The 26 Layers of AI Infrastructure

EDA & Chip IP

Key insight: Synopsys + Cadence global duopoly. Every advanced chip requires their tools.

How it connects: SNPS/CDNS supply design tools to every chip company. Without EDA software, no chip designs exist. Every custom ASIC for AI requires EDA. ARM licenses CPU architecture used in virtually every chip.

Semiconductor Equipment

Key insight: ASML is a single point of failure for all advanced chips. 2-year lead times on EUV machines.

How it connects: ASML supplies EUV lithography exclusively to TSMC, Samsung, Intel. AMAT, LRCX, KLAC supply etch, deposition, and inspection equipment. Every chip manufactured in L04 depends on this equipment. ENTG (L03) supplies materials consumed by this equipment.

Process Materials & Specialty Chemicals

Key insight: Ultra-pure materials with long lead times. Entegris near-monopoly in key filtration.

How it connects: Materials flow into L02 equipment and L04 foundries. Disruption here halts fab production within weeks. New fabs (TSMC Arizona, Intel Ohio) each require parallel materials supply chains.

Foundries & Fabrication

Key insight: TSMC manufactures 90%+ of advanced AI chips. Single point of failure for the entire AI supply chain.

How it connects: TSMC takes chip designs from L01, uses equipment from L02 and materials from L03 to manufacture silicon wafers. Every GPU, CPU, and AI accelerator in L06 comes through this layer. TSMC also performs advanced packaging (L05) integrating GPU dies with HBM memory.

Advanced Packaging & Test

Key insight: TSMC CoWoS sold out through 2026. Packaging is where GPU meets HBM — the new constraint.

How it connects: This is where GPU dies (from L04) get integrated with HBM memory stacks. TSMC CoWoS handles the most advanced packaging. Amkor and ASE handle lower-tier packaging and testing. Output flows to L06 as finished chips.

Compute & AI Silicon

Key insight: $300B+ hyperscaler capex flows through this layer. NVIDIA 80%+ AI accelerator share.

How it connects: Every chip here depends on TSMC (L04) for manufacturing, ASML (L02) for lithography, HBM (L08) for memory, and packaging (L05). These finished chips flow to server assembly (L13) where they become part of complete servers.

Data Storage Systems

Key insight: Storage capacity can be added incrementally. AI training datasets growing faster than any other data category.

How it connects: SSDs and hard drives plug into the server motherboard alongside GPUs (L06) and system memory (L08). AI training requires rapid random access to massive datasets. Storage connects to GPU clusters through network switches (L15).

System DRAM & Memory IP

Key insight: HBM is inside the GPU package (L05). System DRAM sits on the motherboard separately. SK Hynix 62%+ HBM share.

How it connects: System DRAM from Micron, Samsung, SK Hynix sits on the server motherboard separately from HBM (which is inside the GPU package). RMBS licenses memory interface IP used by every memory chip. Memory connects to CPU and GPU through the motherboard.

Voltage Regulators & Power Management ICs

Key insight: 48V rack power transition driven by physics. GaN replacing silicon. MPWR dominant in AI server voltage regulators.

How it connects: Voltage regulators sit on the server motherboard next to GPUs and CPUs, converting power supply output to the 50+ precise voltages each chip needs. MPWR, TXN, ADI supply these chips. Without them, GPUs cannot function.

Component-Level Thermal Management

Key insight: Liquid cooling mandatory for AI GPUs. Cold plates sit directly on GPU packages. Distinct from building-level HVAC (L19).

How it connects: Cold plates and liquid cooling loops sit directly on GPU packages inside the server. This is chip-level thermal management — distinct from building-level HVAC and chillers in L19. As GPU power consumption rises past 1000W per chip, component-level cooling becomes critical.

Optical Components & Transceivers

Key insight: NVIDIA mandates 1.6T optical for every GB300 NVL72 rack. EML laser supply pre-allocated through 2027.

How it connects: Optical transceivers convert electrical signals from GPUs to light for fiber optic transmission. Coherent and Lumentum make the transceiver modules. MACOM supplies InP photodetectors. These modules plug into network switches (L15) and connect servers across the data center.

Timing & Clock Generation

Key insight: SiTime MEMS technology is superior for high-vibration AI data center environments. Only US-listed pure play.

How it connects: Timing chips sit on the server motherboard and provide clock signals that synchronize all other components — GPU, CPU, memory, network interfaces. Without precise timing, components cannot communicate. SiTime MEMS timing is displacing legacy quartz oscillators.

Server OEMs & Systems Assembly

Key insight: Multiple competitors. GPU supply from NVIDIA is the real constraint — not assembly.

How it connects: Server OEMs take finished GPUs (L06), system memory (L08), storage (L07), power regulators (L09), cooling (L10), timing (L12), and transceivers (L11) and solder them onto motherboards. Dell, HPE, SMCI compete for NVIDIA GPU allocation.

Server Power Supplies

Key insight: Key manufacturers are Taiwan/Japan-listed. Bel Fuse is the primary US-listed player. Growing bottleneck as AI rack power exceeds 100kW.

How it connects: Server power supplies convert electricity from the data center's power distribution (L23) into DC power usable by the server motherboard. The PSU feeds voltage regulators (L09) which then deliver precise power to each chip. As AI racks exceed 100kW, PSU design becomes critical.

Network Switching & Routing

Key insight: Multiple credible competitors. Transition to 800G/1.6T Ethernet driving complete network refresh.

How it connects: Network switches from Arista and Cisco connect servers within the data center. Optical transceivers (L11) plug into switch ports. Fiber cables (L18) carry the signals between switches. These switches route AI training traffic between thousands of GPUs.

Data Center Construction & Engineering

Key insight: General labor available but specialized DC construction expertise concentrated.

How it connects: Construction depends on power (L20-L23) being secured first. Contractors need specialized electrical subs (L17) and cooling systems (L19). Design-build firms like Fluor and AECOM handle the entire data center construction process.

Electrical & Mechanical Contractors

Key insight: 100,000+ licensed electrician shortage. EMCOR $13.25B backlog.

How it connects: Electrical contractors install power distribution equipment from L23 (switchgear, PDUs) using materials from L18 (wire, cable, connectors). The binding constraint is LABOR — licensed electricians are scarce nationwide.

Cables, Fiber & Connectors

Key insight: Corning dominates fiber. Amphenol dominates high-speed connectors. Copper availability a concern.

How it connects: Cables and connectors physically link everything inside the data center — servers, networking, power distribution. Corning fiber connects to optical transceivers (L11). Amphenol/TE connectors go into servers (L13) and power systems (L23).

Building-Level HVAC & Cooling Systems

Key insight: Mandatory liquid cooling for AI racks. Vertiv backlog growing 100%+ YoY. Distinct from chip-level cooling (L10).

How it connects: Building-level cooling systems — chillers, cooling towers, CDUs — remove heat from the entire facility. This is distinct from chip-level cold plates (L10) which sit directly on GPUs. Vertiv, Carrier, and Trane supply these building-scale systems.

Natural Gas Pipelines & Fuel Supply

Key insight: Pipeline capacity constrains where gas-fired power plants can be built. Permitting for new pipelines takes years.

How it connects: Natural gas pipelines deliver fuel to gas-fired power plants (L21). Without fuel, turbines don't turn. KMI transports 40% of US LNG export supply. WMB's Southeast Supply Enhancement serves data center power demand. Pipeline routes constrain where new data center power can be built.

Utilities, Nuclear & Gas Turbines

Key insight: Gas turbine lead times 4-5 years. Nuclear restarts 3-7 years. No fast path to new power.

How it connects: Power plants burn gas piped through L20, split uranium, or capture renewables to generate electricity. Output flows to L23 (Deliver the Power) for distribution to data centers. Hyperscaler PPAs are driving unprecedented power demand. Nuclear (CEG, OKLO, SMR) and gas turbines (GEV) are the primary paths to new capacity.

Energy Storage, On-Site Generation & Grid Independence

Key insight: Data centers need grid independence. Batteries, fuel cells, and on-site generation solve reliability and peak demand.

How it connects: These companies solve grid reliability — what happens when the grid isn't enough or isn't reliable. Bloom Energy generates power on-site via fuel cells. Fluence provides battery storage. Stem optimizes when to draw from grid vs storage. All serve the same customer need: power security for data centers.

Transformers, Switchgear & Electrical Distribution

Key insight: Large transformer lead times 2-3 years. Custom switchgear 12-18 months.

How it connects: Grid equipment delivers power from generators (L21) to the data center. Transformers step down voltage. Switchgear distributes and protects the flow. UPS systems provide backup. Every megawatt of data center capacity requires corresponding electrical distribution infrastructure.

Data Center Operators & Powered Campus Developers

Key insight: Former Bitcoin miners pivoting to AI infrastructure. Stranded power assets being repurposed for Tier 3/4 AI data centers.

How it connects: These operators develop powered campuses — they acquire land, secure power, build or lease facilities, and provide ready-to-use infrastructure to hyperscalers and AI companies. NUAI delivers NNN leased powered sites. IREN, WULF, CORZ pivot stranded Bitcoin mining power to AI compute. They depend on power (L20-L23), construction (L16-L19), and server components (L07-L15).

Data Center REITs & Colocation

Key insight: Vacancy near zero in top markets. Rental rates rising 15-30% YoY.

How it connects: REITs own and operate finished data center buildings. They lease space to hyperscalers and enterprises. Equinix and Digital Realty dominate colocation. They depend on construction (L16-L19) and power (L20-L23) to deliver ready facilities. Hyperscalers are both their customers and increasingly their competitors (self-build).

Hyperscalers & AI Cloud

Key insight: Capital is not the constraint — physical infrastructure is. $300B+ annual hyperscaler capex looking for places to land.

How it connects: These are the DEMAND SIGNAL. When Microsoft announces $80B capex, every company in L01 through L25 feels it. Capital flows from here backward through the entire chain. Amazon, Microsoft, Google, Meta, and CoreWeave are the ultimate customers pulling the entire AI infrastructure supply chain.