🖥️ Steelman analysis

Compute is the substrate on which alignment work actually runs. If the US lead is 6-18 months and training compute is growing 4-5x annually, the only way responsible labs stay at the frontier --- where interpretability research has access to the actual systems that will matter --- is to expand compute faster than less cautious actors. Ceding the compute race means ceding alignment leverage. Build the datacenters, run the experiments, keep the lead that lets careful actors set deployment norms.

Protein structure, target discovery, and trial simulation scale directly with compute. NCDs are 74% of global deaths and mental/neurological disorders drive 15-16% of YLDs --- these are tractable numerator terms in the suffering calculus and the pipeline that attacks them runs on frontier inference. More compute means more candidates screened, more mechanisms mapped, shorter discovery cycles. Every quarter of compute delay is averted therapeutic that did not arrive.

Compute is the binding variable. Capability scales with it, training runs are crossing $100M, and the civilizational trend line --- life expectancy 31 to 73, extreme poverty 44% to 8.5% --- is what happens when you let the substrate compound. The brake has to justify itself. Expand compute until physics says stop; every delayed GW is delayed flourishing.

Scale produces the feedback loop. You cannot align what you have not built and deployed at the frontier; interpretability on toy models does not generalize. Compute expansion is the precondition for the only alignment signal that matters --- behavior of actually-deployed frontier systems under real load. Build it, ship it, learn from it.

The US lead is 6-18 months and the fabs are in Taiwan. Compute expansion inside US jurisdiction is national-security infrastructure, not a commercial preference. Enterprise and government absorption already lags by years; without domestic compute scale, the integration gap becomes a strategic gap. Build the compute or cede the order.

Compute is the root-cause substrate. Every downstream suffering-reduction intervention --- drug discovery, mental health triage, biomedical acceleration --- bottlenecks on it. The civilizational trend is real: mortality halved, life expectancy doubled, but 4.9M under-5 deaths annually and a 15% YLD mental-health burden remain. Compute is the thing that compounds against those numerators. Expand it.

The pipeline from frontier compute to averted DALYs runs through drug discovery, diagnostic triage, and care-navigation --- all compute-bound. Sub-Saharan Africa carries 3x the DALY burden of high-income East Asia; compressing discovery cycles by a year for a TB or malaria therapeutic is worth more suffering-averted than any near-term frontier capability race. Build the compute, point it at the pipeline.

80 billion land animals and 1-3 trillion aquatic animals annually is the largest numerator term in any honest suffering calculus. Alternative-protein development --- strain engineering, scaffolding optimization, cost-down modeling --- is compute-bound. Expand compute and the alt-protein displacement curve bends earlier. Every year compute expansion is delayed is another 80B slaughter-cycle at full intensity.