The AI infrastructure trade has moved past GPUs. Power delivery, grid capacity, and data center wiring are now the binding constraints, and this episode maps the investment landscape around that shift. US grid interconnection queues run 5 years deep, making conventional utility power a dead end for hyperscalers that need capacity now.
The episode names specific companies and explains why each matters: Bloom Energy for on-site fuel cells that bypass grid delays, Lumentum and Corning for optical networking that replaces copper as data transfer volumes exceed electrical limits, Marvell for power delivery silicon inside AI hardware, and neocloud operators Nebius, CoreWeave, and IREN as the businesses renting this infrastructure at scale. The technical walkthrough of why light replaces copper at timestamp 11:51 is worth reading in full, not just the conclusion.
The framing underneath all of it is that compute is becoming a hard asset class, not a service. The final segment, starting at 29:47, sketches what the hosts call the substrate layer: the physical and electrical foundation that AI model performance will depend on for the next decade. If you have any position in AI hardware or cloud infrastructure, this is the context you are missing.
[WATCH ON YOUTUBE →]