ASRock Rack Blackwell lineup: 4U8X-EGS2 and DLC HGX systems.

Most of the attention on Blackwell systems goes to the largest OEMs and their flagship rack-scale designs. ASRock Rack occupies a different and, for a certain buyer, more useful position. It builds the same NVIDIA HGX B200 8-GPU platform as the majors, but packages it in a compact 4U chassis, offers an air and a direct-liquid-cooled variant of the same design, and leaves the component decisions open in a way that suits integrators and white-label resellers rather than locking everything to one vendor configuration. For a cost-sensitive buyer sourcing dense HGX nodes, that combination of small footprint, cooling optionality, and a lower price point is the whole appeal.

This is an OEM spotlight, not a partnership announcement. ASRock Rack is a qualified member of the OEM ecosystem Rillor lists and can source against, the same way Rillor lists Supermicro, Gigabyte, Dell, HPE, and Lenovo ISG systems. The point of profiling them here is practical: if you are weighing where to place a forward order for HGX B200 or HGX B300 capacity, ASRock Rack's lineup is a real option, and understanding what it actually is helps you price it on the marketplace.

The 4U8X-EGS2 and its DLC sibling

The center of ASRock Rack's Blackwell lineup is the 4U8X-EGS2, an HGX B200 8-GPU server, and its direct-liquid-cooled counterpart, the 4U8X-EGS2/DLC SYN B200. Both house an NVIDIA HGX B200 8-GPU baseboard with NVIDIA NVSwitch, which is the part that matters most: the baseboard and its NVLink topology are NVIDIA's reference design, so the GPU subsystem is identical to what every other qualified OEM ships. What differs between vendors is the chassis, the cooling, the CPU and memory tree around the baseboard, and the I/O.

The headline number is the form factor. The 4U8X-EGS2/DLC measures 870 by 446 by 175 mm (34.3 by 17.6 by 6.9 inches), which is a genuine 4U enclosure holding eight 1000W Blackwell GPUs. For comparison, several competing HGX B200 air-cooled designs land at 6U or 8U. ASRock Rack's own air-cooled HGX B200 platform, the 8U8X-GNR2, is an 8U chassis built for sites without liquid-cooling infrastructure. Getting the same eight-GPU baseboard into 4U is what makes the EGS2 line dense, and density per rack unit is exactly what a cost-sensitive buyer is optimizing when colocation space is the constraint.

The DLC variant is not a different machine. It is the same 4U8X-EGS2 design with direct-to-chip liquid cooling, which is how you keep eight 1000W GPUs and dual high-core-count CPUs thermally honest in 4U. Choosing between them is a facility decision, not a performance one. If your colo can accept direct-liquid cooling, the DLC variant gives you the tightest thermal envelope and the most headroom for sustained load. If it cannot, the air-cooled path exists at a larger U-height. We walk through that decision in detail in air versus direct-liquid cooling for Blackwell systems, and it is worth reading before you commit a delivery month, because cooling drives lead time as much as it drives unit cost.

What the HGX B200 baseboard actually delivers

Because the baseboard is NVIDIA's reference design, the GPU-subsystem specs are fixed regardless of which OEM wraps a chassis around it. Each HGX B200 GPU carries 180 GB of HBM3e at roughly 7.7 TB/s of memory bandwidth, with fifth-generation NVLink providing 1.8 TB/s of GPU-to-GPU interconnect, at 1000W per GPU. Across the 8-GPU baseboard that is 1.44 TB of HBM3e in a single coherent NVLink domain. That is the asset a Rillor forward references when you trade RIL-GX-B200-2T: a complete OEM HGX B200 system, not a loose tray of accelerators.

The numbers that distinguish one OEM's HGX B200 from another's sit around the baseboard, not on it. That is where ASRock Rack's design philosophy shows.

Component-level flexibility is the real product

ASRock Rack's positioning is built for buyers who integrate. Rather than shipping one fixed configuration, the platform leaves the surrounding components open, which is precisely what an integrator or a white-label reseller wants when they are building to a customer spec rather than buying a finished SKU.

On the CPU side, the 4U8X-EGS2/DLC supports dual Socket E (LGA 4677) 5th and 4th Gen Intel Xeon Scalable processors. ASRock Rack's broader HGX B200 platform extends that to Intel Xeon 6 (the 6700P, 6500P, and 6700E lines), with 8-channel memory in a 2DPC arrangement that yields 16 DDR5 DIMMs per CPU and 32 DIMMs total. The same Blackwell platform family also exists in an AMD EPYC Turin variant: ASRock Rack's MGX-based 4UXGM-GNR2 (Intel) and 4UXGM-TURIN2 (AMD EPYC Turin) servers demonstrate that the company carries both CPU paths across its Blackwell designs rather than committing to one vendor. For a head-node CPU choice, that flexibility matters more than it looks, and we compare the two paths directly in Granite Rapids versus EPYC Turin for GPU server head nodes.

Memory and storage follow the same open pattern. The 4U8X-EGS2/DLC carries 16+16 DDR5 RDIMM / RDIMM-3DS slots in a 2DPC layout, populated with parts like Samsung DDR5-5600 RDIMM. Storage is eight hot-swap 2.5-inch PCIe 5.0 x4 NVMe bays, which pair naturally with a U.2 Gen5 drive such as the Micron 9550 PRO. The 9550 PRO reaches up to 14.0 GB/s sequential read and 10.0 GB/s sequential write in capacities up to 30.72 TB, which is the kind of local-scratch bandwidth a training node wants for checkpoint and dataset staging. Power is handled by 4+4 80 PLUS Titanium 3000W CRPS supplies, giving the redundancy and efficiency a production GPU node needs.

The point of leaving these open is not novelty for its own sake. A buyer who already has a standard for NVMe vendor, a memory qualification, and a preferred head-node CPU can build to that standard instead of accepting a tier-1 OEM's bundled choices and paying for the parts of the package they do not need. That is where the lower price point comes from. You are buying engineering and a chassis, and supplying the integration yourself or through a partner.

Networking and the B200-to-B300 path

The fabric story follows the standard NVIDIA progression, which keeps an ASRock Rack node interoperable with whatever cluster you drop it into. On HGX B200 baseboards, the platform uses NVIDIA ConnectX-7 NICs, with the 400G NDR generation as the common rear-I/O choice and an optional BlueField-3 DPU for offload and isolation. ASRock Rack's launch lineup spanned ConnectX-7 at 200 and 400 Gb/s, so the NIC tier scales with what your fabric needs. The InfiniBand side terminates into NVIDIA Quantum-2 (the QM9700 generation), which is the standard IB fabric for HGX B200 clusters.

The forward path to Blackwell Ultra runs through the same vendor. ASRock Rack has shown an NVIDIA HGX B300 server with liquid-cooled designs, and the HGX B300 generation moves the NIC to the NVIDIA ConnectX-8 SuperNIC. ConnectX-8 is the meaningful jump: it delivers 800 Gb/s networking and is the first SuperNIC to integrate a PCIe Gen6-capable switch (48 lanes of PCIe Gen6), the configuration NVIDIA uses across HGX B300 and GB300 NVL72 systems. So the practical networking rule for an ASRock Rack buyer is straightforward. A B200 build pairs with ConnectX-7 at 400G NDR over Quantum-2. A B300 build steps up to ConnectX-8 at 800G. If you want the full fabric comparison before you decide, ConnectX-7, ConnectX-8, and NVLink5 fabric explained lays out exactly what changes between the two generations and why it matters for cluster scale.

The same ConnectX-8 platform also appears in ASRock Rack's MGX-based 4UXGM line, which is built on the NVIDIA MGX PCIe Switch Board with the ConnectX-8 SuperNIC and ships in both Intel and AMD EPYC Turin variants. That is a different topology from the HGX baseboard designs, aimed at PCIe-attached Blackwell configurations, but it reinforces the same theme: ASRock Rack carries the current NVIDIA fabric generation across multiple chassis and both CPU vendors.

What genuinely differentiates ASRock Rack

Set against the majors, ASRock Rack is not trying to be the OEM that bundles the deepest services or the broadest rack-scale catalog. Its differentiation is narrower and, for the right buyer, sharper.

The first axis is density and footprint. Fitting a full HGX B200 8-GPU baseboard into a 4U chassis, with a DLC variant of the same design, is a real engineering result. For a buyer measuring cost per usable rack unit in a constrained colocation footprint, a compact 4U node changes the economics of the whole row.

The second axis is integrator and white-label friendliness. By leaving CPU, memory, storage, and NIC choices open across Intel Xeon 6 and AMD EPYC Turin, ASRock Rack lets system integrators and resellers build to a customer specification without paying for a fixed bundle. That open posture is the reason a price-sensitive integrator reaches for ASRock Rack over a tier-1 vendor whose value is in the services the integrator is already providing themselves.

The third axis is cooling optionality on a single design. Offering air and direct-liquid-cooled versions of the same chassis means a buyer is not forced to redesign their rack to match the cooling their facility can support. They pick the variant that fits the site and keep everything else constant.

What ASRock Rack does not bundle as deeply as the largest OEMs is the global services-and-support wrap and the broadest rack-scale GB200 and GB300 NVL72 portfolio. For a buyer who wants a turnkey, fully serviced rack-scale deployment, a tier-1 OEM may be the better fit, and Rillor lists those too. For a buyer who integrates their own nodes and wants dense, flexible, well-cooled HGX systems at a keener price, ASRock Rack is exactly the kind of qualified ecosystem member that makes the forward book deep. This is also why single-vendor concentration is a risk worth managing across a buildout, a subject we cover in avoiding single-OEM lock-in across a multi-rack GPU buildout.

How an ASRock Rack node maps to a Rillor forward

The reason vendor breadth helps you as a buyer is that a Rillor SKU references the standardized NVIDIA platform, not a single OEM's part number. That keeps the forward contract liquid across suppliers.

An ASRock Rack 4U8X-EGS2 or 4U8X-EGS2/DLC, carrying the HGX B200 8-GPU baseboard, maps to RIL-GX-B200-2T. An ASRock Rack HGX B300 build maps to RIL-GX-B300-2T. Whether the underlying box ships from ASRock Rack, Supermicro, Gigabyte, Dell, HPE, or Lenovo ISG, the contract is the same standardized forward: a 10 percent deposit at execution with the balance due at delivery, an independent escrow agent holding the deposit, a seller performance bond backing delivery, NVIDIA channel compliance enforced inside the contract with the end customer of record captured, and the ability to transfer the contract to another KYC'd buyer before delivery with Rillor and OEM approval. Plausible indicative pricing for a complete air-cooled HGX B200 8-GPU node sits in the low-to-mid $400,000s, with the DLC variant and a B300 build carrying their own forward curves; the live indicative figure for any SKU is what you would pull from a quote.

That standardization is the whole point. You are not negotiating a bespoke supply agreement with each vendor's channel team in isolation. You are trading a contract whose terms are fixed, against a published reference price, with a delivery month you choose. The difference between that and a one-off purchase order is the difference we lay out in standardized forwards versus bespoke supply agreements. You can see the full set of contracts on the SKU catalog, and if you are sourcing dense nodes specifically, the for-buyers overview explains how the contract flow works end to end.

The reference price that makes this comparison possible is the Rillor Compute Index, a 30-day rolling-blend forward price per SKU computed from active Rillor contracts. It is owned and controlled by Rillor and licensed as a settlement feed and API to exchanges, funds, and researchers, which is why a quote on an ASRock Rack node can be benchmarked against a neutral number rather than taken on faith.