The AI Stack
Jensen's five-layer cake — Energy → Chips → Infrastructure → Models → Applications
The AI build-out, layer by layer — from the power that feeds it, to the chips that compute it, to the models and apps that capture the value. Each layer's key public companies, with the supply-chain chokepoints flagged. Curated structure; live delayed prices from public data. Not investment advice.
Value flows up — power & compute are the base; models & apps capture the higher multiple · tap a layer
Energy — Power & Cooling
The hard physical floor: power generation, electrical distribution gear (transformers, switchgear, UPS), and liquid cooling for the 50–130 kW racks of an AI factory. Power availability caps how much intelligence a region can produce.
Cleanest public pure-play on AI data-center power management + liquid cooling.
Broad electrical portfolio (power distribution, UPS/PDU) + liquid cooling; data-center orders surging.
Supplies the electrons — gas turbines + grid electrification gear for AI campuses.
Builds + connects the grid/substation power infrastructure feeding new AI data centers.
Largest US merchant/nuclear generator; signs power deals directly with hyperscalers.
Independent power producer (gas + nuclear) selling firm capacity to data-center load growth.
Nuclear-heavy IPP; pioneered co-located data-center-at-the-reactor power deals.
Chips — Compute Silicon
ChokepointThe pick-and-shovel bottleneck: merchant GPUs and custom AI ASICs, plus the upstream nobody can route around — TSMC’s leading-edge foundry + ~90% of CoWoS advanced packaging, ASML’s EUV-litho monopoly, and the sold-out HBM oligopoly. Supply, not demand, gates accelerator volume here.
Dominant merchant AI GPU + CUDA + NVLink; the reference platform for the whole stack.
Instinct MI GPUs + EPYC CPUs; the credible second source in training/inference accelerators.
Co-designs custom AI ASICs (XPUs) for the hyperscalers; ~70% of the custom-accelerator design market.
Other custom-ASIC house (Trainium, Maia) + data-center networking/optical silicon.
THE chokepoint: sole leading-edge foundry + ~90% of CoWoS packaging that gates every accelerator.
EUV-lithography monopoly — no advanced AI logic chip exists without its machines.
Only US-listed HBM supplier; HBM is a sold-out 3-player oligopoly gating GPU memory bandwidth.
CPU instruction-set IP under NVIDIA Grace + most data-center/edge CPUs — a royalty toll on compute.
Largest wafer-fab equipment maker; arms the foundries expanding leading-edge + packaging capacity.
Etch/deposition leader critical to HBM stacking + advanced-node and packaging buildout.
Infrastructure — The AI Factory
Wiring tens of thousands of chips into one machine: GPU server/rack integration, in-rack connectivity silicon, optical interconnect that beats the “copper wall” between racks, and the switch fabric. A single degraded interconnect port stalls a whole training job.
Leading GPU-server / liquid-cooled rack integrator assembling silicon into deployable AI-factory racks.
Tier-1 AI server + storage integrator shipping full GPU rack systems to hyperscalers + enterprises.
AI servers + Cray supercomputing systems and the orchestration to run them at scale.
Ethernet switching leader for AI back-end fabrics — the merchant alternative to InfiniBand.
In-rack connectivity silicon pure-play (PCIe/CXL retimers + fabric controllers) wiring GPUs to CPUs/memory.
Active Electrical Cables + SerDes connecting GPUs within/between racks; rack-scale connectivity pure-play.
Optical transceivers + silicon photonics connecting servers across the data center (the copper wall).
Optical components / lasers (EMLs) for 1.6T transceivers; named supplier in next-gen interconnect.
Coherent optical / DWDM systems for campus-to-campus data-center interconnect; record backlog.
Models — Foundation Models
The foundation models across language, biology, physics, and robotics. A breakthrough at the top drives demand all the way down the stack. The frontier leaders are largely private — public exposure comes via their compute + cloud partners.
Owns Gemini + DeepMind + the TPU stack — the cleanest public pure-frontier-model proxy.
Open-weight Llama family + massive in-house AI compute; the public open-model proxy.
Deep OpenAI partnership + in-house models + Azure model hosting — public exposure to frontier models.
Applications — Value Capture
Where the economic value is captured: copilots, agents, autonomy, and applied AI on enterprise data. Higher-multiple than the layers below — value here depends on AI-budget share, not a physical chokepoint.
Copilot across M365/Dynamics + GitHub Copilot — the broadest enterprise AI-app distribution.
Einstein / Agentforce — agentic AI embedded in the dominant CRM workflow.
Now Assist — AI agents automating IT / enterprise service workflows.
AIP operationalizes models against enterprise/government data — an applied-AI deployment platform.
Data + AI app platform (Cortex) where enterprises build/run AI on their own data.
Physical AI — FSD autonomy + Optimus humanoid robotics (Huang’s self-driving + robots example).
Observability for AI apps — the picks-and-shovels of running AI in production.
Curated structure · live delayed prices from public data · for reference only · not investment advice