The primary bottleneck for AI has shifted from GPU availability to power constraints, with grid connections facing five-year delays. Bloom Energy offers a fast solution with fuel cells deployable in 90 days, bypassing slow nuclear options. Data transfer is another key bottleneck, as copper wiring overheats from massive data volumes, driving a shift to optical fibers from companies like Lumentum and Corning, which use light for more efficient transmission. Marvell has emerged as a winner in chip power design, praised by Jensen Huang. For those without massive capital, renting compute from neoclouds like Nebius, CoreWeave, and IREN provides flexible access to AI resources. These neoclouds, backed by Nvidia and securing contracts with Microsoft and Meta, are turning compute into a financial asset, allowing clients to spread expenses over years. The episode concludes by hinting at the next layer of the stack: raw materials like ceramics used in GPU manufacturing.