UK’s trusted IT infrastructure partner since 2003
Servnet
FinanceToolsConfiguratorGet in Touch
AI & GPU servers · 8-GPU NVIDIA HGX B200/B300 · UK supply & support

AI server buyer’s guide 2026 (UK)

The enterprise AI training node has settled on eight NVIDIA HGX GPUs in one chassis — but Dell, HPE, Supermicro and Lenovo wrap very different density, cooling, support and pricing around the same boards. This guide shows what actually separates them and how to choose.

4
mainstream platforms
8
GPUs per node
3–8U
density range
B200/B300
Blackwell boards
Size GPUs for your model →
The landscape

The four 8-GPU platforms, at a glance

Every mainstream enterprise AI server builds on NVIDIA’s HGX baseboard, so a given board (B200, B300 or H200) performs the same in any of them. What differs is the host system around it — and, above all, the support and tooling.

Flagship 8-GPU node — rack height by platformSYS-822GS-NBRTSupermicro8UXE9680Dell6UXD685HPE5USR675 V3Lenovo3ULower is denser per rack unit; taller chassis typically ease airflow and serviceability.
Rack height for each platform’s flagship air-cooled 8-GPU node. Lenovo packs eight GPUs into 3U; Supermicro’s air-cooled B200 node is 8U. Liquid-cooled variants differ.
Supermicro
Supermicro 8U B200/B300 SuperServer
Form
8U air-cooled (4U/2-OU liquid-cooled B300 also offered)
GPU
NVIDIA HGX B200 or B300, 8× SXM
CPU
Dual Intel Xeon 6
Memory
8 TB DDR5 (32 DIMM)

Maximum GPU density and configurability; B300 integrates ConnectX-8 on the baseboard.

Dell
Dell PowerEdge XE9680
Form
6U air-cooled (XE9680L liquid-cooled, XE9685L for B200/B300)
GPU
HGX H100/H200/H20, MI300X or Gaudi 3, 8×
CPU
Dual Intel Xeon (4th/5th Gen)
Memory
4 TB DDR5

The default for Dell-standardised fleets; broadest accelerator-board choice in one chassis.

XE9680 data sheet →
HPE
HPE ProLiant Compute XD685
Form
5U direct-liquid-cooled or 6U air
GPU
HGX H200, B200, B300 or AMD MI355X, 8×
CPU
Dual AMD EPYC 9005 (up to 96c)
Memory
3 TB DDR5 (6400 MT/s)

HPE’s Cray-derived DLC lineage; strong for liquid-cooled B200/B300 at rack scale.

XD685 data sheet →
Lenovo
Lenovo ThinkSystem SR675 V3 / SR680a V3
Form
3U (SR675 V3) or 8U B200 (SR680a V3)
GPU
8× double-wide, or 8× NVIDIA B200 (SR680a V3)
CPU
AMD EPYC 9004/9005
Memory
6 TB DDR5

The densest 3U eight-GPU option; Neptune liquid cooling across the V3 line.

SR675 V3 data sheet →
How to choose

What actually separates the platforms

The GPUs are a constant; the decision rides on density, cooling, host memory, tooling and support. These are the dimensions that move a real procurement.

AI-server platforms — how they compareDellHPESupermicroLenovoRack density (fewest U)Rack-scale liquid coolingAccelerator boards in one chassisGlobal next-business-day onsiteMature fleet toolingRack-scale / OCP density
Qualitative comparison across the factors that decide most AI-server procurements — ●=strong, ◔=partial, —=typically reseller/region-dependent.
Which 8-GPU platform fitsYou already standardise on Dell PowerEdgeDell XE9680 / XE9685LYou need eight GPUs in the least rack spaceLenovo SR675 V3 (3U)You want rack-scale direct liquid coolingHPE XD685Max configurability and GPU densitySupermicro B200/B300You want 8× NVIDIA B200 in a dense chassisLenovo SR680a V3
A starting point, not a rule — we size the exact board, host CPU, memory and cooling against your workload and data-centre constraints.
Compare by vendor

Cross-shop a specific platform

Already leaning one way? These guides benchmark each vendor against the field on cost, density, cooling and support.

Get an 8-GPU AI server shortlisted and quoted

Tell us the model you’re training, the GPU board and rough node count. We shortlist Dell, HPE, Supermicro and Lenovo against it, size the fabric and cooling, and quote new or refurbished.

GPU pooling calculator →
FAQ

Choosing an AI server: common questions

What is the best 8-GPU AI server in 2026?

There is no single best — the four mainstream platforms (Dell PowerEdge XE9680, HPE ProLiant Compute XD685, Supermicro HGX B200/B300 and Lenovo ThinkSystem SR675 V3 / SR680a V3) all build on the same NVIDIA HGX boards, so the GPUs match. The right one follows the fleet you run, the cooling you can deliver, the rack density you need and your support and pricing model.

How many GPUs are in a standard AI training server?

The mainstream enterprise training node carries eight SXM GPUs on one NVIDIA HGX baseboard — B200 or B300 (Blackwell) or H200 (Hopper) — linked by NVSwitch. Rack-scale reference designs such as NVIDIA GB200/GB300 NVL72 and Supermicro SuperClusters group many of these nodes into a single fabric.

Do I need liquid cooling for Blackwell AI servers?

Not always. Air-cooled 8-GPU B200 systems ship today (for example Dell XE9680, Supermicro 8U), but the densest Blackwell deployments move to direct liquid cooling — HPE’s Cray-derived ProLiant Compute XD and Supermicro’s liquid-cooled B300 lines are built for it. The deciding factors are rack power and your data centre’s cooling capability.

Should I buy AI servers outright or on consumption?

Both models exist. Dell, Lenovo and Supermicro are commonly bought as straight capex; HPE markets GreenLake pay-per-use heavily for AI. Owning and depreciating the hardware suits a steady training workload; consumption can suit bursty or uncertain demand. Model both before committing.

Can Servnet supply and support AI servers in the UK?

Yes. We configure, quote and support Dell PowerEdge XE, HPE ProLiant Compute XD, Supermicro HGX and Lenovo ThinkSystem GPU servers, size the GPU count, cluster fabric and cooling for your model, and advise on new-versus-refurbished and third-party maintenance.

Keep reading

AI-server guides, data sheets and tools

Sources

Talk to a UK specialist

Get expert advice or a no-obligation quote — servers, storage, networking, maintenance, finance and cloud. We reply the same working day.

or call 0800 987 4111