AI Infrastructure

What it holds

AI Infrastructure follows the money through the AI bill of materials: the parts it takes to build it, and the companies that earn their revenue supplying them. Each company is read from its own reports, so a holding reflects what it actually sells into the theme, not what it says about it.

How it breaks down

Shares of the strategy as of the data date.

Largest companies

The market

Whole-segment generic vendors disagree about 5x for 2025 (USD 32B to 158B), probably due to scope, while Dell'Oro's 2030 capex above USD 3 trillion is broader, annual, and includes non-AI spending. Part figures overlap and differ in scope, so they should not be summed.

Questions

What is AI Infrastructure?

The whole AI data center build: chips, memory, networking, power, cooling and facilities.

What does AI Infrastructure hold?

Listed companies that earn their revenue from the parts it takes to build AI Supercycle: Graphics Processing Unit (GPU), Video Memory (VRAM), Data Processing Unit (DPU), Cooling System, Voltage Regulator Module (VRM), Motherboard Interface. 86 companies as of 2026-10-28.

What is not in AI Infrastructure?

Cold plates, vapor chambers, thermal interface materials and server level liquid loops: Cooling System, not Power and Cooling Systems; CDUs, rack manifolds, rear door exchangers, chillers and CRAH units: Power and Cooling Systems, not Cooling System Standard and RDMA NIC adapters and controllers: Network Interface and Controller, not Networking Equipment; programmable SmartNICs and DPUs with onboard compute cores: Data Processing Unit (DPU), not Network Interface and Controller Pluggable optical transceivers, AOCs, DACs and AECs: Interconnection Interface, not Networking Equipment or Racks and Cable Management; long haul and DCI optical transport systems: Networking Equipment; passive fiber trunks, patch panels and structured cabling: Racks and Cable Management

Sources

Every part of the bill of materials is researched from published industry sources, and every company from its own annual reports and exchange data. Nothing is estimated that cannot be sourced.

Industry research: acecampai.substack.com/p/superpods-are-not-bigger-clusters · ai-stack.us/products/2-248-nvidia-connectx-8-supernic-900-9x · akashborate.substack.com/p/understanding-how-to-bring-up-nvi · arxiv.org/pdf/2311.02651 · blog.barrack.ai/nvidia-b300-1400w-data-center-requirements/ · builtin.com/articles/optimize-data-backup-recovery-costs · capitalandcompute.net/memory-prices/hbm/ · convergedigest.com/delloro-ai-back-end-switch-market-to-near · dev.to/gentic_news/nvidia-feynman-gpu-power-semi-content-hit · docs.nvidia.com/enterprise-reference-architectures/nvl72-ai- · electronics.alibaba.com/buyingguides/42u-server-rack-price-g · epoch.ai/data-insights/b200-cost-breakdown · finance.biggo.com/news/X4vnUp4B2jrwCtglf11J · finance.yahoo.com/news/cooling-system-single-nvidia-blackwel · finance.yahoo.com/real-estate/articles/data-center-construct · finance.yahoo.com/technology/ai/articles/nvidia-already-make · fourweekmba.com/ai-nvidias-huang-pegs-software-at-about-2000 · fourweekmba.com/hbm-the-memory-that-controls-45-of-gpu-econo · fourweekmba.com/nvidia-compute-revenue/ · github.com/bmsdebug/chinascrape/issues/23

What it is not

Where an item could belong to two parts, these rules decide which owns it; items outside every part are not counted.

• Cold plates, vapor chambers, thermal interface materials and server level liquid loops: Cooling System, not Power and Cooling Systems; CDUs, rack manifolds, rear door exchangers, chillers and CRAH units: Power and Cooling Systems, not Cooling System
• Standard and RDMA NIC adapters and controllers: Network Interface and Controller, not Networking Equipment; programmable SmartNICs and DPUs with onboard compute cores: Data Processing Unit (DPU), not Network Interface and Controller
• Pluggable optical transceivers, AOCs, DACs and AECs: Interconnection Interface, not Networking Equipment or Racks and Cable Management; long haul and DCI optical transport systems: Networking Equipment; passive fiber trunks, patch panels and structured cabling: Racks and Cable Management
• Finished Ethernet and InfiniBand switch systems: Networking Equipment, not Interconnection Interface; merchant switch ASICs and InfiniBand silicon sold as chips: Interconnection Interface, not Networking Equipment
• Board level VRMs, power stages and PMICs: Voltage Regulator Module (VRM), not Power and Cooling Systems
• Server PSUs, rack power shelves and busbars: Power and Cooling Systems, not Servers and IT Equipment, Voltage Regulator Module (VRM) or Racks and Cable Management; power semiconductors inside them counted only within the PSU price
• Server motherboards and GPU baseboards: Motherboard Interface, not Servers and IT Equipment
• ODM rack scale assembly and test value: Servers and IT Equipment, not Racks and Cable Management or Motherboard Interface; rack enclosure hardware itself: Racks and Cable Management
• High speed copper backplanes, scale up cable cartridges and high speed connectors: Interconnection Interface, not Racks and Cable Management
• NVLink and other scale up switch chips: Interconnection Interface, not Graphics Processing Unit (GPU)
• PCIe and CXL retimers and switches: Motherboard Interface, not Interconnection Interface; Ethernet and optical DSPs and standalone SerDes IP: Interconnection Interface
• Custom accelerator design services and IP embedded in AI ASICs: Graphics Processing Unit (GPU), not Interconnection Interface; standalone SerDes IP licensing: Interconnection Interface
• HBM stacks including TSV stacking and base die: Video Memory (VRAM), not Graphics Processing Unit (GPU); interposer and 2.5D integration: Graphics Processing Unit (GPU)
• Server DRAM modules: Video Memory (VRAM), not Servers and IT Equipment
• CPUs including those bundled in GPU superchip modules: Servers and IT Equipment, not Graphics Processing Unit (GPU)
• General purpose server OS, virtualization and cluster orchestration: Datacenter Software, not Software (AI hardware performance); GPU drivers, CUDA, ROCm, AI compilers and inference runtimes: Software (AI hardware performance)
• SSD controllers and RAID or NVMe Storage Devices controllers: Storage Devices, not Data Processing Unit (DPU) or Motherboard Interface
• Backup, replication and recovery software and appliances: Backup and Disaster Recovery Solutions, not data center security software or Storage Devices
• Separately licensed DCIM and monitoring software: Datacenter Software, not Power and Cooling Systems

Not investment advice. Figures are dated and sourced; ThematIQ does not estimate figures it cannot source.