DGX is NVIDIA's own complete AI system. HGX is the GPU baseboard NVIDIA sells to server makers such as Dell, Supermicro and Lenovo, who build their own servers around it. A configured DGX B300 listed at $646,878.46 at Broadberry on 21 September 2026; the live tables below show what the same GPUs rent for by the hour instead.
This article lines up DGX systems from 2016 through B300 against their prices, explains what actually separates DGX from HGX, and uses live tables for renting H100, H200, B200 and B300 GPUs instead.
The DGX lineup, DGX-1 to DGX B300
The table below covers selected multi-GPU DGX systems through B300. System names, GPU counts and historical dates follow the NVIDIA documentation and named reporting cited in each row. Total GPU memory is each GPU's memory from the site's spec table, multiplied by the documented GPU count, derived, except for DGX B200 and DGX B300, where NVIDIA states system totals directly. The V100 rows use different memory configurations; they are not interchangeable.
Prices are for whole systems and are dated: launch prices from NVIDIA or the press for older systems, and configured reseller listings or an analyst estimate for current ones. They are not a like-for-like market range, so get a quote for your configuration before budgeting.
| System | When | GPUs | Total GPU memory | System purchase price |
|---|---|---|---|---|
| DGX-1 (Pascal) | 2016 announcement | 8x Tesla P100 | 128GB | $129,000 at launch (TechCrunch, 5 Apr 2016) |
| DGX-1 (Volta refresh) | May 2017 refresh | 8x Tesla V100 | 128GB | $149,000 at refresh (Tom's Hardware, May 2017) |
| DGX-2 | 2018 announcement | 16x Tesla V100 | 512GB | $399,000 at announcement (TechPowerUp, 27 Mar 2018) |
| DGX A100 | 2020 announcement | 8x A100 | 320GB | From $199,000 at launch (NVIDIA, 14 May 2020) |
| DGX H100 | 2022 announcement (NVIDIA) | 8x H100 | 640GB | about $270,000, analyst estimate (SemiAnalysis, 29 May 2023) |
| DGX H200 | In use by January 2025 (Oregon State University) | 8x H200 | 1,128GB | $467,311.30, configured reseller listing (Broadberry, 21 Sep 2026) |
| DGX B200 | Reported shipping October 2024 (wccftech) | 8x B200 | 1,440GB | $562,154.09, configured reseller listing (Broadberry, 21 Sep 2026) |
| DGX B300 | Mar 2025 announcement (NVIDIA) | 8x B300 | 2.1TB / 2.3TB, conflicting NVIDIA totals | $646,878.46, configured reseller listing (Broadberry, 21 Sep 2026) |
NVIDIA's own pages disagree on DGX B300's total GPU memory. Its product page retrieved 28 September 2026 states 2.1TB; its user guide retrieved 21 September 2026 and its original March 2025 announcement state 2.3TB for an eight-GPU configuration. The site's spec table puts a single B300 at 262GB to 288GB. Eight times those endpoints gives 2,096GB and 2,304GB, or approximately 2.1TB and 2.3TB in decimal units, derived. Confirm the memory configuration in your quote.
NVIDIA gives no launch date for the DGX H200; the table uses Oregon State University's announcement of one, dated 30 January 2025.
DGX vs HGX: who builds what
NVIDIA's HGX and DGX B200 pages illustrate why the names can be confused: both describe systems using Blackwell GPUs. Its HGX page describes a baseboard for server makers with eight Rubin, Blackwell or Blackwell Ultra SXM GPUs; its DGX B200 page describes NVIDIA's complete system around the board.
| Type | Who builds the complete system | What is included | Dated system-price examples | Who it suits |
|---|---|---|---|---|
| DGX | NVIDIA | NVIDIA's system documentation and support SKUs describe GPUs, CPUs, storage, NVSwitch and networking, with support terms tied to the SKU | About $270,000 for DGX H100 (SemiAnalysis analyst estimate, 29 May 2023); $646,878.46 for a configured DGX B300 (Broadberry reseller listing, 21 Sep 2026) | Buyers who want one company responsible for the whole box |
| HGX | An OEM such as Dell, Supermicro or Lenovo | NVIDIA's GPU baseboard inside an OEM server; Dell and Supermicro specifications document the CPU, memory, storage and networking configurations | $306,672 for an eight-GPU Supermicro HGX H100 server (Newegg last listed price, retrieved 21 Sep 2026); $785,606.50 for an eight-GPU B300 ThinkSystem SR680a V4 (Lenovo-reported usual customer sale price, 15 Jun 2026) | Buyers who want to choose their own configuration or already work with that OEM |
These examples do not establish one option as reliably cheaper. Lenovo reported a usual customer sale price of $785,606.50 for its eight-GPU B300 ThinkSystem SR680a V4 as of 15 June 2026. That is higher than Broadberry's configured DGX B300 listing of $646,878.46, retrieved 21 September 2026. The dates and configurations differ, so this is a comparison of those two published figures, not proof of a general DGX discount. Ask for comparable configurations before choosing.
Software is part of the difference too: Viperatech's DGX H100 listing, retrieved 28 September 2026, describes the system as "pre-loaded with NVIDIA AI Enterprise." Check the software and support terms in each quote alongside the hardware.
What is inside a current DGX
NVIDIA's DGX B300 user guide lists eight Blackwell Ultra GPUs (the conflicting system memory totals are discussed above), two Intel Xeon 6776P CPUs, and 2TB of system memory, expandable to 4TB. The guide also lists eight ConnectX-8 SuperNICs at up to 800Gb/s plus two BlueField-3 DPUs. The guide specifies 14.5kW power consumption and a 15kW system maximum, with twelve 3.2kW power supplies or busbar power. The site's spec table lists B300 at 18,000 TFLOPS sparse FP4 and 9,000 TFLOPS sparse FP8. Multiplying each by eight and dividing by 1,000 gives derived totals of 144.0 PFLOPS sparse FP4 and 72.0 PFLOPS sparse FP8. These are sums of per-GPU peak specifications, not measured system throughput.
SuperPOD, Cloud, Station and Spark
DGX SuperPOD
DGX SuperPOD is not a single box. NVIDIA describes it as a turnkey reference architecture that stacks multiple DGX systems, or GB200 and GB300 NVL72 racks, with NVIDIA-specified networking, storage and management software, and sells the cluster with support rather than as parts. NVIDIA publishes separate DGX B300 reference designs for Quantum-X800 InfiniBand and for Spectrum-4 Ethernet with busbar power. Use the reference design for your chosen fabric and power arrangement when planning the cluster. NVIDIA's DGX B300 page, retrieved 28 September 2026, cites Eli Lilly's DGX SuperPOD with DGX B300 systems for drug discovery.
DGX Cloud
NVIDIA's DGX Cloud page, retrieved 28 September 2026, describes an internal environment it uses to develop models such as Nemotron and Cosmos and run production workloads.
NVIDIA announced DGX Cloud Lepton on 18 May 2025 as a separate platform with a compute marketplace connecting developers to partner GPUs. Its launch announcement named CoreWeave, Crusoe, Lambda and Nebius among the partners. Those were its launch partners.
DGX Station
NVIDIA's DGX Station product page, retrieved 28 September 2026, describes a single-GPU desktop. It lists one Blackwell Ultra GPU, one Grace CPU with 496GB of its own memory, a 900GB/s NVLink-C2C link, and 748GB of combined CPU and GPU memory, in a 1,600W tower requiring a 20A circuit. The combined-memory figure is a system specification, not GPU memory alone.
NVIDIA's page showed no price on 28 September 2026. Tom's Hardware reported on 23 August 2026 that Exxact listed the Valence VWS-158270643 DGX Station configuration from $94,930. That is a reseller's starting price, not an NVIDIA list price.
DGX Spark and the rack scale systems
DGX Spark is a smaller, separate product again: a single desktop unit built around NVIDIA's GB10 chip, per NVIDIA's own technical documentation, rather than a data center GPU. DGX Spark vs cloud GPU covers what it costs and how many rented GPU-hours the same money buys. At the opposite end, NVIDIA's documentation brands the liquid-cooled GB200 NVL72 and GB300 NVL72 racks as "DGX Grace Blackwell rack scale systems" too, connecting 72 GPUs into one NVLink domain. GB200 and GB300 NVL72 explained covers what is inside those racks and what they cost.
Renting the same eight GPUs
A DGX H100, H200, B200 or B300 holds eight GPUs. This is what each of those GPUs rents for right now:
Supermicro's own store listed its fixed-configuration eight-GPU HGX B300 server, part number AS-8126GS-NB3RT-01-G2, at $662,845.95 on 21 September 2026. The table below turns that budget into GPU-hours at each GPU's cheapest live price instead:
Divide the GPU-hours by eight to get how long a full eight-GPU node would run at those per-GPU prices. It is a budget comparison, not a quote for a node with the same interconnect and contract terms. For clusters bigger than one node, or for reserved and multi-month pricing instead of on-demand, see GPU cluster pricing.
Buy a DGX, buy an HGX server, or rent
Buy a DGX when you want one company answering for the whole box, GPUs, CPUs, networking and support from NVIDIA in a single SKU, and you are building a fixed, always-on cluster large enough to justify it. Buy an HGX-based server from an OEM such as Dell, Supermicro or Lenovo when you already run your own infrastructure and want to choose the CPU, storage and network fabric yourself, and check the OEM's price against the equivalent DGX rather than assuming it is cheaper. Rent when the workload is not fixed and always on: a training run, a short-term spike, or a project that has not yet proven it needs dedicated hardware for years.
Run your own numbers through the rent versus buy GPU calculator before committing to either purchase. NVLink vs PCIe vs SXM explains the interconnect inside whichever box you choose, NVIDIA H100 price and NVIDIA GPU prices cover the per-card end of this same market, and B300 vs B200 compares these two Blackwell GPU families directly.
Choose DGX for a unified system purchase, an HGX-based OEM server for a configuration tailored to your infrastructure, and rental for work that does not justify a permanent hardware commitment.
Sources
- TechCrunch: Nvidia announces a supercomputer aimed at deep learning and AI, 5 April 2016
- Tom's Hardware: Nvidia Volta V100 DGX-1, HGX-1, May 2017
- TechPowerUp: Nvidia announces the DGX-2 system, 27 March 2018
- NVIDIA: DGX A100 launch press release, 14 May 2020
- NVIDIA: DGX H100 announcement, 22 March 2022
- NVIDIA DGX H100 user guide, retrieved 21 September 2026
- Microway: DGX A100 throughput and hardware summary, retrieved 28 September 2026
- Dell PowerEdge XE9680, retrieved 21 September 2026
- Supermicro SYS-821GE-TNHR, retrieved 21 September 2026
- SemiAnalysis: AI server cost analysis, memory is the..., 29 May 2023
- Oregon State University: Announcing NVIDIA DGX H200, 30 January 2025
- Broadberry: NVIDIA DGX H200, retrieved 21 September 2026
- wccftech: Nvidia Blackwell DGX B200 price, 12 October 2024
- Broadberry: NVIDIA DGX B200, retrieved 21 September 2026
- NVIDIA DGX B200 product page, retrieved 21 September 2026
- CDW: DGX B200 system with support, retrieved 21 September 2026
- NVIDIA: Blackwell Ultra, DGX SuperPOD supercomputer AI factories, 18 March 2025
- NVIDIA DGX B300 product page, retrieved 28 September 2026
- NVIDIA DGX B300 user guide, retrieved 21 September 2026
- Broadberry: NVIDIA DGX B300, retrieved 21 September 2026
- NVIDIA HGX platform page, retrieved 21 September 2026
- Viperatech: NVIDIA DGX H100 Deep Learning Console, retrieved 28 September 2026
- Newegg: Supermicro eight GPU H100 HGX server, retrieved 21 September 2026
- Lenovo Press: on premise vs cloud generative AI TCO, 2026 edition, updated 24 July 2026
- Supermicro: 8U GPU Gold Series AS-8126GS-NB3RT-01-G2, retrieved 21 September 2026
- NVIDIA DGX SuperPOD FAQ, retrieved 28 September 2026
- NVIDIA DGX SuperPOD reference architecture, B300 XDR, retrieved 28 September 2026
- NVIDIA DGX SuperPOD reference architecture, B300 Ethernet, retrieved 28 September 2026
- NVIDIA DGX Cloud page, retrieved 28 September 2026
- NVIDIA: DGX Cloud Lepton announcement, 18 May 2025
- NVIDIA DGX Station product page, retrieved 28 September 2026
- Tom's Hardware: Nvidia's GB300-powered DGX Station listed for nearly $100,000, 23 August 2026
- NVIDIA DGX GB200 user guide, retrieved 28 September 2026
- NVIDIA GB200 NVL72 product page, retrieved 28 September 2026
- NVIDIA GB300 NVL72 product page, retrieved 28 September 2026
- NVIDIA DGX Spark hardware documentation, retrieved 28 September 2026