# Supply Chain Chokepoint A chokepoint is a place where a very large flow passes through a very small opening. The compute supply chain has several, and they have the useful property of being identifiable in advance: find the step performed by the fewest firms at the fewest sites, and that is where the next shortage will originate. Knowing them converts the buildout from a mystery into a schedule, and a schedule is something a country can act on. > [!note] The one-sentence version > Compute capacity equals wafer starts times yield times packaging slots times memory allocation times megawatts delivered; a shortage at any one term starves everything downstream. ## Why they exist Each step in making advanced computing demands capital measured in billions, specialist knowledge accumulated across decades, and a tolerance for defects measured in atoms. Those conditions reward concentration: the firm that gets there first accumulates the learning, and the learning compounds faster than a competitor can buy it. Concentration delivers remarkable capability and a narrow aperture at the same time. The honest description is a chain of extraordinary industrial achievements, each performed by very few hands. ## The principal chokepoints **Lithography tools.** Leading-edge exposure systems come from one supplier; the world's access to the finest transistors flows through a single order book. See [[wiki/Extreme Ultraviolet Lithography|Extreme Ultraviolet Lithography]]. **Leading-edge fabrication.** More than ninety percent of the world's most advanced chips are manufactured by one company, overwhelmingly at sites on one island. Bringing a new fab to mature yield takes years, which is why additional capacity is a multi-year answer rather than a multi-month one. See [[wiki/Foundry|Foundry]] · [[wiki/TSMC|TSMC]]. **Advanced packaging.** When demand spiked in 2023 and 2024, the industry could start enough wafers and could not bond dies to memory fast enough; the constraint sat in assembly capacity. See [[wiki/Advanced Packaging|Advanced Packaging]]. **Stacked memory.** High-bandwidth memory comes from three manufacturers, and its generations set accelerator generations. Allocation is negotiated years ahead. See [[wiki/High-Bandwidth Memory|High-Bandwidth Memory]]. **Large power transformers and switchgear.** Grid-scale transformers are built by a small number of factories to order, with lead times that have stretched into years; this constraint now governs the pace of data-center energization as firmly as silicon does. See [[wiki/Transformer Supply|Transformer Supply]]. **Gas turbines and generation equipment.** Order books for large turbines are committed years out, which shapes how quickly behind-the-meter generation can be fielded. See [[wiki/Turbine Supply|Turbine Supply]]. **The trades.** Licensed electricians, pipefitters, and controls technicians are the scarcest input in several regions, and a workforce cannot be ordered with a lead time. See [[wiki/Electrical Trades|Electrical Trades]] · [[wiki/Data Center Workforce|Data Center Workforce]]. ## How the constraint migrates The pattern is consistent and worth internalizing. In the 2000s the binding limit was transistor density. Through the 2010s scaling slowed and the limit became die area and power, which produced [[wiki/Chiplet|chiplets]]. In the middle of this decade it moved outward to packaging throughput and memory supply. It is now moving again, into electrical interconnection, transformers, turbines, and labor. Each generation of engineering dissolves one wall and reveals the next scarcest thing; the constraint never disappears, it relocates. Reading it correctly a year early is most of what strategic advantage consists of. ## What loosens a chokepoint Capacity added before it is needed; a second qualified supplier for every step; geographic distribution of identical capability; standards that let parts from different vendors interoperate; domestic manufacture of the long-lead electrical and mechanical equipment; and apprenticeship pipelines started years before the buildings. Every one of these is something a nation can decide to do, which is the practical reason to map chokepoints at all. ## Relationships - **constrains:** [[wiki/Accelerator|Accelerator]] · [[wiki/Utilization|Utilization]] · [[wiki/Gigawatt Campus|Gigawatt Campus]] - **located in:** [[wiki/Extreme Ultraviolet Lithography|Extreme Ultraviolet Lithography]] · [[wiki/Advanced Packaging|Advanced Packaging]] · [[wiki/High-Bandwidth Memory|High-Bandwidth Memory]] · [[wiki/Transformer Supply|Transformer Supply]] · [[wiki/Turbine Supply|Turbine Supply]] - **addressed by:** [[wiki/Semiconductor Export Controls|Semiconductor Export Controls]] · [[wiki/Critical Minerals|Critical Minerals]] · [[wiki/Apprenticeship|Apprenticeship]] · [[wiki/National Compute Capacity|National Compute Capacity]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** single-supplier leading-edge lithography; one company manufacturing above ninety percent of the most advanced chips, concentrated in Taiwan; packaging as the binding constraint during the 2023 to 2024 demand spike; three suppliers of stacked memory; multi-year fab bring-up timelines. - **Established (datestamped):** as of October 2026, large power transformer and gas turbine lead times are quoted in years across the industry. - **Strongly indicated:** skilled electrical trades as a binding regional constraint; staffing shortfalls are widely reported by builders and utilities and are not centrally measured. - **Plausible:** published estimates of a very large world-economy loss from a severe disruption to Taiwanese chip supply; these are macroeconomic models rather than measured outcomes. ## Sources - [How AI Is Pushing the Semiconductor Supply Chain to the Limit](https://www.bloomberg.com/) — Bloomberg Originals, October 2026; foundry concentration, sole-supplier lithography, fab bring-up timelines, conflict-loss estimate. - [How the Electrical Grid Is Being Rebuilt for AI](https://www.bloomberg.com/) — Bloomberg Primer, October 2026; equipment, supply-chain, and workforce atrophy during the flat-demand decades.