NVIDIA MGX
This article is about the NVIDIA server and rack reference architecture. For the Abu Dhabi state-backed investment company, see MGX.
NVIDIA MGX is a modular reference architecture that NVIDIA publishes so that server makers and contract manufacturers can build accelerated systems around NVIDIA GPUs, CPUs, DPUs and networking without engineering each machine from scratch. NVIDIA unveiled it as the "NVIDIA MGX server specification" at Computex in Taipei on May 28, 2023, presenting it as a way for manufacturers to assemble more than 100 server variations for AI, high performance computing and Omniverse workloads. [1] What began as a chassis and module specification for 1U, 2U and 4U servers has since become the rack-level standard behind NVIDIA's largest systems, including GB200 NVL72, GB300 NVL72 and the Vera Rubin platform, which NVIDIA calls the third generation of MGX rack-scale systems. [3][8]
MGX is a specification rather than a product NVIDIA sells. Partners build to it, and NVIDIA's own rack-scale systems are built to it, which is what makes it useful as a common denominator across the industry. Since late 2025 it has also become the mechanical and supply-chain half of NVLink Fusion, the program through which non-NVIDIA accelerators and CPUs enter NVIDIA racks. [12][13]
What MGX specifies
Manufacturers start from a basic system architecture for their chassis and then choose the GPU, CPU and DPU that suit a workload. [1] At launch the specification covered 1U, 2U and 4U chassis in air-cooled and liquid-cooled forms, the then-current NVIDIA GPU portfolio (H100, L40 and L4), the Grace CPU Superchip, the GH200 Grace Hopper Superchip and x86 CPUs, plus BlueField-3 DPUs and ConnectX-7 network adapters. [1]
The envelope has widened considerably. NVIDIA's MGX product page, as of September 2026, describes support for rack-scale and PCIe GPU solutions including Blackwell and Rubin; Arm CPUs such as Grace and Vera alongside x86 and other Arm-based processors; networking spanning NVLink for scale-up, Spectrum-X Ethernet, Quantum InfiniBand and BlueField DPUs; rack-scale AI-native storage; and NVIDIA Groq 3 LPX racks that scale 256 LPU accelerators in a single liquid-cooled MGX rack. [10]
NVIDIA has attached the same headline economics to MGX since 2023: building to the specification can cut development costs by up to three-quarters and shorten development time by two-thirds, to roughly six months. [1][2] A May 2025 NVIDIA technical blog put more specific numbers on it, claiming a reduction of $2 million to $4 million in research and development per platform from shared reference designs, factory pre-integration of roughly 80 percent of rack components (busbars, cold plates and power whips among them), and ODM deployment timelines falling from about 12 months to under 90 days. [4] These are vendor figures rather than independently measured ones.
MGX compared with HGX and DGX
NVIDIA drew the MGX-versus-HGX line in the original announcement. MGX offers flexible, multi-generational compatibility so that system builders can reuse existing designs and adopt next-generation parts without an expensive redesign; HGX, by contrast, is built around an NVLink-connected multi-GPU baseboard aimed at the highest-end AI and HPC systems. [1] The two are not alternatives in practice: an HGX baseboard can sit inside a chassis designed to MGX rules.
| Platform | What NVIDIA supplies | Who builds the finished system |
|---|---|---|
| MGX | A modular reference architecture and design files for chassis, racks, power, cooling and connectors | OEMs and ODMs, and NVIDIA for its own rack-scale systems [1][10] |
| HGX | A multi-GPU baseboard and platform design carrying SXM GPU modules | Server makers integrating the baseboard into their own systems [1] |
| DGX | A complete, NVIDIA-branded and NVIDIA-supported system | NVIDIA |
Generations and milestones
| Date | Milestone |
|---|---|
| May 28, 2023 | MGX server specification announced at Computex. ASRock Rack, ASUS, GIGABYTE, Pegatron, QCT and Supermicro named as adopters; SoftBank Corp. planned MGX-based data centers in Japan. [1] |
| February 2024 | Ahead of GTC 2024, NVIDIA previewed an MGX Pavilion in which system makers would show more than 500 servers across 18 racks, all built on GH200 Grace Hopper Superchips. [5] |
| June 2, 2024 | MGX extended to Blackwell, including the GB200 NVL2 platform. NVIDIA reported more than 90 systems from over 25 partners released or in development, up from 14 systems from six partners a year earlier. AMD and Intel said they would deliver their own CPU host processor module designs for MGX, covering the AMD Turin platform and the Intel Xeon 6 processor with P-cores. [2] |
| October 15, 2024 | NVIDIA contributed key portions of the GB200 NVL72 electro-mechanical design to the Open Compute Project. The press release stated that GB200 NVL72 is based on the MGX modular architecture. [3] |
| May 16, 2025 | NVIDIA published a technical overview of the MGX rack system around Computex 2025, citing more than 200 ecosystem partners adopting MGX components. [4] |
| October 13, 2025 | At the OCP Global Summit NVIDIA detailed the MGX-generation Vera Rubin open architecture rack servers, with more than 50 MGX partners preparing for them, and said it planned to contribute the upgraded rack and compute tray designs to OCP as an open standard. [6] |
| March 16, 2026 | The Vera CPU rack, integrating 256 liquid-cooled Vera CPUs, was announced as an MGX-based design supported by 80 ecosystem partners. NVIDIA cited more than 80 MGX ecosystem partners behind the Vera Rubin platform. [7][9] |
| May 31, 2026 | NVIDIA said Vera Rubin marks the third generation of MGX rack-scale systems and is ramping with hundreds of supply-chain partners, 150 of them in Taiwan, across more than 350 factories in 30 countries. [8] |
| September 10, 2026 | d-Matrix announced it would integrate its Raptor inference XPUs into the NVIDIA MGX rack architecture through NVLink Fusion. [12][13] |
The partner counts above are not directly comparable. The 2024 figure counts released or in-development MGX systems and the partners shipping them, the 2025 figure counts companies making MGX components, and the 2026 figures count supply-chain participants for a specific rack generation.
Inside an MGX rack
An MGX rack is organised around two kinds of module: compute trays, which carry the CPU and GPU combinations that do the work, and NVLink switch trays, which provide the scale-up fabric linking those trays into a single domain. [4] Around them sits a specified set of mechanical, electrical and cooling parts, which is where most of the standardisation value lies. NVIDIA's May 2025 description names a modular rack with a power shelf bracket and slide rails; MGX 54V and 1400A busbars, a 33 kW power shelf, MGX power whips and high-speed cables on the electrical side; and MGX cold plates, a 44RU manifold and two quick-disconnect types (the NVIDIA NVQD and the Universal UQD) on the liquid-cooling side. [4]
The same document ties the design to thermal reality: Blackwell-generation racks draw up to 120 kW, and NVIDIA says the MGX liquid-cooled busbars and manifolds hold a coolant temperature differential below 15 degrees Celsius even under 1400 A loads. [4] Because these parts are specified rather than improvised per vendor, standard components can be pre-installed at the factory and connected on site.
The rack-scale era
The shift from "MGX server" to "MGX rack" is the most consequential change in the architecture's history. GB200 NVL72, the liquid-cooled rack that joins 36 Grace CPUs and 72 Blackwell GPUs into one 72-GPU NVLink domain, is built on the MGX modular architecture, and the same MGX rack footprint carries GB300 NVL72 and the Vera Rubin generation. [3][6]
For Vera Rubin, NVIDIA redesigned the MGX compute tray around a central printed circuit board midplane that replaces cable-based connections, with modular expansion bays for ConnectX-9 networking and Rubin CPX for long-context inference. The rack design uses 45 degree Celsius liquid cooling, a liquid-cooled busbar and, by NVIDIA's account, 20 times more energy storage to steady power delivery. [6] Vera Rubin is also a multi-rack platform rather than a single rack type: NVIDIA's May 2026 description unifies Vera Rubin NVL72 systems, the Vera CPU rack, NVIDIA Groq 3 LPX, Vera BlueField-4 STX storage and Spectrum-6 SPX Ethernet racks, all on MGX. [8][9]
Looking further out, NVIDIA's Kyber rack generation, which rotates compute blades vertically to reach 576 Rubin Ultra GPUs, is being prepared by the same MGX-aligned partner ecosystem alongside a move to 800 VDC power distribution. [6]
Open standards and the Open Compute Project
MGX was described as compatible with Open Compute Project and Electronic Industries Alliance rack specifications from its first announcement. [1] NVIDIA's contributions to OCP have deepened since. In October 2024 the company shared key portions of the GB200 NVL72 electro-mechanical design with the OCP community, specifically the rack architecture, compute and switch tray mechanicals, liquid-cooling and thermal environment specifications, and NVLink cable cartridge volumetrics; NVIDIA noted it had already contributed the HGX H100 baseboard design specification in an earlier generation. [3] At the October 2025 OCP Global Summit NVIDIA said it planned to contribute the upgraded Vera Rubin rack and its compute tray innovations to OCP as an open standard, and Vertiv introduced an 800 VDC MGX reference architecture for power and cooling. [6] By 2026 NVIDIA was describing the Vera Rubin ramp as resting on a "proven, open source MGX design", and an NVIDIA developer blog referred to the "OCP MGX rack-scale server architecture". [8][14]
The partner ecosystem
MGX system partners listed on NVIDIA's product page as of September 2026 include Aetina, ASRock Rack, ASUS, Cisco, Compal, GIGABYTE, HPE, Ingrasys, Inventec, Lanner, Lenovo, MiTAC, MSI, Pegatron, Quanta Cloud Technology, Supermicro, Wistron and Wiwynn. [10] Taiwan is the centre of gravity: NVIDIA said in May 2026 that more than 1 million MGX rack components for Vera Rubin infrastructure come together in Taiwan across 25 factory sites, drawing on an ecosystem of more than 500 partners. [11]
QCT and Supermicro were first to market in 2023, with Supermicro's ARS-221GL-NR using the Grace CPU Superchip and QCT's S74G-2U using the GH200 Grace Hopper Superchip. [1]
MGX and third-party silicon
MGX is the physical half of NVIDIA's answer to custom accelerators. NVLink Fusion licenses the NVLink scale-up fabric to outside silicon; MGX gives that silicon a rack, a cooling loop, a power architecture and a supply chain it does not have to invent. NVIDIA's own framing is that by standardising on a common rack, a data centre can be built once and then support GPUs, CPUs and XPUs without a separate rack architecture for each processor type. [13]
Two worked examples exist. At AWS re:Invent on December 2, 2025, Amazon Web Services said it is designing Trainium4 to integrate with NVLink 6 and the MGX rack architecture, the first step in a multi-generational NVLink Fusion collaboration. [14] On September 10, 2026, d-Matrix announced that its next-generation Raptor inference XPUs, a follow-on to Corsair, will plug into the MGX rack alongside NVIDIA Vera CPUs, NVLink switches, BlueField-4 DPUs, ConnectX-9 SuperNICs and Spectrum-X Ethernet, in a rack of modular cable-free trays built on the MGX ecosystem and supply chain. Initial availability of Raptor XPUs in an MGX rack is expected in the fourth quarter of 2027. [12]
The commercial logic is the same in both cases. The accelerator is the differentiated part; everything around it is expensive, slow to qualify and not a source of advantage. MGX turns that surrounding work into something a partner can buy into.
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10NVIDIA. "NVIDIA MGX Gives System Makers Modular Architecture to Meet Diverse Accelerated Computing Needs of World's Data Centers." NVIDIA Newsroom, May 28, 2023. nvidianews.nvidia.com/...-mgx-server-specification
- ^1 ^2NVIDIA. "Computer Industry Joins NVIDIA to Build AI Factories and Data Centers for the Next Industrial Revolution." NVIDIA Newsroom, June 2, 2024. nvidianews.nvidia.com/...ai-factories-data-centers
- ^1 ^2 ^3 ^4NVIDIA. "NVIDIA Contributes Blackwell Platform Design to Open Hardware Ecosystem, Accelerating AI Infrastructure Innovation." NVIDIA Newsroom, October 15, 2024. nvidianews.nvidia.com/...infrastructure-innovation
- ^1 ^2 ^3 ^4 ^5NVIDIA. "Building the Modular Foundation for AI Factories With NVIDIA MGX." NVIDIA Technical Blog, May 16, 2025. developer.nvidia.com/...-factories-with-nvidia-mgx
- ^NVIDIA. "Rack 'n' Roll: NVIDIA Grace Hopper Systems Gather at GTC." NVIDIA Blog, February 27, 2024. blogs.nvidia.com/...mgx-accelerated-systems-gtc
- ^1 ^2 ^3 ^4 ^5NVIDIA. "NVIDIA, Partners Drive Next-Gen Efficient Gigawatt AI Factories in Buildup for Vera Rubin." NVIDIA Blog, October 13, 2025. blogs.nvidia.com/...tt-ai-factories-ocp-vera-rubin
- ^NVIDIA. "NVIDIA Launches Vera CPU, Purpose-Built for Agentic AI." NVIDIA Newsroom, March 16, 2026. nvidianews.nvidia.com/...pose-built-for-agentic-ai
- ^1 ^2 ^3 ^4NVIDIA. "NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide." NVIDIA Newsroom, May 31, 2026. nvidianews.nvidia.com/...uction-agentic-ai-factory
- ^1 ^2NVIDIA. "NVIDIA Vera Rubin Opens Agentic AI Frontier." NVIDIA Newsroom, March 16, 2026. nvidianews.nvidia.com/...nvidia-vera-rubin-platform
- ^1 ^2 ^3NVIDIA. "MGX Platform for Modular Server Design." NVIDIA product page, accessed September 15, 2026. nvidia.com/...mgx
- ^NVIDIA. "Taiwan's Industry Titans Turbocharge World's AI Infrastructure Buildout With NVIDIA." NVIDIA Blog, May 31, 2026. blogs.nvidia.com/...an-ecosystem-ai-infrastructure
- ^1 ^2 ^3d-Matrix. "d-Matrix Adopts NVIDIA NVLink Fusion Rackscale Infrastructure for Ultra-Low Latency AI Inference." d-Matrix newsroom, September 10, 2026. d-matrix.ai/...d-matrix-rackscale-nvidia
- ^1 ^2 ^3NVIDIA. "d-Matrix Adopts NVIDIA NVLink Fusion for Rack-Scale XPU Deployment." NVIDIA Blog, September 10, 2026. blogs.nvidia.com/...d-matrix-nvlink-fusion
- ^1 ^2NVIDIA. "AWS Integrates AI Infrastructure With NVIDIA NVLink Fusion for Trainium4 Deployment." NVIDIA Technical Blog, December 2, 2025. developer.nvidia.com/...n-for-trainium4-deployment
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
1 revision · v2 · 2,186 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independently fact-checked against 46 cited and primary sources (193 claims). 28 defects found, 7 material, all corrected.
Cite this page: AI Wiki. "NVIDIA MGX." aiwiki.ai, updated 15 Sept 2026, fact-checked 15 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/nvidia_mgx