Picks & shovels
The bottlenecks behind the trade.
Three frontier domains, fifteen structural bottlenecks, the companies that own each layer. Click any point on a blueprint to read the editorial take and see who supplies it.
AI · Bottlenecks
AI Compute Stack
Five layers, each with its own monopoly economics.
The AI build is not one trade. It is a stack of five distinct supply chains, each with different incumbents and different bottleneck dynamics. Capital moves up and down this stack as constraints shift. Today, GPU compute and memory bandwidth are the binding constraints; in 2026 the constraint moves to grid interconnect and cooling.
ROBOTICS · Bottlenecks
Humanoid Robotics
30+ degrees of freedom. Five places it can break.
Humanoid robotics is now a credible category — Figure, Tesla Optimus, 1X, Agility, Apptronik. Each company is building the same product against the same bottlenecks: actuators, vision, on-board compute, battery, and the foundation model that ties them together. Hardware is converging. Software is the divergent axis.
QUANTUM · Bottlenecks
Quantum Computing
Three architectures racing. None has won yet.
Quantum is the most architecturally unsettled of the three frontiers. Superconducting (IBM, Google, Rigetti), trapped-ion (IonQ, Quantinuum), and photonic (PsiQuantum, Xanadu) are all credible. Error rates are still 100-1000x where they need to be. The bottleneck is not qubit count — it's logical qubits after error correction, which still requires ~1000 physical qubits each.