HGX B300 servers compared
Series 3 · GPU server catalog · HGX B300 8-GPU servers · catalog as of 7 Nov 2026
Almost every server maker now sells an eight-GPU machine built around NVIDIA's HGX B300 board. The GPUs, their memory and the 800-gigabit network ports are the same in all of them; what differs is the host around the board — Intel or AMD processors, air or liquid cooling, eight to ten rack units or a slim node for a liquid-cooled rack — and that is what moves the price and decides where the server can be hosted.
Key facts
- Compute Census lists 22 HGX B300 8-GPU server systems from 9 manufacturers (ASUS, Cisco, Dell Technologies, GIGABYTE (Giga Computing), Hewlett Packard Enterprise, Inventec, Lenovo, QCT (Quanta Cloud Technology), Supermicro), 22 of them shipping, as of 7 Nov 2026.
- Dated public listings put one server at $518,139–$827,308, median $599,995 across 7 listings checked 2026-11-07.
- Renting the same 8 GPUs around the clock at the median on-demand rate ($9.50 per GPU-hour) costs about $55,480 a month.
Who builds them
The catalog lists 22 HGX B300 systems from 9 manufacturers, all of them shipping. That breadth is the point of the HGX model: NVIDIA sells the baseboard with eight Blackwell Ultra GPUs, the NVLink switches between them and, on this generation, eight ConnectX-8 network adapters mounted on the board itself, and each maker designs the chassis, the host and the cooling around it. A buyer comparing two HGX B300 servers is therefore comparing everything except the GPUs.
Supermicro has the widest range, with air-cooled 8U systems, 4U liquid-cooled systems and a liquid-cooled node for 21-inch OCP racks, each in an Intel Xeon 6 and an AMD EPYC version. Dell sells the PowerEdge XE9780 (Intel) and XE9785 (AMD) as 10U air-cooled servers and the XE9780L and XE9785L as liquid-cooled nodes for its IR7000 rack. Gigabyte mirrors the same split with 8U and 4U models on both processor families. HPE, Lenovo, ASUS, QCT, Cisco and Inventec each list one or two systems, mostly on Intel Xeon 6 hosts.
The host processor is a real choice, not a detail. The Intel versions use two Xeon 6 processors with performance cores, typically up to 350 W each; the AMD versions use two EPYC 9005 processors, with up to 192 cores each on Dell’s XE9785. For training jobs that live almost entirely on the GPUs the difference is small; for data preparation, inference serving with heavy pre- and post-processing, or software that is licensed per core, it changes both the price and the result.
Air, liquid or a rack node
Three physical shapes cover the whole platform. The air-cooled server is the familiar one: eight to ten rack units, rows of fans, and a power draw that many colocation halls can only supply for one or two servers per rack. The liquid-cooled server carries cold plates on the GPUs and processors, shrinks to four or five units and moves most of its heat into a water loop; Supermicro’s 4U model, for instance, is sold only with a complete rack and onsite service. The third shape is the rack node — Supermicro’s 2OU OCP system, Dell’s 3OU nodes, QCT’s 3U MGX node — designed to slide into a liquid-cooled rack that is fed by a busbar and power shelves rather than individual power supplies. Supermicro states up to 18 of its nodes, or 144 GPUs, in one rack.
The cooling column in Table 1 is taken from each maker’s own wording. 10 systems are explicitly liquid-cooled and 8 explicitly air-cooled; for the rest the product page does not say, and we do not guess from the model number.
Power and networking
Manufacturers of HGX B300 servers publish the power supplies, not the maximum draw: Supermicro’s 8U system carries six 6,600 W supplies in a three-plus-three arrangement, Dell’s XE9780 twelve 3,200 W supplies, Lenovo’s SR680a V4 eight. Redundancy hides the real figure, so the planning number comes from NVIDIA, which states about 14 kW for its own eight-GPU DGX B300. Lenovo specifies each B300 at 1,100 W and Dell’s liquid-cooled XE9785L at 1,200 W per GPU, which explains why the liquid versions exist: the heat of eight such parts is hard to move with air.
Networking has become uniform. Almost every system carries eight ConnectX-8 adapters at up to 800 Gb/s, one per GPU, for the cluster fabric — InfiniBand or Ethernet, chosen at order time — while the host network for storage and management is separate and varies by maker.
What one costs, and what renting it costs
Dated public listings of complete HGX B300 servers run from $518,139 to $827,308, with a median of $599,995 across 7 listings. The spread follows the configuration: an AMD host with two BlueField-3 processors, preconfigured NVMe storage or a larger memory build lists well above a minimal Intel starting configuration. Renting the same eight GPUs around the clock at the median on-demand rate costs about $55,480 a month, so the median listed server equals roughly 10.8 months of rent — before the electricity, cooling, rack space and network fabric that ownership adds and rental includes.
That arithmetic is why most HGX B300 buyers are clouds, AI labs with sustained training load, or enterprises with data that cannot leave their own sites. For shorter projects, reserved rental of the same GPUs is covered on the B300 price page.
Every HGX B300 8-GPU server system in the catalog
Table 1HGX B300 8-GPU servers: systems by manufacturer, shipping first
| System | GPUs | Form factor | Cooling | CPU | Networking | Max power | Status |
|---|---|---|---|---|---|---|---|
| ASUS XA NB3I-E12Up to 1:1 GPU-to-NIC ratio | 8 | 9U | not stated | 2x Intel Xeon 6700P, up to 350W TDP | Embedded ConnectX-8 SuperNICs (8x CX8); BlueField-3 DPU support | not stated | shipping |
| Cisco UCS C880A M8Based on NVIDIA HGX reference architecture | 8 | 10RU | Air | 2x Intel Xeon 6776P or 6767P | 8x GPU-board integrated ConnectX-8 (E/W) | not stated | shipping |
| Dell Technologies PowerEdge XE9780B300 config: 8x HGX B300 NVL8 270GB 1100W SXM6 | 8 | not stated | Air | 2x 6th Gen Intel Xeon (Xeon 6), up to 86 cores each | 8x ConnectX-8 OSFP (embedded, B300 config) | not stated | shipping |
| Dell Technologies PowerEdge XE9780LLiquid-cooled CPUs, GPUs and NVLink switches | 8 | 3OU (IR7000 ORv3 rack) | Liquid (DLC) | 2x Intel Xeon 6, up to 72 cores each | not stated | not stated | shipping |
| Dell Technologies PowerEdge XE9785AMD EPYC host; alternative AMD Instinct MI355X configuration | 8 | 10U | Air | 2x 5th Gen AMD EPYC 9005, up to 192 cores each | 8x ConnectX-8 OSFP (embedded, B300 config) | not stated | shipping |
| Dell Technologies PowerEdge XE9785LDirect-to-chip liquid-cooled CPUs and GPUs | 8 | 3OU (IR7000 ORv3 rack) | Liquid (DLC) | 2x 5th Gen AMD EPYC 9005, up to 192 cores each | not stated | not stated | shipping |
| GIGABYTE (Giga Computing) G4L4-SD3-LAX7Liquid-cooled HGX B300 (DLC) | 8 | 4U | Liquid (DLC) | 2x Intel Xeon 6700/6500-series | 8x 800 Gb/s OSFP XDR InfiniBand or dual 400 Gb/s Ethernet via onboard ConnectX-8 SuperNIC | not stated | shipping |
| GIGABYTE (Giga Computing) G4L4-ZD3-LAX7Liquid-cooled HGX B300 (DLC), AMD EPYC host | 8 | 4U | Liquid (DLC) | 2x AMD EPYC 9005/9004 | 8x 800 Gb/s OSFP XDR InfiniBand or dual 400 Gb/s Ethernet via onboard ConnectX-8 SuperNIC | not stated | shipping |
| GIGABYTE (Giga Computing) G894-SD3-AAX7Onboard ConnectX-8 SuperNICs | 8 | 8U | not stated | 2x Intel Xeon 6700/6500-series, up to 350W TDP | 8x 800 Gb/s OSFP XDR InfiniBand or dual 400 Gb/s Ethernet via onboard ConnectX-8 SuperNIC | not stated | shipping |
| GIGABYTE (Giga Computing) G894-ZD3-AAX7AMD EPYC host for HGX B300 | 8 | 8U | not stated | 2x AMD EPYC 9005/9004, cTDP up to 500W | 8x 800 Gb/s OSFP XDR InfiniBand or dual 400 Gb/s Ethernet via onboard ConnectX-8 SuperNIC | not stated | shipping |
| Hewlett Packard Enterprise HPE Compute XD69010U single-node air-cooled chassis | 8 | 10U | Air | 2x Intel Xeon 6 with P-cores (e.g. 6776P, 6767P) | 8x ConnectX-8 (InfiniBand XDR or Spectrum-X Ethernet) | not stated | shipping |
| Hewlett Packard Enterprise HPE ProLiant Compute XD685 (HGX B300 DLC)Requires XD685 DLC High Power server chassis (S6Y07A) | 8 | 5U | Liquid (DLC) | 2x 5th Gen AMD EPYC 9005 | not stated | not stated | shipping |
| Inventec P5800G7 (LC)Liquid-cooled HGX B300 NVL8 in 5U | 8 | 5U | Liquid (DLC) | Intel Xeon 6700/6500-series P-cores or 6700-series E-cores | 8x 800Gb OSFP (ConnectX-8 SuperNICs) | not stated | shipping |
| Inventec P9000G7 (AC)Decoupled 54V GPU and 12V host power domains | 8 | 10U | Air | Intel Xeon 6700/6500-series P-cores or 6700-series E-cores | 8x 800Gb OSFP (E-W) | not stated | shipping |
| Lenovo ThinkSystem SR680a V4Eight B300 1100W GPUs, 288 GB HBM3e each | 8 | 8U | Air | 2x Intel Xeon 6700P-series (up to 86 cores, 350W) | 8x integrated ConnectX-8 800Gb/s (OSFP) | not stated | shipping |
| QCT (Quanta Cloud Technology) QuantaGrid D75H-10UFront-access PSUs and SSDs | 8 | 10U | not stated | 2x Intel Xeon 6, up to 350W TDP | 8x ConnectX-8 OSFP ports, 800G Ethernet or InfiniBand | not stated | shipping |
| QCT (Quanta Cloud Technology) QuantaGrid D75L-3U19-inch NVIDIA MGX 1.1 rack node | 8 | 3U (MGX rack) | Liquid (DLC) | 2x Intel Xeon 6700 series (P-cores), up to 350W | 8x ConnectX-8 SuperNICs, up to 800Gb/s | not stated | shipping |
| Supermicro AS -4126GS-NB3RT-LCCAMD EPYC host, 4U DLC | 8 | 4U | Liquid (DLC) | 2x AMD EPYC 9005/9004 series, up to 500W | 8x ConnectX-8 SuperNIC (8x OSFP 800GbE); up to 2x BlueField-3 DPU | not stated | shipping |
| Supermicro AS -8126GS-NB3RTAMD EPYC host for HGX B300 | 8 | 8U | Air | 2x AMD EPYC 9005/9004 series | 8x ConnectX-8 SuperNIC (8x OSFP 800GbE); up to 2x BlueField-3 DPU | not stated | shipping |
| Supermicro SYS-222GS-NB3OT-ALC21-inch OCP ORv3 node; up to 18 nodes / 144 GPUs per rack | 8 | 2OU (OCP ORv3) | Liquid (DLC) | 2x Intel Xeon 6700 series (P-cores), up to 350W | 8x ConnectX-8 SuperNIC (8x OSFP 800 Gb/s InfiniBand) | not stated | shipping |
| Supermicro SYS-422GS-NB3RT-LCCDirect-to-chip cold plate liquid cooling | 8 | 4U | Liquid (DLC) | 2x Intel Xeon 6700 series (P-cores), up to 350W | 8x ConnectX-8 SuperNIC (8x OSFP 800 Gb/s) | not stated | shipping |
| Supermicro SYS-822GS-NB3RTAir-cooled, front-I/O HGX B300 system | 8 | 8U | Air | 2x Intel Xeon 6700 series (P-cores), up to 350W | 8x ConnectX-8 SuperNIC (8x OSFP 800GbE); 2x 10GbE RJ45 | not stated | shipping |
Specifications as published by each manufacturer on its product page or datasheet; "not stated" means the manufacturer does not publish the figure. Product pages are listed on Sources, GPU server catalog.
Table 2HGX B300 8-GPU servers: purchase prices from dated public listings, USD
| Configuration | Listings | Low | Median | High | Seller type | Checked |
|---|---|---|---|---|---|---|
| Eight-GPU server (HGX) e.g. Supermicro SYS-822GS-NB3RT HGX B300 8-GPU 8U air-cooled · Supermicro SYS-822GS-NB3RT-01-G2 Gold Series HGX B300 8-GPU 8U | 7 | $518,139 | $599,995 | $827,308 | reseller, system builder | 2026-11-07 |
List prices of starting configurations on the dates shown, before volume discounts; tax as listed by each seller (often not stated). Sellers are recorded with their page and a quote of the price but not named or linked: final configuration, delivery date and terms differ for each order, so the form below returns current offers.
Table 3One server: owning against renting the same GPUs
| Line | Value | Basis |
|---|---|---|
| GPUs per server | 8 | platform definition |
| Rent at the median on-demand rate, per month | $55,480 | $9.50 per GPU-hour × 8 × 730 h |
| Rent at the lowest on-demand rate, per month | $34,398 | $5.89 per GPU-hour |
| Rent at the 12-month reserved rate, per month | $42,165 | $7.22 per GPU-hour |
| Median listed purchase price | $599,995 | 7 dated listings, tax as listed |
| Months of median-rate rent equal to the price | 10.8 | excludes power, cooling, hosting, networking, resale value |
730 hours = one month of continuous use. Rental rates: GPU rental price index, model page B300.
Table 4HGX B300 8-GPU servers: power as stated by each manufacturer
| System | Max power | Power supplies or shelves |
|---|---|---|
| ASUS XA NB3I-E12 | not stated | 5+5 3200W 80 PLUS Titanium |
| Cisco UCS C880A M8 | not stated | 12x 3.2kW 50V MCRPS (N+N) |
| Dell Technologies PowerEdge XE9780 | not stated | 12x 3200W Titanium, hot-swap redundant |
| Dell Technologies PowerEdge XE9780L | not stated | 33kW power shelf (6x 5500W AC PSUs) |
| Dell Technologies PowerEdge XE9785 | not stated | 12x 3200W Titanium, hot-swap redundant |
| Dell Technologies PowerEdge XE9785L | not stated | 33 kW power shelf (6x 5500W AC PSUs) |
| GIGABYTE (Giga Computing) G4L4-SD3-LAX7 | not stated | 5+5 3000W 80 PLUS Titanium redundant |
| GIGABYTE (Giga Computing) G4L4-ZD3-LAX7 | not stated | 5+5 3000W 80 PLUS Titanium redundant |
| GIGABYTE (Giga Computing) G894-SD3-AAX7 | not stated | 12x 3000W 80 PLUS Titanium redundant |
| GIGABYTE (Giga Computing) G894-ZD3-AAX7 | not stated | 12x 3000W 80 PLUS Titanium redundant |
| Hewlett Packard Enterprise HPE Compute XD690 | not stated | 12x 3200W (N+N) |
| Hewlett Packard Enterprise HPE ProLiant Compute XD685 (HGX B300 DLC) | not stated | 12x HPE 3000W CRPS Titanium (default for B300 configs) |
| Inventec P5800G7 (LC) | not stated | 3300W 54V ATS PSUs (4+1 or 4+2) + 2000W 12V PSUs (1+1) |
| Inventec P9000G7 (AC) | not stated | 3300W 54V ATS PSUs (5+1) + 2000W 12V PSUs (1+1) |
| Lenovo ThinkSystem SR680a V4 | not stated | 8x hot-swap: 3200W AC (N+1) or 3800W HVAC (N+N), 80 PLUS Titanium |
| QCT (Quanta Cloud Technology) QuantaGrid D75H-10U | not stated | 12x 3200W (N+N) |
| QCT (Quanta Cloud Technology) QuantaGrid D75L-3U | not stated | MGX rack 1400A busbar with power shelves |
| Supermicro AS -4126GS-NB3RT-LCC | not stated | 4x 6600W Titanium (2+2) |
| Supermicro AS -8126GS-NB3RT | not stated | 6x 6600W Titanium (3+3) |
| Supermicro SYS-222GS-NB3OT-ALC | not stated | Rack busbar (1400A) with power shelves |
| Supermicro SYS-422GS-NB3RT-LCC | not stated | 4x 6600W Titanium (2+2) |
| Supermicro SYS-822GS-NB3RT | not stated | 6x 6600W Titanium (3+3) |
Max power only where the manufacturer states a maximum draw; power-supply capacity is shown separately because redundant supplies (N+N) overstate the real load.
Table 5HGX B300 8-GPU servers: cooling of the catalogued systems
| Cooling | Systems |
|---|---|
| Liquid (DLC) | 10 |
| Air | 8 |
| Not stated | 4 |
As published by each manufacturer. Liquid-cooled systems need direct liquid cooling in the hosting hall.
Questions and answers
- How much does an HGX B300 server cost?
- Dated public listings of complete eight-GPU HGX B300 servers range from about $518,000 to $827,000 as listed (most sellers do not state the tax treatment), with a median near $600,000. These are starting configurations; memory, storage, network adapters and support terms push a configured quote higher.
- Is an HGX B300 server air-cooled or liquid-cooled?
- Both exist. Air-cooled versions take eight to ten rack units; liquid-cooled versions put the same eight GPUs into four or five units, or into a slim node for an OCP or MGX rack, and need direct liquid cooling in the hall.
- What is the difference between HGX B300 and DGX B300?
- HGX B300 is NVIDIA's eight-GPU board that server makers build their own systems around. DGX B300 is NVIDIA's own complete system on the same GPUs, sold with NVIDIA's software and support; it is usually listed above HGX servers.
- How much power does an HGX B300 server need?
- Most makers publish only their power-supply configuration, not a maximum draw. NVIDIA states about 14 kW for its own eight-GPU DGX B300, which is a sound planning figure for an air-cooled HGX B300 server.
Sources: manufacturer pages behind every row are listed under Sources, GPU server catalog. All platforms: GPU server catalog.
Quote for HGX B300 servers
Tell us how many servers, which host processor and cooling, and the delivery site. We return current offers from manufacturers' partners with delivery dates; hosting can be quoted alongside.