AI Wiki

Explore AI Wiki

Citation and evidence

NVIDIA DGX

24 min full readUpdated 62 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision historyDiscussion

Browse categories

AI HardwareAI InfrastructureNVIDIA

Cite this article

NVIDIA DGX is NVIDIA's line of integrated artificial-intelligence supercomputers: systems that package the company's highest-end data-center GPUs, CPUs, high-speed NVLink and NVSwitch interconnects, networking, storage and a preconfigured software stack into a single product that runs deep-learning training and inference out of the box. NVIDIA unveiled the first system, the DGX-1, at its GPU Technology Conference on 5 April 2016, calling it "the world's first deep learning supercomputer." [3] Over the following decade the brand grew from a single 8-GPU server into a portfolio spanning rack-scale machines such as the DGX GB300 and DGX Vera Rubin NVL72, scale-out reference architectures (DGX SuperPOD), a cloud service (DGX Cloud) and deskside systems (DGX Spark and DGX Station). [11][12][43][54] The DGX-1 became a symbol of the modern AI era after NVIDIA chief executive Jensen Huang hand-delivered the first production unit to OpenAI in San Francisco in August 2016. [24][25] On 24 September 2026 NVIDIA marked the line's tenth anniversary with a retrospective video, "10 Years of NVIDIA DGX: From One System to AI Factories." [22]

Overview

The defining idea behind DGX is the "AI supercomputer in a box": rather than asking customers to assemble GPUs, networking and software themselves, NVIDIA sells a turnkey system that is validated and supported as a unit. NVIDIA's 2016 launch release described the DGX-1 as "a turnkey system" that "comes fully integrated with hardware, deep learning software and development tools for quick, easy deployment." [3] DGX systems are built around NVIDIA's highest-end GPUs of each generation and use the company's proprietary NVLink interconnect, and from 2018 the NVSwitch (NVLink Switch) fabric, to let the GPUs exchange data at bandwidths far higher than standard PCIe. [3][29] They ship with NVIDIA's software stack: the 2016 DGX-1 bundled the DIGITS training system, the cuDNN library and optimized builds of Caffe, Theano and Torch, while current systems ship with NVIDIA DGX OS, NVIDIA AI Enterprise and NVIDIA Mission Control. [3][40] DGX hardware is sold to enterprises, research labs and cloud providers, directly and through partners. [7][54]

NVIDIA's chief executive framed the original system's purpose in sweeping terms. "The DGX-1 is easy to deploy and was created for one purpose: to unlock the powers of superhuman capabilities and apply them to problems that were once unsolvable," Huang (then known as Jen-Hsun Huang) said at the 2016 launch. [3]

DGX and HGX

DGX should be distinguished from NVIDIA's related HGX platform. HGX is a modular multi-GPU board and platform that server makers integrate into their own systems; DGX is NVIDIA's own fully assembled, NVIDIA-branded product, designed to be operational without additional scaffolding. DGX and HGX products of the same generation share GPU specifications (for example HGX B300 and DGX B300 both use Blackwell Ultra GPUs). [2] NVIDIA has also used DGX as a reference and blueprint for its partners. In August 2016 TOP500 described the DGX-1 as "a reference platform of sorts" for the Pascal-generation Tesla GPUs until those GPUs became broadly available to OEMs, and NVIDIA said in 2022 that its Eos DGX SuperPOD would "serve as a blueprint for advanced AI infrastructure from NVIDIA, as well as its OEM and cloud partners." [4][8] In 2023 NVIDIA likewise said it intended to provide the DGX GH200 design "as a blueprint to cloud service providers and other hyperscalers." [36]

History and generations

DGX-1 and the OpenAI delivery (2016)

The original DGX-1 was announced on 5 April 2016 at NVIDIA's GPU Technology Conference. It combined eight Pascal-generation Tesla P100 GPUs (16 GB each) in an NVLink "hybrid cube mesh", a 7 TB SSD cache, dual 10GbE and quad 100 Gb InfiniBand networking in a 3U, 3,200 W chassis. NVIDIA rated it at up to 170 teraflops of half-precision (FP16) performance and claimed it delivered "the equivalent throughput of 250 x86 servers." [3] The list price was 129,000 US dollars. [4][26] NVIDIA said general availability in the United States would begin in June 2016, with other regions following from the third quarter. [3]

In August 2016 Huang personally hand-delivered what NVIDIA called "the first production DGX-1" to OpenAI's office in San Francisco. [24] Elon Musk, then an OpenAI co-founder and backer, thanked "@nvidia and Jensen for donating the first DGX-1 AI supercomputer to @OpenAI" in a post on 9 August 2016, and a photograph released by NVIDIA shows Huang and Musk with the machine, which Huang had signed: "To Elon and the OpenAI Team! To the future of computing and humanity. I present you the world's first DGX-1!" [25][26] NVIDIA publicized the delivery on 15 August 2016, saying Huang had made it "last week." [24][26] "I thought it was incredibly appropriate that the world's first supercomputer dedicated to artificial intelligence would go to the laboratory that was dedicated to open artificial intelligence," Huang said. [24] In the same NVIDIA account, OpenAI research scientist Ilya Sutskever called the DGX-1 "a huge advance," and Andrej Karpathy described using the added compute to train chatbot-style models on years of Reddit conversations rather than a month. [24] Huang also said building the DGX-1 took 3,000 people three years, and that "if this is the only one ever shipped, this project would cost $2 billion." [24] Fortune reported that NVIDIA described the machine as donated and Data Center Knowledge called it a gift, while TOP500 noted at the time that it was not clear whether OpenAI had paid for it. [4][26][61]

Volta: DGX-1 refresh, DGX Station and DGX-2 (2017 to 2018)

On 10 May 2017 NVIDIA announced Volta-based DGX systems. [5] The refreshed DGX-1 used eight Volta-generation Tesla V100 GPUs, raising peak FP16 Tensor Core performance to about 960 teraflops, at a list price of 149,000 US dollars; NVIDIA said buyers of the Pascal version could upgrade to V100 boards for free. [27][28] The same announcement introduced the first DGX Station, a liquid-cooled deskside workstation with four V100 GPUs priced at 69,000 US dollars, which NVIDIA called "the world's first personal supercomputer for AI development." [5][27][28]

NVIDIA followed on 27 March 2018 with the DGX-2, which Huang introduced as "the world's largest GPU." [31][32] It combined sixteen 32 GB V100 GPUs on two boards [6], each carrying eight GPUs and six first-generation NVSwitch chips, so that all sixteen GPUs shared a unified memory space. [29][30] NVSwitch let the GPUs communicate at 2.4 TB/s, which Tom's Hardware described as bisection bandwidth. [29][31] NVIDIA called the DGX-2 "the first single server capable of delivering two petaflops of computational power." It listed at 399,000 US dollars, drew up to about 10 kW and filled a 10U chassis. [29][30]

Ampere and Hopper (2020 to 2023)

The third-generation DGX A100 launched on 14 May 2020 with eight Ampere-architecture A100 GPUs, 320 GB of total GPU memory, six NVSwitch chips, nine ConnectX-6 200 Gb/s network interfaces and 5 petaflops of AI performance, at a starting price of 199,000 US dollars. Unlike earlier DGX systems, which used Intel Xeon processors, it used two 64-core AMD EPYC 7742 CPUs. The first order went to the U.S. Department of Energy's Argonne National Laboratory for COVID-19 research. [7][28][32][33] On 16 November 2020 NVIDIA added a 640 GB DGX A100 model built with 80 GB A100 GPUs, and the DGX Station A100, a four-GPU "AI data-center-in-a-box" rated at 2.5 petaflops with up to 320 GB of GPU memory. [34]

The fourth-generation DGX H100 was announced on 22 March 2022. It used eight Hopper-architecture H100 GPUs with 640 GB of total GPU memory and delivered 32 petaflops at the new FP8 precision, which NVIDIA said was six times the prior generation. NVIDIA's announcement specified eight ConnectX-7 400 Gb/s InfiniBand adapters and two BlueField-3 data-processing units; NVIDIA's user guide for the shipping system lists eight single-port ConnectX-7 cluster adapters plus two dual-port ConnectX-7 cards for storage and management. The system comes in an 8U chassis with two 56-core Intel Xeon 8480C processors. [8][35] NVIDIA's DGX H100/H200 user guide also covers a DGX H200 variant with 1,128 GB of GPU memory from eight H200 GPUs. [35]

On 28 May 2023 NVIDIA announced the DGX GH200, which used the NVLink Switch System to join 256 GH200 Grace Hopper Superchips into what NVIDIA described as a single GPU with 144 TB of shared memory and 1 exaflop of performance. NVIDIA said Google Cloud, Meta and Microsoft were among the first expected to gain access to it. [36]

Blackwell and Blackwell Ultra (2024 to 2026)

With the Blackwell architecture NVIDIA split its data-center DGX line into two families, both announced on 18 March 2024: an air-cooled 8-GPU server and a liquid-cooled rack. [37] NVIDIA called the DGX B200 "the sixth generation of air-cooled, traditional rack-mounted DGX designs." It pairs eight Blackwell GPUs with two Intel Xeon Platinum 8570 processors and provides 1,440 GB of GPU memory, 144 petaflops of FP4 and 72 petaflops of FP8 Tensor Core performance (NVIDIA quotes both with sparsity; dense FP4 is 72 petaflops). [37][38]

The DGX GB200 is a full rack: each system contains 36 GB200 Superchips, meaning 36 Grace CPUs and 72 Blackwell GPUs joined by fifth-generation NVLink, and a DGX SuperPOD is built from eight or more such systems. [37] The underlying GB200 NVL72 design exposes 13.4 TB of HBM3E GPU memory and 1,440 petaflops of FP4 Tensor Core compute with sparsity (720 petaflops dense), with the NVLink Switch System providing 130 TB/s of GPU-to-GPU bandwidth inside the rack. [9]

On 18 March 2025 NVIDIA announced two Blackwell Ultra systems. [39] The liquid-cooled DGX GB300 is again a rack of 72 Blackwell Ultra GPUs and 36 Grace CPUs, now with 72 ConnectX-8 SuperNICs at up to 800 Gb/s; NVIDIA lists 20 TB of GPU memory and 1,440 petaflops of FP4 compute with sparsity (1,080 petaflops dense), matching the GB300 NVL72 rack design. [10][11][39] The DGX B300 is a separate 8-GPU server in a 10U chassis with Intel Xeon 6776P processors, eight ConnectX-8 ports, about 14 kW of power draw and 144 petaflops of sparse FP4 compute (108 petaflops dense). NVIDIA's March 2025 announcement gave it 2.3 TB of HBM3e, while its current specification lists 2.1 TB of total GPU memory. [39][40] As of September 2026 NVIDIA's product page describes DGX B300 systems as shipping. [40] The same March 2025 announcement introduced NVIDIA Mission Control software and NVIDIA Instant AI Factory, a managed service in which Equinix would host DGX GB300 and DGX B300 SuperPODs. [39]

Vera Rubin (2026)

At CES on 5 January 2026 NVIDIA announced the Vera Rubin platform and two Rubin-generation DGX systems for DGX SuperPOD: the rack-scale DGX Vera Rubin NVL72 and the 8-GPU DGX Rubin NVL8. [41][42] NVIDIA's preliminary specifications give the DGX Vera Rubin NVL72 72 Rubin GPUs, 36 Vera CPUs, 20.7 TB of HBM4 and 3,600 petaflops of NVFP4 inference compute, and the DGX Rubin NVL8 eight Rubin GPUs, two Intel Xeon 6776P processors, 2.3 TB of GPU memory and 400 petaflops of NVFP4 inference compute. [43][44] NVIDIA said DGX SuperPODs with either system would be available in the second half of 2026. [42] At GTC on 16 March 2026 NVIDIA said Vera Rubin-based products would be available from partners from the second half of 2026, and on 31 May 2026 it said Vera Rubin was ramping into full production, with "production shipments" set to begin in the fall. [45][46]

Generation summary

SystemAnnouncedAcceleratorsPeak AI performance (NVIDIA figure)GPU memoryNotable details
DGX-1 (Pascal)5 Apr 20168x Tesla P100170 TFLOPS FP16128 GB129,000 USD list; first production unit delivered to OpenAI [3][4]
DGX-1 (Volta)10 May 20178x Tesla V100~960 TFLOPS FP16 (Tensor Core)128 GB149,000 USD; P100-to-V100 swap offered [27][28]
DGX Station10 May 20174x Tesla V100~480 TFLOPS64 GBLiquid-cooled deskside; 69,000 USD [27][28]
DGX-227 Mar 201816x Tesla V100 32 GB2 PFLOPS512 GBFirst NVSwitch system; 399,000 USD [29][30]
DGX A10014 May 20208x A1005 PFLOPS320 GB (640 GB model from Nov 2020)From 199,000 USD; AMD EPYC CPUs [7][33][34]
DGX Station A10016 Nov 20204x A1002.5 PFLOPSUp to 320 GBDeskside workgroup server [34]
DGX H10022 Mar 20228x H10032 PFLOPS FP8640 GB8x ConnectX-7 400 Gb/s (2x BlueField-3 in 2022 announcement); H200 variant has 1,128 GB [8][35]
DGX GH20028 May 2023256x GH200 Superchip1 EFLOPS144 TB sharedNVLink Switch System spanning 256 superchips [36]
DGX B20018 Mar 20248x Blackwell144 PFLOPS FP4 sparse (72 dense)1,440 GBAir-cooled; Intel Xeon Platinum 8570 [37][38]
DGX GB20018 Mar 202472x Blackwell + 36x Grace1,440 PFLOPS FP4 sparse (720 dense)13.4 TB HBM3ELiquid-cooled NVL72 rack; 130 TB/s NVLink [9][37]
DGX Spark6 Jan 2025 (as Project DIGITS)GB10 SuperchipUp to 1 PFLOPS FP4128 GB unified3,999 USD at launch, 4,699 USD from Feb 2026 [53][55]
DGX Station (GB300)18 Mar 2025Blackwell Ultra + 72-core GraceUp to 20 PFLOPS FP4 sparse748 GB coherent (252 GB HBM3e + 496 GB LPDDR5X)Partner orders opened around GTC 2026 [20][57]
DGX B30018 Mar 20258x Blackwell Ultra144 PFLOPS FP4 sparse (108 dense)2.1 TBXeon 6776P; ConnectX-8 [39][40]
DGX GB30018 Mar 202572x Blackwell Ultra + 36x Grace1,440 PFLOPS FP4 sparse (1,080 dense)20 TBLiquid-cooled rack; 72 ConnectX-8 SuperNICs [11][39]
DGX Rubin NVL85 Jan 20268x Rubin400 PFLOPS NVFP4 inference2.3 TBPreliminary spec; x86 host [42][44]
DGX Vera Rubin NVL725 Jan 202672x Rubin + 36x Vera3,600 PFLOPS NVFP4 inference20.7 TB HBM4Preliminary spec; rack-scale [42][43]

Expanded article table

NVIDIA's headline performance figures mix precisions and, from Blackwell onward, often quote sparse throughput, so the numbers in the table are not directly comparable across generations.

DGX SuperPOD

The DGX SuperPOD is NVIDIA's scale-out reference architecture for connecting many DGX systems into a single supercomputer. Rather than a single product, it is a validated blueprint covering DGX compute nodes, InfiniBand and Ethernet networking, management nodes and storage, which NVIDIA markets as "a ready-to-run, turnkey AI supercomputer." [12][13] The design is modular: nodes are grouped into scalable units (SUs). NVIDIA's DGX H100 reference architecture describes four SUs with 128 DGX nodes (32 per SU) and says the design can scale "up to and beyond 64 SU with 2000+ DGX H100 nodes." [13] As of September 2026 NVIDIA offers DGX SuperPOD with Rubin and Blackwell compute options and says it scales "to tens of thousands of NVIDIA GPUs." [12]

NVIDIA has repeatedly built its own SuperPODs:

SystemYearCompositionNotes
DGX SuperPOD (DGX-2H)201996 DGX-2H servers, 1,536 V100 GPUs9.444 petaflops on HPL; No. 22 on the June 2019 TOP500 list [47]
DGX A100 SuperPOD2020140 DGX A100 systems700 petaflops of AI compute, built for internal research [7]
Selene2020DGX A100 systems27.58 petaflops Rmax at No. 7 in June 2020; after expansion, 63.46 petaflops Rmax at No. 5 in November 2020 on TOP500 [48]
Eos2023576 DGX H100 systems, 4,608 H100 GPUs18.4 exaflops FP8; No. 9 on TOP500 with 121.40 petaflops Rmax [49][50]
DGX Vera Rubin SuperPODAnnounced 202614 DGX Vera Rubin NVL72 systems, 1,008 Rubin GPUs50.4 exaflops FP4 and 1,046 TB of fast memory (NVIDIA figures) [42]

Expanded article table

NVIDIA's Eos is the reference example of a Hopper-generation SuperPOD. It was announced in March 2022, revealed at the SC23 conference in November 2023 and described in detail in February 2024. It is built from 576 DGX H100 systems (4,608 H100 GPUs) on NVIDIA Quantum-2 400 Gb/s InfiniBand and delivers 18.4 exaflops of FP8 AI performance. On the double-precision LINPACK benchmark it recorded an Rmax of 121.40 petaflops, placing it ninth on the November 2023 TOP500 list. NVIDIA noted that a separate "sister" Eos SuperPOD with 10,752 H100 GPUs was used for its November 2023 MLPerf training runs. [8][49][50]

Customer SuperPODs include LillyPod, which Eli Lilly brought online in Indianapolis in February 2026. NVIDIA describes it as the first DGX SuperPOD built with DGX B300 systems, with 1,016 Blackwell Ultra GPUs and more than 9,000 petaflops of AI performance, assembled in four months. [51] For the Rubin generation, NVIDIA's reference SuperPOD uses either 14 DGX Vera Rubin NVL72 racks (1,008 GPUs) or 64 DGX Rubin NVL8 systems (512 GPUs). [42]

DGX Cloud

DGX Cloud is a service that rents DGX infrastructure rather than selling the hardware outright. NVIDIA launched it on 21 March 2023, saying enterprises could access "their own AI supercomputer using a simple web browser" and rent clusters monthly. The service was hosted by partner clouds, starting with Oracle Cloud Infrastructure, with Microsoft Azure expected to follow the next quarter and Google Cloud later. Each instance had eight H100 or A100 80 GB GPUs, and customers managed workloads with NVIDIA Base Command Platform and NVIDIA AI Enterprise software. [15] Tom's Hardware reported the launch price as 36,999 US dollars per instance per month, a premium over hyperscaler GPU rentals. [16]

NVIDIA later adjusted its cloud strategy. On 18 May 2025, at Computex, it announced DGX Cloud Lepton, a compute marketplace connecting developers to GPU capacity from NVIDIA Cloud Partners including CoreWeave, Crusoe, Lambda, Nebius and SoftBank, which Tom's Hardware described as making NVIDIA "less of a rival and more of an aggregator." [16][52][62] In September 2025 Tom's Hardware, citing The Information, reported that NVIDIA was using most DGX Cloud capacity for internal research and had stopped positioning the service as a direct competitor to Amazon Web Services and Microsoft Azure. [16] As of September 2026 NVIDIA's DGX Cloud page describes the service as "NVIDIA's AI proving ground," says its Nemotron open models are trained on it, and offers DGX Cloud with AWS, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure. [17]

Personal DGX systems

Alongside the data-center line, NVIDIA has sold smaller deskside DGX systems since 2017. The original DGX Station (2017) had four V100 GPUs, and the DGX Station A100 (2020) had four A100 GPUs. [5][34] In the Blackwell generation NVIDIA revived and expanded this category.

At CES on 6 January 2025 NVIDIA unveiled Project DIGITS, a personal AI computer built on the new GB10 Grace Blackwell Superchip, which MediaTek helped design. [53] At GTC on 18 March 2025 NVIDIA renamed it DGX Spark. [54] DGX Spark pairs a 20-core Arm CPU (10 Cortex-X925 and 10 Cortex-A725 cores) with a Blackwell GPU, 128 GB of coherent LPDDR5X memory, ConnectX-7 200 Gb/s networking and up to 4 TB of storage, and NVIDIA rates it at up to 1 petaflop of FP4 AI performance. NVIDIA says it can run inference on models of up to 200 billion parameters and fine-tune models of up to 70 billion parameters. [18][19] NVIDIA and its partners began shipping DGX Spark in the week of 13 October 2025, with orders on NVIDIA.com from 15 October. [18] The launch echoed 2016: Huang hand-delivered the first DGX Spark to Musk at SpaceX's Starbase in Texas and presented another to OpenAI chief executive Sam Altman, with president Greg Brockman looking on. [60] The Founders Edition launched at 3,999 US dollars; in February 2026 NVIDIA raised its list price to 4,699 US dollars, citing "worldwide constraints in memory supply." [55][56]

The same GTC 2025 announcement introduced a new DGX Station built on the GB300 Grace Blackwell Ultra Desktop Superchip, which combines a 72-core Grace CPU and a Blackwell Ultra GPU over a 900 GB/s NVLink-C2C link, with ConnectX-8 networking at up to 800 Gb/s. [20][54] NVIDIA's March 2025 announcement gave it 784 GB of coherent memory and said partners would offer it "later this year." [21][54] The system actually reached the market about a year later: NVIDIA's partners began taking orders at GTC in March 2026, with ASUS, Dell, Gigabyte, HP, MSI and Supermicro among the vendors offering systems, and the shipping specification lists 748 GB of coherent memory (252 GB of HBM3e plus 496 GB of LPDDR5X), which ServeTheHome attributed to a partially enabled Blackwell Ultra GPU. [20][57] NVIDIA rates it at up to 20 petaflops of FP4 compute with sparsity and says it supports models of up to 1 trillion parameters. [20] Most vendors did not publish list prices; in August 2026 Tom's Hardware found an Exxact DGX Station configuration listed from 94,930 US dollars. [57][58] NVIDIA also markets a DGX Station for Windows variant for enterprise desks. [20]

Software

DGX systems ship with a software stack that NVIDIA maintains alongside the hardware. NVIDIA DGX OS is a customized installation of Ubuntu Linux with system-specific drivers, optimizations and diagnostic and monitoring tools; DGX OS 7 is based on Ubuntu 24.04. [59] NVIDIA Base Command software manages AI development on DGX SuperPOD infrastructure and was the management layer for DGX Cloud at launch. [8][15] NVIDIA AI Enterprise supplies frameworks, pretrained models and libraries. [15][40] NVIDIA Mission Control, announced in March 2025, handles data-center operations and orchestration for Blackwell-based DGX systems and, per NVIDIA's DGX B300 specification, incorporates Run:ai technology. [39][40]

Tenth anniversary

On 24 September 2026 NVIDIA published "10 Years of NVIDIA DGX: From One System to AI Factories," a 3 minute 31 second retrospective video on its YouTube channel (by 25 September the watch page also displayed the title "Decade of NVIDIA DGX: From A Single System to AI Factories"). NVIDIA's description says Huang introduced the DGX-1 "to help researchers tackle AI challenges once thought out of reach" and "delivered the first DGX-1 to OpenAI, which started the AI revolution," and that "What began as a single system has evolved into the blueprint for the modern AI factory." The description invites viewers to register for DGX sessions at GTC Berlin. [22] NVIDIA's GTC Berlin 2026 conference runs 20 to 22 October 2026, with Huang's keynote at the Tempodrom on 21 October. [23]

Significance

DGX systems have been closely tied to the growth of deep learning. OpenAI researchers said in 2016 that they planned to use the donated DGX-1 for generative modeling and training on large conversational datasets, and NVIDIA's own tenth-anniversary material credits that delivery with starting "the AI revolution." [22][24] The line set a template of integrated GPUs, interconnect and software that NVIDIA also offers to partners through HGX, and NVIDIA has built DGX SuperPODs for its own research with several GPU generations. [2][7][47][49] By extending the brand downward to DGX Spark and the Blackwell-era DGX Station, NVIDIA has sought to put the same software environment used in its largest data-center systems onto individual developers' desks. [18][54]

References

  1. Nvidia DGX. Wikipedia. en.wikipedia.org/...Nvidia_DGX
  2. ^1 ^2Decoding Nvidia's Blackwell Products: how do B200s, GB200s, HGX/DGX systems, and NVL supercomputers differ? Modal. modal.com/...nvidia-blackwell
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8NVIDIA Launches World's First Deep Learning Supercomputer. NVIDIA Newsroom, 5 April 2016. nvidianews.nvidia.com/...ep-learning-supercomputer
  4. ^1 ^2 ^3 ^4NVIDIA Delivers DGX-1 Supercomputer in a Box to OpenAI. TOP500, 15 August 2016. top500.org/...x-1-supercomputer-in-a-box-to-openai
  5. ^1 ^2 ^3NVIDIA Advances AI Computing Revolution with New Volta-Based DGX Systems. NVIDIA Newsroom, 10 May 2017. nvidianews.nvidia.com/...w-volta-based-dgx-systems
  6. ^NVIDIA Announces the DGX-2 System - 16x Tesla V100 GPUs, 30 TB NVMe Memory for $400K. TechPowerUp, 2018. techpowerup.com/...-30-tb-nvme-memory-for-usd-400k
  7. ^1 ^2 ^3 ^4 ^5NVIDIA Ships World's Most Advanced AI System, NVIDIA DGX A100, to Fight COVID-19; Third-Generation DGX Packs Record 5 Petaflops of AI Performance. NVIDIA Newsroom, 14 May 2020. nvidianews.nvidia.com/...taflops-of-ai-performance
  8. ^1 ^2 ^3 ^4 ^5NVIDIA Announces DGX H100 Systems: World's Most Advanced Enterprise AI Infrastructure. NVIDIA Newsroom, 22 March 2022. nvidianews.nvidia.com/...erprise-ai-infrastructure
  9. ^1 ^2GB200 NVL72. NVIDIA. nvidia.com/...gb200-nvl72
  10. ^Designed for AI Reasoning Performance & Efficiency: NVIDIA GB300 NVL72. NVIDIA. nvidia.com/...gb300-nvl72
  11. ^1 ^2 ^3DGX GB300: AI Factory Infrastructure for Enterprises. NVIDIA. nvidia.com/...dgx-gb300
  12. ^1 ^2 ^3DGX SuperPOD: AI Infrastructure for Enterprise Deployments. NVIDIA. nvidia.com/...dgx-superpod
  13. ^1 ^2DGX SuperPOD Architecture (NVIDIA DGX SuperPOD Reference Architecture Featuring DGX H100). NVIDIA Docs. docs.nvidia.com/...dgx-superpod-architecture
  14. NVIDIA Reveals Eos Supercomputer: 4,600 H100 GPUs for 18 AI Exaflops. Inside HPC, 2024. insidehpc.com/...4600-h100-gpus-for-18-ai-exaflops
  15. ^1 ^2 ^3NVIDIA Launches DGX Cloud, Giving Every Enterprise Instant Access to AI Supercomputer From a Browser. NVIDIA Newsroom, 21 March 2023. nvidianews.nvidia.com/...ercomputer-from-a-browser
  16. ^1 ^2 ^3Nvidia steps back from DGX Cloud, stops trying to compete with AWS and Azure. Tom's Hardware, 12 September 2025. tomshardware.com/...nvidia-steps-back-from-dgx-cloud
  17. ^NVIDIA's AI factory in the cloud (NVIDIA DGX Cloud). NVIDIA. nvidia.com/...dgx-cloud
  18. ^1 ^2 ^3NVIDIA DGX Spark Arrives for World's AI Developers. NVIDIA Newsroom, 13 October 2025. nvidianews.nvidia.com/...-for-worlds-ai-developers
  19. ^Personal AI Supercomputer Powered by Blackwell, NVIDIA DGX Spark. NVIDIA. nvidia.com/...dgx-spark
  20. ^1 ^2 ^3 ^4 ^5Personal AI Supercomputer: NVIDIA DGX Station. NVIDIA. nvidia.com/...dgx-station
  21. ^Nvidia unveils DGX Station workstation PCs with GB300 Blackwell Ultra inside. Tom's Hardware, 19 March 2025. tomshardware.com/...s-gb300-blackwell-ultra-inside
  22. ^1 ^2 ^310 Years of NVIDIA DGX: From One System to AI Factories. NVIDIA (YouTube), 24 September 2026. youtube.com/watch
  23. ^GTC Berlin AI Conference, 20-22 October 2026. NVIDIA. nvidia.com/...gtc
  24. ^1 ^2 ^3 ^4 ^5 ^6 ^7NVIDIA CEO Delivers World's First AI Supercomputer in a Box to OpenAI. NVIDIA (via WebWire), 15 August 2016. webwire.com/ViewPressRel.asp
  25. ^1 ^2Elon Musk on X: "Would like to thank @nvidia and Jensen for donating the first DGX-1 AI supercomputer to @OpenAI in support of democratizing AI technology". 9 August 2016. x.com/...763096729134575617
  26. ^1 ^2 ^3 ^4Nvidia Just Gave A Supercomputer to Elon Musk-backed OpenAI. Fortune, 15 August 2016. fortune.com/...ligence-openai-nvidia-supercomputer
  27. ^1 ^2 ^3 ^4NVIDIA Volta-Powered DGX-1 And DGX Station AI Supercomputers Debut At GTC 2017. HotHardware, May 2017. hothardware.com/...upercomputers-debut-at-gtc-2017
  28. ^1 ^2 ^3 ^4 ^5Big Bang For The Buck Jump With Volta DGX-1. The Next Platform, 19 May 2017. nextplatform.com/...1632810
  29. ^1 ^2 ^3 ^4 ^5NVIDIA Boosts World's Leading Deep Learning Computing Platform, Bringing 10x Performance Gain in Six Months. NVIDIA Newsroom, 27 March 2018. nvidianews.nvidia.com/...rmance-gain-in-six-months
  30. ^1 ^2 ^3Nvidia Ups Hardware Game with 16-GPU DGX-2 Server and 18-Port NVSwitch. HPCwire, 27 March 2018. hpcwire.com/...th-16-gpu-dgx-2-server-and-nvswitch
  31. ^1 ^2Inside The World's Largest GPU: Nvidia Details NVSwitch. Tom's Hardware, 2018. tomshardware.com/...lds-largest-gpu-nvswitch,37661
  32. ^1 ^2The Nvidia DGX-2 Is the World's Largest GPU, and It's Made for AI. Design News, 2018. designnews.com/...largest-gpu-and-it-s-made-for-ai
  33. ^1 ^2Introduction to the NVIDIA DGX A100 System. NVIDIA DGX A100 User Guide. docs.nvidia.com/...introduction-to-dgxa100
  34. ^1 ^2 ^3 ^4NVIDIA DGX Station A100 Offers Researchers AI Data-Center-in-a-Box. NVIDIA Newsroom, 16 November 2020. nvidianews.nvidia.com/...s-ai-data-center-in-a-box
  35. ^1 ^2 ^3Introduction to NVIDIA DGX H100/H200 Systems. NVIDIA DGX H100/H200 User Guide. docs.nvidia.com/...introduction-to-dgxh100
  36. ^1 ^2 ^3NVIDIA Announces DGX GH200 AI Supercomputer. NVIDIA Newsroom, 28 May 2023. nvidianews.nvidia.com/...gx-gh200-ai-supercomputer
  37. ^1 ^2 ^3 ^4 ^5NVIDIA Launches Blackwell-Powered DGX SuperPOD for Generative AI Supercomputing at Trillion-Parameter Scale. NVIDIA Newsroom, 18 March 2024. nvidianews.nvidia.com/...erative-ai-supercomputing
  38. ^1 ^2DGX B200: The Foundation for Your AI Factory. NVIDIA. nvidia.com/...dgx-b200
  39. ^1 ^2 ^3 ^4 ^5 ^6 ^7NVIDIA Blackwell Ultra DGX SuperPOD Delivers Out-of-the-Box AI Supercomputer for Enterprises to Build AI Factories. NVIDIA Newsroom, 18 March 2025. nvidianews.nvidia.com/...upercomputer-ai-factories
  40. ^1 ^2 ^3 ^4 ^5 ^6An AI Factory for AI Reasoning: NVIDIA DGX B300. NVIDIA. nvidia.com/...dgx-b300
  41. ^NVIDIA Kicks Off the Next Generation of AI With Rubin: Six New Chips, One Incredible AI Supercomputer. NVIDIA Newsroom, 5 January 2026. nvidianews.nvidia.com/...platform-ai-supercomputer
  42. ^1 ^2 ^3 ^4 ^5 ^6NVIDIA DGX SuperPOD Sets the Stage for Rubin-Based Systems. NVIDIA Blog, 5 January 2026. blogs.nvidia.com/...dgx-superpod-rubin
  43. ^1 ^2 ^3Gigascale AI Training & Inference Platform: NVIDIA DGX Vera Rubin NVL72. NVIDIA. nvidia.com/...dgx-vera-rubin-nvl72
  44. ^1 ^2Infrastructure for Agentic AI at Scale: NVIDIA DGX Rubin NVL8. NVIDIA. nvidia.com/...dgx-rubin-nvl8
  45. ^NVIDIA Vera Rubin Opens Agentic AI Frontier. NVIDIA Newsroom, 16 March 2026. nvidianews.nvidia.com/...nvidia-vera-rubin-platform
  46. ^NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide. NVIDIA Newsroom, 31 May 2026. nvidianews.nvidia.com/...uction-agentic-ai-factory
  47. ^1 ^2NVIDIA DGX SuperPOD Delivers World Record Supercomputing to Any Enterprise. NVIDIA Technical Blog, 2019. developer.nvidia.com/...-supercomputing-enterprise
  48. ^Selene: NVIDIA DGX A100, AMD EPYC 7742 64C 2.25GHz, NVIDIA A100, Mellanox HDR Infiniband. TOP500. top500.org/...179842
  49. ^1 ^2 ^3NVIDIA Eos Revealed: Peek Into Operations of a Top 10 Supercomputer. NVIDIA Blog, 15 February 2024. blogs.nvidia.com/...eos
  50. ^1 ^2Eos NVIDIA DGX SuperPOD: NVIDIA DGX H100, Xeon Platinum 8480C 56C 3.8GHz, NVIDIA H100, Infiniband NDR400. TOP500. top500.org/...180239
  51. ^Now Live: Lilly AI Factory for Pharmaceutical Discovery and Development. NVIDIA Blog, 26 February 2026. blogs.nvidia.com/...lilly-ai-factory-live
  52. ^NVIDIA Announces DGX Cloud Lepton to Connect Developers to NVIDIA's Global Compute Ecosystem. NVIDIA Newsroom, 18 May 2025. nvidianews.nvidia.com/...-global-compute-ecosystem
  53. ^1 ^2NVIDIA Puts Grace Blackwell on Every Desk and at Every AI Developer's Fingertips. NVIDIA Newsroom, 6 January 2025. nvidianews.nvidia.com/...-ai-developers-fingertips
  54. ^1 ^2 ^3 ^4 ^5 ^6NVIDIA Announces DGX Spark and DGX Station Personal AI Computers. NVIDIA Newsroom, 18 March 2025. nvidianews.nvidia.com/...ion-personal-ai-computers
  55. ^1 ^22/23/2026 Price Change Announcement. NVIDIA Developer Forums, February 2026. forums.developer.nvidia.com/...361713
  56. ^Nvidia DGX Spark gets $700 price hike as memory shortages bite: Founders Edition price jumps 18% to $4,699, up from $3,999. Tom's Hardware, 27 February 2026. tomshardware.com/...-now-usd4-699-up-from-usd3-999
  57. ^1 ^2 ^3NVIDIA DGX Station Systems Available At Last GB300 and GB200 Workstations For Your Desktop. ServeTheHome, 20 March 2026. servethehome.com/...-workstations-for-your-desktop
  58. ^Nvidia's GB300-powered DGX Station desktop tower listed for nearly $100,000 online. Tom's Hardware, 23 August 2026. tomshardware.com/...mere-mortals-with-lots-of-cash
  59. ^About DGX OS 7. NVIDIA DGX OS 7 User Guide. docs.nvidia.com/...introduction
  60. ^Elon Musk Gets Just-Launched NVIDIA DGX Spark: Petaflop AI Supercomputer Lands at SpaceX. NVIDIA Blog, 13 October 2025. blogs.nvidia.com/...live-dgx-spark-delivery
  61. ^Why Nvidia Gifted Elon Musk's AI Non-Profit Its Latest Supercomputer. Data Center Knowledge, 19 August 2016. datacenterknowledge.com/...-openai-a-supercomputer
  62. ^Connect Developers to Global GPU Compute: NVIDIA DGX Cloud Lepton. NVIDIA. nvidia.com/...dgx-cloud-lepton

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

4 revisions · v5 · 4,894 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: xg08 I1 audit + V6 independent verification (52 sources; NVIDIA releases, spec pages, TOP500, 2016 posts); 10 fixes applied

Cite this page: AI Wiki. "NVIDIA DGX." aiwiki.ai, updated 25 Sept 2026, fact-checked 25 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/nvidia_dgx

Suggest edit