AI Supply Chain
The full value flow from EDA, chips and fabrication through to cloud, foundation models and end applications
The AI supply chain strings together everything needed to train and run large models into a single value chain. Upstream sits EDA tooling, semiconductor equipment and materials, AI accelerator chips and high-bandwidth memory. Midstream covers wafer foundry, advanced packaging, AI servers, optical modules, network interconnect, data centers and cloud compute platforms. Downstream are the foundation models, AI application software and on-device/edge deployment. Value propagates bottom-up under the pull of compute demand: a surge in applications and models drives cloud and server expansion, which in turn pulls chips, memory, equipment and materials. The more exclusive a chokepoint (EUV lithography, leading-edge process, the CUDA ecosystem, HBM), the more excess profit it captures.
EDA & Semiconductor IP
EDA (electronic design automation) is the digital blueprint factory that takes a chip from concept to tape-out; logic synthesis, place-and-route and verification for every leading-edge chip depend on it. IP cores are reusable design building blocks, and AI chips broadly license architectures such as Arm. The big three EDA vendors form a de facto chokepoint: without their software you cannot complete a 5nm/3nm-and-below design, and the moat runs deep.
Number one in global EDA share, with design tools spanning the full digital/analog flow plus DesignWare high-speed interface IP; core infrastructure for AI chip design.
Number two in global EDA share, leading in analog/mixed-signal and package simulation; Virtuoso is the standard tool for analog chip design, and the company also supplies interface/memory IP.
The dominant CPU/NPU instruction-set architecture licensor; the vast majority of mobile AI chips and a large share of data-center inference chips are built on Arm.
Its Siemens EDA unit (formerly Mentor) is one of the big three, supplying PCB, IC packaging and digital-twin verification, with a strong hold in system-level and automotive semiconductor verification.
The clear domestic-Chinese EDA leader and the only Chinese vendor offering a full-flow EDA tool system for analog circuit design, covering four platforms: analog, digital, flat-panel display and RF.
China's largest semiconductor IP licensing and chip-customization platform, developing its own GPU/NPU/DSP IP and taking on AI chip design services; a key chokepoint on the path to a domestic substitute for Arm/Cadence IP.
Semiconductor Manufacturing Equipment
The core machines that turn silicon wafers into chips, spanning lithography, etch, thin-film deposition and metrology, and directly determining which process node is reachable. AI drives leading-edge nodes (3nm/2nm) and HBM capacity expansion, keeping equipment demand hot; the EUV lithography machine is supplied by ASML alone, the single strongest chokepoint.
The world's only EUV lithography maker; there is no way around it at 7nm and below, and it also dominates high-end DUV, leaving leading-edge capacity rigidly dependent on it.
The largest semiconductor equipment company by revenue, covering thin-film deposition, CMP, ion implantation and metrology; the single largest supplier by fab procurement share.
Number one in global etch share (about 45%), nearly irreplaceable in 3D NAND multi-layer etch and high-aspect-ratio structures; HBM stacking capacity directly drives demand.
Leader in global process control (metrology and inspection), at roughly 50% share; defect detection grows exponentially harder at leading-edge nodes, and KLA's share of spend on every advanced line keeps rising.
Japan's largest semiconductor equipment maker and number three globally, strong in coat/develop, plasma etch and atomic-layer deposition; indispensable in EUV ancillary steps and the new GAA process.
The leader in single-wafer cleaning equipment globally, a key supplier of wet-process tools for front-end cleaning and coat/develop.
China's broadest semiconductor equipment platform, covering etch, deposition, diffusion/oxidation and cleaning; a core pillar for full-chain localization at domestic fabs.
China's etch and MOCVD equipment leader, whose CCP/ICP etchers have entered top-tier fab supply chains including TSMC and SMIC; the domestic equipment maker closest to the international frontier.
The leader in single-wafer atomic-layer deposition (ALD) globally, also making epi, CVD and diffusion front-end deposition equipment; a key supplier for leading logic and GAA transistors.
A specialist in ion-implantation equipment; the Purion platform spans power, memory and logic, with a strong hold in high-energy implant for power semiconductors.
A near-monopoly in EUV photomask inspection, owning ACTIS, the world's only commercial actinic (13.5nm same-wavelength) mask inspection system.
The global leader in wafer dicing, grinding and polishing equipment; a key supplier of thinning and singulation tools for back-end and advanced packaging.
The leader in advanced-packaging die-attach and hybrid-bonding equipment; the core assembly supplier for AI/HBM/Chiplet 2.5D-3D integration.
A back-end lithography coat/develop, wafer bonding and photomask cleaning equipment maker, focused on advanced packaging, MEMS and power devices, with over 6,000 tools installed worldwide.
A front-end and advanced-packaging process-control (metrology + inspection) equipment maker; its Dragonfly platform holds a strong position in advanced-packaging defect detection.
Semiconductor Test Equipment
Verifies electrical performance, speed and burn-in for bare die and packaged parts before volume production: the final gate on AI accelerator and HBM yield and reliability. Automated test equipment (ATE) is highly concentrated in the Advantest/Teradyne duopoly.
The global leader in SoC and memory testers and a key supplier for AI chip volume test; the core provider of final-test equipment for leading GPUs and HBM.
One of the two ATE giants, led by its UltraFLEX platform, covering SoC and memory (including HBM) test, with a foothold in collaborative robots via Universal Robots.
The world's number one supplier of wafer probe cards, leading in advanced probe cards (over 85% of the total probe-card market).
A designer and maker of non-memory/SoC probe cards, the world's second-largest probe-card vendor, supplying mainly top tech and foundry customers.
The global leader in back-end test handlers, also making test interfaces/contactors, thermal-control subsystems and inspection/metrology automation.
A test-and-measurement instrument maker covering system-level test (SLT), power-device test and semiconductor/electronics measurement, tied to AI server and power-semiconductor test demand.
Semiconductor Materials
The raw-material base of wafer fabrication, spanning silicon wafers, photoresist, specialty gases, CMP slurry and photomasks, and directly determining yield, process window and final performance. Material categories are many and fragmented; Japanese firms have long held oligopolies in high-end silicon wafers, photoresist and developers, while domestic-Chinese substitution is still early but accelerating.
Number one in global semiconductor silicon-wafer share (about 30%), one of the two 300mm large-wafer giants with SUMCO; AI compute capacity directly drives wafer demand.
Number two in global semiconductor silicon-wafer share (about 26%), focused on 300mm polished and epitaxial wafers; with Shin-Etsu it controls roughly 56% of the world market.
One of the world's three major photoresist suppliers, with volume capability in both ArF and EUV resists and a roadmap toward next-generation process resists.
Number one in global photoresist share and a leading EUV/ArF supplier; taken private by the Japanese government-backed fund JIC and operated as a strategic national material asset.
Number one in global CMP slurry share, also supplying ultra-pure chemicals, filtration/purification systems and advanced-packaging materials; leading-edge nodes and 3D stacking directly drive demand.
The world's third-largest semiconductor silicon-wafer supplier; a leading provider of 300mm/200mm substrates upstream of logic and memory chips.
China's largest semiconductor silicon-wafer maker, focused on localizing 300mm large wafers; a core path for domestic fabs such as SMIC to reduce reliance on Japan.
China's CMP slurry leader, covering copper/tungsten/dielectric applications and moving into functional wet chemicals; a core substitute for Entegris/Cabot.
China's electronic specialty-gas leader, whose specialty gases serve over 90% of domestic 8-inch-and-above fabs, with some products qualified into 14nm/7nm/5nm leading-edge lines.
An integrated gas provider, entering semiconductor materials/fabrication/packaging via ultra-pure ammonia, high-purity nitrous oxide, electronic-grade TEOS and high-purity carbon dioxide, and pushing into on-site bulk carrier-gas generation.
A fluorinated electronic specialty-gas leader under CSSC (Peric), ranking number one globally in both NF3 and 6N-grade WF6 capacity; customers include TSMC, Samsung, Micron, SMIC and YMTC.
One of the two global industrial-gas powers, supplying fabs with electronic bulk gases (N2/H2/He), electronic specialty gases, precursors and on-site air-separation units; the on-site gas mainstay for leading-edge capacity expansion.
The world's largest industrial-gas company, whose electronics business supplies fabs with bulk gases, electronic specialty gases and on-site air separation; tied to new leading-edge capacity such as TSMC Arizona.
Merck's Electronics segment (including the specialty-gas and ultra-pure chemical lines acquired with Versum), supplying a full set of semiconductor materials across electronic specialty gases, CVD/ALD precursors, photoresist and wet chemicals; one of the few platform-scale material players.
One of the two mask-blank giants, with over 75% of EUV mask-blank shipments, and currently the only supplier with a qualified High-NA EUV mask-blank product; unavoidable at leading logic of 2nm and below.
Number one in global mask-blank share, with over 59% of total blanks; together with Hoya it holds roughly 93% in a duopoly, supplying ultra-pure low-expansion glass substrates for DUV and EUV masks.
One of the leaders in high-end ABF/FC-BGA package substrates, specializing in high-layer, high-density-interconnect substrates for AI/HPC; a core supplier of carriers for top processors such as NVIDIA GPUs and Intel.
Number one in global ABF substrate share (about 22-28%), capable of substrates up to 32 layers, taking roughly 30% of NVIDIA AI accelerator carrier orders, with high-end capacity full under the inflow of large AI ASIC orders.
One of Taiwan's three major IC substrate makers, supplying ABF/FC-BGA and FC-CSP package substrates, with roughly 15% of NVIDIA AI accelerator carrier orders.
A Formosa Plastics Group ABF/FC-BGA substrate maker; ABF products are about 60% of revenue, and it holds roughly 10% of NVIDIA AI accelerator carrier orders, among the names with the highest pricing leverage in this AI substrate cycle.
One of the leaders in high-end ABF/FC-BGA substrates and lead frames, at roughly 12-13% global ABF share, supplying mainly HPC/server/automotive high-end substrates; taken private and delisted by Japan's JIC in June 2025.
A first-tier global commercial photomask maker and the only US-based commercial mask shop with high-end process capability. The photomask is the master for transferring the pattern of every lithography layer; the more layers at leading edge and the more frequent the AI-chip tape-outs, the more masks each design consumes, making it a key consumable on the fabrication side.
The global shipment leader in rigid copper-clad laminate (CCL), supplying base material for high-multilayer PCBs in AI servers, switches and routers. High-speed/high-frequency CCL sets signal transmission rate and loss; AI compute is driving a surge in demand for low-loss material (M6/M7/M8 grades), and the company's high-end high-speed products have passed qualification with leading global AI server makers.
A maker of high-multilayer, high-speed PCBs for AI servers/HPC and high-speed network switches, processing high-speed CCL into GPU/switch motherboards and other high-layer boards. Already designed into multiple North American major-customer supply chains; one of the core PCB suppliers for high-end AI server boards such as NVIDIA's GB200.
The global leader in high-speed/high-frequency copper-clad laminate (CCL), one of NVIDIA's three AI-server CCL suppliers, supplying low-loss base material for GPU boards, midplanes and 800G/1.6T switches, with high-end grades such as M8/M9 in volume. Note: per late-2025 reports, qualification of its next-generation Rubin platform GB300 hit setbacks, and exclusive Rubin CCL may shift to Doosan, leaving its next-generation share uncertain.
AI Accelerator Chips / GPU
The heart of AI compute, handling the parallel math of large-model training and inference. The GPU is the mainstream form, while custom ASICs (XPUs) are rising fast as the core path for hyperscalers to cut cost. The moat comes from ecosystem lock-in (software stack, interconnect protocols) and exclusive access to leading-edge capacity.
The undisputed king of GPU compute, dominating the global AI training market; the CUDA ecosystem forms the deepest software moat.
The largest custom AI ASIC (XPU) supplier, designing custom chips for hyperscalers such as Google TPU and Meta MTIA; also the largest networking/switch chip supplier.
Huawei's chip design arm; the Ascend line is China's strongest homegrown GPU alternative, paired with the MindSpore software stack.
The number two in GPU compute; the Instinct MI line enters large-model inference on high HBM capacity, with the ROCm stack steadily chasing CUDA.
A supplier of custom ASICs and AI interconnect chips, designing custom XPUs such as Amazon Trainium and also making optical-interconnect DSP.
A flagship of China's homegrown AI chips; the MLU line of training/inference chips serves mainly domestic cloud and government/enterprise compute demand.
A supplier of domestic x86-compatible CPUs plus DCU (deep computing units); the DCU supports full-precision AI compute and is ROCm-compatible, serving mainly finance and telecom.
A wafer-scale AI chip (WSE) maker, using an enormous single-die area to minimize AI inference latency and take on hyperscale inference compute orders.
A supplier of the Gaudi line of AI accelerators, leaning on its IDM model for advanced packaging (EMIB/Foveros) and competing on value to enter the open ecosystem.
A maker of LPU (language processing unit) architecture chips, built for very low inference latency, commercialized through the GroqCloud inference-as-a-service model.
A domestic AI compute chip designer on a full-function GPU path, founded by NVIDIA's former China GM; products span AI compute, graphics acceleration and intelligent SoCs.
A GPGPU designer founded by ex-AMD talent, focused on general-purpose AI training/inference compute; its Xiyun line targets NVIDIA's high-end training cards.
A domestic high-end GPGPU designer and the first mainland-China GPU maker to list in Hong Kong, focused on general-purpose AI compute solutions.
A domestic general-purpose GPU company, supplying GPU chips, accelerator cards and custom AI compute solutions (the Tianhai/Zhikai lines).
Memory Chips (HBM / DRAM / NAND)
The memory subsystem of the AI accelerator. HBM uses 3D stacking to connect DRAM dies directly to the GPU/ASIC, delivering several TB/s of bandwidth, and is the key bottleneck for large-model training and inference. The market is highly concentrated among three firms, where process barriers and scarce capacity drive excess profit.
The global HBM leader, the sole volume supplier of HBM3E to flagship AI accelerators, roughly 1-2 generations ahead of peers.
The world's largest DRAM and NAND maker and an HBM supplier, with logic-chip foundry capability as well.
The only US full-stack DRAM/NAND maker, supplying HBM3E to flagship AI accelerators and rapidly closing the HBM share gap.
China's only volume DRAM maker, focused on DDR4/LPDDR5 localization; a core supplier for domestic AI server memory.
China's largest NAND flash maker, focused on consumer storage and domestic data-center SSDs; export controls limit its access to leading-edge technology.
One of the world's NAND flash leaders (formerly Toshiba Memory, BiCS FLASH 3D NAND); rather than DRAM-type HBM, it enters AI storage via high-bandwidth flash (HBF) and GPU-direct high-speed SSDs as a complement/alternative to HBM.
Wafer Foundry
Foundries take orders from fabless chip designers and run hundreds of process steps to turn circuit patterns into working chips. Leading-edge nodes (3nm/5nm and below) are intensely capital-heavy with very high moats; only TSMC and Samsung have volume capability. The mid-range is split among Intel, SMIC, UMC and GlobalFoundries.
The undisputed monopoly at leading-edge nodes, also dominating CoWoS/SoIC advanced packaging; the sole volume source of AI compute chips.
One of only two foundries with 3nm GAA volume capability, with the added advantage of vertical integration across memory and logic.
Mainland China's foundry leader and the domestic mainstay at 14nm and above; a core vehicle for the self-sufficiency strategy.
A specialist foundry for mature nodes (22-28nm), focused on embedded memory, OLED driver IC, RF and other specialty processes.
The world's third-largest independent foundry, focused on specialty processes such as GaN, RF, SiGe and embedded memory; a strategic US/European domestic supplier.
Mainland China's second-largest and the world's seventh-largest foundry, focused on specialty mature processes (power discretes, embedded non-volatile memory/eNVM, analog and power management), with China's largest power-discrete capacity.
The only US-based IDM with a leading-edge process roadmap, transitioning to foundry and carrying the strategic intent of the US CHIPS Act.
A specialty IC foundry roughly 27% owned by TSMC, focused on differentiated 8-inch mature processes such as high-voltage, ultra-high-voltage, BCD, power management and display driver.
Advanced Packaging & OSAT
After wafer fabrication, chips must be packaged and tested before they ship; advanced packaging (CoWoS, HBM stacking, Fan-Out, Chiplet) integrates multiple dies at high density and is the key path for AI compute to break past Moore's-Law physical limits. TSMC monopolizes high-end packaging, while traditional OSATs dominate the high-volume mid-to-low end.
The de facto monopoly in advanced packaging: CoWoS (2.5D) and SoIC (3D) are the only large-scale volume path for high-density integration of AI GPUs and HBM, with capacity in chronic shortage.
The world's largest independent OSAT and the leader in both packaging and test; its USI unit covers system-in-package (SiP).
The world's second-largest independent OSAT, with advanced-packaging capacity in Vietnam, Korea and Portugal; an important partner for major-customer SiP modules.
Mainland China's largest OSAT; its acquisition of STATS ChipPAC put it in the global top three, with steadily upgrading advanced-packaging (Bumping/Fan-Out) capability.
The world's largest memory OSAT, focused on DRAM/NAND/MCP packaging and test, and developing advanced packaging such as TSV and fan-out panel-level packaging (FOPLP) to enter high-density AI memory.
The world's largest dedicated semiconductor test house, focused on wafer test, final test and burn-in, with a deep foothold in testing high-compute AI GPU/ASIC chips.
AMD's strategic packaging/test partner, handling its CPU/GPU packaging and test and actively building out Chiplet interconnect packaging.
An up-and-coming OSAT focused on copper-wire bonding and advanced QFN/DFN packaging, targeting the automotive and industrial semiconductor packaging/test niche.
Analog, Power, Power Semiconductors & Passives
The physical foundation running alongside the digital-compute spine: it provides the base devices for powering AI and its signal chain, including power management (PMIC/multiphase/vertical power delivery), power semiconductors (SiC/GaN), the analog signal chain, and the passives (MLCC/inductors) whose per-box use multiplies with AI servers. A single GPU now draws over a thousand amps, pushing power-delivery efficiency and power density to a value high-ground second only to the compute chip; passive-component use multiplies as the number of power rails grows, an invisible must-have in the AI server BOM. Chokepoints here are relatively fragmented and more substitutable than exclusive nodes such as EUV/HBM, but leaders hold pricing power through proprietary processes and automotive-grade reliability.
A leader in high-performance analog and signal chain, supplying AI data centers with precision power control, thermal management, monitoring and optical-module signal chain; in 2026 it acquired Empower for $1.5 billion to enter GPU bypass integrated voltage regulation (IVR) and silicon-capacitor vertical power delivery, attacking the power-density bottleneck directly.
A leader in analog and embedded processing; its power management (PMIC/DC-DC) and signal chain are base devices for AI server power delivery and board-level power management, covering the multi-stage power tree from rack to processor.
The global leader in power semiconductors and a core supplier of AI server main power (PSU) and board-level power; combining Si/SiC/GaN, it ships 8kW/12kW high-density PSUs, co-develops the 800V HVDC rack power architecture with NVIDIA, and takes a large share of Blackwell-platform power management.
A power-semiconductor and SiC (EliteSiC) maker, supplying AI data centers with the full power tree from substation high-voltage AC/DC to processor-level regulation; it works with NVIDIA on the 800V HVDC architecture and acquired Qorvo SiC JFET and Aura Vcore technology to strengthen its AI power portfolio.
A core maker of board-level DC-DC/multiphase/vertical power and a key NVIDIA GPU board-power supplier: the primary supplier of the 12V-to-1V voltage regulation modules (VRMs) on the GB200 Bianca board; it focuses on 48V architecture and a proprietary BCD process, holding the GPU last-inch power chokepoint.
A maker of high-density power modules and vertical power delivery (VPD); its proprietary Factorized Power Architecture (FPA) places current multipliers directly beneath the processor, delivering over 1000A to the GPU while sharply cutting PDN loss, supporting OAM and custom AI accelerators.
A leader in SiC substrates/devices, supplying upstream substrates and materials for 800V HVDC high-voltage power and SiC devices in AI data centers; it leads 200mm SiC substrate volume production and is the key upstream material supplier in the SiC power chain.
A maker of GaN (GaNFast) and SiC (GeneSiC) power devices, whose technology was selected by NVIDIA in May 2025 to support its 800V HVDC data-center power architecture, targeting 1MW IT racks and Kyber rack-scale systems (powering GPUs such as Rubin Ultra).
The global leader in MLCC (multilayer ceramic capacitors), supplying AI server power and signal chain with vast quantities of high-capacitance/high-end MLCC: a single AI server uses roughly 8x the MLCC of a conventional server, one of the cost items in the AI server BOM second only to GPU/memory.
A major passives maker whose inductors (VLBUC/ERUC) and MLCC handle voltage conversion and noise filtering in AI server power; a single AI server contains thousands of MLCC and inductors, and TDK lists AI data centers as a core investment direction.
A major passives maker, the world's largest chip-resistor and tantalum-capacitor vendor and third-largest in MLCC and inductors, supplying AI server power networks with MLCC, tantalum capacitors, inductors and resistors, benefiting from multiplying per-box use.
A low-power FPGA maker handling board-level power sequencing/power-up management, platform management and hardware root of trust in AI servers; devices such as MachXO5-NX TDQ provide secure boot, firmware resilience (NIST SP 800-193) and post-quantum cryptography.
AI Servers & ODM
Integrates core chips such as GPUs and CPUs into deliverable AI compute nodes, handling rack-level system design, thermal/power management and supply-chain integration; the primary procurement target for hyperscalers and GPU neoclouds. The business is low-margin, high-turnover contract manufacturing, but the more custom the work and the deeper the customer lock-in, the more pricing room there is.
Hon Hai's A-share platform; the single largest entity in global AI server contract manufacturing by scale.
Taiwan's ODM leader; a core contract manufacturer of cloud servers and AI racks.
A leader in AI server design and direct sales, the earliest to volume-produce rack-scale liquid cooling.
The world's largest enterprise IT distributor, taking AI servers into both the cloud-vendor and enterprise lanes.
A supplier of enterprise AI servers and HPC clusters, carrying the Cray supercomputing heritage.
A leader in AI servers and ODM, delivering customized full systems to hyperscale CSPs via ThinkSystem-brand machines, Wentian localization and ODM+.
China's AI server leader, supplying general/AI servers, storage and liquid-cooled full systems, with both branded machines and JDM custom delivery.
An AI server and edge-compute ODM with a leading manufacturing footprint in India.
A cloud ODM-Direct maker spun off from Wistron, directly supplying high-power AI servers and racks to hyperscale data centers.
A server ODM and AI infrastructure contract manufacturer, a long-term strategic supplier to major customers.
Optical Communications (Modules · Components · DSP/CPO)
Core components for interconnect inside AI clusters and backbone transport between data-center campuses; expanding compute scale directly drives demand for 400G/800G/1.6T optical modules. The technical barrier lies in high-speed electro-optical conversion and packaging, and leaders carry strong pricing power.
Number one in global optical-module shipments and the primary supplier of 800G-and-above high-speed products.
A second-tier high-speed optical-module maker, ramping 400G/800G capacity fast.
A vertically integrated optical-component and module maker; after merging with II-VI it covers the full stack from laser to system.
A key supplier of EML/VCSEL lasers and optical-module components.
A specialist supplier of passive optical components (optical connectors, FA assemblies), sitting upstream in the optical-module value chain.
An optical-module and component supplier; the only A-share player with a vertically integrated optical-chip-plus-optical-module IDM across the full chain.
A vertically integrated optical-module/component maker, self-producing laser chips through to finished modules; its 800G data-center modules have won hyperscale volume orders, 1.6T is in development, and it aggressively adopts the Microsoft-led LPO (linear pluggable optics) architecture.
An upstream optical-module laser-chip maker, supplying mainly DFB/EML/CW and high-power silicon-photonics light sources, holding the most upstream position in the optical-module chain.
A precision-manufacturing service provider for optical-network products, taking on outsourced production for the likes of Lumentum and Coherent.
The undisputed leader in optical-interconnect DSP (from its Inphi acquisition); its PAM4 DSP/TIA/laser drivers are the digital heart of 400G/800G/1.6T optical modules, and it also offers coherent DCI DSP and CPO.
A leader in coherent optical transport systems, with in-house WaveLogic coherent DSP/modems (latest WL6), dominating data-center interconnect (DCI) backbone and metro/long-haul coherent links: the core carrier of AI cross-data-center traffic.
A high-performance analog/optical-semiconductor supplier, providing laser drivers, TIAs (transimpedance amplifiers) and photonic devices for 100G-1.6T high-speed optical links, covering both intra-data-center interconnect and DCI.
An up-and-comer in silicon-photonics interconnect and optical compute interconnect, using the Passage optical-interconnect platform for AI-cluster scale-up: M1000 is a 3D photonic interposer super-chip (claimed roughly 114 Tbps total optical bandwidth), and L200 is a CPO platform co-packageable with next-generation XPUs/switch chips, replacing copper with light to break the bandwidth bottleneck of 100,000+ accelerator clusters.
An up-and-comer in optical I/O chiplets (in-package optical interconnect), putting optical transceivers directly inside the chip package as close to the GPU/XPU as possible for electro-optical conversion, to break the copper-interconnect bandwidth and power bottleneck of AI scale-up. In April 2025 it launched the industry's first UCIe-standard optical-interconnect chiplet, targeting in-package optical I/O for GPUs/XPUs.
Networking & High-Speed Interconnect
As AI clusters scale from ten-thousand to hundred-thousand GPUs, the interconnect bandwidth and latency between GPUs, and between GPUs and storage, become the bottleneck on overall compute efficiency, driving rapid growth in specialized markets such as Ethernet switches, PCIe/CXL retimers and optical-interconnect DSP.
The de facto standard in AI/HPC data-center Ethernet switching, a leader in high-speed Ethernet and AI-cluster interconnect, and a founding member of the Ultra Ethernet Consortium.
The global leader in networking infrastructure; a vendor of integrated data-center networking and security solutions.
A leader in networking and high-speed interconnect (its H3C unit), supplying data-center switches, campus networks and full-stack ICT for GPU servers.
A specialist in PCIe/CXL retimers and smart fabric switching, whose products are standard devices on AI server boards.
A supplier of high-speed serial interconnect chips, whose active electrical cables (AEC) and optical DSP are widely used for inter-rack interconnect in hyperscale data centers.
Its networking unit (from the Mellanox acquisition) supplies Quantum InfiniBand, Spectrum-X Ethernet and NVLink interconnect, turning intra-rack GPU interconnect on GB200/GB300 into a scale-up/scale-out compute fabric, bundled vertically with GPUs as full racks.
The undisputed leader in commercial Ethernet switch/routing chips; Tomahawk (TH6 at 102.4T) builds scale-out spine-leaf, Jericho handles long-haul routing, and it is the standard chip for white-box switches and hyperscale AI networking; it also co-packages optical engines into the switch ASIC via silicon-photonics CPO.
The world's largest white-box (open networking) switch ODM, contract-manufacturing 100G/400G/800G AI/ML switches for hyperscalers who load their own software stacks such as SONiC.
A contract manufacturer of white-box network hardware and AI-optimized full racks for hyperscalers, having won multiple 1.6T switch-platform projects (including system-level design of liquid-cooled full systems); its high-value HPS service business carries the load.
A leader in MEMS precision timing (oscillators/clocks), supplying reference clocks, Super-TCXO and clock generators (the Chorus clock SoC) for AI servers, Ethernet switches and optical modules; the Elite 2 Super-TCXO provides sub-nanosecond time synchronization across AI clusters, directly improving GPU utilization and compute efficiency, an invisible must-have for high-speed interconnect and network synchronization.
A leader in AI high-speed interconnect connectors and active/passive copper cables (AEC/DAC), customizing NVLink spine connector cards and backplane/midplane connectors for intra-rack scale-up copper connections on NVIDIA GB200/GB300 NVL72; a physical-connection layer complementary to SerDes/retimers, with per-rack copper-cable counts reaching the thousands.
A leader in high-speed connectors/interconnect and sensing, supplying data-center high-speed backplanes, power connections and interconnect; it provides end-to-end 112G/800G AEC and other high-speed solutions, serving AI server interconnect as NVIDIA's second source for NVLink copper connections.
Power & Energy Supply
AI data-center power draw is swelling at the pace of megawatt-scale racks, and power supply (shortages, interconnect queues, a nuclear revival) has replaced chip capacity as the new bottleneck on compute expansion. This layer delivers power to the data-center door: generators/independent power producers (IPPs) and nuclear/SMRs provide baseload, while heavy electrical gear such as gas turbines, transformers and switchgear plus grid engineering get the power out; furthest upstream are uranium mines and enriched fuel. Value concentrates heavily in scarce capacity: dispatchable nuclear/gas units, gas turbines booked past 2030, transformers with stretched lead times, and the few firms controlling independent Western uranium supply, all holding pricing power through long-term contracts and capacity scarcity. Note this layer is generation and transmission/distribution, distinct from the rack-level UPS/liquid cooling inside data centers (see the "Data Centers, Power & Cooling" layer).
The largest US nuclear operator (about 22.1GW, 21 units), having signed data-center nuclear long-term contracts with Microsoft and others (including the Three Mile Island restart); the flagship name for the AI pure-nuclear-baseload narrative.
One of the largest US competitive power producers, with a dispatchable fleet of mainly Texas gas plus nuclear units, having signed over 2GW of nuclear long-term contracts with Meta and others to supply data centers.
A leader in heavy gas turbines and a supplier of grid equipment and nuclear services; its combined gas-and-electrification backlog totals roughly 100GW, with turbine slots booked past 2030.
A leader in electrification equipment; its power-management unit supplies transformers, low/medium-voltage switchgear and distribution gear, holding the power side of the grid-to-data-center service entrance (distinct from Vertiv's in-rack UPS/cooling).
An independent power producer whose core asset is Pennsylvania's Susquehanna nuclear plant (about 2.2GW), having signed a nuclear PPA with Amazon AWS through 2042; a model for the behind-the-meter direct-supply data-center approach.
One of the world's three heavy-gas-turbine giants, also supplying transformers, substations and fluorine-free switchgear; 2025 turbine sales jumped sharply, with orders driven by data-center demand.
The world's second-largest uranium producer and the largest independent Western supplier, controlling the world's highest-grade uranium ore; a core chokepoint for fuel security under the nuclear revival (also a Westinghouse stakeholder).
One of the world's three electrification and grid-equipment giants, supplying medium/high-voltage switchgear, transformers, distribution automation and data-center power solutions; tight T&D equipment capacity and long certification cycles give it high order visibility.
A leader in energy management and electrification, supplying low/medium-voltage switchgear, transformers and distribution systems, covering the power side from grid service entrance to data-center primary distribution.
North America's largest power-infrastructure EPC contractor, building high-voltage transmission, substations and generation-interconnection projects; a core builder for tying generation capacity into the grid and easing interconnect queues.
An advanced small modular reactor (SMR/microreactor) developer using a build-own-operate, sell-power model, having signed multi-year power intent with hyperscale customers; targeting first reactor in 2028.
The sole supplier of US Navy nuclear reactors and nuclear fuel/components, also making SMR pressure vessels and building defense-grade uranium enrichment; a pick-and-shovel play on the nuclear and SMR revival.
A large US integrated generation and retail power company, focused on gas and retail customers, now entering data-center power demand with dispatchable gas capacity.
A US SMR design pioneer; its 77MWe module is currently the only small modular reactor with NRC design certification, leading peers on technical progress.
The only licensed US producer of HALEU (high-assay low-enriched uranium), a key chokepoint for localizing and de-risking advanced-reactor/SMR fuel supply, taking on DOE and commercial enrichment orders.
One of the world's three heavy-gas-turbine giants (Mitsubishi Power); its GTCC combined-cycle units are core gear for data-center captive/grid-tied power, and it is building out hydrogen gas turbines.
Hitachi Energy is a global leader in transformers/high-voltage equipment/digital grid, a key supplier for data-center grid-to-rack power (including support for the NVIDIA 800VDC architecture); as a wholly owned subsidiary of Hitachi Ltd that is not separately listed, it trades under the parent at 6501.TSE.
A solid-oxide fuel-cell (SOFC) maker, supplying AI data centers with rapidly deployable, grid-bypassable site-level/grid-tied baseload power (behind-the-meter) that skips the interconnect queue; its products already support 800V DC.
Data Centers, Power & Cooling
The physical carrier layer for AI compute, spanning campus power infrastructure (data-center REITs/operators), critical power and cooling gear (UPS/liquid cooling/precision air conditioning), and compute-as-a-service GPU cloud platforms (neoclouds). As training power jumps from hundreds of kilowatts to megawatt-scale racks, liquid cooling and high-density power delivery become the core chokepoints.
The leader in data-center critical power and liquid-cooling infrastructure; a direct beneficiary of rising AI compute density.
The world's largest neutral data-center REIT, with 260+ data centers spanning interconnect hubs across 33 countries.
The world's second-largest data-center REIT and the primary supplier of hyperscale whole-building leases.
The leader in data-center cooling, offering a full stack from air cooling (3DVC vapor chambers) to liquid cooling (cold plates/manifolds/quick disconnects).
A leading domestic supplier of data-center precision liquid cooling and chilled-water systems.
A leading third-party (carrier-neutral) IDC operator, building and running high-availability data centers in China's tier-one economic hubs, offering hosting/managed-hosting/hybrid-cloud services.
A data-center liquid-cooling maker, offering a full liquid-cooling solution including cold plates, CDUs and quick disconnects.
An electrical-protection and connection maker; its Systems Protection division supplies data-center racks/protective enclosures, liquid cooling (cold plates/CDU/manifold), smart PDUs and switchgear, holding the rack-level power-plus-liquid-cooling position.
China's earliest and a leading carrier- and cloud-neutral IDC provider (formerly 21Vianet), offering colocation, managed hosting and interconnect across 30+ cities, with wholesale IDC growing fast on AI demand.
A REIT that began in document storage and information management, with its global data-center division as the high-growth engine, offering colocation (racks/cages/suites plus power, cooling and interconnect) and taking on hyperscale and enterprise leases.
The global leader in AI server power (PSU) and power systems, covering 800V HVDC, power shelves (power shelf/capacitor shelf) and on-board 48V DC-DC, and also making cooling fans and liquid-cooling CDUs; it has supplied qualified in-rack CDUs for NVIDIA GB200 NVL72, a core supplier of full-system AI power.
A data-center liquid-cooling and thermal-management supplier; through its Airedale business it offers CDUs (coolant distribution units, including a 1MW model), cold plates and air-to-liquid heat exchange plus cooling-control systems, serving the liquid-cooling and precision-cooling needs of high-power AI racks.
Neocloud / AI Compute Rental
New AI clouds (neoclouds) focused on renting out GPU compute, distinct from integrated hyperscalers and traditional IDCs: they build or long-lease high-power campuses, stack NVIDIA GPUs at scale, and wholesale compute to large-model labs and hyperscalers on multi-year take-or-pay contracts. Capex is extremely heavy and highly sensitive to GPU depreciation and utilization, and many are converted bitcoin miners.
The world's largest GPU-dedicated neocloud, focused solely on renting AI training/inference compute.
A full-stack AI cloud (neocloud) reorganized from former Yandex assets, building large GPU clusters and offering a cloud platform with developer tools, supplying GPU compute to hyperscalers and enterprises.
A former bitcoin miner (Iris Energy) turned AI infrastructure player, with large self-built data centers and power, both renting GPU cloud compute to AI customers and taking on hyperscale contracts (a multi-year GB300 compute deal with Microsoft).
An AI data-center developer-operator whose main model is the AI landlord: building high-power campuses and leasing them long-term to neoclouds/hyperscalers (a 15-year deal with CoreWeave); it also has a smaller AI cloud-services division (slated for spin-off).
A former bitcoin miner turned AI/HPC data-center operator, converting roughly 1.3GW of mining infrastructure into AI compute hosting, with CoreWeave as the main customer (about 590MW across six sites); CoreWeave's $9 billion all-stock takeover was terminated in October 2025 after a shareholder vote, keeping it independently listed.
An AI-factory cloud-infrastructure player, building data centers on cheap energy such as associated gas and stranded power and offering Crusoe Cloud GPU compute; one of the core builders of OpenAI Stargate (the Abilene campus).
Cloud Compute Platforms / Hyperscalers
Provides large-scale compute, storage and networking infrastructure for AI model training and inference, outputting compute resources to external customers and its own AI business via on-demand or reserved instances, forming a scale-effect moat.
AWS is number one in global cloud market share, with Trainium/Inferentia in-house chips strengthening its AI-compute competitiveness.
Azure is OpenAI's core cloud partner and one of the world's largest AI compute lessors.
Google Cloud offers external compute rental, with in-house TPU chips building a differentiated AI infrastructure.
Alibaba Cloud is the largest cloud provider in China's public-cloud market, with AI compute demand driving a strategic pivot in its cloud business.
OCI has become a rising star in the AI compute rental market on the back of high-density GPU clusters.
Tencent Cloud is China's second-largest cloud provider, forming a differentiated compute entry point off its WeChat/gaming ecosystem.
Foundation Model Labs
Takes the output of the compute layer and focuses on pretraining foundation models, alignment tuning and API commercialization, forming a technical barrier through differentiated model capability and outputting intelligence to the application layer; the business model is mainly per-API-call billing, enterprise deployment licensing or product integration. Most leading labs are still private, but all are key nodes on the chain.
The GPT model family and ChatGPT have the world's largest user base; the most widely deployed conversational AI product.
The Claude model family uses safety alignment as its differentiator, with enterprise API customers as the main revenue source.
Gemini is one of the world's largest-scale multimodal foundation models, with DeepMind continuing to produce foundational research breakthroughs.
The Llama open-source models are the most-downloaded open-weight foundation models in the world, setting the open-ecosystem standard; it builds its own hyperscale compute clusters to support R&D.
One of China's most technically advanced open-source large models, shaking up global AI-compute valuation logic with extremely low training cost.
The Grok models form unique training and distribution advantages off the X platform's data and user entry point.
Europe's leading open-source large-model lab, entering the enterprise private-deployment market with efficient small models and an open-source strategy.
The Qwen (Tongyi Qianwen) open-source models stay ahead on global open-source leaderboards; the core engine of Alibaba Cloud's AI strategy.
Ernie is one of the earliest commercialized domestic-Chinese large models, with Baidu Search providing a native distribution channel.
With a Tsinghua technical background, the GLM family is one of the domestic large models with higher procurement rates among Chinese government and enterprise customers.
Kimi uses an ultra-long context window as its core differentiator, with relatively high share in China's knowledge-worker market.
One of China's large-model "six tigers," strong in multimodal and video generation.
A domestic foundation-model lab integrating self-built compute (the Lingang AIDC), the SenseNova multimodal large model and applications in one.
The Baichuan model family has strong penetration in vertical scenarios such as healthcare.
ByteDance's in-house Doubao large model, building a consumer AI assistant and enterprise API (Volcano Engine) off the super-traffic entry points of Douyin and Toutiao.
Huawei's Pangu large model targets government/enterprise and industry vertical scenarios, bundled with Ascend AI compute and the MindSpore framework into a domestic integrated hardware-software solution.
The in-house Hunyuan large model is embedded in WeChat/QQ/Tencent Meeting and advertising, and also output externally via Tencent Cloud.
AI Applications · Enterprise Software & Data Platforms
Embeds AI into enterprise workflows, data infrastructure and vertical scenarios, monetizing through subscription, seat or consumption-based billing. AI deepens automation and raises per-seat value, pushing customers from tool procurement toward platform lock-in, and net revenue retention (NRR) becomes the core moat.
The world's largest CRM SaaS, using Agentforce to embed AI agents across the full sales and service workflow.
An enterprise IT and business-process automation platform; Now AI closes the loop across tickets, approvals and knowledge base.
An AI data operating system; the AIP platform connects large models to core decision data for government and enterprises.
A cloud data platform; Cortex AI lets enterprises run LLM inference and vector search directly inside the data warehouse.
A cloud-native observability platform; its LLM Observability module is built specifically for AI application monitoring.
China's leader in speech and cognitive AI; the iFlytek Spark large model outputs to vertical scenarios in education, healthcare and government.
China's largest office-software platform; WPS AI embeds large-model capability across documents, spreadsheets and presentations.
An AI-driven language-learning platform; generative AI drives the cost of personalized conversation practice toward zero.
A document-database leader; its managed cloud service Atlas has built-in vector search (Atlas Vector Search), keeping operational data, vectors and metadata in one database, the operational-data foundation for AI applications/RAG and agents, so developers need no separate dedicated vector store.
A real-time data-streaming platform built on Apache Kafka, adding Flink stream processing and governance to continuously feed AI applications/agents with live event streams and real-time context; the data pipeline and freshness foundation for RAG/agentic AI. Note: as of March 2026 it was acquired by IBM for roughly $11 billion ($31/share, all cash) and folded into watsonx.data.
A search and observability platform; Elasticsearch provides hybrid vector-plus-keyword retrieval and reranking, positioned as an AI context platform, serving as the retrieval and recall layer for RAG/agentic workflows by connecting an enterprise's existing search/log data to large models.
A high-performance all-flash storage maker (note: a hardware vendor positioned as the physical layer of the data foundation); DirectFlash/FlashBlade provide the high-throughput, low-latency data lake and storage foundation needed for AI training/inference; FlashBlade//EXA targets AI/HPC, and it won a Meta hyperscale design.
A data lakehouse platform unifying data engineering, analytics and AI/ML training; with MosaicML for model training and the launch of Lakebase and Agent Bricks, it is one of the leaders in integrated enterprise AI data and training/agent platforms.
A dedicated managed vector database, with a serverless architecture built for large-scale similarity search; the representative pure vector store for RAG retrieval infrastructure, built to inject real-time, relevant context into large models and reduce hallucination.
AI Dev Tools & Code Agents
Weaves large models into the full software-development workflow, from code completion and AI editors to autonomous agents that take end-to-end tickets. Mostly per-developer-seat subscriptions, with the underlying models bought in; value capture sits in the distribution channel (IDE/code hosting), workflow-data accumulation and switching cost rather than the model itself, and many AI-native players are still private.
GitHub Copilot is the most-installed AI coding assistant, embedded in the world's largest code-hosting and IDE ecosystem.
An integrated DevSecOps platform; the GitLab Duo Agent Platform weaves AI agents across the code-CI/CD-security workflow.
The leader in Jira/Confluence team collaboration; the Rovo agent and Rovo Dev embed AI into project management and knowledge-retrieval workflows.
The AI-native code editor Cursor, rapidly eroding the traditional IDE-plus-plugin paradigm with an agentic coding experience.
A browser-based cloud dev environment plus agent, letting non-specialists vibe-code deployable apps directly in natural language.
The autonomous AI software engineer Devin, aiming to take end-to-end ticket-level coding tasks rather than just completion.
AI Content Generation
Uses generative AI to produce images, video, design and marketing creative, billed mostly by subscription or per-generation credits. Leaders build moats through model aesthetic tuning, copyright compliance and creative-workflow lock-in; those running their own video/image models must carry training and inference compute themselves, with margins constrained by GPU cost, and a number of well-regarded players are still private.
Uses the Firefly generative model to embed generative AI across Creative Cloud and the Document/Marketing Cloud suite.
Kling, incubated by short-video platform Kuaishou, is a commercially leading AI video-generation model with dual API-plus-subscription tracks.
A leader in imaging and design SaaS, achieving a productivity pivot by rebuilding photo/video/e-commerce design with AI (Meitu, DesignKit, Wink).
An online design-creation platform; the Magic Studio generative suite brings AI design to non-specialist mass users and enterprise teams.
An AI video-generation pioneer (the Gen series), offering generation and editing toolchains for film/advertising professional creators.
A top AI image-generation service, building a strong word-of-mouth barrier in the creator community on high aesthetic quality.
AI Search & Advertising
AI is reshaping the entry points for information distribution and ad monetization: generative answers rewrite the search results page, and deep learning optimizes ad bidding and conversion. Incumbents embed AI into existing search/social/advertising cash cows, while AI-native answer engines and independent programmatic platforms enter from the flank; value capture sits in the traffic entry point, data and ad ROI.
AI Overviews/Gemini reshape the search results page, embedding generative answers into the world's largest search-advertising entry point.
Drives feed-ad automation with Advantage+ generative ad tools, paired with the open-source Llama ecosystem.
Bing Search plus Microsoft 365 Copilot embed OpenAI models into the search and office entry points.
The Ernie large model rebuilds Baidu Search, also reshaping search advertising and intelligent cloud with AI.
The largest independent demand-side platform (DSP) on the open internet, doing programmatic bidding and optimization across CTV/display/audio for advertisers via the Kokai AI engine.
A mobile-advertising AI platform; the AXON engine uses deep learning to sharply improve ad bidding and conversion efficiency.
An AI-native answer engine, directly challenging the traditional search-results-page paradigm with cited generative Q&A.
AI Security
AI is both the thing being protected and the defensive weapon: zero-trust/SASE, endpoint XDR and firewall vendors embed AI into threat detection and the autonomous SOC, using AI to fight AI-driven attacks. Mostly high-margin subscription (ARR) models, where vast telemetry trains detection models to form a data flywheel and a high-renewal moat.
The cybersecurity leader, using a platformization strategy plus Precision AI/Cortex to embed AI across its three platforms: network, cloud and SecOps.
The cloud-native endpoint/cloud-security leader, doing AI-driven detection and autonomous response with Charlotte AI plus the Falcon platform.
The cloud-native zero-trust (Zero Trust Exchange) leader, positioned as the online traffic-security proxy layer for the AI/agent era.
A global edge-network-plus-security platform, doing web security/Zero Trust while offering serverless inference at the edge via Workers AI: the platform that agents run on.
An AI-native endpoint and XDR platform, advancing the autonomous SOC with Purple AI (Athena): agents that investigate and remediate threats automatically.
A cybersecurity and firewall leader, using the FortiAI suite to embed AI into threat detection, security operations and LLM-application protection, while pivoting to SASE.
On-Device & Edge AI
Builds AI inference capability into consumer-electronics, automotive and robotics hardware on the device side, constructing an integrated hardware-software moat through privacy protection, low latency and offline capability; the outlet where the AI value chain reaches its vast base of end users.
Embeds on-device LLM inference across the iPhone/Mac product line via Apple Intelligence, with the in-house chip's Neural Engine as the execution carrier.
FSD self-driving software plus the Optimus humanoid robot; one of the largest producers of real-world AI training data.
The leader in on-device AI SoCs; the Snapdragon platform provides the on-device inference compute foundation for flagship Android and AI PCs.
The shipment leader in on-device AI SoCs; Dimensity covers mid-to-high-end phones, extending into custom ASIC/edge AI chips.
A supplier of autonomous-driving vision-perception chips and systems; the EyeQ series of SoCs is the world's most widely deployed volume ADAS platform.
A leader in the phone-plus-AIoT ecosystem; on-device AI (HyperOS plus HyperAI) spans phones, home, automotive and other devices.
A Chinese autonomous-driving AI chip designer; the Journey series of SoCs is built specifically for in-vehicle intelligent-driving inference.
Puts an on-device NPU (about 49 TOPS) plus on-device inference into Galaxy phones/foldables/AI glasses via Galaxy AI.
An edge AI vision-SoC maker, supplying low-power on-device inference chips (the CVflow architecture) for security/automotive/robotics.
A leader in automotive and industrial edge chips, pushing AI inference down to the in-vehicle/industrial/IoT edge with i.MX processors plus the eIQ Neutron NPU and eIQ Agentic framework.
A humanoid and service-robot maker (the Walker S series), putting embodied intelligence/AI inference into physical robots for industrial and service scenarios; the world's first listed company built on a humanoid-robot motherboard.