The capex numbers matter, but not because "bigger budget wins."

Big spend creates its own problems: fixed commitments, debt cycles, pressure to keep utilization high. When the bill is hundreds of billions, you start treating GPUs, DRAM/HBM, and fab slots like strategic assets.

The pieces in play aren't just "models." They're chip capacity, memory packaging, data center power, network topology, and routing decisions about where inference gets served.

The labs that win make the whole stack resilient under scarcity.

This is also why decentralized compute matters: it doesn't beat hyperscalers on everything, but it prevents a single chokepoint from becoming critical.