone else, cannot be universalized without contradiction. If every safety-conscious institution adopted 'do not expand compute,' the frontier moves to actors who reject the duty entirely. Willing the end (aligned frontier AI) requires willing the means (compute for the alignment work).

Leverage 0.85, enterprise-absorption friction 0.9 --- the market pays for this bet whether or not it reaches the suffering-reduction thesis. The downside is capex and public friction; the upside is the substrate on which every subsequent suffering-reduction intervention runs. Even weighted by the probability that compute gets pointed at tax software instead of protein folding, the optionality value dominates. It is dead money that either pays off civilizationally or doesn't, and the don't-build-it branch has no payoff at all.