The DGX B300 is NVIDIA's own eight-GPU Blackwell Ultra server: eight B300 GPUs with 2.1 TB of GPU memory, two Intel Xeon 6776P processors and eight 800 Gb/s network ports in an air-cooled 10 RU chassis that draws 14.5 kW. NVIDIA publishes no price for it. Reuters put an eight-GPU B300 server at about $550,000 in the United States in April 2026. The same GPUs also come on NVIDIA's HGX B300 board inside servers from other makers, which is the kind QuantaCloud builds to order.
DGX B300 specs#
The DGX B300 is one fixed configuration with two power options. These figures are from NVIDIA's October 2026 datasheet, its product page and its user guide, read on 2026-11-12.
| Spec | DGX B300 |
|---|---|
| GPUs | 8 Blackwell Ultra SXM, the B300 |
| GPU memory | 2.1 TB, with 62 TB/s of HBM3E bandwidth |
| FP4 Tensor Core | 144 PFLOPS sparse, 108 dense |
| FP8/FP6 Tensor Core | 72 PFLOPS sparse |
| NVLink | Two fifth-generation NVLink switches, 14.4 TB/s in total |
| CPUs | 2 Intel Xeon 6776P |
| System memory | 2 TB, upgradable to 4 TB |
| Cluster network | 8 ConnectX-8 ports, up to 800 Gb/s InfiniBand or Ethernet each |
| Storage and management network | 2 dual-port BlueField-3 DPUs, up to 400 Gb/s |
| Drives | 2 x 1.9 TB NVMe for the operating system, 8 x 3.84 TB NVMe for data |
| Power | 14.5 kW on a busbar, 15.1 kW with its own power supplies |
| Power input | 12 supplies on 200 to 240 V AC, or a 54 V DC busbar |
| Cooling | Air, 10 to 35°C at the inlet |
| Size and weight | 10 RU, 17.4 x 19 x 35.74 in, 168 kg with power supplies or 123 kg without |
| Software | DGX OS included. Ubuntu, Red Hat Enterprise Linux and Rocky supported |
| Support | Three years of NVIDIA business-standard hardware and software support |
Two figures differ between NVIDIA's own documents. The datasheet and product page give 2.1 TB of GPU memory, which is eight GPUs at 270 GB, while the user guide says 8 x 288 GB = 2.3 TB. NVIDIA puts the Blackwell Ultra chip's maximum at 288 GB, and a B300 reports 275,040 MiB to nvidia-smi, which is 268.6 GiB (our calculation), so 2.1 TB is the figure I plan on. Power is about 14 kW on the product page and 14.5 kW in the user guide, which also gives a system maximum of 15 kW and a heat output of 49,476 BTU per hour.
Against the DGX B200, NVIDIA's datasheet claims 1.5x the dense FP4 and 2x the attention performance. The B300 vs B200 comparison covers what else changed, including the FP64 and INT8 the B300 gave up.
Power and placement#
The DGX B300 is air-cooled, which the Rubin systems that follow it are not. It still asks a lot of one rack position: 14.5 kW and 1,350 to 1,500 CFM of airflow. The AC version has twelve power supplies and twelve power cords. NVIDIA says the system needs six supplies to run at full performance, so the other six are the redundancy. The DC version has no supplies of its own: it clips to a 54 V busbar in an MGX rack, and power shelves in the rack do the conversion. Network cables and the boot drives are reached from the front. The eight ConnectX-8 ports sit on the GPU tray and connect to the GPUs over PCIe Gen6, which is how each GPU gets a full 800 Gb/s to the cluster network.
How the DGX B300 is sold, and what it costs#
The DGX B300 is sold by quote, not from a price list. NVIDIA's page lists it as shipping and available now, and its buttons lead to a contact form and a where-to-buy page rather than a price. NVIDIA also runs programs for colocation facilities that host DGX systems and for firms that operate them on a customer's behalf. One large US distributor shows the pattern: CDW listed the DGX B200 with "Request Pricing" in place of a price on 2026-11-12, and its search returned no DGX B300 listing.
| Date | Source | What the price covers | Price |
|---|---|---|---|
| 2026-04-30 | Reuters, citing two industry sources | A B300 server with eight GPUs, in the United States | About $550,000 |
| 2026-04-30 | Same report | The same server late in 2025 | Around $500,000 |
| 2026-06-24 | Tom's Hardware, in a report on Chinese black-market prices | The DGX B300 at US retail, with no source given | Nearly $400,000 |
Reuters does not say whose server that is, so read it as the price of an eight-GPU B300 system, not as NVIDIA's price for the DGX. The only figure I found for the DGX-branded system itself is a passing line in that Tom's Hardware report, which gives no source for it and sits about $150,000 below Reuters' figure. I cannot reconcile the two from public information, so treat both as dated estimates. When you compare quotes, check what each one includes against the datasheet: DGX OS is included, NVIDIA AI Enterprise and Mission Control are listed as recommended rather than included, and three years of business-standard support come with the system.
DGX vs HGX vs MGX#
DGX is a finished NVIDIA system, HGX is the GPU board other makers build servers around, and MGX is the chassis and rack standard both can use.
| DGX B300 | HGX B300 server | MGX | |
|---|---|---|---|
| What it is | NVIDIA's own branded system | NVIDIA's eight-GPU baseboard with NVLink switches, built into a server by another maker | NVIDIA's open modular reference architecture for servers and racks |
| Who you buy from | NVIDIA's DGX sellers, by quote | The server maker or the builder | Nobody: it is a design, not a product |
| GPUs and GPU figures | 8 Blackwell Ultra SXM, 2.1 TB, 144 and 108 PFLOPS of FP4 | The same | Whatever the system built on it carries |
| CPU, memory, storage | Fixed: two Xeon 6776P, 2 TB of RAM with one upgrade to 4 TB, set drives | Chosen by whoever builds the server | More than 100 combinations, per NVIDIA |
| Support | NVIDIA, three years included | The builder's own terms | |
| Software | DGX OS included | Your choice of Linux, with the same NVIDIA driver and CUDA |
That split has four practical consequences. On configuration, a DGX is the same machine for every buyer, while an HGX server can be specified around the workload: CPU, memory, storage, network cards and chassis. On support, a DGX comes with NVIDIA's own contract, and an HGX server comes with whatever its builder commits to in writing. On price, neither has a list price, so the comparison only exists between two written quotes for the same GPU count. On lead time, both depend on the same Blackwell Ultra GPUs, and I have found no published figure for either, so that belongs in the quote too.
MGX is the odd one out, because it is not an alternative to the other two. The DGX B300 is offered in an MGX-compatible chassis, and NVIDIA's NVL72 racks are MGX racks.
DGX B200, DGX GB300, the Rubin systems and SuperPOD#
The DGX name covers servers and whole racks. These are the current ones, from NVIDIA's datasheets.
| System | What it is | GPUs and GPU memory | Power and cooling |
|---|---|---|---|
| DGX B200 | The previous eight-GPU server | 8 Blackwell GPUs, 1,440 GB | About 14.3 kW, air-cooled, 10 RU |
| DGX B300 | The eight-GPU Blackwell Ultra server | 8 Blackwell Ultra GPUs, 2.1 TB | 14.5 kW, air-cooled, 10 RU |
| DGX GB300 | NVIDIA's branded GB300 NVL72 rack | 72 Blackwell Ultra GPUs and 36 Grace CPUs, 20 TB | 100% liquid-cooled. A full GB300 NVL72 rack can require up to 142 kW |
| DGX Rubin NVL8 | The eight-GPU Rubin server | 8 Rubin GPUs, 2 TB | About 24 kW, liquid-cooled, 2 RU |
| DGX Vera Rubin NVL72 | NVIDIA's branded Vera Rubin rack | 72 Rubin GPUs and 36 Vera CPUs, 20.7 TB | 100% liquid-cooled, 48 RU |
| DGX SuperPOD | A whole cluster rather than one machine: DGX systems with NVIDIA's networking, storage and management software, validated together | Any of the above, up to tens of thousands of GPUs | Depends on the systems |
For rack prices, the GB200 vs GB300 comparison has the dated estimates. The desk-side DGX Spark is a different class of machine, covered in the DGX Spark price guide.
The HGX B300 route with QuantaCloud#
QuantaCloud's route is HGX. We order and build HGX B300 servers to your spec, as one dedicated GPU server or as a GPU cluster of several joined over 800 Gb/s InfiniBand or Ethernet. You send a brief with the GPU count, CPU, memory, storage, network, location, start date and term, and QuantaCloud replies in writing with a configuration, a lead time and commercial terms. The B300 page has the GPU's specs, what fits in 270 GB and the CUDA and driver versions it needs.
Lead time depends on GPU supply when the order is placed, and it is confirmed in writing before you commit.
DGX B300 questions#
Is the DGX B300 air-cooled?
Yes. It is an air-cooled 10 RU system with an inlet range of 10 to 35°C, and NVIDIA's user guide gives its airflow as 1,350 to 1,500 CFM at 70% fan speed, depending on the power version. The Rubin generation's DGX Rubin NVL8 is liquid-cooled.
How much power does a DGX B300 need?
Plan for 14.5 kW per system, and 15.1 kW for the version with its own power supplies, according to the datasheet. The AC version takes twelve cords on 200 to 240 V.
What is the difference between DGX B300 and HGX B300?
The GPUs are the same eight Blackwell Ultra modules. The DGX B300 is NVIDIA's complete system with a fixed CPU, memory and storage and NVIDIA's support. HGX B300 is the GPU board, and the server around it comes from whoever builds it.
What is a DGX SuperPOD?
It is NVIDIA's validated design for a cluster of DGX systems, with the networking, storage and management software specified. NVIDIA offers it with each of the DGX systems in the table above, from the DGX B200 to the DGX Vera Rubin NVL72.
The rule I follow: buy the DGX B300 when you want NVIDIA's fixed configuration and NVIDIA's support contract on your own floor, and build on HGX B300 when you want to choose the server around the GPUs, or want a cluster sized to one workload. Either way, get the configuration, the lead time and the support terms in writing before you compare prices. For the HGX route, start with a brief on the reserved capacity page.
Plan an HGX B300 build