HPE Blackwell lineup: ProLiant Compute XD685 and GB300 systems.

Most OEMs arrived at Blackwell-class thermal density by extending enterprise server lines. HPE arrived from the other direction. Its supercomputing portfolio is built on the industry's first 100 percent fanless direct-liquid-cooling architecture, the same Cray lineage that powers exascale systems, and that engineering now anchors how HPE delivers Blackwell. For buyers sizing a multi-megawatt training hall, where the cooling plant is the real constraint and not the GPU count, that pedigree is the reason to put HPE on the shortlist.

This is a vendor spotlight, not a partnership announcement. HPE is one of the qualified OEMs whose systems Rillor lists and can source on forward contract. We profile the ProLiant Compute XD685 (HGX B200 and B300, air and direct-liquid-cooled) and the NVIDIA GB300 NVL72 by HPE, walk the CPU, memory, and fabric pairings a buyer actually has to choose between, and map both to the Rillor SKUs you would reference on a quote.

ProLiant Compute XD685: one chassis, the full accelerator menu

The XD685 is HPE's purpose-built 8-way GPU node. The point of interest for a buyer is breadth: a single platform that covers the Blackwell generation top to bottom and reaches across to AMD. Per HPE's store specs, the XD685 supports eight NVIDIA Blackwell Ultra (B300 HGX) GPUs, eight NVIDIA Blackwell (B200) GPUs, eight NVIDIA H200 Tensor Core GPUs, or eight AMD Instinct MI355X accelerators. That means a single approved platform and one operational model whether you are standing up B200 capacity now, stepping up to B300 for the denser FP4 attention workloads, or running a mixed estate that includes the previous Hopper generation. If you are still weighing the generational step, our breakdowns of B200 versus B300 at the system level and where the MI355X fits cover that decision in depth.

The chassis comes in two thermal builds, and the choice is not cosmetic. The XD685 ships in a 6U form factor for air cooling and a 5U form factor for direct liquid cooling, with both options offered across the GPU configurations. The DLC variant is the one to read closely if your hall is power-dense. Pulling a U out of the rack while removing the per-node fan wall is exactly the kind of trade that matters when you are budgeting cooling power against IT power. The general air-versus-DLC tradeoff is its own subject, covered in air versus direct-liquid cooling for Blackwell.

CPU, memory, and the head-node decision

The XD685 pairs dual AMD EPYC 9005 Series processors (Turin) with 24 DDR5-6400 RDIMMs, running 12 channels per CPU at up to 6400 MT/s with one DIMM per channel and ECC throughout. That is a high-bandwidth, high-core-count host designed to keep eight Blackwell GPUs fed rather than to bottleneck them. The HGX baseboard carries HBM3e across the GPU stack; the host DRAM is conventional DDR5. EPYC Turin is a sensible default for an 8-way GPU box because of memory channel count and PCIe lane budget, though Intel Xeon 6 (Granite Rapids) remains the alternative head-node CPU across the broader Blackwell field. We compare the two directly in Granite Rapids versus EPYC Turin for GPU server head nodes, and the HBM3e side is covered in HBM3e capacity and bandwidth across the Blackwell line.

Fabric: ConnectX and Quantum-2 InfiniBand

An 8-way node is only useful at scale if its east-west fabric keeps pace. The XD685 is built around the standard NVIDIA networking stack: ConnectX-7 (400G NDR) or ConnectX-8 (800G) adapters at the node, with Quantum-2 QM9700 InfiniBand switching (64 ports of 400Gb/s NDR in 1U) or Spectrum-X SN5600 Ethernet at the spine. The fabric choice, rail-optimized InfiniBand versus Spectrum-X Ethernet, is the single largest determinant of cluster scaling efficiency past a few hundred GPUs. If you are designing the spine, start with ConnectX-7, ConnectX-8, and NVLink5 explained and the Quantum-X800 versus Spectrum-X comparison.

NVIDIA GB300 NVL72 by HPE: rack-scale with Cray cooling

Where the XD685 is a node, the GB300 NVL72 by HPE is a rack. It integrates 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs interconnected with fifth-generation NVLink, forming a 72-GPU NVLink domain that behaves as a single massive Blackwell Ultra GPU for models above one trillion parameters. The NVLink domain is the whole reason this form factor exists: at rack scale the interconnect is no longer a network you tune around, it is shared memory fabric. Per NVIDIA's specs, the platform delivers 130 TB/s of total NVLink bandwidth (1.8 TB/s per GPU), 20 TB of GPU memory, 1.5x more dense FP4 Tensor Core FLOPS, and 2x higher attention performance than standard Blackwell, for up to a 50x AI-factory output increase over Hopper.

The HPE-specific contribution sits in the cooling plant. The GB300 NVL72 by HPE incorporates HPE direct-liquid-cooling technologies, capturing roughly 90 percent of rack heat to liquid and 10 percent to air. That matters because the rack runs at 132 kW nominal thermal design power with a peak electrical design power around 155 kW. At those numbers, the cooling architecture is not an accessory; it is the system. This is where the Cray heritage earns its place: HPE's supercomputing portfolio is built on the industry's first 100 percent fanless DLC architecture, which reduces cooling power use by up to 90 percent, drawing on the same exascale engineering lineage. For the AI fabric, GB300 NVL72 deployments use Quantum-X800 InfiniBand or Spectrum-X Ethernet with ConnectX-8 SuperNICs at 800 Gb/s per GPU.

If you are choosing between rack-scale generations, our GB200 NVL72 versus GB300 NVL72 piece and the power and thermal budgets per rack primer pair well with this section.

What actually differentiates HPE

Every qualified OEM in the field can integrate the same NVIDIA HGX baseboards and the same NVLink racks. The differentiation lives in the parts NVIDIA does not supply. For HPE that is three things.

The first is thermal engineering. The fanless DLC architecture is not a marketing line; it is a measured reduction in cooling power use of up to 90 percent, and it comes from a supercomputing organization that has been cooling exascale machines for years. For a buyer whose buildout is power-bound rather than capital-bound, that is the headline number.

The second is supercomputing integration. The XD685 and GB300 NVL72 by HPE do not arrive as loose nodes. They arrive inside a portfolio that knows how to wire, cool, and operate a coherent machine at hall scale, with the management and interconnect practices that exascale work demands. That reduces the integration risk that otherwise lands on the buyer's own engineering team.

The third is global support through HPE Pointnext, which matters when a fleet has to stay up across regions and the SLA is the difference between a training run finishing and a quarter slipping.

None of this changes the underlying GPU. It changes the reliability, the power bill, and the operational burden around it, which for a serious buyer is most of the total cost.

How the HPE systems map to Rillor SKUs

Rillor standardizes contracts on complete OEM GPU systems, so the SKU you reference is the system, not the bare accelerator. The HPE Blackwell lineup maps cleanly:

• ProLiant Compute XD685 with eight B200 GPUs maps to RIL-GX-B200-2T.
• ProLiant Compute XD685 with eight B300 (Blackwell Ultra) GPUs maps to RIL-GX-B300-2T.
• NVIDIA GB300 NVL72 by HPE maps to RIL-NVL72-GB300.

You can browse the full SKU catalog at /skus and see live indicative levels on the marketplace. A contract on any of these is a bilateral OTC forward with intent of physical delivery, never cash-settled: 10 percent deposit at execution, balance at delivery, an independent escrow agent holding funds, a seller performance bond standing behind the seller's obligation, and NVIDIA channel compliance with the end-customer-of-record captured throughout. Pre-delivery transfer to another KYC'd buyer is permitted with Rillor and OEM approval. The all-in take rate is 2 percent (1 percent buyer, 1 percent seller) plus roughly 1,000 dollars in escrow per contract. If you are new to the mechanics, start with what KYC and onboarding look like for a verified buyer and the case for why serious buyers need a forward market, not a waitlist.

Indicative pricing gives a sense of scale rather than a quote. An XD685 B300 node sits in the high six figures depending on configuration and delivery window, and a GB300 NVL72 by HPE rack runs into the multiple millions once the full cooling and fabric build is counted. Both move with the forward curve that forms from real contracts rather than from a list price, which is the point of pricing them on a forward market in the first place. For the buyer-side mechanics, see /for-buyers.

The same contract data feeds the Rillor Compute Index, a 30-day rolling-blend forward price per SKU, computed from active Rillor contracts and owned and controlled by Rillor. Exchanges, funds, and researchers license it as a settlement feed and API, and third-party venues build cash-settled products against it. Rillor never operates those venues and never cash-settles its own contracts. That separation, physical-delivery forwards on one side and a licensed reference rate on the other, is covered at /for-markets.

Where the XD685 and GB300 fit a buildout

The practical read for a buyer is straightforward. If you are scaling out flexible 8-way capacity, want one approved platform across B200, B300, and the prior generation, and value the option to add AMD Instinct without changing your operational model, the XD685 is the node to spec, and the 5U DLC variant is the one to spec if your hall is power-dense. If you are building a coherent trillion-parameter training machine and the binding constraint is the cooling plant rather than the silicon, the GB300 NVL72 by HPE puts Cray-grade thermal engineering directly under the NVLink domain. Both belong on the shortlist for any buyer who treats cooling and integration as first-class engineering problems rather than afterthoughts.

To put indicative forward levels and a delivery window against a real configuration, request a quote on the XD685 B300 SKU below.