What Nvidia makes for data centers

Seen in real machines

Nvidia is named in 13 of the rack teardowns and supercomputer breakdowns on this site. Each line below is a piece of that machine.

• Nvidia GB300 NVL72 · rack teardown
  • Rack architecture and reference design (NVDA)
  • 18× compute trays (Bianca board: 2× Grace + 4× B300)
  • 9× NVLink switch trays (2× NVSwitch each)
  • and 1 more on the Nvidia GB300 NVL72 page
• JUPITER · supercomputer
  • Booster compute node (4x GH200 superchips, 4x NDR200 InfiniBand adapters, 960 GB NVMe boot disk; 5,884 nodes in about 125 racks) (superchip and InfiniBand adapters)
  • GH200 Grace Hopper Superchip (Hopper GPU plus 72-core Grace CPU joined by NVLink-C2C)
  • System memory (Booster 4x 120 GB LPDDR5X per node; Cluster 768 GB DDR5 per Turin node, 512 GB DDR5 per Rhea1 node) (the Booster LPDDR5X sits on the Grace Hopper module)
  • and 2 more on the JUPITER page
• Alps · supercomputer
  • GH200 compute blade and node (2 nodes per blade, 4 Grace Hopper modules and 4 NICs per node; 7 of 8 blade slots filled per chassis; 2,688 nodes) (Grace Hopper modules fitted four to a node)
  • GH200 Grace Hopper Superchip (72-core Grace CPU plus Hopper GPU with 96 GB HBM3, joined by NVLink-C2C; 10,752 superchips)
  • NVIDIA A100 partition (144 nodes, one AMD EPYC host and 4 A100 GPUs per node, 576 GPUs) and AMD Instinct MI250X partition (24 nodes, 4 MI250X per node, 96 GPUs; HPE Cray EX235A blades) (A100 GPUs, 4 per node)
• LUMI · supercomputer
  • LUMI-D visualization nodes (eight nodes with 2 x AMD EPYC 7742 and 8 x NVIDIA A40 each; 64 A40 GPUs in total) (A40 visualization GPUs; the docs say their main purpose is visualization, not GPU computing)
• Leonardo · supercomputer
  • A100 GPU (custom 64 GB SXM4 part, four per Booster node; GA100 die fabricated on TSMC 7nm N7) (A100 Tensor Core GPU, NVLink 3.0 inside the node)
  • Network adapters (ConnectX-6 HDR100: two dual-port cards per Booster node, one single-port card per Data Centric node) (Mellanox ConnectX-6 InfiniBand adapters)
  • InfiniBand fabric (QM8700 HDR switches, 823 in total, Dragonfly+ across 23 cells with an all-to-all link between every cell pair) (Quantum QM8700 HDR switches (the former Mellanox line); the paper's description of the four Ethernet-to-InfiniBand gateways cites Nvidia's Skyway datasheet)
  • and 1 more on the Leonardo page
• Isambard-AI · supercomputer
  • Compute node (4x GH200 superchips with 4x Slingshot NICs; 42 Phase 1 nodes plus 1,320 Phase 2 nodes, 1,362 in total) (GH200 superchips on each node)
  • GH200 Grace Hopper Superchip (Hopper GPU plus 72-core Grace CPU joined by NVLink-C2C, 5,448 in the full system)
  • System memory (4x 120 GB LPDDR5X per node, 460 GB usable, on the Grace Hopper modules) (the LPDDR5X sits on the Grace Hopper module, per Nvidia's architecture write-up)
• Venado · supercomputer
  • Grace Hopper compute blades and nodes (2,560 GH200 Superchips, direct liquid cooled; TOP500 names the blade HPE Cray EX254n) (GH200 Grace Hopper Superchips, 2,560 in the system)
  • GH200 Grace Hopper Superchip (Arm-based Grace CPU plus Hopper H100 GPU joined by NVLink-C2C; 2,560 superchips)
  • Grace CPU Superchip partition (920 CPU-only superchips, each two 72-core Grace CPUs, 144 Arm cores, linked by NVLink-C2C; up to 960 GB LPDDR5X per superchip) (Grace CPU Superchip, the first large-scale deployment of the part in the United States per LANL)
• Perlmutter · supercomputer
  • NVIDIA A100 accelerator (Ampere GA100, 40 GB HBM2 or 80 GB HBM2e, fabricated on TSMC 7nm; four per GPU node) (designer of the A100 GPU)
• Eagle (Microsoft Azure) · supercomputer
  • NVIDIA H100 Tensor Core GPU (SXM, 80 GB, 8 per VM; 14,400 in the TOP500 run) (H100 GPUs, NVSwitch and NVLink 4.0 baseboard fabric)
  • InfiniBand adapters (ConnectX-7, 400 Gb/s NDR, one per GPU, 8 per VM) (Quantum-2 ConnectX-7 InfiniBand adapters)
  • In-node GPU fabric (NVSwitch and NVLink 4.0, 3.6 TB/s bisectional bandwidth between the 8 GPUs of each VM) (NVSwitch and NVLink 4.0, part of the H100 SXM platform)
  • and 1 more on the Eagle (Microsoft Azure) page
• Eos · supercomputer
  • System design and integration (NVIDIA DGX SuperPOD built from 576 DGX H100 systems, designed and assembled by NVIDIA for its own use) (designs the DGX SuperPOD architecture and built Eos from its own DGX H100 systems; NVIDIA also states Eos serves as a blueprint for its OEM and cloud partners)
  • DGX H100 compute system (8U server with eight H100 GPUs on an NVSwitch baseboard, two Xeon hosts, 640 GB of HBM3 and 2 TB of DDR5, up to 10.2 kW) (designs and sells the DGX H100 system; Eos is made of 576 of them)
  • H100 GPU and in-system NVLink/NVSwitch (900 GB/s between any two GPUs in a DGX H100) (H100 Tensor Core GPU, fourth-generation NVLink and third-generation NVSwitch inside each DGX H100; NVIDIA states there is no NVSwitch fabric between nodes on Eos)
  • and 3 more on the Eos page
• MareNostrum 5 · supercomputer
  • H100 GPU (Hopper, 64 GB HBM2e, four per accelerated-partition node) (H100 GPUs; BSC's slides list Nvidia as a technology provider)
  • In-node GPU fabric (NVLink between the four H100 GPUs, 300 GB/s bidirectional between pairs; x16 PCIe Gen5 to the host CPUs) (NVLink on the H100 modules)
  • InfiniBand adapters (ConnectX-7 NDR200, one per GPU in the accelerated nodes; one per two nodes in the general-purpose partition) (ConnectX-7 InfiniBand adapters, named in BSC's paper; BSC's slides list Nvidia as a technology provider)
  • and 1 more on the MareNostrum 5 page
• Horizon · supercomputer
  • Grace Blackwell compute nodes (GB200 NVL4: 1 Grace CPU and 2 Blackwell GPUs per node; 2,000 nodes, 4,000 GPUs) (Blackwell GPUs and Grace CPUs on the GB200 NVL4 platform)
  • Vera CPU nodes (4,750 planned, two 88-core Vera sockets each, 72 nodes per rack; due early 2027) (Vera CPUs)
  • InfiniBand fabric (NVIDIA XDR switches in two non-blocking fat trees joined by a third level; Quantum-X800) (XDR InfiniBand switches (slides name QM3400 and QM3200 models) and Quantum-X800 platform; the storage back-end network is designed and supplied by VAST)
• Doudna (NERSC-10) · supercomputer
  • Compute platform (NVIDIA Vera Rubin CPU-GPU platform: Rubin GPUs and Vera CPUs) (Rubin GPUs and Vera CPUs; NVIDIA's June 2026 release names Doudna as a Dell system powered by Vera Rubin)
  • InfiniBand network (NVIDIA Quantum-X800, with quality-of-service for data streamed in over ESnet) (Quantum-X800 InfiniBand networking platform)
  • Early-access system Cech (72 Grace CPUs, 144 Blackwell GPUs, Quantum-X800 InfiniBand; Dell IR7000-series direct-liquid-cooled rack with an in-row coolant distribution unit and rear-door heat exchanger) (Grace CPUs, Blackwell GPUs and Quantum-X800 InfiniBand)

Disclosed customers

From the blog

• Nine supercomputers, six countries, one foundry 2026-10-05
• Every company inside a GB300 NVL72 rack, by exchange 2026-09-02
• Introducing the catalog: every part of an AI data center, and who makes it 2026-08-28

Latest feed

• Nvidia eyes converting old nuclear plants into massive AI data centers — Tom's Hardware, 2026-10-25
• Nvidia unveils OOI CPO prototype for post-2028 AI needs — DigiTimes, 2026-10-25
• Nvidia rejects near-term memory glut fears, leaves CXMT cooperation open — DigiTimes, 2026-10-24
  Nvidia's CEO stated that he does not anticipate a surplus in memory supply in the near term and maintained an open dialogue regarding collaboration with China's CXMT.
• MSI XpertStation WS300 Review NVIDIA DGX Station GB300 — ServeTheHome, 2026-10-23
  The article reviews the MSI XpertStation WS300, an AI solution built upon the NVIDIA DGX Station platform.
• Illegal Nvidia Blackwell Ultra B300 GPUs found inside state-owned China Telecom's network — Tom's Hardware, 2026-10-23
  A security firm discovered Nvidia Blackwell Ultra B300 AI chips operating within the China Telecom network, although the physical location of the restricted components remains unconfirmed.
• Lenovo mounts AI comeback with Nvidia Vera Rubin NVL72 deliveries and US plant expansion — DigiTimes, 2026-10-23
• Nvidia B300 chips reportedly surface on China Telecom-linked network, raising export-control questions — DigiTimes, 2026-10-23
  An Israeli security firm reportedly found 32 Nvidia Blackwell Ultra B300 GPUs on a network linked to China Telecom, raising concerns about US export controls.
• Analysis: Nvidia's Vera CPU takes on Intel and AMD with a platform play — DigiTimes, 2026-10-23
  Nvidia secured a CPU supply order from Modal, challenging established CPU manufacturers like Intel and AMD by expanding its hardware offerings beyond accelerators.
• Tesla, Nvidia reportedly invest KRW3.7 trillion to secure Samsung Electro-Mechanics ABF capacity — DigiTimes, 2026-10-23
  Tesla and Nvidia are reportedly investing combined funds into Samsung Electro-Mechanics to expand ABF substrate production capacity in South Korea.
• Analysis: Nvidia's DSX POD reshapes AI factory supply chains — DigiTimes, 2026-10-23
  Nvidia introduced its DSX POD reference design for AI data centers, which involves integrating and validating equipment within containerized modules for on-site deployment.