chokepoints.ai
SUBSCRIBE
10 layers580 nodes2,376 dependencies95 chokepoints6,500+ companiesnode size = companies identified

THE STACK

The AI compute stack, layer by layer.

The chain runs from raw materials at the bottom to applications at the top. Between extraction and the data centre it splits into a silicon track and an energy track. The diagram shows the full flow; each layer opens into its own detail.

L0Raw Materials & Extractionchokepoint

Every chip, cable and magnet in this atlas begins as a rock or a chemical pulled from the ground somewhere. This layer covers the mining, refining and processing of the physical inputs the whole chain depends on: rare earths for magnets, silicon and specialty gases for fabs, copper for wiring, and the high-purity quartz that is essentially a single-source dependency for semiconductor production. The chokepoints are not in the mines but in the refining plants, most of which sit in China.

L5Data-Centre Physicalchokepoint

Once the power arrives, it still has to be distributed safely across a building full of servers and then removed again as heat. This layer covers the physical data-centre shell: site selection, construction, electrical distribution, and cooling. AI racks run far hotter than conventional servers, forcing a shift from air cooling to direct liquid cooling throughout the facility. Capacity is being added at historic speed, but liquid-cooling hardware, medium-voltage electrical gear and specialist commissioning teams are all supply-constrained.

L6Cloud & Infrastructure Software

Physical servers become useful only when software turns them into something a developer or business can actually call. This layer covers the platforms and tooling that do that work: the hyperscaler clouds (AWS, Azure, Google Cloud) offering on-demand GPU clusters, the GPU-first neoclouds like CoreWeave serving overflow demand, and the orchestration, inference engines, data pipelines and monitoring tools that keep workloads running. Unlike the layers below, this one is vigorously competitive.

L7Models & Foundation Labs

The large language models that power most AI products today are produced by a small number of organisations running some of the most expensive training runs ever attempted. This layer covers the frontier closed labs (OpenAI, Anthropic, Google DeepMind), the open-weight releases that challenge them (Meta's Llama, Mistral), specialist fine-tuning providers, training-data suppliers, and the growing evaluation and safety ecosystem that regulators in the US and EU are beginning to mandate. Capability here is what ultimately determines the value of every layer below.

L8Applications & Inference

This is where the AI supply chain meets paying customers. It covers the products and platforms that put AI in front of end users: consumer apps like ChatGPT, vertical software for healthcare, legal and finance, autonomous agents that complete multi-step tasks, and AI features embedded into incumbent platforms like Microsoft 365 and Salesforce. Everything above this layer was built in anticipation of demand here, which makes this layer the one investors across the whole chain are watching most closely.

LXCross-Cutting Enablers & Services

INTERACTIVE ATLAS

Each layer in detail

For each layer: who controls it, the thesis, current signals, the companies we track, and the supply chain feeding it.

L2 · CHOKEPOINT

Compute Hardware

This is the layer most people picture when they think about the AI supply chain: the chips, servers, networking gear and memory that actually do the computing. NVIDIA's GPU dominance is the headline, but the full picture includes custom accelerators built by Google and Amazon, high-bandwidth memory from SK Hynix and Micron that stacks inside each GPU package, and the optical interconnects that wire racks together at speed. Each of those components has its own tight supply chain.

L2INPUTSL2BUYERSCOMPUTE HARDWARESUPPLY SCHEMATIC · L2

WHY IT'S A CHOKEPOINT

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.

Signals

  • SK Hynix targets ~200k wafers/month of HBM by 2026, with output fully sold out (TrendForce).
  • HBM4 prices ~29% above HBM3E and HBM4e ~61% higher, reflecting severe supply tightness (TrendForce).
  • TSMC CoWoS reaches ~125-130k wafers/month by end-2026 (toward ~220k by 2028), yet a 15-20% supply gap persists (TrendForce).
  • Marvell holds ~70% of 800G optical DSPs and ~50% of 1.6T, the interconnect feeding the GPU (TrendForce).

The investment angle

HBM is structurally supply-constrained through at least 2027, giving SK Hynix and Micron pricing power; in-rack power delivery (48V/800V) is the overlooked bottleneck, with exposure via Vertiv, Eaton and Infineon.

Dominant playerNVIDIA
Concentration~80–85% of accelerator revenue in 2026 (peaked ~87% in 2024)
Key metricHBM and optics are the binding constraints behind the GPU
GeographyUS / Korea / Taiwan

Companies we track

NVIDIA
dominant GPU; controls CoWoS + HBM allocation
NVDA · US
Broadcom
custom AI ASICs; networking + optics silicon
AVGO · US
SK Hynix
~62% HBM share; 2026 capacity sold out
000660.KS · KR
Micron Technology
HBM3E ramp; only US-based HBM supplier
MU · US
AMD
GPU challenger; MI300X / MI355X
AMD · US
Marvell Technology
custom silicon + optical interconnect DSPs
MRVL · US

Supply chain

Raw inputs

SK Hynix HBM3EKR
dominant HBM supplier
Coherent / LumentumUS
lasers & optical components
TSMC CoWoSTW
sole AI-grade packaging

Key suppliers

NVIDIAUS
~80–85% (2024 peak ~87%)
AMDUS
~10% and rising (MI300X/MI355X)
BroadcomUS
custom XPU / switch fabric

Buyers

HyperscalersUS
largest GPU buyers
Neoclouds (CoreWeave, Nebius)US/NL
GPU-first operators