Compute already has a large, measured market: a GPU loan book is priced on who signs the offtake. The futures everyone talks about trade almost nothing, and the return sits with whoever packages and protects GPU credit.

Banks lend against signed contracts at SOFR + 150 bp. Developers borrowing against similar contracts, before anything cash-flows, pay three to four times that, and a USD 1bn fund paid to wait in that gap nets about 6.6% a year (modeled) after a deliberately harsh loss charge.

Part 1 mapped what the incumbents are building for. This piece looks for the holes they leave. The door is free and the meter is going free, so I'd back what agents buy and the risk of what they do, at prices a realistic exit can repay.

The agent-payment protocols are free, and the open ones carry almost no money. The return sits with whoever issues the agent's credential and writes the rule for who eats a bad payment, mostly card networks and issuers, and with whoever owns the shopper's attention.

A ground-truth read on the fuel, the generation, and the equipment behind the AI build-out: what's physically available, where the constraints exist, and where capital can earn a return

A follow-up on Conrad, my autonomous GTM agent. This round he learned to grade his own work, run his own A/B tests, and rewrite his own playbook when the numbers slip. Here is what I built, what broke, and what I want to try next.