Key Suppliers

Key suppliers of AI tech

The AI cycle is shaped by a few chokepoints. Compute, memory, networking, optics, power, cooling, and advanced packaging set the pace. Use this map to understand who makes what and where supply risk can appear.

Compute silicon

• Accelerators Nvidia, AMD, Intel. Focus areas include tensor cores, interconnect bandwidth, software stacks such as CUDA and ROCm, and total board power.
• CPUs and DPUs or IPUs Intel, AMD, Nvidia, Broadcom, Marvell. Useful for orchestration, security, storage offload, and network processing near the GPU.
• Switch and NIC silicon Broadcom for Tomahawk and Jericho lines, Marvell, Nvidia for InfiniBand and Ethernet NICs.

Foundry and advanced packaging

• Leading edge wafers TSMC with focus on 3 nanometer and below nodes. Samsung Foundry and Intel Foundry Services as additional sources.
• Advanced packaging TSMC CoWoS and SoIC, Intel Foveros, Samsung X Cube. These determine how many accelerator boards can ship each quarter.
• Substrates and ABF materials Ibiden, Unimicron, Nan Ya, Shinko Electric, Ajinomoto Fine Techno for ABF films. Substrate layers often gate ramp speed.

Memory and storage

• HBM stacks SK hynix, Samsung, Micron. Watch bit growth, HBM3E mix, and known good stack rates since HBM often binds supply.
• DDR and LPDDR Samsung, SK hynix, Micron. Important for servers and edge devices that pair with accelerators.
• Enterprise storage Kioxia, Samsung, Western Digital, Seagate. NVMe throughput and endurance influence training data pipelines.

Networking and optics

• Switch systems and optics Arista, Cisco, Juniper for systems. Coherent, Lumentum, Innolight, Accelink for optical modules and transceivers.
• Interconnect choices Ethernet at 400 G and 800 G for training and inference fabrics, InfiniBand for some training clusters. Optics availability and switch silicon roadmaps drive cluster build speed.
• Cables and connectors Amphenol, Molex, TE Connectivity for high speed copper and fiber assemblies.

Servers and integrators

• Server makers Supermicro, Dell, HPE. Differentiation in thermal design, board layout, and serviceability matters for dense racks.
• Original design manufacturers Quanta, Wiwynn, Inventec, Foxconn. These build many cloud designs and set lead times for volume builds.
• Global integrators Accenture, Deloitte, Infosys, TCS. Useful for enterprise rollout, data plumbing, and change management.

Power, cooling, and data center build

• Power delivery and backup Eaton, Schneider Electric, Vertiv, ABB, Generac, Cummins. Track utility interconnect lead times and substation upgrades near cloud regions.
• Thermal management Vertiv, CoolIT, Asetek, ZutaCore, GRC, Submer, Iceotope, LiquidStack. Air to liquid transitions and immersion options define achievable rack density.
• Racks and distribution Legrand, Panduit, Rittal. Cable management and airflow influence reliability at scale.

Chip equipment and materials

• Lithography and deposition ASML for EUV systems. Applied Materials, Lam Research, KLA, Tokyo Electron, SCREEN for deposition, etch, and inspection.
• Packaging tools Besi, ASMPT, Kulicke and Soffa. These enable bumping, bonding, and advanced substrates for CoWoS class builds.
• Chemicals and photoresists JSR, Shin Etsu, Sumitomo Chemical, DuPont, Tokyo Ohka Kogyo. Materials availability can gate both wafer and packaging output.

EDA and IP

• Design tools Synopsys, Cadence, Siemens EDA. Tool availability and tape out timing affect new accelerator cycles.
• CPU and GPU IP Arm for CPU cores, Imagination and others for graphics and compute IP in custom SoCs.

Cloud capacity and orchestration

• Cloud platforms AWS, Azure, Google Cloud, Oracle Cloud. Capacity, regional availability, and managed inference platforms shape access for builders.
• Software stacks CUDA, ROCm, Triton inference server, Kubernetes with GPU scheduling, Ray, and popular agent frameworks that coordinate tool use.

Where bottlenecks usually appear

• Advanced packaging and substrate capacity often limit accelerator shipments even when wafer starts are healthy.
• HBM stack yields can bind server builds and influence price for full systems.
• Optical modules for 800 G links can delay cluster turn up and reduce utilization.
• Power and cooling upgrades near cloud regions can extend lead times for new capacity.

How to monitor supply risk

1. Track supplier commentary on capacity, yield, and mix during earnings calls and industry events.
2. Watch regional build activity and utility announcements for power access and permitting timelines.
3. Follow roadmaps for memory, packaging, and switch silicon since these dictate feasible cluster sizes and cost.
4. Maintain second sources for key parts when possible and align contracts to expected ramp windows.