Physical rack hardware to buy and own

GPU Servers Across Five PCIe Racks & Two Frontier Systems

The seven-system rack and frontier range provides dedicated GPU capacity rather than hourly GPU hosting. Credible sustained operation depends on the accelerator plane, host platform, airflow, power and management path.

Inspect five PCIe racks Model power and cooling

Three-quarter supplier render of a 4U OEM multi-GPU rack server — OEM platform reference render showing the rack-server form and external service access.

Five PCIe racks · two frontier systems

Seven rack and frontier systems. Three infrastructure classes.

A rack GPU server is a physical infrastructure purchase. GPU density matters, but so do PCIe topology, per-card memory, cooling, power delivery, remote management and the room around the machine.

01
Accelerator plane

How many GPUs, with what per-card VRAM and lane allocation?
Exact cards, slots, lanes and execution pattern recorded
02
Host platform

Will CPU, ECC memory, storage and network feed the workload?
Named bill of materials and workload path
03
Thermal path

Can the chassis and room remove heat under sustained load?
Burn-in conditions, temperatures and site checks
04
Power path

Are circuits, protection, connectors and UPS policy suitable?
Electrician or facilities confirmation where required
05
Service path

Can the system be monitored, recovered and physically maintained?
Management access, spares and handover ownership

Complete rack and frontier range

Five PCIe rack systems and two tightly coupled frontier routes.

PCIe rack 01 / 05 Value Rack 128 Four 32GB workers in a remotely managed 4U platform. PCIe rack 02 / 05 Value Rack 256 Eight 32GB workers for independent endpoints or queued work. PCIe rack 03 / 05 Enterprise 384 Four 96GB workers where per-card memory is the constraint. PCIe rack 04 / 05 Enterprise 768 Eight 96GB workers at the top of the RTX PRO PCIe range. PCIe rack 05 / 05 H200 1.1TB Eight H200 NVL GPUs in a high-memory PCIe reference system. Frontier 01 / 02 Frontier Native 2.3TB Eight B300 GPUs on a tightly coupled HGX NVLink/NVSwitch platform. Frontier 02 / 02 Frontier Rack 20TB A liquid-cooled GB300 NVL72 rack for facility-scale programmes.

Open 4U OEM GPU server chassis showing passive GPUs, cooling fans, processors and memory slots — OEM supplier render showing one possible internal layout. Components vary with the ordered build. OEM supplier reference image.

Hot-swap GPU server cooling fan module with a pull handle and protective grille — Supplier component image showing the removable cooling fan design used for service access. OEM supplier reference image.

Diagram of cool air entering a rack server, heat leaving it and checks for the circuit, room and meter — Power, airflow and room conditions form one site-readiness decision rather than three separate afterthoughts. GPU Servers technical illustration.

Specification notes

Hardware decisions that remain visible after the GPU headline.

Each note helps compare the practical fit of the available hardware, beyond the headline GPU specification.

01 / 04

What makes a GPU server different?

A serious multi-GPU server combines a server CPU platform, ECC system memory, appropriate PCIe lanes, high-airflow cooling, remote management, storage and power designed for sustained load. It is not simply a mining chassis with newer cards.

The reference platforms support multiple GPU workers and, where the software and model support it, tensor or pipeline-parallel inference. The execution pattern must be tested rather than inferred from aggregate VRAM.

4U rack form factor with front-to-rear airflow
AMD EPYC server platforms and ECC memory
10GbE baseline and IPMI remote management
Hot-swap, redundant power on the larger platforms

02 / 04

GPU memory is the first sizing conversation

Two 24GB GPUs are not automatically the same as one 48GB GPU. Four 32GB GPUs are not automatically one 128GB GPU.

Per-GPU VRAM constrains what can run on a single card. Aggregate VRAM matters only when the runtime can divide a workload effectively or when several independent services use separate cards.

Context length, KV cache, batch size and concurrent requests also consume memory. That is why a model name alone is not enough to size a server.

03 / 04

Power, heat and facility scale are part of the specification

The catalogue spans user-adjacent towers, forced-air PCIe racks, an approximately 14kW HGX system and a full-rack GB300 route documented at up to 142kW. Actual consumption depends on the chosen system and workload. A site pre-flight must confirm circuits, rack, cooling, room conditions, network, noise tolerance and UPS policy.

If the server cannot be housed responsibly, a tower workstation, colocation or cloud service is the better answer.

04 / 04

Configured to order around the workload

Model, GPU, memory, storage, networking and site needs are selected together, then presented as one complete configuration.

The customer receives the agreed system rather than a near-enough build hidden behind a generic product title.

Questions answered

Straight answers to common questions

Do you rent GPU servers?

GPU Servers supplies physical hardware for the customer to own. Optional marketplace hosting is a separate, opt-in assessment for genuinely spare capacity and is never guaranteed.

Can a GPU server sit in an ordinary office?

Some smaller workstations may, but the rack range has meaningful heat and noise. Value Rack 128 and Enterprise 384 should be treated as server-room or colocation equipment unless an acoustic and thermal solution is designed.

Why not reuse an old mining server?

Mining systems often have limited CPU, RAM and PCIe bandwidth. They can still be useful for independent workers, rendering or batch tasks, but should not be presented as modern coherent multi-GPU LLM platforms without evidence.

Continue the decision

Useful next steps

Resolve the hardware question

Match the chassis and room to the workload.

Tell us about the workload and facility so we can prepare the specification, price and test plan.

Inspect five PCIe racks Model power and cooling

Decision check

GPU Server: Operating Fit & Evidence

When evaluating the GPU server, start with the workload, operating boundary, evidence and credible alternative.

Any recommendation for this decision should record what is known, what still needs testing and when a different route is more appropriate. Include rack installation and multi-GPU operation where those factors change the decision.

Accelerator plane

Host platform

Thermal path

Power path

Service path