Intelligence BuildoutMethodology
← All SystemsSystem 07 / computeReviewed Aug 29, 2026

Compute & Racks

Accelerators become a useful computer only with CPUs, memory, switches, power shelves, busbars, firmware, serviceability, and cooling.

Understand the system

How it works

The rack is emerging as the unit of AI system design, coupling silicon performance to power, networking, mechanics, operations, and thermal limits.

Equipment dependencies

What makes a rack operate

Select a capability to explore its role and connections.

Selected: AI accelerator

AI accelerator

Packaged compute and memory device optimized for training and inference workloads.

Explore AI accelerator in Atlas

Documented connections

Selected documented dependencies. Arrows retain their Atlas direction; they do not represent quantities or a complete engineering process.

An accelerator headline hides the rest of the computer. Host CPUs feed work, DPUs manage data, local memory holds state, scale-up links coordinate devices, NICs reach the cluster, storage supplies datasets, and firmware keeps the whole assembly observable. As densities rise, rack busbars, power shelves, coolant manifolds, cable routing, floor loading, and maintenance clearance become architectural constraints.

The heterogeneous node

AI servers combine accelerators, host processors, memory, storage, NICs, management controllers, and power conversion. Bottlenecks emerge at the boundaries between those components.

  • Expose data paths through the whole server
  • Separate accelerator power from platform power
Supporting evidence · 2
  1. [1] A DGX H200 system contains eight H200 GPUs and has a documented maximum input power of 10.2 kW; the system manual specifies front-to-back airflow.
  2. [2] NVIDIA specifies 141 GB of HBM3E capacity and 4.8 TB/s of memory bandwidth for the H200 accelerator.

Scale-up inside the machine

High-bandwidth, low-latency fabrics let accelerators exchange tensors and memory traffic as one system. Topology and collective behavior determine how close scaling comes to linear.

  • Collectives follow the physical topology
  • Synchronization consumes usable bandwidth
Supporting evidence · 2
  1. [1] UALink 1.0 defines 200 Gb/s-per-lane links and a scale-up fabric designed for as many as 1,024 accelerators.
  2. [2] PCI Express 6.0 doubles the raw data rate to 64 GT/s and supports up to 256 GB/s bidirectionally on an x16 link.

Rack-scale power delivery

At high density, upstream transformers, switchgear, busways, power shelves, backup strategy, and monitoring must be designed as one conversion chain from grid to package.

  • Every voltage conversion introduces losses
  • Rated and measured demand must remain distinct
Supporting evidence · 2
  1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
  2. [2] Open Compute Project workstreams are addressing future AI rack power envelopes from 250 kW toward 1 MW alongside liquid-cooling standards.

Serviceability is availability

Cables, blind-mate connectors, coolant lines, firmware, spares, and safe maintenance procedures affect mean time to repair. Dense compute is valuable only while the cluster can schedule it.

  • Replaceable units need safe maintenance paths
  • Repair time reduces effective utilization
Supporting evidence · 3
  1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
  2. [2] CUDA and ROCm each combine programming models, runtimes, compilers, libraries, debugging tools, and hardware interfaces into an accelerator software stack.
  3. [3] NVIDIA documents a maximum of 64 framebuffer page retirements for legacy GPUs and up to 512 framebuffer row-remapping entries for Ampere and later generations.

Featured evidence

Cumulative 100%-nameplate rack energy scenario

A derived upper-bound-style scenario using the documented approximate NVL72 rack power continuously through one year.
Cumulative nameplate scenarioGWh
View chart values
Cumulative 100%-nameplate rack energy scenario — underlying values in GWh
CategoryCumulative nameplate scenario
Month 00 GWh
Month 30.26 GWh
Month 60.53 GWh
Month 90.79 GWh
Month 121.05 GWh

DERIVED SCENARIO — assumes continuous 100% of the documented approximate rack envelope; not metered load, forecast, or expected annual energy.

Key indicators

Key measures & constraints

  1. 01

    NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.

    Class
    vendor spec
    Geography
    Global
    Period
    DGX GB200 platform
    Confidence
    high
    Trace evidence
  2. 02

    A DGX H200 system contains eight H200 GPUs and has a documented maximum input power of 10.2 kW; the system manual specifies front-to-back airflow.

    Class
    vendor spec
    Geography
    Global
    Period
    DGX H200 system specification
    Confidence
    high
    Trace evidence
  3. 03

    Holding NVIDIA's approximately 120 kW DGX GB rack specification at 100 percent for 8,760 hours yields a transparent upper-use scenario of 1.0512 GWh per year; it is not a measured average and excludes facility overhead.

    Class
    estimate
    Geography
    Global scenario
    Period
    One 8,760-hour year using March 2026 specification
    Confidence
    high
    Trace evidence
  4. 04

    NVIDIA documents a maximum of 64 framebuffer page retirements for legacy GPUs and up to 512 framebuffer row-remapping entries for Ampere and later generations.

    Class
    vendor spec
    Geography
    Global
    Period
    Documentation reviewed August 2026
    Confidence
    high
    Trace evidence

System map

Connections across the system

Explore inputs, outputs, and shared capabilities in the Atlas. Open the register for every documented relationship.

Across the system boundary Selected documented relationships · arrows follow the Atlas

Enters this system 8

  1. Advanced packaging→ AI acceleratorTrace relationship
  2. Memory fabrication→ High-bandwidth memoryTrace relationship

Leaves this system 6

  1. Advanced packaging← High-bandwidth memoryTrace relationship
  2. Accelerator architecture← Runtime & frameworksTrace relationship
System 07Compute & Racks12 internal relationships

Shared capabilities can belong to more than one System. These links carry materials, energy, information, capital, or permissions; they describe dependencies, not quantities or a complete process model.

Explore 10 capabilities and operating boundaries
Atlas nodes / 10
All documented relationships · 26
  1. criticaloutput
    High-bandwidth memoryAdvanced packaging

    Qualified HBM stacks are bonded beside logic in advanced packages.

    Trace relationship
    Supporting evidence · 2
    1. [1] Micron specifies more than 1.2 TB/s of memory bandwidth per HBM3E stack.
    2. [2] TSMC's CoWoS platform integrates logic chiplets and high-bandwidth memory on an interposer within an advanced package.
  2. criticalinput
    Advanced packagingAI accelerator

    Assembly and test turn logic and HBM into a qualified accelerator module.

    Trace relationship
    Supporting evidence · 2
    1. [1] TSMC's CoWoS platform integrates logic chiplets and high-bandwidth memory on an interposer within an advanced package.
    2. [2] NVIDIA specifies 141 GB of HBM3E capacity and 4.8 TB/s of memory bandwidth for the H200 accelerator.
  3. criticalinternal
    High-bandwidth memoryAI accelerator

    Memory capacity and bandwidth bound accelerator workload performance.

    Trace relationship
    Supporting evidence · 2
    1. [1] Micron specifies more than 1.2 TB/s of memory bandwidth per HBM3E stack.
    2. [2] NVIDIA specifies 141 GB of HBM3E capacity and 4.8 TB/s of memory bandwidth for the H200 accelerator.
  4. criticalinternal
    AI acceleratorAI rack system

    Accelerator modules are integrated with power, cooling, network, and service systems.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
    2. [2] A DGX H200 system contains eight H200 GPUs and has a documented maximum input power of 10.2 kW; the system manual specifies front-to-back airflow.
  5. criticalinternal
    Runtime & frameworksAI rack system

    Drivers, kernels, frameworks, and scheduling convert hardware into useful work.

    Trace relationship
    Supporting evidence · 2
    1. [1] CUDA and ROCm each combine programming models, runtimes, compilers, libraries, debugging tools, and hardware interfaces into an accelerator software stack.
    2. [2] PyTorch documents data, fully sharded data, tensor, and pipeline parallelism as complementary approaches for distributed model training.
  6. criticalinput
    Memory fabricationHigh-bandwidth memory

    Qualified DRAM dies from memory fabs become the active layers of HBM stacks.

    Trace relationship
    Supporting evidence · 2
    1. [1] SK hynix described its March 2025 12-layer HBM4 customer sample as a 36 GB stack; the announcement documented sample and certification status rather than qualified volume production.
    2. [2] Micron lists HBM3E configurations of 24 GB in an eight-high stack and 36 GB in a twelve-high stack.
  7. criticalinput
    Package test & qualificationAI accelerator

    Final qualification determines whether integrated accelerator packages can enter systems.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA documents a maximum of 64 framebuffer page retirements for legacy GPUs and up to 512 framebuffer row-remapping entries for Ampere and later generations.
    2. [2] SIA's high-level semiconductor value chain runs from research and development through chip design, front-end wafer fabrication, back-end assembly, test and packaging, and finally circuit-board integration into products.
  8. criticalinternal
    Server & ODM integrationAI rack system

    Server and tray integration turns qualified components into serviceable rack-scale compute.

    Trace relationship
    Supporting evidence · 2
    1. [1] A DGX H200 system contains eight H200 GPUs and has a documented maximum input power of 10.2 kW; the system manual specifies front-to-back airflow.
    2. [2] Open Compute Project workstreams are addressing future AI rack power envelopes from 250 kW toward 1 MW alongside liquid-cooling standards.
  9. criticalinternal
    Rack power deliveryAI rack system

    Rack power conversion and distribution deliver usable electrical capacity to compute trays.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
    2. [2] Open Compute Project workstreams are addressing future AI rack power envelopes from 250 kW toward 1 MW alongside liquid-cooling standards.
  10. importantoutput
    Runtime & frameworksAccelerator architecture

    Software capabilities and workload behavior feed back into hardware design.

    Trace relationship
    Supporting evidence · 1
    1. [1] CUDA and ROCm each combine programming models, runtimes, compilers, libraries, debugging tools, and hardware interfaces into an accelerator software stack.
  11. importantinternal
    Host CPU & DPUAI accelerator

    Host processors dispatch work and manage data around accelerator execution.

    Trace relationship
    Supporting evidence · 1
    1. [1] A DGX H200 system contains eight H200 GPUs and has a documented maximum input power of 10.2 kW; the system manual specifies front-to-back airflow.
  12. importantinternal
    Host CPU & DPUAI rack system

    Host and infrastructure processors complete the rack compute platform.

    Trace relationship
    Supporting evidence · 1
    1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
  13. importantinput
    Scale-up fabricAI rack system

    Scale-up links couple accelerators within the rack-scale computer.

    Trace relationship
    Supporting evidence · 2
    1. [1] UALink 1.0 defines 200 Gb/s-per-lane links and a scale-up fabric designed for as many as 1,024 accelerators.
    2. [2] PCI Express 6.0 doubles the raw data rate to 64 GT/s and supports up to 256 GB/s bidirectionally on an x16 link.
  14. importantoutput
    AI rack systemScale-out network

    Rack endpoints attach to the cluster's scale-out switching fabric.

    Trace relationship
    Supporting evidence · 2
    1. [1] Broadcom specifies 102.4 Tb/s of switching capacity for a single Tomahawk 6 Ethernet switch chip.
    2. [2] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
  15. importantoutput
    Runtime & frameworksScale-out network

    Parallelism and collective libraries shape network traffic and topology needs.

    Trace relationship
    Supporting evidence · 1
    1. [1] PyTorch documents data, fully sharded data, tensor, and pipeline parallelism as complementary approaches for distributed model training.
  16. importantinput
    AI storageRuntime & frameworks

    Datasets and checkpoints feed training and recovery workflows.

    Trace relationship
    Supporting evidence · 1
    1. [1] Industry presenters working through SNIA describe AI storage as needing to sustain dataset ingest while absorbing checkpoints and supporting recovery without leaving accelerators idle.
  17. importantoutput
    AI rack systemData-center campus

    Rack power, mass, cooling, and service envelopes constrain facility design.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
    2. [2] Open Compute Project workstreams are addressing future AI rack power envelopes from 250 kW toward 1 MW alongside liquid-cooling standards.
  18. importantinput
    Liquid coolingAI rack system

    Liquid loops capture and transport heat from high-density rack components.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
    2. [2] Open Compute Project workstreams are addressing future AI rack power envelopes from 250 kW toward 1 MW alongside liquid-cooling standards.
  19. importantinput
    Export controlsHigh-bandwidth memory

    Export rules can restrict advanced memory products and related technology.

    Trace relationship
    Supporting evidence · 1
    1. [1] The December 2024 U.S. package added controls on 24 categories of semiconductor manufacturing equipment, three software-tool categories, and HBM, plus 140 Entity List additions.
  20. importantinternal
    AI rack systemHardware lifecycle

    Refresh and retirement send systems into reuse, parts recovery, or recycling.

    Trace relationship
    Supporting evidence · 2
    1. [1] The world generated 62 million tonnes of electronic waste in 2022, while 22.3 percent was documented as formally collected and recycled.
    2. [2] UNITAR projects global electronic waste to reach 82 million tonnes in 2030 under current trends.
  21. importantinternal
    Server & ODM integrationAI rack system

    Server & ODM integration supplies a distinct operating capability within its broader Atlas system.

    Trace relationship
    Supporting evidence · 2
    1. [1] A DGX H200 system contains eight H200 GPUs and has a documented maximum input power of 10.2 kW; the system manual specifies front-to-back airflow.
    2. [2] Open Compute Project workstreams are addressing future AI rack power envelopes from 250 kW toward 1 MW alongside liquid-cooling standards.
  22. importantinternal
    Rack power deliveryAI rack system

    Rack power delivery supplies a distinct operating capability within its broader Atlas system.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA's DGX GB200 documentation describes a rack-scale system with 72 GPUs, 36 Grace CPUs, and an approximately 120 kW power envelope, with liquid cooling alongside air-cooled components.
    2. [2] Open Compute Project workstreams are addressing future AI rack power envelopes from 250 kW toward 1 MW alongside liquid-cooling standards.
  23. importantinternal
    System memory hierarchyHost CPU & DPU

    System memory hierarchy supplies a distinct operating capability within its broader Atlas system.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA specifies 141 GB of HBM3E capacity and 4.8 TB/s of memory bandwidth for the H200 accelerator.
    2. [2] CXL 3.1 supports fabric-attached memory use cases including memory sharing and pooling across hosts and devices.
  24. importantinternal
    Firmware & rack managementAI rack system

    Firmware & rack management supplies a distinct operating capability within its broader Atlas system.

    Trace relationship
    Supporting evidence · 2
    1. [1] NVIDIA documents a maximum of 64 framebuffer page retirements for legacy GPUs and up to 512 framebuffer row-remapping entries for Ampere and later generations.
    2. [2] CUDA and ROCm each combine programming models, runtimes, compilers, libraries, debugging tools, and hardware interfaces into an accelerator software stack.
  25. importantinput
    Repair, reuse & recyclingHardware lifecycle

    Repair, reuse & recycling supplies a distinct operating capability within its broader Atlas system.

    Trace relationship
    Supporting evidence · 3
    1. [1] U.S. EPA guidance places reuse and redeployment of functional electronics ahead of recycling, with refurbishment, repair, component harvesting, and secure data handling used to retain equipment value where practicable.
    2. [2] The world generated 62 million tonnes of electronic waste in 2022, while 22.3 percent was documented as formally collected and recycled.
    3. [3] UNITAR projects global electronic waste to reach 82 million tonnes in 2030 under current trends.
  26. enablingoutput
    Hardware lifecycleCritical minerals

    Recovery can return selected metals to material supply and reduce virgin demand.

    Trace relationship
    Supporting evidence · 1
    1. [1] The world generated 62 million tonnes of electronic waste in 2022, while 22.3 percent was documented as formally collected and recycled.

Named companies + institutions

Organizations & their roles

  1. principal · consortium

    CXL Consortium

    Compute Express Link Consortium

    Industry consortium maintaining the Compute Express Link cache-coherent interconnect standard.

    Roles, locations & evidence
    Role
    standards setter
    Products / service
    coherent memory interconnect standards
    Documented activity
    World
    Headquarters
    Not cataloged
    Why this organization belongs in the System
    • System memory hierarchy

      Primary-source evidence connects this institution to coherent memory interconnect standards in the specified Atlas capability.

  2. principal · nonprofit

    Open Compute Project

    Open Compute Project Foundation

    Industry foundation developing open specifications and practices for scalable computing infrastructure.

    Roles, locations & evidence
    Role
    standards setter
    Products / service
    open rack and AI-cluster infrastructure specifications
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • AI rack system

      Primary-source evidence connects this institution to open rack and ai-cluster infrastructure specifications in the specified Atlas capability.

  3. principal · standards body

    PCI-SIG

    PCI Special Interest Group

    Industry standards body maintaining the PCI Express interconnect specification.

    Roles, locations & evidence
    Role
    standards setter
    Products / service
    PCI Express component interconnect standards
    Documented activity
    World
    Headquarters
    Not cataloged
    Why this organization belongs in the System
    • Host CPU & DPU

      Primary-source evidence connects this institution to pci express component interconnect standards in the specified Atlas capability.

  4. material · public company

    Arista Networks

    Arista Networks, Inc.

    Cloud-networking company supplying high-speed switching systems, routing platforms, and network software.

    Roles, locations & evidence
    Role
    supplier
    Products / service
    Network operating software
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • Runtime & frameworks

      Arista Networks's filing-backed operating portfolio includes network operating software; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

All other organizations · 12
  1. material · public company

    Cisco

    Cisco Systems, Inc.

    Networking and security company supplying switching, routing, optical, observability, and data-center infrastructure.

    Roles, locations & evidence
    Role
    supplier
    Products / service
    Network operating and observability software
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • Runtime & frameworks

      Cisco's filing-backed operating portfolio includes network operating and observability software; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  2. material · public company

    Eaton

    Eaton Corporation plc

    Power-management company supplying electrical distribution, switchgear, UPS, and power-quality systems.

    Roles, locations & evidence
    Role
    supplier
    Products / service
    Power distribution and conversion
    Documented activity
    World
    Headquarters
    Ireland
    Why this organization belongs in the System
    • Rack power delivery

      Eaton's filing-backed operating portfolio includes power distribution and conversion; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  3. material · public company

    HPE

    Hewlett Packard Enterprise Company

    Enterprise infrastructure company supplying servers, supercomputers, networking, storage, and services.

    Roles, locations & evidence
    Role
    supplier · manufacturer
    Products / service
    Integrated rack-scale systems · AI and high-performance computing systems
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • AI rack system

      HPE's filing-backed operating portfolio includes integrated rack-scale systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • Server & ODM integration

      HPE's filing-backed operating portfolio includes ai and high-performance computing systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  4. material · nonprofit

    PyTorch Foundation

    Open-source foundation supporting the PyTorch machine-learning framework and distributed-computing ecosystem.

    Roles, locations & evidence
    Role
    standards setter
    Products / service
    distributed machine-learning software
    Documented activity
    World
    Headquarters
    Not cataloged
    Why this organization belongs in the System
    • Runtime & frameworks

      Primary-source evidence connects this institution to distributed machine-learning software in the specified Atlas capability.

  5. material · research institution

    UNITAR

    United Nations Institute for Training and Research

    United Nations institute publishing the Global E-waste Monitor with international partners.

    Roles, locations & evidence
    Role
    community party
    Products / service
    global electronics-lifecycle evidence
    Documented activity
    World
    Headquarters
    Not cataloged
    Why this organization belongs in the System
    • Hardware lifecycle

      Primary-source evidence connects this institution to global electronics-lifecycle evidence in the specified Atlas capability.

  6. representative · public company

    AMD

    Advanced Micro Devices, Inc.

    Semiconductor company that designs CPUs, GPUs, adaptive computing products, and associated AI software.

    Roles, locations & evidence
    Role
    designer · supplier
    Products / service
    Instinct data-center accelerators · EPYC server processors · ROCm software platform
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • AI accelerator

      AMD's filing-backed operating portfolio includes instinct data-center accelerators; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • Host CPU & DPU

      AMD's filing-backed operating portfolio includes epyc server processors; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • Runtime & frameworks

      AMD's filing-backed operating portfolio includes rocm software platform; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  7. representative · public company

    Dell Technologies

    Dell Technologies Inc.

    Enterprise technology company supplying servers, storage, networking, and integrated AI infrastructure.

    Roles, locations & evidence
    Role
    supplier · manufacturer
    Products / service
    Rack-scale AI infrastructure · AI servers and integrated systems
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • AI rack system

      Dell Technologies's filing-backed operating portfolio includes rack-scale ai infrastructure; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • Server & ODM integration

      Dell Technologies's filing-backed operating portfolio includes ai servers and integrated systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  8. representative · public company

    Intel

    Intel Corporation

    Integrated semiconductor company spanning processor design, fabrication, packaging, systems, and foundry services.

    Roles, locations & evidence
    Role
    designer
    Products / service
    Xeon server processors
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • Host CPU & DPU

      Intel's filing-backed operating portfolio includes xeon server processors; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  9. representative · public company

    Micron

    Micron Technology, Inc.

    Memory and storage manufacturer supplying DRAM, HBM, NAND, and related products for AI systems.

    Roles, locations & evidence
    Role
    manufacturer
    Products / service
    High-bandwidth memory · HBM advanced-packaging facility under construction
    Documented activity
    World · Singapore
    Headquarters
    Boise
    Why this organization belongs in the System
    • High-bandwidth memory

      Micron's filing-backed operating portfolio includes high-bandwidth memory; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • High-bandwidth memory

      Micron's project updates directly locate the HBM packaging buildout in Singapore and retain its contribution as future, not operating, capacity.

  10. representative · public company

    NVIDIA

    NVIDIA Corporation

    Computing-platform company that designs AI accelerators, interconnects, systems, and the software stack that operates them.

    Roles, locations & evidence
    Role
    designer · supplier
    Products / service
    Data-center GPU accelerators · Rack-scale AI systems · CUDA software platform
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • AI accelerator

      NVIDIA's filing-backed operating portfolio includes data-center gpu accelerators; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • AI rack system

      NVIDIA's filing-backed operating portfolio includes rack-scale ai systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • Runtime & frameworks

      NVIDIA's filing-backed operating portfolio includes cuda software platform; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  11. representative · public company

    Supermicro

    Super Micro Computer, Inc.

    Server and systems manufacturer supplying modular, rack-scale, and liquid-cooled AI infrastructure.

    Roles, locations & evidence
    Role
    manufacturer
    Products / service
    Rack-scale AI systems · AI servers and rack integration
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • AI rack system

      Supermicro's filing-backed operating portfolio includes rack-scale ai systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

    • Server & ODM integration

      Supermicro's filing-backed operating portfolio includes ai servers and rack integration; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

  12. representative · public company

    Vertiv

    Vertiv Holdings Co

    Critical-digital-infrastructure company supplying power, thermal management, and service systems for data centers.

    Roles, locations & evidence
    Role
    supplier
    Products / service
    Rack and facility power systems
    Documented activity
    World
    Headquarters
    United States
    Why this organization belongs in the System
    • Rack power delivery

      Vertiv's filing-backed operating portfolio includes rack and facility power systems; no subregional operating place is asserted, and headquarters is not used as a proxy. Representative status does not imply market rank.

Affected policies

Relevant policies

Legal status is separated from policy objective. Every record keeps its jurisdiction, mechanism, affected nodes, and verification date.

  1. effective

    Advanced-semiconductor export controls, as amended

    Restrict specified capabilities used to produce advanced semiconductors and advanced computing systems in China.

    Mechanism, scope & sources

    Combines controlled equipment, software, HBM, entity, end-use, and license requirements with later transaction-specific exceptions and case-by-case review policies in the current EAR framework.

    United States export administrationOpen full policy record

Read this System in the field guide

Continue in the Volumes

Start with three selected readings, or explore the complete reading list.

Start here
  1. 01
    Volume IV · MachineNode anatomy
  2. 02
    Volume IV · MachineRack power delivery
  3. 03
    Volume IV · MachineHeat and sustained performance
Complete reading list
Volume IV

Machine

13 spreads
  1. IV · 01From accelerator to node
  2. IV · 02Node anatomy
  3. IV · 03Host interconnect and memory expansion
  4. IV · 04Firmware, management, and security
  5. IV · 05Tray, chassis, and rack
  6. IV · 06Scale-up fabric
  7. IV · 07Rack power delivery
  8. IV · 08Cabling, materials, and mechanics
  9. IV · 09Heat and sustained performance
  10. IV · 10The rack liquid loop
  11. IV · 11Reliability and serviceability
  12. IV · 12Build, ship, install, accept
  13. IV · 13Refresh, reuse, and retirement
Related cross-System reading · 27 spreads
Volume I

Matter

3 spreads
  1. I · 01Conductors and joints
  2. I · 02Critical minerals
  3. I · 03Power and thermal materials
Volume II

Precision

1 spreads
  1. II · 01Compute architectures
Volume III

Package

4 spreads
  1. III · 01Capacity and bandwidth
  2. III · 02Power and heat begin in the package
  3. III · 03Inspection, test, and qualification
  4. III · 04From package to module
Volume V

Fabric

12 spreads
  1. V · 01The real workload
  2. V · 02Distance has a price
  3. V · 03Scale up
  4. V · 04Scale out
  5. V · 05Endpoints
  6. V · 06Routes and patching
  7. V · 07Congestion and topology
  8. V · 08The data pipeline
  9. V · 09Memory as fabric
  10. V · 10Observe and recover
  11. V · 11Programming stacks
  12. V · 12Parallelism
Volume VI

Plant

6 spreads
  1. VI · 01The data hall
  2. VI · 02Voltage path
  3. VI · 03Prove the plant
  4. VI · 04Heat path
  5. VI · 05Air and liquid
  6. VI · 06Coolant distribution
Volume VII

Permission

1 spreads
  1. VII · 01Training, inference, and utilization

Evidence & review

currentEvidence health35 active claims · next review Oct 13, 2026
Active claims
35
Sources
48
High volatility
7
Next review
Oct 13, 2026
Due soon
0
Overdue
0
Superseded
0
Open complete evidence ledger

Behind the claim

Evidence, in context.

Opening the evidence record…