GS-VNDVendor index
Who actually designs NVIDIA AI servers.
Every company on this index engineers at least part of the machine itself: the motherboard or baseboard carrier, the liquid-cooling loop, the rack, the firmware. Pure integrators who assemble other people's barebones are deliberately excluded. Each profile states what the vendor designs, which NVIDIA platforms it ships, and cites its sources.
Admission criterion
A vendor is listed when we can document genuine in-house hardware engineering around NVIDIA silicon: board or baseboard design, thermal and liquid-cooling engineering, rack-scale integration engineering, or platform firmware. Resellers and integrators who build on Supermicro, GIGABYTE or ASUS barebones are useful businesses, but they belong in a channel guide, not here. Every profile carries a verification date and a source list.
§ 01
Global OEMs
Full-line server makers with their own board, chassis and cooling engineering, sold as turnkey systems with global support.
| Vendor | HQ | NVIDIA platforms shipped | Designs in-house | |
|---|---|---|---|---|
| Supermicro | San Jose, California, USA | HGX B200, HGX B300, GB200 NVL72, GB300 NVL72 | DLC-2: second-generation in-house direct liquid cooling with own in-rack CDU (250kW) and in-row CDU (1.8MW); cold plates on CPU, GPU, memory, PCIe switches and VRMs, up to 98% heat capture, 45C inlet liquid | Profile → |
| Dell Technologies | Round Rock, Texas, USA | GB200 NVL72, GB200 NVL4, HGX H100/H200/B200, HGX B200 | Integrated Rack 7000 (IR7000): own 21-inch ORv3-based rack platform purpose-built for liquid cooling, up to 480kW per rack, near 100% heat capture | Profile → |
| HPE | Spring, Texas, USA | HGX H100/H200, GB200 NVL72, GB300 NVL72, PCIe | Cray supercomputing heritage (acquired 2019): the source of HPE's HPC and DLC engineering stack | Profile → |
| Lenovo | Beijing, China and Morrisville, North Carolina, USA (dual HQ) | GB300 NVL72, GB200 NVL4, HGX B200, HGX B300 | Lenovo Neptune: proprietary warm-water direct liquid cooling, now 6th generation; 5th gen claimed up to 40% lower power vs air | Profile → |
| Cisco | San Jose, California, USA | HGX B300 NVL8, HGX H100/H200, PCIe | UCS C-Series dense GPU servers (C885A M8, C880A M8) built on NVIDIA HGX reference architecture with Cisco-designed board and NIC integration and Cisco BMC | Profile → |
| Fujitsu | Kawasaki, Japan | HGX B200, HGX B300, HGX H100 (4-GPU), HGX A100 (8-GPU) | Current flagship Blackwell server PRIMERGY GX2570 M8s (HGX B200) is a Supermicro OEM product: Fujitsu wraps SupportDesk maintenance and Infrastructure Manager tooling around a Supermicro design | Profile → |
| Bull (ex Eviden) | Les Clayes-sous-Bois, France | GH200 (Blackwell roadmap), GH200, HGX H100/H200, GB200 NVL72 | Genuine in-house design house: BullSequana XH3500 (Nov 2025) ships fan-less 5th-generation direct liquid cooling, an iterative in-house DLC lineage | Profile → |
| Hitachi Vantara | Santa Clara, California, USA | HGX H100/H200, later HGX B200, MGX, HGX B200, GB300 NVL72 | Core engineering strength is storage and data software (VSP One, Hitachi Content Software for File) plus the iQ Studio AI-orchestration layer, not GPU server hardware | Profile → |
| IBM | Armonk, New York, USA | storage attach (HGX, GB200 NVL72), PCIe / HGX-class cloud, Power9 + V100 NVLink | IBM does not build NVIDIA HGX/MGX/NVL72 GPU compute servers and is not an NVIDIA systems builder in the Dell/HPE/Supermicro sense | Profile → |
| NEC | Tokyo, Japan | PCIe, HGX A100 (Supermicro-built) | NEC's NVIDIA-Certified Systems are all PCIe GPU rack servers (Express5800, NX7700x), not HGX baseboard, MGX or NVL72 platforms; no Blackwell-generation NEC system found as of mid-2026 | Profile → |
§ 02
ODM system designers
The design manufacturers who engineer the boards, baseboards and rack systems themselves, for hyperscalers, for NVIDIA, and increasingly under their own brand.
| Vendor | HQ | NVIDIA platforms shipped | Designs in-house | |
|---|---|---|---|---|
| Foxconn / Ingrasys | New Taipei City, Taiwan | GB200 NVL72, GB300 NVL72, HGX B200 / HGX B300 / MGX, Vera Rubin NVL72 | Ingrasys spans component-level design (chassis, boards) through L10 system integration to L12 rack-and-cluster integration, plus liquid cooling | Profile → |
| Quanta / QCT | Taoyuan, Taiwan | GB200 NVL72, MGX, HGX B200 / HGX B300 | Designs its own server boards and compute trays; builds complete GB200 NVL72 racks from 18x QuantaGrid D75B-1U trays | Profile → |
| Wistron | Taipei, Taiwan | HGX B200 / HGX B300, GB200 NVL72 / GB300 NVL72, MGX | Self-described world leader in AI baseboard and GPU platform manufacturing; NVIDIA HGX Partner Program member with early reference-architecture access | Profile → |
| Wiwynn | Taipei, Taiwan | HGX B200 / HGX B300, HGX B300 NVL16, GB200 NVL72 / GB300 NVL72, GB300 NVL72 | ODM-direct model: end-to-end from board-level components through complete L11 rack integration; OCP Platinum member | Profile → |
| Inventec | Taipei, Taiwan | HGX B200, HGX B300, GB200 NVL72 / GB300 NVL72, Vera Rubin NVL72 | Builds full racks, not just boards: Artemis rack line for GB200/GB300 NVL72, MGX 1U-architecture compliant, performing motherboard assembly and full-rack integration (L10/L11) | Profile → |
| Pegatron | Taipei, Taiwan | GB200 NVL72, Vera Rubin NVL72, HGX B200, GB300 NVL72 | In-house-designed RA4401-72N1 GB200 NVL72 rack, listed on the Open Compute Project | Profile → |
| Compal Electronics | Taipei, Taiwan | HGX Rubin NVL8 (Vera Rubin), GB200 NVL72, HGX B200 | SG231-2-L1: in-house-designed 2U, 8x NVIDIA Rubin GPU direct-liquid-cooled server on HGX Rubin NVL8, unveiled at GTC 2026 (~24kW, 400 PFLOPS NVFP4) | Profile → |
| MiTAC Computing (TYAN) | Taoyuan, Taiwan (US office: Newark, California) | HGX H200 NVL, MGX (Grace/Vera CPU), MGX | In-house TYAN-brand motherboard and server engineering heritage dating to 1989 | Profile → |
| GIGABYTE (Giga Computing) | New Taipei City, Taiwan | HGX B300 NVL16, HGX B200/B300, HGX B300, GB200 NVL72 | Decades of in-house motherboard and server engineering heritage predating the 2023 spin-off | Profile → |
| ASUS | Taipei, Taiwan | GB300 NVL72, HGX B300, HGX B200, PCIe | In-house server design lineage since 1995; ESC GPU-server line and XA rack-scale line designed by its own Infrastructure Business Group | Profile → |
| ASRock Rack | Taipei, Taiwan | HGX B200, HGX B300 NVL16, HGX B300, GB200 NVL72 | Designs its own server motherboards, barebones and full rack systems in-house; R&D-centric with manufacturing outsourced to the Pegatron group footprint | Profile → |
| MSI | New Taipei City, Taiwan | MGX, DGX Spark | Server unit branded Enterprise Platform Solutions, repositioning from consumer vendor to AI infrastructure provider | Profile → |
| AIC Inc. | Taoyuan, Taiwan | PCIe, PCIe edge, BlueField DPU storage | Nearly 30 years of in-house high-density storage-server and server-barebone engineering, OEM/ODM and own brand | Profile → |
§ 03
Hyperscale design-manufacturers
US and international firms that design and build AI racks to order, hyperscaler-style, outside the classic OEM channel.
| Vendor | HQ | NVIDIA platforms shipped | Designs in-house | |
|---|---|---|---|---|
| Hyve Solutions | Fremont, California, USA | HGX B200, MGX | Fully vertically integrated design manufacturer: design, US-based PCBA/SMT manufacturing, thermal engineering, rack integration and deployment | Profile → |
| Aivres | Milpitas, California, USA | HGX B300, HGX H200, GB200 NVL72 / GB300 NVL72, Vera Rubin NVL72 | Joint-design-manufacturing model with in-house product design and engineering teams; claims 2,000+ full-lifecycle qualification tests | Profile → |
| KAYTUS | Singapore | GB200 NVL72, GB300 NVL72, HGX B300, MGX | Claims 3 factories, 2 R&D centers and a joint-design-manufacturing model; no factory or R&D addresses published | Profile → |
| Celestica | Toronto, Canada | Ethernet switch (non-NVIDIA), GB200 NVL72 | Joint-design-manufacturing model for hyperscalers, designing from silicon level up rather than build-to-spec only | Profile → |
| ZT Systems | Secaucus, New Jersey, USA | GB200 NVL72, PCIe | The split, precisely: AMD acquired ZT for $4.9B (closed March 2025), kept the ~1,000-engineer design business for AMD Instinct rack-scale systems, and divested manufacturing plus the ZT Systems brand to Sanmina (closed Oct 2025) | Profile → |
| Flex | Austin, Texas, USA (Singapore-incorporated) | Vera Rubin (power), Vera Rubin NVL72 (power), MGX (cooling), PCIe (cooling) | Vertically integrated grid-to-chip datacenter model: compute integration plus power distribution plus liquid cooling as complete rack-level solutions, racks rated to 120kW with a stated path to 300kW | Profile → |
§ 04
Edge & specialist designers
Designers of short-depth, rugged and edge NVIDIA systems: MGX servers, Jetson platforms and defense-grade GPU boxes.
| Vendor | HQ | NVIDIA platforms shipped | Designs in-house | |
|---|---|---|---|---|
| Aetina | New Taipei City, Taiwan | MGX, Jetson Thor, Jetson Orin, PCIe (MXM) | Designs its own Jetson carrier boards in-house (ACE-N510/N622, AN110) and offers custom carrier-board engineering from concept to mass production | Profile → |
| Advantech | Taipei, Taiwan | MGX, Jetson Thor, Jetson Orin | World's largest industrial-PC supplier with in-house design and manufacturing | Profile → |
| ADLINK | Taoyuan, Taiwan | Jetson Thor, Jetson (IGX Thor), Jetson Orin, PCIe (MXM) | Designs its own embedded modules, carrier boards and systems in-house: COM Express, MXM GPU modules, robotic controllers, smart cameras | Profile → |
| One Stop Systems | Escondido, California, USA | PCIe, SXM A100 | PCIe expansion and fabric design is the core in-house specialty; PCI-SIG-active engineering, consistently first to each PCIe generation (currently Gen 5) | Profile → |
§ 05
China domestic (export-restricted)
Server designers a Western buyer generally cannot procure NVIDIA systems from, documented because their spinoffs and their history shape the market.
| Vendor | HQ | NVIDIA platforms shipped | Designs in-house | |
|---|---|---|---|---|
| Inspur | Jinan, Shandong, China | HGX, PCIe | Only server vendor holding membership in all three open-hardware bodies: OCP (Platinum), ODCC and Open19, with the OCP-accepted NF8380M5 four-socket platform contributed | Profile → |
| H3C | Hangzhou / Beijing, China | HGX / PCIe, HGX | Two decades of in-house hardware engineering lineage from the original Huawei-3Com networking and server joint venture | Profile → |
| xFusion | Zhengzhou, Henan, China | domestic accelerator, PCIe | Inherited Huawei's x86/GPU server engineering team and IP when spun out in September 2021 | Profile → |
| Sugon | Beijing, China | domestic accelerator, PCIe | Spin-off of Chinese Academy of Sciences supercomputing research, with ongoing joint R&D programs | Profile → |
Index changelog
2026-07-15Added the US manufacturing tracker, the statistics page and head-to-head vendor comparisons.
2026-07-15Index launched: 37 vendor profiles across 5 tiers, each sourced and dated. Notable calls at launch: Fujitsu's Blackwell flagship documented as a Supermicro OEM product; Eviden recorded under its new Bull identity and French-State ownership; Celestica and Flex profiled with their non-NVIDIA compute roles stated plainly; Aivres and KAYTUS carry explicit Entity-List caveats.
NOTE — Profiles are engineering references, not endorsements or offers. GPU Smith is an independent firm and is not affiliated with NVIDIA or any vendor listed. The assessment quotes specific vendors and pricing for a defined build.
Request an assessment