DGX B300 vs HGX B300: What the Price Difference Actually Buys

◆ THE DISTINCTION

DGX is a product, HGX is a part

DGX B300 is a system NVIDIA designs, builds, ships and supports as one unit: eight B300 GPUs on the HGX baseboard, two Intel Xeon 6776P CPUs, 2TB of system memory, eight ConnectX-8 SuperNICs at up to 800 Gb/s each, two BlueField-3 DPUs, NVMe storage, all in a 10U air-cooled chassis with DGX OS, NVIDIA AI Enterprise, Base Command and Mission Control preinstalled. You cannot change the CPU, the memory, the NIC count or the cooling. That is the point of it.

HGX B300 is the baseboard inside that system: the eight GPUs, the NVLink Switch fabric and the PCIe connectivity, sold by NVIDIA to OEMs who put it in their own server. Supermicro, Dell, Lenovo, HPE and others each build an HGX B300 product line with their own choice of CPU, memory, storage, NICs, chassis and cooling. Some of those are air-cooled and look much like a DGX. Others are liquid-cooled, denser, or built for a specific rack standard.

So every DGX B300 contains an HGX B300. Not every HGX B300 is a DGX. When a search query says "DGX B300 price" the buyer usually means "8-GPU B300 server price," and the answer depends entirely on which of the two they end up buying.

◆ DGX B300 VS HGX B300, SIDE BY SIDE

Specifications from NVIDIA DGX B300 documentation as compiled by GPUSmith and NADDOD. Prices from published reseller and OEM listings, June to October 2026, excluding VAT. NVIDIA publishes no DGX list price. Rental band from the GPUaaS.com October 2026 benchmark.

◆ WHAT THE INVOICE DIFFERENCE BUYS

Integration, software and one throat to choke

The surprising thing in the published numbers is that the DGX is not the expensive option. Reseller quotes for a configured DGX B300 cluster around $400,000 to $500,000, while OEM HGX B300 builds from Lenovo and others are listing above $785,000. That inverts the usual assumption, and the reason is what the OEM wraps around the baseboard: higher-core CPUs, more system memory, more NVMe, liquid cooling loops, and an OEM support contract priced for enterprise accounts. The HGX baseboard itself is the same.

What the DGX buys that an HGX does not is the validated stack. NVIDIA has tested that exact combination of firmware, drivers, NICs, OS and orchestration together, ships it that way, and supports all of it under one contract. For a team without infrastructure engineers who want to run NVIDIA's reference software, that is worth real money in time not spent. For a team with its own platform, its own orchestration and its own preferred CPU, it is money spent on things they will replace.

The B300 node price breakdown works through where the money goes inside an 8-GPU node and lands on the same conclusion: the GPUs are the smaller half of the bill on a high-end HGX build, and the spread between the cheapest and most expensive node with the same eight GPUs is 2.6x.

◆ COOLING AND FACILITY

The air-cooled DGX is a colocation problem, the liquid HGX is a facility problem

DGX B300 is air-cooled, which sounds like the easy option until the number is read. It draws up to 14.5 kW and moves around 1,500 CFM of air through one 10U chassis. Most legacy colocation racks are provisioned at 10 to 20 kW for the whole rack. One DGX B300 per rack is feasible in a lot of facilities; two is a conversation with the data center; four is a new contract. Hot-aisle containment stops being optional, and facilities running warm supply-air setpoints for efficiency need to re-check them before swapping a B300 into a slot that held a B200.

HGX B300 in a liquid-cooled OEM chassis solves the density problem and creates a different one: the facility needs a coolant distribution unit, manifolds and a liquid loop. Where that exists, four to eight B300 nodes per rack is routine. Where it does not, the capex to add it is counted in the hundreds of thousands and the lead time in months. This is why the HGX liquid-cooled build costs more and why it is the configuration most hyperscaler and neocloud operators actually deploy.

For a buyer who is renting rather than owning, none of this is their problem, which is the strongest argument in this post for renting B300 in 2026. The provider has already solved power and cooling, and the committed band prices the GPU-hour, not the chassis.

◆ FOR RENTAL

Buyers quote the GPU and the fabric, not the badge

Across requests through the GPUaaS.com network, almost nobody specifies DGX or HGX. They specify B300, a GPU count, a term, a region, and whether the inter-node fabric is InfiniBand or Ethernet. The committed band this month is $5.00 to $6.50 per GPU-hour on either chassis, and where a quote lands inside that band is set by term length, cluster size, interconnect and region. The October benchmark covers how to read the band.

That has an implication for anyone buying nodes with the intention of renting them out. The DGX premium for software and single-vendor support is invisible to a rental buyer, who brings their own stack and never calls NVIDIA. The HGX premium for liquid cooling and density is invisible too, except that it lets a provider put more GPUs in the same facility footprint, which shows up in their margin rather than in the quote. A provider buying for rental should optimise for cost per GPU-hour over three years, which the node price post puts between $3.06 and $8.05 depending on purchase price and facility cost.

The B300 vs B200 comparison covers whether B300 is the right part at all, and for teams who want a single B300 rather than an 8-GPU node, packet.ai handles self-serve access without the chassis question arising.

◆ FAQ

Frequently asked questions