AI Servers: The 2026 Data Study

Almost every high-end AI server is built around NVIDIA’s HGX 8-GPU SXM baseboard: eight GPUs on one carrier, wired all-to-all by NVLink and NVSwitch rather than PCIe, so they behave like one large accelerator.^[src] SXM is what enables that mesh; cheaper PCIe add-in cards are used for inference and smaller deployments but can only bridge GPUs in pairs.^[src] Around the GPUs sits a dual-socket CPU host with 1–2 TB of RAM,^[src] and a high-speed east-west fabric — InfiniBand (deterministic latency, in-network reductions) or 400/800GbE Ethernet (broader ecosystem) — to scale training across many nodes.

NVLink bandwidth is the defining metric of the platform, and it has doubled with Blackwell:

Workload/infrastructure split per industry analysis of training vs inference server design.

Every figure is from the manufacturer’s datasheet. These are static hardware specs (no pricing), so they don’t go stale.

Servnet supplies and configures the Dell, HPE and Lenovo platforms above — including the 8-GPU Lenovo ThinkSystem SR780a V3. Build one to your spec in the Lenovo configurator, or explore the ThinkSystem range.

The single biggest practical shock for buyers is power density. A traditional enterprise rack is provisioned for perhaps 6–15 kW. A single 8-GPU AI node can consume that on its own, and a rack-scale Blackwell system is an order of magnitude beyond it — which is why NVIDIA’s Blackwell generation is designed around direct-liquid cooling.^[src]

Practical read: many Hopper (H100/H200) servers still run air-cooled and fit high-density enterprise or colo racks; Blackwell largely does not. Plan power, cooling and rack space before the hardware.

Acquisition prices for AI servers and data-centre GPUs are almost entirely quote-gated and volatile, so we don’t publish a headline “street price” — anyone who does is guessing. What is knowable is the shape of the decision and the current cloud rental rates.

The durable rule: rent for short, bursty or uncertain workloads; own (or colocate) when utilisation is high and sustained over years. At continuous use, an H100 rents for roughly the price of buying one within a year — so the more steadily you run GPUs, the more owning pays. The exact break-even depends on your utilisation, GPU choice and current cloud pricing — model your own case in the AI GPU calculator (it carries its own cited cloud-vs-own comparison), and finance the capex over 2–5 years with the IT finance calculator if you buy.

Figures from named analysts (McKinsey, Gartner). Enterprise “AI repatriation” to on-prem/colo is a real trend but is largely evidenced by vendor-sponsored surveys, so we cite only the named-analyst anchors here and treat the sponsored numbers as directional.

Hardware specifications are taken directly from manufacturer datasheets (NVIDIA, AMD) and OEM product pages (Dell, HPE, Lenovo, Supermicro, NVIDIA), and were cross-checked in an adversarial fact-check pass — during which one figure (the H100 PCIe NVLink bridge) was corrected to 900 GB/s. Market figures are attributed to named analysts (McKinsey, Gartner). Cloud prices are dated and labelled indicative because they change frequently; we publish no hardware “street prices” because those are quote-gated and volatile. Last updated 5 July 2026.