The money is rotating into the second layer of the AI value chain. Layer one was $NVDA and the hyperscalers. That money's been made. Layer two is the picks-and-shovels: memory, optics, power, test, storage, foundry. Most trade at half the GPU multiple. The stack: Data: $P — all-flash for the AI pipeline, revenue +38%, just crushed earnings. AI is a data problem before it's a compute problem. Memory: $PENG (earnings Tuesday), $RMBS, The bottleneck nobody can print. Optical/fabric: $AAOI $CIEN $CRDO $MXL $SMTC $ANET $FN. My fabric-shift thesis in one basket. Power: $MPWR. Power density rises every generation and they tax all of it. Test/materials: $AEHR $AXTI $ACLS $ENTG $ONTO. The tollbooths before the fab. Foundry: $GFS. Regionalized supply chains give this strategic value. None of these need AI to work in the abstract. They get paid when racks get built. That's capex you can track — not a narrative.