6 segments · 66 companies · as of Sep 18, 2026
AI Servers & Racks
The companies that turn accelerators into servers, racks and working clusters
Where it sits
Fed by: Compute Silicon · Memory & Storage · Networking & InterconnectFeeds: Cloud & AI Capital
AI servers combine accelerators, processors, networking, power and cooling into deployable systems. Branded vendors sell supported platforms, Taiwanese ODMs design hyperscale hardware, and manufacturing and integration specialists assemble and validate complete racks.
Why now
GB300 deployments overlap the 2026 Vera Rubin production ramp. AMD Helios brings Open Rack Wide designs into the market, with volume deployments expected in the second half. Integration capacity, liquid-loop quality and power validation increasingly determine delivery schedules.
Server OEMs & rack designs
Supported GPU systems and rack reference designs set the platform that customers deploy. Share depends on accelerator access, cooling engineering, service coverage and the ability to deliver validated systems.
PowerRack with XE9812 servers ships Vera Rubin NVL72 to CoreWeave; PowerEdge XE also covers GB300
leaderDLC-2 liquid-cooled GPU servers and SuperClusters; GB300 and Vera Rubin rack platforms
leaderNvidia AI racks and AMD Helios systems with 72 MI455X GPUs, Venice CPUs and HPE Juniper networking
majorThinkSystem GPU servers and Neptune liquid-cooled GB300/Vera Rubin rack platforms
majorDGX GB300 and Vera Rubin NVL72 systems; MGX supplies the modular rack reference architecture
majorGiga Computing G894-SD3-AAX7 8U servers integrate eight Nvidia HGX B300 GPUs for AI training
majorXA NB3I-E12 HGX B300 servers and XA GB721-E2 GB300 NVL72 liquid-cooled racks target AI factories
majorNF5688G7 eight-GPU HGX servers package accelerator compute and thermal engineering for AI buyers
majorHelios MI455X/Venice Open Rack Wide reference design; retained ZT design team enables partner systems
challengerAI PODs use UCS C845A/C885A M8 GPU servers; also named as a Vera Rubin system partner
challengerMiTAC Computing GPU servers incorporate the former TYAN business under a unified MiTAC brand
challenger4U16X-GNR2/DLC B300 servers integrate eight HGX B300 GPUs with direct-to-chip liquid cooling
challengerCG480-S5063 4U servers support eight RTX PRO 6000 or H200 NVL GPUs for enterprise AI workloads
challengerPRIMERGY GX2570 M8s servers package eight HGX B200 GPUs with air- or liquid-cooled configurations
challengerCB502-NM 5U eight-GPU systems and barebones servers; also participates in Rubin infrastructure
nicheGodi Blackwell Ultra GPU servers support B300-based AI systems for regional cloud and enterprise buyers
nichePower11 S1122 rack servers pair with Spyre accelerators for on-premises enterprise AI inference
nichecotomi Generative AI Appliance Server combines GPU hardware with local enterprise inference software
nicheSKY-642E3 4U servers support eight RTX PRO GPUs and EPYC processors for enterprise and edge AI
nicheVia Hitachi Vantara, Hitachi iQ HGX B300 and MGX RTX PRO servers package supported enterprise AI systems
nichePro6000 servers ship to Digital Dynamic in Japan; September filing reports about $38M delivered
emerging
Sources: nvidianews.nvidia.com, superx.gcs-web.com, ibm.com, jp.fujitsu.com, servers.asus.com, backendnews.net
ODM-direct systems
Cloud buyers source customized compute trays and racks directly from these manufacturers. IDC puts ODM Direct collectively at 50.2% of worldwide server revenue in 1Q26, down from 64.1% a year earlier. Design, factory capacity and rack qualification determine which programs scale.
Builds GB300 and Vera Rubin cloud racks; management estimates over 40% of the broader AI-server market
leaderQCT designs and builds liquid-cooled Nvidia GPU racks for hyperscale cloud customers
leaderIngrasys cloud and AI server production within a separately listed Foxconn manufacturing business
majorVera Rubin NVL72 and AMD Helios ORW racks; brings open-compute reference designs into production
majorAccelerator server assemblies and rack platforms in Nvidia's Vera Rubin production ecosystem
majorHyperscale AI server and rack manufacturing; named Vera Rubin system production partner
majorHyve Solutions co-designs compute boards, servers and rack platforms directly for hyperscale cloud buyers
majorCustom cloud compute hardware design and manufacturing; Helios work initially centers on scale-up switches
majorCo-designs custom cloud compute nodes and rack architectures through its cloud infrastructure engineering business
majorGPU server and rack manufacturing for the Vera Rubin ecosystem alongside broader electronics ODM work
challengerCloud server ODM expands into Vera Rubin system manufacturing and AI compute platforms
challengerDesigns and manufactures AI servers and switches for internet customers' training and inference clusters
challengerCo-designs, assembles and validates complete AI racks, including XPU sleds and power and thermal subsystems
challengerGiga Computing provides customized server design and manufacturing in addition to its branded GPU systems
challengerMiTAC Computing supplies customized cloud server design and manufacturing using its integrated TYAN engineering
challengerJ-422G platform and custom server engineering provide cloud buyers with design-to-production system services
challengerOEM/ODM services customize server boards, chassis, BIOS and complete GPU or storage systems
nicheODM/JDM server engineering covers motherboards, BIOS/BMC, thermal design and complete L10 systems
niche
Sources: focustaiwan.tw, wiwynn.com, usiglobal.com, backendnews.net, tech-insider.org, maktinta.com
EMS & system manufacturing
Contract manufacturers turn customer designs into repeatable board, node and rack production. Complex assembly, thermal testing and regional manufacturing capacity matter more than the end-system brand.
Owns ZT Systems manufacturing; preferred US NPI partner for AMD cloud rack and cluster-scale AI designs
leaderAssembles and tests customer-designed AI compute boards and nodes through its global cloud manufacturing operations
majorBuilds and tests cloud server boards and J-422G systems, with Mikros cold-plate assembly for liquid-cooled nodes
majorManufactures hyperscale compute hardware and develops Helios ORW scale-up switch systems with AMD
majorBuilds accelerator server boards and system assemblies for large AI platform programs
majorHigh-volume server and cloud hardware assembly, including Ingrasys operations and overseas factories
majorGlobal AI rack assembly and manufacturing services; exposure includes its listed manufacturing subsidiaries
majorCustom server and switch manufacturing for internet buyers, including large AI training and inference systems
challengerPre-validates AI rack assemblies combining compute interconnects, power and liquid-cooling hardware
challengerAssembles, tests and validates HPE's high-power liquid-cooled AI server blades in US manufacturing programs
nicheHPC board design, complex PCB assembly and higher-level integration support computing OEMs
nicheManufactures server motherboards, backplanes and system assemblies through its dedicated server EMS offering
nicheManufactures HPC boards and computing subassemblies, extending assembly and test through complete storage racks
nicheManufactures AI infrastructure cooling and power products; FY2025 segment revenue is RMB943M
emergingHPC electronic assemblies for AWS form a distinct program alongside optical manufacturing, per FY2026 management
emerging
Sources: ir.sanmina.com, ir.bench.com, usiglobal.com, zollner-electronics.com, intel.com, fool.com
China server systems
Domestic server vendors integrate locally available accelerators and processors for Chinese cloud, operator and enterprise buyers. Software compatibility, domestic supply and service relationships shape competition.
NF5688G7 HGX GPU servers and other AI systems; product availability remains subject to export rules
leaderAtlas systems and CloudMatrix384 supernodes integrate Ascend 910C NPUs with Kunpeng host CPUs
leaderThinkSystem GPU and AI infrastructure sold to Chinese enterprise and cloud customers
majorH3C UniServer R5500 G7 X supports eight domestic OAM GPUs in a modular AI training server
majorDawning servers and domestic-processor clusters integrate AI accelerators for Chinese research and enterprise buyers
majorFusionServer G8600 V7 eight-GPU systems and G5500 V7 accelerator servers; IPO review continues
majorHGX and OAM eight-GPU server configurations address operator and enterprise AI training and inference
majorCustom AI server platforms target Chinese internet customers and large training and inference deployments
challengerKunTai A722 systems pair Kunpeng 920 CPUs with Ascend Atlas acceleration for enterprise inference
challengerNingchang X640 G40 systems support eight GPUs with configurable accelerator interconnects for AI training
challengerTiangong AT800 servers combine Kunpeng CPUs with Ascend accelerators for domestic AI training and inference
challengerPR8908E 8U servers support eight-GPU SXM5 or OAM modules for training and HPC clusters
nicheZhaohan computing products combine Kunpeng CPUs and Ascend AI processors for domestic infrastructure
nicheChaoqiang A800I inference servers and A900P systems integrate Kunpeng and Ascend domestic compute
nicheQingtian GF7280 V5 AI servers support up to 12 double-width GPUs for domestic compute deployments
niche
Sources: ieisystem.com, h3c.com, xfusion.com, nettrix.com.cn, powerleader.com.cn, backendnews.net
Rack integration & deployment
Integrated racks leave the factory with compute, networking, power and cooling tested together. Buyers value repeatable acceptance tests, liquid-loop commissioning and the ability to deploy many racks consistently.
Factory-integrated PowerEdge AI racks combine compute, networking and liquid cooling for large deployments
leaderL11-tested SuperCluster racks integrate DLC-2 cooling, networking and GPU compute before shipment
majorTurnkey Nvidia and AMD Helios racks combine system validation, HPE services and deployment support
majorNeptune liquid-cooled ThinkSystem racks combine factory integration and enterprise deployment services
majorZT manufacturing adds hyperscale rack assembly, test and NPI capacity for cloud and AMD AI systems
majorHyve Solutions integrates and validates custom hyperscale AI racks across its manufacturing network
majorDesigned, built and deployed SK Telecom's Haein AI cluster; provides ongoing cluster integration expertise
majorIntegrates compute, power shelves and JetCool liquid cooling into validated customer AI rack configurations
majorIntegrates DGX GB200/GB300 and SuperPOD platforms through its Advanced Technology Center and AI Proving Ground
majorFactory rack integration installs compute, networking, power and cooling, then tests the assembled cloud rack
majorRack Integration Services assemble AI systems, cable and validate whole racks before customer deployment
majorOwns complete rack assembly and validation for co-designed AI systems with accelerator sleds
challengerAI Rack offering combines thermal, power and interconnect hardware into a pre-validated rack configuration
challengerRacks, cables, configures and tests OEM server systems under long-term integration programs
nicheIntegration centers perform rack assembly, cabling, burn-in and shipping preparation for enterprise deployments
nicheSolutions Integration Centers build multi-rack systems with cabling, configuration and pre-shipment validation
nicheConfiguration centers rack servers, cable equipment and run burn-in tests before shipping complete installations
nicheStages, assembles and quality-checks site-ready server racks before coordinated delivery to customer locations
nicheData Center Factory assembles and validates high-density GPU racks with integrated power, cooling and cabling
niche
Sources: celestica.com, jabil.com, careers.computacenter.com, insight.com, cdw.com, shi.com
Specialist AI & HPC systems
Sovereign supercomputers, regional enterprise appliances and rugged GPU systems require different packaging and support from hyperscale racks. Architecture expertise and deployment constraints define these specialists.
Cray supercomputing and liquid-cooled AI systems serve national laboratories and sovereign compute programs
leaderBuilds BullSequana systems for European HPC and AI programs; now owned directly by the French state
majorPRIMEHPC systems and research-cluster integration complement PRIMERGY GPU nodes for scientific computing
majorCustom AI/HPC clusters combine hardware design, deployment and operating software for sovereign buyers
majorThinkSystem SC750 V4 Neptune direct-water-cooled nodes support dense research and AI supercomputing clusters
majorPowerEdge XE9680 cluster designs combine GPU nodes, fabric validation and HPC deployment services
majorMulti-node BigTwin and GPU systems support custom HPC clusters with dense compute and direct liquid cooling
majorIntegrates domestic scientific-computing clusters with parallel software, scheduling and HPC system engineering
majorCAE and scientific-computing cluster solutions combine compute nodes, parallel storage and workload scheduling
majorDGX SuperPOD reference clusters specify GPU fabrics, storage and management beyond the individual DGX server
majorDesigns and integrates custom GPU clusters around Godi nodes for research and regional compute operators
nicheSX-Aurora TSUBASA vector servers and C401-8 systems target simulation and scientific HPC clusters
nicheRugged GPU servers and GPUltima rack clusters support defense, vehicle and industrial AI workloads
nicheDesigns and builds liquid-cooled supercomputers with cluster integration, ClustSafe power control and service
niche
Sources: lenovopress.lenovo.com, supermicro.com, megware.com, ieisystem.com, jpn.nec.com, docs.nvidia.com
Glossary
- ODM
- An original design manufacturer develops hardware that a cloud buyer or another brand sells or deploys.
- EMS
- Electronics manufacturing services assemble and test hardware designed by customers or partners.
- NVL72
- A rack-scale Nvidia system connecting 72 GPUs through NVLink; generation names specify the processor platform.
- Open Rack Wide
- An OCP rack design with a wider compute tray; AMD Helios uses this open mechanical architecture.
- Rack integration
- Installing, wiring, filling and testing servers, network equipment, power and cooling as a complete rack.
Heat, spend, exposure and shares are editorial estimates. How to read the map ↗