Server rack system integration
Mechanics, power, cooling, compute and interconnects converge in rack integration, followed by factory testing, delivery and site acceptance.
Explore this role on the mapSuggest a correctionBack to contents
SUPPLY CHAIN HANDBOOK
Connect rack mechanics, power, cooling and interconnects with factory integration and site acceptance. Rack delivery depends on subsystems working together; a server or component specification does not describe the complete system.
Rack delivery combines compute, switching, power, cooling and management into one system. Mechanical interfaces enable installation, connection and service. Individually working devices can still encounter shared power, cooling or interconnect problems under combined load, requiring factory integration and site acceptance. Rails, busbars, cables and management systems matter because they support repeatable assembly, sustained operation and fault handling, rather than directly adding chip compute.
Frames carry the load, trays hold boards and cooling parts, and rails provide installation and withdrawal interfaces. Their dimensions must accommodate rear power, network and coolant connections.
Power shelves and busbars feed modules, interconnects move data, and cooling and management sustain operation. Integrators combine them into a compatible configuration and validate shared loading.
Factory tests check configuration, wiring and operation. Site work connects facility power, cooling and networks, then completes software setup and tests within suitable weight and space limits.
Compare at the same validation boundary. PSU nameplate power is not redundant-system usable power, and individual tray tests cannot replace whole-rack acceptance under specified facility conditions.
Mechanics, power, cooling, compute and interconnects converge in rack integration, followed by factory testing, delivery and site acceptance.
Explore this role on the mapSuggest a correctionBack to contents
Defines how compute, switching, mechanics, power and cooling fit together and provides a common design basis for system partners.
NVIDIA MGX is a modular reference architecture implemented by OEMs and ODMs for specific platforms and customer requirements. Mechanical fit alone does not ensure power, cooling, interconnect or firmware compatibility.
Distinction: MGX, HGX and NVL72 are not interchangeable terms, and AI servers do not all use the same rack design.
Scope of these company examples
MGX and NVL compute and switching architecture.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Supports trays, power equipment and plumbing with defined mounting positions and service access.
Chenbro has expanded from chassis into rack solutions; Delta and Vertiv also publish rack infrastructure products. Structure must accommodate equipment loads, layout, transport and installation.
Distinction: Supplying the rack structure does not establish responsibility for GPUs, software and the complete cluster.
Scope of these company examples
2026 expansion from chassis into rack structures.
Product / role sourceMGX racks with power and cooling infrastructure.
Product / role sourceMechanical, power, cooling and preintegration solutions.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Chassis and trays locate boards and thermal components and interface mechanically with the rack and rear connections.
Chenbro publishes MGX chassis in several form factors. Board, connector and cooling-loop changes affect space, tolerance and service requirements across generations.
Distinction: An empty tray, a populated compute module and a complete rack are different deliverables.
Scope of these company examples
Mechanical products for multiple form factors and platforms.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Supports equipment installation and pull-out servicing in coordination with the chassis, rack posts and mounting hardware.
King Slide offers customized server rails, REPON publishes MGX-compatible solutions and Accuride offers server slides. Design challenges include limited side space, loaded deflection, operating force and reliable locking.
Distinction: Rails are one mechanical subsystem. King Slide documentation in a specific NVIDIA switch manual does not establish exclusive supply across all racks.
Scope of these company examples
Custom cloud-equipment rails; named installation documentation exists for specific NVIDIA switches.
Product / role sourcePublished GB200/300 and Vera Rubin rail solutions.
Product / role sourcePublished server slides; no specific NVIDIA supply inferred.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
A shelf integrates PSU modules, management and distribution interfaces for rack installation. Rated capacity and usable redundant power must be read separately.
Delta lists four 13.75 kW modules in a 55 kW shelf; Acbel lists six 5.5 kW PSUs in a 33 kW shelf. The LITEON/QCT 110 kW demonstration is also shelf-level.
Server PSU modulesBusbars and power interconnectsRack backup and transient energy storageBoard DC/DC and core regulation
Distinction: Aggregate shelf ratings cannot be compared directly with one PSU. An 800 V architecture does not make every shelf an 800 V-output unit.
Published power-shelf examples. Listed companies are not presumed to share all customers or jointly supply one rack.
Rack power and 800 VDC designs.
Product / role source110 kW power shelf in a joint QCT showcase.
Product / role sourceSix hot-swappable PSUs; distinguish aggregate capacity from redundancy configuration.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Distributes shelf output to modules through busbars, power cables and connectors.
BizLink publishes MGX busbars and server power interconnects; Delta also offers rack distribution structures. Contact quality, temperature rise, space and interfaces affect reliability.
Distinction: A busbar distributes power; it is different from a power supply that converts voltage.
Scope of these company examples
Published MGX busbars and connection products.
Product / role sourceRack distribution and related infrastructure.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Batteries and capacitors address different time scales, supporting interruption ride-through or transient load management.
Delta publishes rack backup products; NVIDIA describes rack energy storage for smoothing rapid load changes. Capacity, controls and fault handling must match the purpose.
Distinction: BBUs, capacitors and facility UPS systems differ; one component rating cannot establish facility backup duration.
Scope of these company examples
Published battery backup shelf products.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Connects compute and switch modules with designs matched to speed, reach, density and assembly.
Amphenol publishes AI interconnect products; the Rubin design uses preintegrated and validated rear copper cable cartridges. Product capability is distinct from a confirmed platform supply relationship.
Within and between racks, compare DAC/ACC/AEC, AOC, ALC and transceivers with passive fiber. These are choices governed by rate, reach and interfaces, not mandatory sequential components.
Passive fiber cables and patch cordsActive optical cables (AOC)Micro-emitter optics / ALCHigh-speed copper: DAC / ACC / AEC
Distinction: Simplified tray wiring does not eliminate rack copper cabling, and CPO does not remove every short electrical link.
Scope of these company examples
AI interconnect products; capability is not a confirmed NVL72 order.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Integrates processors, memory, boards, cooling and management into a rack-mountable compute unit.
System builders implement platform-specific boards and modules. Chip and HBM packaging are different stages from server assembly and can be traced through cross-category links.
Distinction: A compute tray is not one GPU; accelerator count, memory and cooling vary by system.
Scope of these company examples
Foxconn subsidiary Ingrasys publishes GB300 NVL72 solutions.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Lets accelerators in multiple compute modules exchange data within a scale-up fabric.
NVLink switch trays work with compute trays in NVIDIA NVL systems; external fabrics use Ethernet or InfiniBand according to network design.
Distinction: NVLink modules and Ethernet switches serve different roles. Ethernet vendors are not automatically NVLink suppliers.
Scope of these company examples
Defines in-rack accelerator connectivity and switching.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Management processors and firmware monitor equipment and support remote maintenance and fault handling.
ASPEED supplies BMC processors, while system builders integrate firmware, networks and management tools with boards, sensors and operations.
Distinction: A BMC is not a GPU. Listing ASPEED products does not establish a particular chip in a specific NVIDIA rack.
Scope of these company examples
Remote management processors; no specific rack design-in asserted.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
System builders combine compute, switching, mechanics, power, cooling and management into a customer-specific rack.
QCT, Ingrasys, Wistron/Wiwynn and Supermicro publish rack-scale products or integration capabilities. Design, manufacturing, brand and site-service responsibilities vary by contract.
LITEON and QCT jointly demonstrated a Vera Rubin NVL72 solution at GTC 2026, combining a LITEON 110 kW power shelf with QCT system integration. This named demonstration does not establish contracts among other companies in this map.
Distinction: These companies offer different solutions; the list does not imply that they all assemble the same rack together.
Scope of these company examples
Quanta subsidiary QCT jointly demonstrated a rack platform with LITEON.
Product / role sourceFoxconn subsidiary Ingrasys GB300 NVL72 solution.
Product / role sourceWiwynn publicly names Wistron in its Rubin platform work.
Product / role sourcePublished liquid-cooled rack, power and interconnect integration.
Product / role sourceDesign, assembly, testing and delivery services.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Checks rack configuration and operation before shipment to identify compatibility, wiring and load-related problems.
Supermicro describes multi-vendor interoperability, rack burn-in and test reports. Cooling and power must also meet the design and customer acceptance conditions.
Distinction: Factory testing does not eliminate site checks. Labels such as L10, L11 and L12 must be read with each vendor’s actual test scope.
Scope of these company examples
Published rack testing, performance reports and delivery documentation.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents
Delivers racks to a suitable site, connects power, cooling and networking, then completes project-specific software configuration and testing.
Supermicro covers packing, delivery and deployment; Vertiv offers integrated infrastructure and commissioning. Facility preparation and rack manufacturing can run in parallel but require coordination.
Distinction: Power-on, hardware tests, cluster acceptance and production readiness are separate milestones.
Scope of these company examples
Delivery, software configuration and deployment tests.
Product / role sourcePower, cooling, rack infrastructure and commissioning.
Product / role sourceExplore this role on the mapSuggest a correctionBack to contents