HGX vs DGX: What's the Difference for GPU Rental? (2026)

Every "HGX vs DGX" explainer online is written for someone with a six-figure capex budget deciding what to buy. If you're renting GPUs by the hour instead, that framing doesn't fit, because you were never choosing between the two. Almost every H100, H200, or B200 node you rent from a neocloud is built on an HGX baseboard. Actual DGX systems are boxes NVIDIA sells and ships to buyers who want to own the hardware, not inventory that shows up in a cloud provider's fleet.

That distinction changes what you should expect from a rental: how pricing is structured, what's fixed versus configurable in the node you get, and who you call when something breaks. This post covers the hardware difference and then answers the question that actually matters for renters: what's under the hood of the instance you're booking.

HGX vs DGX at a Glance

If you rent an H100 or an H200 today, you're renting an HGX-spec node. The rest of this post explains why, and what to check before you book one.

DGX: NVIDIA's Fully Integrated System Explained

DGX is NVIDIA's own hardware brand: a complete server that NVIDIA designs, builds, and ships as a single SKU. When you buy a DGX H100, you get eight H100 SXM5 GPUs, the CPUs, the NVMe storage, the networking cards, the cooling, and the chassis, all specified and integrated by NVIDIA, running NVIDIA's own DGX OS on top. There's one vendor, one support contract, and one invoice.

That's the appeal for an enterprise buyer: you're not sourcing components, validating a BOM, or debugging a compatibility issue between a motherboard vendor and a NIC vendor. You call NVIDIA. A complete DGX H100 system typically runs $373,000 to $450,000 depending on configuration and support tier, with additional annual support contracts adding $10,000 to $50,000 on top (source). That price buys the hardware, the software stack, and NVIDIA's direct support relationship, not just eight GPUs.

What's Fixed vs What NVIDIA Controls

Everything in a DGX box is NVIDIA's decision, not the buyer's:

• GPU count and configuration - eight SXM GPUs per system, wired into NVIDIA's own NVSwitch fabric.
• CPU and RAM - a specific, non-negotiable pairing NVIDIA validated for that DGX generation.
• Storage and networking - NVIDIA picks the NVMe drives and the NICs (ConnectX-class InfiniBand or Ethernet).
• Software - DGX OS ships pre-installed, and NVIDIA AI Enterprise licensing is part of the package.
• Support - a direct contract with NVIDIA, not a third-party integrator.

You can't swap in a cheaper CPU or a different NIC to shave cost. That rigidity is the product: NVIDIA is selling a validated, supported system, not a parts list.

Don't Confuse This With DGX Cloud Lepton or DGX Spark

NVIDIA reuses the "DGX" name for two other products that have nothing to do with the enterprise DGX H100/B200 systems described above, and conflating them is a common mistake when researching this topic.

DGX Cloud Lepton is a marketplace and orchestration layer, not hardware. It lets you provision GPU capacity across multiple partner clouds through a single interface. The compute it routes you to is overwhelmingly HGX-spec hardware running at those partner data centers, the same as any other neocloud rental. If you're evaluating Lepton against a direct rental, see our breakdown of DGX Cloud Lepton alternatives.

DGX Spark is a small desktop AI computer NVIDIA launched for local development, built around a single Grace Blackwell superchip rather than eight data-center GPUs. It shares a name with the enterprise DGX line and nothing else about its architecture, pricing, or use case. We cover how it fits alongside cloud GPU workflows in our DGX Spark and GPU cloud pipeline guide.

HGX: The OEM Reference Platform Behind Most Cloud GPU Nodes

HGX is NVIDIA's 8-GPU baseboard and NVLink/NVSwitch interconnect fabric, sold to OEMs rather than end customers. NVIDIA builds and validates the baseboard itself, GPU placement, power delivery, and the NVSwitch wiring that gives all eight GPUs high-bandwidth, all-to-all connectivity. Everything around that baseboard, the CPU, RAM, storage, NICs, chassis, and cooling, is the OEM's choice.

This split emerged during the shift from the Pascal (P100) generation to Volta (V100). Before that, OEMs assembled individual GPUs into servers themselves. NVIDIA standardized the entire 8x SXM GPU platform into a pre-validated baseboard OEMs could buy as a unit and build around, without risking a botched thermal or electrical integration (source). That single change is why an HGX H100 server from Supermicro and one from Dell can run the identical GPU and NVLink configuration while differing entirely in CPU, storage, and price.

Who Builds on HGX

NVIDIA licenses the HGX reference design to a specific set of server OEMs, who then compete on everything outside the baseboard:

• Dell - PowerEdge XE-series servers built on HGX baseboards.
• HPE - Cray and ProLiant lines with HGX configurations.
• Supermicro - one of the highest-volume HGX server builders, widely used by neoclouds.
• Lenovo - ThinkSystem servers with HGX options.
• Gigabyte - HGX-based servers popular with GPU cloud operators for cost efficiency.
• Inspur - a major HGX OEM in Asian markets.

A cloud provider doesn't have to buy a finished server from one of these OEMs either. Some, including large neoclouds, build directly to NVIDIA's HGX reference design themselves and skip the OEM markup entirely.

HGX GPU Configurations Across Generations

NVIDIA's current HGX product page lists the Blackwell (B200), Blackwell Ultra (B300), and Rubin-generation NVL8 baseboards, and every one of them is an 8-GPU configuration (source). The Hopper-generation HGX H100 and HGX H200 boards, which are the two configurations powering most of the H100 and H200 capacity you'd actually rent today, followed the same 8-GPU baseboard structure; NVIDIA's page has simply moved on to marketing the newer generations. What changes generation to generation is the GPU silicon and memory, not the 8-GPU baseboard structure:

If you're weighing SXM against other form factors for a specific generation, we've broken that down separately for H100 NVL vs SXM5 vs PCIe and H200 NVL vs SXM5. For what comes after single-baseboard HGX nodes, rack-scale systems like GB300 NVL72 vs GB200 NVL72 link 72 GPUs into one NVLink domain instead of eight.

Why Almost Every Neocloud Rents You HGX Nodes, Not DGX

The economics only work one way. DGX bundles NVIDIA's fixed hardware, DGX OS, and a direct support contract into one six-figure purchase aimed at a buyer who wants a single vendor relationship and is willing to pay for it. A cloud provider renting GPUs by the hour to thousands of customers has the opposite requirements: they need to control cost per node, tune the CPU-to-GPU ratio and networking for their specific customer mix, and run their own orchestration and monitoring stack instead of NVIDIA's.

HGX fits that model. An operator buys or builds to the HGX reference design, pairs it with whatever CPU, storage, and NIC configuration fits their cost and workload targets, and runs their own software on top. CoreWeave's public HGX H100/H200 product listing is explicit about this: it markets the node by its HGX designation, not as a DGX-branded system, because that's what it is (source). Lambda's on-demand H100 instances are the same story: 8x H100 SXM at 80 GB per GPU, sold per GPU-hour off HGX hardware (source).

Jensen Huang described this split directly in a 2026 interview, describing how NVIDIA's own accelerated computing platform reaches the market through two different channels: "We integrate that platform into all the clouds and to all the OEMs, and then we have another platform that's now the data center platform, or the AI factory platform" (source). The HGX baseboard is what ships into "all the clouds and all the OEMs." DGX is the separate, NVIDIA-owned system business.

HGX vs DGX Pricing: Buying a DGX System vs Renting an HGX Node

These two numbers aren't measuring the same thing, but buyers frequently compare them as if they are, so it's worth being explicit about what each one includes.

Buying a DGX H100 outright costs roughly $373,000 to $450,000 for the complete 8-GPU system, support contract included (source). That's a capital purchase: you own the hardware, and the cost is fixed regardless of how many hours you actually run it.

Renting HGX-spec capacity is priced per GPU-hour and varies by provider and by how the node is packaged:

• Lambda Cloud prices its on-demand 8x H100 SXM HGX instances at $2.59/hr per GPU (source).
• CoreWeave lists its 8x H100 HGX node at roughly $49.28/hr total, which normalizes to about $6.16/hr per GPU (source). CoreWeave sells the full 8-GPU node as a unit; there's no single-GPU self-serve option at that tier.
• Spheron rents H100 SXM5 on-demand from $3.38/hr per GPU, by the single GPU or in multi-GPU configurations, with per-minute billing and no node minimum.

A DGX purchase only pencils out if you're running the hardware near-continuously for years and want to own it outright. For most teams running variable training or inference workloads, renting HGX-spec capacity avoids the upfront capital and lets you scale the GPU count up or down with demand. Pricing fluctuates based on GPU availability. The prices above are based on 5 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

What You're Actually Renting When You Book an H100 or H200 Instance

When you book an H100 or an H200 GPU instance from Spheron or any other neocloud, you're renting an HGX-spec 8-GPU baseboard, whether the provider built it themselves or sourced it from an OEM like Supermicro or Dell. The GPU count, memory per GPU, and NVLink/NVSwitch topology are fixed by NVIDIA's HGX spec regardless of which provider you choose. What varies is everything around it: the CPU generation and core count, local NVMe storage, and critically, the networking fabric connecting nodes together.

That networking choice matters more than most renters realize. A DGX system ships with NVIDIA's own fixed networking configuration. An HGX-based rental doesn't, since the operator chose it, and it can be InfiniBand, RoCE-based Ethernet, or something else entirely depending on the provider. If you're running multi-node training where inter-node bandwidth is the bottleneck, that's a real difference between two nodes that both call themselves "H100 SXM5." Our GPU networking guide covering InfiniBand vs RoCE vs Spectrum-X walks through what to check. The NVLink and NVSwitch fabric inside the baseboard itself, which is fixed regardless of provider, is covered in our NVLink interconnect explainer. If you're new to how GPU cloud rental works at all before diving into hardware specifics, our GPU cloud primer is the place to start.

How to Check Whether a Provider's Node Is a Real HGX Reference Build

Before booking a multi-GPU node for anything latency-sensitive, verify a few things directly with the provider rather than assuming "8x H100" means the same thing everywhere:

1. Confirm NVSwitch topology, not just GPU count. An 8-GPU server without full NVSwitch fabric won't give you the same all-to-all bandwidth an HGX reference build does. Ask specifically.
2. Ask about the inter-node network. InfiniBand, RoCE, and standard Ethernet have materially different bandwidth and latency profiles for multi-node jobs.
3. Check GPU memory per unit against the generation's HGX spec. 80 GB for H100, 141 GB for H200, 192 GB for B200, 288 GB for B300. A listing that doesn't match should raise questions.
4. Ask what happens on hardware failure. DGX comes with NVIDIA's direct support contract built in. An HGX-based rental's support model is entirely down to the operator, so ask before you need it, not after.
5. Don't assume "DGX" in a listing means a literal DGX box. Some marketing loosely uses the term for HGX-class hardware. Read the actual spec sheet.

Whether you're scaling a training run or standing up inference capacity, what you need is HGX-spec GPU hardware with the networking and support to back it, not a six-figure DGX purchase.

Rent H100 GPU →