# Chokepoints in the AI Stack A chokepoint is a layer where very few suppliers control something everyone else needs, so whoever controls it can restrict the whole system. The concept is about structure, not nationality. ## Current AI chokepoints - **Lithography.** EUV machines come from a single company, ASML. - **Leading-edge fabrication and packaging.** Concentrated at TSMC, with advanced packaging a separate bottleneck ([[Advanced Packaging MOC]]). - **High-bandwidth memory.** A few suppliers. - **Accelerators and software.** Nvidia hardware plus the CUDA ecosystem ([[Below CUDA - GPU Kernels, PTX and Streaming Multiprocessors]]). - **Power and grid connection.** Local and slow to expand ([[Grid Interconnection Queues]]). - **Frontier model access.** Only a few labs can serve the best models, so API access is a potential chokepoint ([[API Distillation]]). - **Critical materials and refining.** e.g. [[Silver Refining Chokepoint]]. - **Naming and routing.** e.g. [[The Resolver as Sovereignty Chokepoint]]. ## Which layers are not chokepoints Released [[Open Weights]] are the opposite: infinitely copyable, no supplier to pressure. Policy aimed at them after release has little grip. This is why controls end up pushing weight onto hardware, power and access layers. ## The investor question For any company or policy, ask: which chokepoint does it sit on, does it control one, or is it exposed to one? Durable positions tend to sit on or next to a real chokepoint. Related: [[AI Chip Export Controls]], [[Jurisdiction and Control Planes]], [[Defensibility Principles MOC]]