Gigabyte Blackwell lineup: G893, G894, and GB300 rack systems.

Gigabyte, through its Giga Computing server division, has built a reputation for shipping NVIDIA platforms early and pricing them aggressively. For a buyer comparing eight-way HGX chassis from a dozen qualified vendors, that combination matters. The boards are the same NVIDIA HGX baseboards regardless of who screws them into a chassis, so the differentiation lives in the chassis engineering, the component flexibility, the lead time, and the unit price. On all four, Gigabyte tends to land in the value column.

This is a spotlight on the part of Gigabyte's Blackwell lineup that Rillor lists and can source: the G894 8U HGX platforms for B200 and B300, the G893 8U HGX B200 platform, and the GIGAPOD and GB300 NVL72 rack-scale systems. Gigabyte is a member of NVIDIA's qualified Blackwell-platform partner ecosystem, named alongside ASRock Rack, ASUS, Pegatron, and the major OEMs. It is not a signed or announced Rillor partner, and nothing here implies one. It is a vendor whose hardware shows up in our marketplace because buyers ask for it by price.

Where Gigabyte fits in a Blackwell buy

Every eight-GPU Blackwell system on the market is built on one of two NVIDIA HGX baseboards. HGX B200 is an 8-GPU board carrying 1.4TB of total HBM3e GPU memory, with fifth-generation NVLink delivering 1.8TB/s of GPU-to-GPU bandwidth and 14.4TB/s aggregate. HGX B300, the Blackwell Ultra board, keeps the same NVLink fabric but lifts total GPU memory to 2.1TB. Those numbers are fixed by NVIDIA. No OEM changes them.

What an OEM does change is everything around the board: the chassis thermals, the CPU and memory options, the NVMe and PCIe topology, the power delivery, and the price. Gigabyte's pitch across its G894 and G893 lines is high-density chassis engineering paired with broad component flexibility, sold at a unit price that is usually at or below the field. For a buyer who has already decided on B200 or B300 and is now choosing a vendor, that is the entire decision.

For how the two boards diverge at the system level, see our breakdown of B200 versus B300 systems. For the fabric side, ConnectX-7, ConnectX-8, and NVLink5 explained covers the networking choices below.

The G894 line: 8U air-cooled flagships

The G894 is Gigabyte's 8U air-cooled HGX platform, and it is the workhorse of this profile. Two variants matter for Rillor sourcing.

G894-AD1-AAX5 (HGX B200, RIL-GX-B200-2T)

The G894-AD1-AAX5 is an 8U dual-processor HGX B200 server. It pairs the 8-GPU B200 baseboard with two Intel Xeon 6900-series CPUs (Granite Rapids AP, the platform's headline being the Xeon 6980P), 12-channel DDR5 across 24 DIMM slots accepting RDIMM or MRDIMM, and eight 2.5-inch Gen5 NVMe bays. Networking runs on NVIDIA ConnectX-7 NICs at 400G NDR with BlueField-3 DPU compatibility. Power is handled by a 6+6 array of 3000W 80 PLUS Titanium PSUs, which is what an air-cooled eight-way Blackwell box needs to stay redundant under load.

This is the system that maps to Rillor SKU RIL-GX-B200-2T. On the reference tape, a complete G894-class HGX B200 system indicates near 340,000 USD, moving with CPU choice, memory capacity, NVMe fill, and NIC count. You can pull the current curve for that SKU on the marketplace or request a contract quote directly.

G894-SD3-AAX7 (HGX B300, RIL-GX-B300-2T)

Launched in September 2025, the G894-SD3-AAX7 is the 8U air-cooled HGX B300 sibling, built on eight Blackwell Ultra GPUs. It carries the same 1.8TB/s GPU-to-GPU NVLink and 14.4TB/s total bandwidth as the B200 generation, with the memory step up to the B300 board's 2.1TB. Gigabyte rates the platform at roughly seven times the AI performance of a Hopper-class system.

The networking generation moves up with the board. The SD3-AAX7 supports NVIDIA ConnectX-8 SuperNICs, configurable as 8 by 800Gb/s OSFP InfiniBand or 16 by 400Gb/s Ethernet, with BlueField-3 DPU support. CPU pairing is dual Intel Xeon 6700 or 6500-series across 32 DDR5 DIMM slots, and power scales to twelve 3000W 80 PLUS Titanium redundant PSUs to feed the higher-draw Ultra GPUs in air.

This system maps to RIL-GX-B300-2T, with indicative reference pricing near 390,000 USD for a complete configuration. If B300 is your target, the SKU-specific B300 quote page carries the live forward.

The G893 line: a second 8U HGX B200 path

One correction worth stating plainly, because it circulates wrong. The G893-SD1-AAX5 is an 8U platform, not a 5U one. It runs the HGX B200 board with dual 4th or 5th Gen Intel Xeon Scalable processors (up to 350W each), up to 2TB of DDR5 across 32 DIMM slots, eight 2.5-inch Gen5 NVMe bays, and the same 6+6 3000W power topology as the G894-AD1.

If you have seen a 5U HGX B200 attributed to the G893 designation, that is a different chassis. Gigabyte's 5U HGX B200 platform is the G593 series, a separate product line. For procurement, the distinction is real: rack-unit count drives how many systems fit per rack and how your power and thermal budget per rack pencils out, which is its own planning exercise (see power and thermal budgets per rack for Blackwell). We carry the correct chassis spec on the SKU record so a contract never ships against the wrong rack-unit assumption.

The practical read: the G893-SD1 gives a buyer a second 8U HGX B200 option from the same vendor, with an older Xeon Scalable head node rather than Granite Rapids AP. That is useful when a buyer is standardizing head nodes against an existing fleet or wants the slightly lower-cost CPU pairing. On the EPYC side, Gigabyte also offers AMD EPYC Turin pairings across its GPU server families for buyers who run AMD head nodes; if you are weighing the two, Granite Rapids versus EPYC Turin for GPU server head nodes lays out the tradeoff.

Rack-scale: GIGAPOD and GB300 NVL72

Above the single chassis, Gigabyte fields GIGAPOD, its rack-scale reference solution, and the NVIDIA GB300 NVL72 rack. Shown in detail at GTC 2025, GIGAPOD comes in both air-cooled and direct-liquid-cooled variants. The liquid-cooled node uses CoolIT coldplates on the GPUs, the CPUs, and the NVLink Switch, and the design supports HGX H200, HGX B200, and HGX B300 boards, so a buyer can standardize a pod architecture across generations rather than re-architecting per refresh.

The GB300 NVL72 is the rack-scale Blackwell Ultra system, and it is a different animal from the eight-way boxes above. It is a fully liquid-cooled rack uniting 72 Blackwell Ultra GPUs and 36 Grace CPUs (2,592 Arm Neoverse V2 cores), with 20TB of HBM3e GPU memory, 17TB of LPDDR5X CPU memory, 130TB/s of NVLink bandwidth, and ConnectX-8 SuperNICs at 800Gb/s per GPU. NVIDIA rates it at up to a 50x leap in AI-factory output versus Hopper. Air is not an option at this density. The rack is direct-liquid-cooled by design.

For the buyer choosing between rack generations, GB200 NVL72 versus GB300 NVL72 at rack scale and air versus direct-liquid cooling for Blackwell are the two reads that decide the facility-side commitment.

Spec comparison across the Gigabyte Blackwell line

The table below is the at-a-glance view. Treat the prices as indicative reference-tape figures for complete systems, configuration dependent, not quotes.

On switching, the same buyer choice applies as with any Blackwell vendor. NVIDIA Quantum-2 QM9700 carries the InfiniBand path; Spectrum-X with the SN5600 carries the Ethernet path. Both pair cleanly with Gigabyte's NIC options. The decision is a cluster-architecture one, not a vendor one, and we cover it in Quantum-X800 versus Spectrum-X.

Component flexibility, accurately scoped

Part of why Gigabyte lands as a value option is the breadth of components it qualifies. Memory across the line is Samsung-class DDR5-5600 RDIMM, and Gigabyte also validates Micron DDR5-5600 RDIMM for buyers standardizing on that supply. Storage commonly fills the Gen5 NVMe bays with Micron 9550 PRO drives. CPU pairing spans Intel Xeon, from the 6900-series Granite Rapids AP on the newest G894 down to 4th and 5th Gen Xeon Scalable on the G893, and AMD EPYC Turin across the broader GPU server families. That flexibility is real procurement leverage. It lets a buyer match an existing fleet's head node and storage standard instead of accepting a fixed bill of materials.

What we will not do is dress this up as more than it is. Gigabyte holds NVIDIA Qualified status. That is a genuine listing, and a meaningful one, but it is distinct from the stricter NVIDIA-Certified Systems program. NVIDIA maintains a separate Qualified System Catalog precisely because the two designations are not the same. We carry the accurate status on every Gigabyte SKU record so a buyer's compliance and support expectations are set correctly before a contract is signed, not after.

How a Gigabyte system trades on Rillor

When you contract a Gigabyte HGX B200 system through Rillor, you are buying a standardized OTC forward on a complete OEM system with physical delivery. A 10 percent deposit goes into independent escrow at execution, the balance settles at delivery, and the seller posts a performance bond. The all-in take rate is 2 percent (1 percent buyer, 1 percent seller) plus roughly 1,000 USD per contract in escrow cost. NVIDIA channel compliance is handled inside the contract, with the end customer of record captured so the unit clears the channel cleanly.

The reason a buyer comes to a forward rather than chasing spot is timeline. A forward converts a guessing-game delivery window into a contracted delivery month, with deposits that bind both sides. If your facility readiness or your DLC coordination slips, you can transfer the contract pre-delivery to another KYC'd buyer with Rillor and OEM approval, which is its own step-by-step process. The vendor on the other side may be Gigabyte, Supermicro, Dell, or another qualified OEM. The contract structure does not change. That is the point of a standardized forward, and it is why a value-priced, broadly-available vendor like Gigabyte fits the model well. If you are weighing vendors, the broader buyer view and the SKU catalog lay out the full qualified set.

If you want the forward curve and a contract on a Gigabyte-class HGX B200 system, start here.