The AI compute stack, mapped

This page is an interactive atlas of the AI compute supply chain: 580 market nodes and 562 dependencies across ten layers, from raw materials to applications. 9 of those nodes are chokepoints, markets concentrated enough that their owners set the terms for everyone downstream, and 112 more are bottlenecks under strain.

The 9 chokepoints, by name:

• DUV immersion lithography (ArF 193 nm) — ASML
• EUV lithography (13.5 nm wavelength) — ASML Holding NV
• HBM stacking - TSV and mass reflow/thermocompression bonding — SK Hynix
• High-bandwidth memory (HBM) — SK Hynix
• High-purity quartz (HPQ) mining and processing — Sibelco
• Lithography systems — ASML
• Mask blanks — HOYA
• Mask writing (e-beam writers) — JEOL
• Rare earth separation and refining — Lynas Rare Earths

• L0 Raw Materials & Extraction: Refining is the chokepoint, not mining. Rare-earth separation, sintered magnets, gallium, germanium and graphite are all heavily China-concentrated and now under active export controls. High-purity quartz is a near single-source dependency for the entire semiconductor chain. Dominant player: China (refining & separation). Concentration: China ~60–90% of rare-earth refining. 18 bottlenecks in this layer.
• L1 Semiconductor Manufacturing: This is the most concentrated cluster of chokepoints on the map. CoWoS packaging, EUV lithography, actinic mask inspection, HBM stacking, ABF substrates and EDA each run through one or two suppliers worldwide. ASML builds every EUV scanner on the planet. TSMC prints almost every leading-edge chip. Dominant player: TSMC / ASML. Concentration: ASML 100% of EUV; TSMC ~90%+ of sub-5nm. 36 bottlenecks in this layer.
• L2 Compute Hardware: The merchant GPU gets the headlines, but HBM memory and the lasers and optical components feeding the interconnect are the tighter constraints. Behind every GPU order sits the in-rack power-delivery chain, now moving to 48V and 800V. Dominant player: NVIDIA. Concentration: ~80–85% of accelerator revenue in 2026 (peaked ~87% in 2024). 17 bottlenecks in this layer.
• L3 Power Generation: The acute 2026 constraint is electricity itself. Gas-turbine slots carry multi-year backlogs and reservation fees from a three-OEM oligopoly, and the SMR pipeline is gated by HALEU enrichment. Power availability now caps how much compute can actually run. Dominant player: GE Vernova / Siemens Energy / Mitsubishi. Concentration: 3-OEM gas-turbine oligopoly. 6 bottlenecks in this layer.
• L4 Transmission & Grid: High-voltage and generator step-up transformers have the longest lead times of any physical component in the chain, gated by a global shortage of grain-oriented electrical steel. The interconnection queue now decides which data centres get built, not the GPU order book. Dominant player: Hitachi Energy / Siemens Energy. Concentration: HV transformer lead times measured in years. 19 bottlenecks in this layer.
• L5 Data-Centre Physical: AI rewrote the physical spec for data centres faster than the supply chain could follow. Facilities built for air-cooled racks now need liquid cooling throughout, and CDUs, cold plates, manifolds and commissioning capacity are all undersupplied. Permitting timelines add months on top. Dominant player: Vertiv / Schneider Electric. Concentration: Liquid-cooling supply chain is the binding constraint. 8 bottlenecks in this layer.
• L6 Cloud & Infrastructure Software: Cloud is the most competitive layer on the map. Three hyperscalers, a growing bench of GPU-first neoclouds and open broker markets stop any durable lock-in from forming. The constraint on cloud supply is not cloud. It is everything below it. Dominant player: AWS / Azure / Google Cloud. Concentration: Big Three ≈ 63% of cloud infra (AWS ~28%, Azure ~21%, GCP ~14%). 5 bottlenecks in this layer.
• L7 Models & Foundation Labs: Frontier closed labs hold the capability lead; open-weight models are closing the gap and undercutting price. The investable edges sit to the side: licensed training data, plus the evaluation, red-teaming and safety-assurance market that EU and US regulators are now mandating. Dominant player: OpenAI / Anthropic / Google DeepMind. Concentration: Anthropic ~40%, OpenAI ~27%, Google ~21% enterprise spend (Menlo Ventures, 2025). 1 bottleneck in this layer.
• L8 Applications & Inference: This is where the revenue has to show up. A $690B+ build-out in 2026 is being funded ahead of demonstrated end-revenue, and the gap between those two numbers is the question every investor in the layers above has to answer. Dominant player: Vertical & agentic AI. Concentration: Demand fragmented across software incumbents & startups.
• LX Cross-Cutting Enablers & Services: Capital is available; the bottleneck is the firms that deploy it. Project-finance arrangers, PPA offtake structurers and grid-interconnection specialists set the pace for every build above them. Permitting and skilled labour bind in parallel. The rate-limiter is execution capacity, not money. Dominant player: Infra equity & private credit. Concentration: Capital, permitting & skilled labour are binding constraints. 2 bottlenecks in this layer.

Chokepoints publishes one company deep dive per issue on the owners of these constraints, free, at chokepoints.ai/deepdives.