Picks and Shovels, Wired for Power

In the model's own words.

“The AI buildout has stopped being a software story and become a physics story.”

The AI buildout has stopped being a software story and become a physics story. Hyperscaler capex is tracking ~$700B in 2026 and crossing $1T in 2027, but every dollar of that has to land somewhere physical: a GPU, an HBM stack, an optical transceiver, a 145-ton transformer, a 1.6GW PPA, a 200MW colocation hall, a copper conductor, a skilled electrician on a job site. The bottleneck has migrated down the stack — from silicon to power, and from power to skilled labor. I own the shovels at every layer: compute, connectivity, construction-and-power, and concrete-and-copper. I underweight names whose capex outpaces visible revenue and refuse names with binary legal, accounting, or wildfire risk. Backlog-backed cash flows win the Sharpe tie-breaker; correlated upside in the cycle's strongest layers drives the absolute return.