Mineshop now lists four Supermicro HGX 8-GPU AI servers you can order in the EU. Two are built around the NVIDIA HGX B200 platform and two around the newer HGX B300 (Blackwell Ultra) — each in an Intel Xeon and an AMD EPYC variant. Every machine is configured to order with a written B2B quote within one business day, tracked delivery to all 27 EU countries, and invoice payment with EU reverse-charge VAT for registered businesses.
The Supermicro lineup at a glance
| Model | GPU platform | CPU platform | GPU memory |
|---|---|---|---|
| SYS-A22GA-NBRT | HGX B200, 8× NVIDIA B200 | Dual Intel Xeon 6900P | 1.4 TB HBM3e |
| AS-A126GS-TNBR | HGX B200, 8× NVIDIA B200 | Dual AMD EPYC 9005 | 1.4 TB HBM3e |
| SYS-822GS-NB3RT | HGX B300, 8× NVIDIA B300 | Dual Intel Xeon 6700P | 2.3 TB HBM3e |
| AS-8126GS-NB3RT | HGX B300, 8× NVIDIA B300 | Dual AMD EPYC 9005 | 2.3 TB HBM3e |
Specifications are Supermicro's published figures as listed on each product page, checked 2 November 2026. Configurations are made to order, so final specs follow your quote.
HGX B200 pair: the 10U air-cooled workhorses
The SYS-A22GA-NBRT and AS-A126GS-TNBR both put eight NVIDIA B200 SXM GPUs on the HGX B200 board — 1.4 TB of HBM3e in total, with NVLink at 1.8 TB/s per GPU. Both ship in air-cooled 10U chassis with six 5250 W Titanium power supplies in a 3+3 redundant layout, which keeps the rack conversation simpler than liquid-cooled alternatives.
The Intel variant takes dual Xeon 6900P processors with up to 128 cores and 504 MB of cache per CPU, 24 DIMM slots for up to 6 TB of DDR5-6400 RDIMM or DDR5-8800 MRDIMM, ten hot-swap PCIe 5.0 NVMe U.2 drives plus two M.2, and ten PCIe 5.0 x16 low-profile slots — enough for a 1:1 GPU-to-NIC 400G fabric. The AMD variant pairs the same GPU complex with dual EPYC 9005 (or 9004) processors at up to 384 cores in total, eight NVMe bays with optional E1.S when BlueField-3 DPUs are fitted, and eight PCIe 5.0 x16 slots.
HGX B300 pair: Blackwell Ultra with an 800G fabric on board
The SYS-822GS-NB3RT and AS-8126GS-NB3RT step up to eight NVIDIA B300 GPUs (Blackwell Ultra) — 2.3 TB of HBM3e across the board — and add eight NVIDIA ConnectX-8 SuperNICs driving 8× 800G OSFP ports for a 1.6 TB/s AI fabric out of the chassis. Both are 8U machines with six 6600 W Titanium PSUs (3+3).
These two ship as Gold Series presets with an estimated delivery of 6–12 weeks. The AMD Gold preset is the most concretely specified: two EPYC 9575F processors (128 cores in total, 3.3 GHz base), 3 TB of DDR5-6400, and 61 TB of E1.S NVMe storage (8× 7.68 TB) plus dual 1.92 TB boot drives, with two dual-port 200G BlueField-3 options for the fabric. The Intel variant takes dual Xeon 6700P at up to 86 cores per CPU and scales to 8 TB of DDR5-6000 in a 2DPC layout.
Intel or AMD: how the platforms differ
GPU throughput is identical within each pair — the choice is about the CPU complex around it. The Xeon 6900P route offers the highest per-socket core counts and the largest cache, which suits hosts that run heavy CPU-side preprocessing alongside training. The EPYC 9005 route offers more total cores in the B200 pair (up to 384) and slightly different I/O layouts, including the E1.S plus BlueField-3 combination. Memory ceilings, NVMe bay counts and PCIe slot mixes differ model by model, so the practical answer is to match the host platform to the fabric and storage you already run.
B200 or B300?
The honest framing is memory and fabric, not marketing tiers. The B300 machines carry 2.3 TB of HBM3e versus 1.4 TB on the B200 machines — the difference between what fits in GPU memory for the largest models — and integrate the 800G ConnectX-8 fabric natively. The B200 machines remain the mainstream workhorse with a longer field history and, in these listings, more PCIe slot headroom on the Intel side. Neither choice is wrong; the workload decides. If you are unsure, describe your models and timeline in the quote request and let the specialist align the configuration.
Who these servers are for
These are 8U and 10U, multi-kilowatt, export-controlled systems for AI labs, research groups and companies building serious inference or training capacity in Europe. They are not desk machines. If your workload fits a single workstation — large-memory professional GPUs in a quiet tower — our Mineshop AI Workstation line remains the practical answer, and our earlier explainers on the Intel GPU D270 and AMD Instinct MI455X cover where the accelerator market is heading.
What happens after you request a quote
- You send configuration, quantity and timeline on any of the four product pages — or ask for advice if you are unsure.
- A hardware specialist checks availability and replies with a written B2B quote within one business day, excluding VAT.
- End-user details are confirmed — NVIDIA data-centre GPUs are export-controlled, so this step is required before an order can be accepted.
- Order confirmed within 24 hours of quote acceptance, with lead time stated in your quote (Gold Series presets: estimated 6–12 weeks).
- Tracked delivery to all 27 EU countries, with warranty and after-sales support, and EU reverse-charge VAT invoicing for VAT-registered businesses.
Frequently asked questions
Why is there no price on the page?
These servers are configured to order: GPU bin, CPU, memory, storage and fabric all move the number. A written quote within one business day is faster and more accurate than a stale list price.
Can I buy just one?
Yes — single-unit orders are quoted the same way. Quantities and timelines shape lead time, not eligibility.
Do you ship outside the EU?
The listings cover tracked delivery to all 27 EU countries; export control on the GPUs means end-user details are part of the process. For anything beyond that scope, ask in the quote request.
Comments (0)