HGX Vs DGX: Same Blackwell GPUs, Different Systems and Cluster Designs

Summary

HGX B300 and DGX B300 carry the same NVIDIA HGX baseboard: eight Blackwell SXM GPUs linked by NVLink. DGX is NVIDIA's own complete system, with the CPU, networking, software and support fixed by NVIDIA. An HGX server is built around that baseboard by an OEM such as Dell, Supermicro or Lenovo, which chooses the rest. Choose DGX for a single-tenant enterprise cluster on one fixed NVIDIA specification with NVIDIA support, and choose HGX to sell compute to many customers.

An HGX server also fits if you need multi-tenancy, where many separate customers share one cluster, or want your own CPU, cooling and chassis. For a neocloud selling capacity to many customers, that points to HGX.

The HGX baseboard is the board NVIDIA supplies with the GPUs mounted as SXM modules, a socketed GPU form factor. NVLink is NVIDIA's direct GPU-to-GPU interconnect, and an OEM, or original equipment manufacturer, designs and sells the finished server around the board.

NVIDIA leaves the rest to the server builder. In the notes to its HGX Rubin NVL8 specs, the page for NVIDIA's HGX platform, updated September 2026, says "CPU and memory specs are defined by OEM offerings." The same split holds for B300 and B200. This page covers Blackwell 8-GPU systems, B200 and B300, and we leave rack-scale NVL72 systems, a separate category, to another article.

Comparison Table

The GPUs and the NVLink fabric are the same in both columns, and nearly everything around them differs. DGX figures come from NVIDIA's spec page for the DGX B300 as listed in October 2026. HGX figures, including the 1.8 TB/s per-GPU NVLink bandwidth, come from NVIDIA's HGX page.

The B200 generation follows the same pattern. The DGX B200 user guide lists about 14.3 kW maximum in a 10U chassis. HGX B200 carries 1.4 TB of GPU memory per 8-GPU system. For per-generation detail, see the HGX B200 hardware page.

Tenancy: Why Many GPU Clouds Build on HGX

NVIDIA designs its DGX SuperPOD for a single-tenant, multi-user enterprise, while Cisco's published cloud design, compliant with NVIDIA's Cloud Partner (NCP) program, uses HGX servers built for many tenants. For an operator selling compute to third parties, the cluster design matters more than any row in the spec table.

DGX SuperPOD is NVIDIA's cluster-scale product, defined by a reference architecture with a specific bill of materials for hardware and software. It grows in scalable units, the repeatable blocks of systems a cluster is built from. In the current XDR edition of the DGX SuperPOD reference architecture, updated September 2026, each scalable unit contains 72 DGX B300 systems.

NVIDIA states that the design is "for single-tenant, multi-user enterprise environment." Reconfiguring a SuperPOD into a multi-tenant environment for private or public cloud services is at the customer's "own discretion and risk." NVIDIA "does not support these capabilities and performance with the DGX SuperPOD as a product," and enterprise-level support continues for the individual components.

That position covers SuperPOD as a product. It says nothing against a single DGX system, which carries its own three-year NVIDIA support, and component support continues inside a reconfigured SuperPOD. What NVIDIA withholds is support for multi-tenant capabilities and performance across the SuperPOD as a whole.

Cisco's design shows the multi-tenant path. Cisco's cloud reference architecture, published in September 2026, is designed to be compliant with NVIDIA's Cloud Partner (NCP) reference architecture. It "uses NVIDIA-Certified HGX B300 rack servers" and supports multitenancy "to maximize the use of deployed hardware."

For enterprise builds, NVIDIA publishes its own HGX reference design. The HGX AI Factory reference architecture, updated May 2026, builds on HGX B300 systems in 32-, 64- and 128-node designs. Those sizes hold 256, 512 and 1,024 GPUs.

The largest GPU clouds offer the same hardware. CoreWeave made HGX B300 on CoreWeave Cloud generally available in March 2026. Nebius announced HGX B300 clusters on Nebius AI Cloud in October 2026, saying they "are available on Nebius AI Cloud today."

DGX does run public clouds as well. In a January 2026 release about its GB200 NVL72 expansion, Bitdeer AI described an earlier launch. Bitdeer AI's GPU Cloud, the company said, launched "powered by NVIDIA DGX SuperPOD with NVIDIA Hopper GPUs." The DGX route works for a cloud that accepts the support boundary NVIDIA draws around SuperPOD.

Dedicated clusters are a third model. In December 2025, Voltage Park described a dedicated HGX B200 cluster for Cursor as "a customized NVIDIA HGX B200 SuperPOD," engineered with NVIDIA and Dell. Voltage Park calls this HGX build a SuperPOD, so confirm which design a proposal means before you compare quotes.

A cloud selling to many customers earns its revenue from utilization across tenants, so the supported multi-tenant design path matters as much as the GPUs. GPUs sitting idle between customers earn nothing, and a design built to split one cluster across many tenants keeps more of it billed.

An operator whose cluster serves one customer under one offtake contract runs a single-tenant cluster, and either system can fit. The offtake contract is the agreement under which a customer commits to buy the cluster's compute, and GPU financing for neoclouds starts from it.

Performance: the Same GPUs on the Same Baseboard

DGX has no built-in speed advantage over an HGX server. Both carry eight Blackwell GPUs on the HGX baseboard with 1.8 TB/s of NVLink GPU-to-GPU bandwidth per GPU, switched through the NVLink 5 Switch. NVIDIA lists 14.4 TB/s of total NVLink bandwidth per system for HGX B300 and HGX B200 alike.

The differences that do change results sit around the baseboard. The CPU, the network adapters, the cooling design and the cluster fabric that links systems together all vary between a DGX and an OEM build. Two servers with identical GPUs can therefore post different sustained results across a large cluster.

Networking is the clearest case. DGX B300 pairs its GPUs with eight ConnectX-8 network adapters at up to 800 Gb/s. It adds two BlueField-3 DPUs, data processing units that offload networking and storage tasks. On an HGX server, the OEM and the operator choose the adapters and the fabric.

Compare like with like. Set a DGX H100 against an HGX B300 server and you are measuring two GPU generations, Hopper against Blackwell Ultra. NVIDIA's DGX H100/H200 user guide lists eight H100 GPUs with 640 GB of total GPU memory per system, so the 8-GPU DGX format predates Blackwell. Rubin is next, and NVIDIA's HGX page already leads with it.

Configuration and Site: What the OEM Decides

With an HGX server, the OEM's design sets the power, cooling and rack plan, so the server model has to be matched to the site. HGX B300 servers vary by OEM in CPU (Intel Xeon 6 or AMD EPYC), cooling (air or liquid) and height.

Heights run from roughly 4U liquid-cooled to 8–10U air-cooled, where U is one rack unit. Supermicro's HGX B300 systems, for example, come as a 4U liquid-cooled or 8U air-cooled system with dual Intel Xeon 6 or AMD EPYC 9005/9004 processors.

Power per GPU explains the spread. Supermicro's 8U air-cooled system runs eight 1,100 W HGX B300 GPUs, which is 8.8 kW for the GPUs alone before CPUs, networking and fans. How the OEM removes that heat, by air or by liquid, sets the chassis height and the rack density a site can hold.

NVIDIA's certification program helps narrow the list. Through NVIDIA-Certified Systems, NVIDIA certifies partner servers from OEMs.

NVIDIA's certification documentation lists two HGX reference configurations, 2-8-9-800 and 2-8-10-800, that support HGX B200 and B300 systems. Its lists include Lenovo ThinkSystem SR680a V4, HPE ProLiant Compute XD685 and XD690, Supermicro systems and GIGABYTE G894 models. Dell PowerEdge XE9780 appears in its Data Center Servers list. Certification covers the listed models and tested configurations, so check the exact SKU.

DGX B300 is a fixed 10U system at about 14 kW, and USD.AI's B300 reference lists it as air-cooled. With one fixed design, the site plan is known in advance. Before you sign the purchase order for any system, confirm the site's power per rack and cooling type against the exact SKU. Your colocation agreement sets those limits, and one OEM's specs don't carry over to another OEM's HGX server.

Financing: How the Choice Shows up in a Review

In a financing review, the HGX vs DGX choice appears as one line of the equipment schedule: the system model and builder. The equipment schedule is the itemized list of hardware a financing covers, and that one line sets the rest of it. CPUs, memory and storage, networking, power distribution, cooling type, racks, delivery schedule and site all follow from the model you pick.

A financing review asks for the same core documents whichever system is on the schedule:

• System configuration and quantity. The model, builder and server count that define the equipment schedule.
• Supplier and purchase order. For servers already deployed, ownership records take the place of the purchase order.
• Deployment site and colocation agreement. Where the servers will run, and the power and space the site commits to.
• Power and cooling readiness. Evidence that the site matches the SKU's power draw and cooling type.
• Offtake contract or documented on-demand revenue. The revenue that services the financing.

The offtake contract repays the loan, so the system choice matters through what it lets you sell and where it can run.

One published deal shows the scale. In September 2026, a $54.4 million financing through USD.AI funded 1,040 NVIDIA B300 GPUs across 130 servers in Ohio. The deal appears in the USD.AI loan history, and at eight GPUs per server it uses the 8-GPU format this page compares.

When to Choose Which

Your offtake contract and your site decide most of this choice. Three common cases cover many operators.

A multi-tenant GPU cloud selling to many customers. Choose HGX servers on an NCP-compliant cloud design; for enterprise builds, NVIDIA's HGX AI Factory design. Cisco's published NCP-compliant design uses HGX B300 servers and supports multitenancy.

A single-tenant cluster contracted to one enterprise or AI lab. Choose DGX if the customer wants NVIDIA's fixed specification and support, with room to scale to SuperPOD. A dedicated HGX cluster also fits, as Voltage Park's Cursor build shows.

A site with limited cooling or a liquid-cooled build-out. Choose an HGX server whose cooling and height match the facility. USD.AI's B300 reference lists DGX B300 as air-cooled at about 14 kW, so it suits a site with the airflow and power for that draw.

FAQ

What is the difference between HGX and DGX?

HGX is NVIDIA's 8-GPU baseboard platform that OEMs build servers around. DGX is NVIDIA's own complete system on the same baseboard, with the CPU, networking, software and support fixed by NVIDIA.

Is DGX faster than HGX?

There is no inherent speed difference, since both use the same GPUs and the same NVLink. Sustained performance depends on the cooling, networking and cluster fabric of the specific system.

Who makes HGX servers?

OEMs and system builders make them, including Dell, HPE, Lenovo, Supermicro and GIGABYTE. NVIDIA's certified-systems program lists the models it has tested.

Can you run a public GPU cloud on DGX?

Yes, and some providers have: Bitdeer AI says it launched its GPU Cloud on DGX SuperPOD with Hopper GPUs. NVIDIA designs DGX SuperPOD for single-tenant enterprise use and doesn't support multi-tenant or public-cloud reconfiguration of SuperPOD as a product, though component support continues. Cisco's NCP-compliant cloud design uses HGX servers.

Conclusion

The GPUs are the same in both systems. The choice comes down to who sets the rest of the system and how the cluster is meant to be used. For an operator selling compute to many customers, Cisco's published NCP-compliant design points to HGX.

Bring the equipment schedule, the site and the offtake contract to a financing conversation. GPU financing through USD.AI covers HGX servers and DGX systems alike, for new deployments and to refinance GPUs already in service.

Every financing through USD.AI is scored by Compute IQ, which assesses the operator, the customer's credit and the offtake contract.

Talk to USD.AI

Editor Notes

• Server-types page. The equipment-schedule fields come from usd.ai/hardware/server-types, which isn't linked because it isn't in the Internal links sitemap.
• Borrower name. The borrower behind the September 2026 B300 financing is withheld under the Borrower names rule unless USD.AI has published it on usd.ai.
• Per-loan rate. The rate on the September 2026 financing is withheld under the brief and the Loan terms rule.
• Deal term. The outline asked for the three-year term of the September 2026 financing; it is withheld because the brief bars tenor anywhere in this article.
• Loan terms. Standard terms (loan-to-value, equity, rates, tenor, reserve, minimum size) are excluded per the reviewer's decision.
• DGX eligibility. Samuel confirmed on Nov 8, 2026 that USD.AI finances either system; the conclusion now says so.
• Server price. The optional Q4 2026 figure of about $600,000 for a complete 8-GPU B300 server is omitted. Next to the $54.4 million, 130-server deal it would let readers infer an advance rate.
• October 2026 financing. The October 2026 financing of 1,728 HGX B300 GPUs in Finland isn't used. The reviewer chose the September deal, and the borrower name and term would be withheld anyway.
• B300 memory per GPU. Per-GPU B300 memory is excluded because NVIDIA, Dell and Supermicro publish different figures (270 GB versus 288 GB). OEM totals of 2.3 TB are excluded for the same reason.
• DGX price and value. No DGX price is given, and the article makes no claim that either system finances better or holds value better.
• NVL72 article. The scope line points readers to USD.AI's NVL72 article; link it once published (Internal links rule: unpublished posts get an editor note).