Eight GPUs in 4U.
Built for your server room, not a hyperscale facility.
RM-4U8G is a purpose-built integrated-loop liquid-cooled AI server platform. Eight GPU slots in a four-rack-unit envelope, the cooling loop closed inside the chassis, and the rest of the build - motherboard, CPUs, RAM, storage - assembled from off-the-shelf enterprise parts so the system has no vendor lock-in. Designed in Slovenia by the EK engineering team. Engineered to deploy in standard server-room conditions, not just industrial datacentres.
4U
Rack envelope
8
GPU slots
Liquid
Integrated cooling
Standard
Server-room ambient
Four design decisions that make this server different.
01
4U for 8 GPUs
The physical density most other 8-GPU servers need 6U or 8U to achieve. The savings compound across a rack - twice the deployment density per rack unit, with no compromise on thermal headroom.
02
Off-the-shelf internals · no motherboard lock-in
Standard EATX server motherboards from your vendor of choice. CPU sockets, RAM, NVMe, NICs are all standard parts. Three years from now you replace the motherboard, not the server. Most 8-GPU servers don't allow that.
03
Integrated liquid cooling - no CDU required
The radiator and pump live inside the chassis. There is no external coolant distribution unit, no rear-door heat exchanger, no facility plumbing required. The server runs in a standard server-room rack the day it lands.
04
Higher ambient tolerance
Liquid extracts heat from die to radiator with order-of-magnitude better efficiency than air. The chassis tolerates the elevated ambient temperatures that standard server rooms - not industrial datacentres - actually run at.
A category the market wasn't serving.
The AI server market sells one of three things. Hyperscale-grade datacentre racks - proprietary motherboards, custom power, external coolant distribution units, multi-six-figure entry, built for facilities that already have rear-door heat exchangers and 30 kW+ per rack. Air-cooled mid-market boxes - vendor-locked or consumer-grade cooling, 4–6 GPUs in any sane thermal envelope, throttling in standard server rooms during summer. DIY workstations on rack rails - what most "edge AI" deployments actually look like, and where most of the things that go wrong, go wrong.
RM-4U8G sits in a fourth bucket the market doesn't currently have a name for. A purpose-built integrated-loop liquid-cooled 8-GPU server using off-the-shelf internals. Every part of that sentence is doing work. The chassis, the power distribution, and the cooling are engineered as a single system around standard enterprise parts - so you get the density and thermal envelope of a hyperscale rack with the procurement and serviceability profile of a workstation build.
The customer this server is for is the enterprise IT team or research lab that needs production AI capacity inside their existing facilities - not in a hyperscale datacentre, not on someone else's cloud. That deployment shape barely existed in the vendor catalogue until now. We built the platform that fits it.
Twenty-five years of liquid cooling, transferred to enterprise rack scale.
The cooling system in this chassis - and the components throughout - is the work of the EK engineering team. EK has been the de-facto reference for high-performance liquid cooling since the early 2000s. Materials discipline, pressure envelope, manifold geometry, O-ring chemistry - the principles that made EK the standard among workstation builders are the same principles that govern the cooling loop inside RM-4U8G.
The difference is the chassis those principles now sit in. The transfer from prosumer and workstation cooling to enterprise rack cooling is not a brand stretch - it's the same engineering discipline, applied at the scale where modern AI workloads actually run.
// EK engineering legacy
2003 - first-generation copper water blocks
2010s - workstation cooling reference
2020s - sister brand LM TEK incorporated
today - enterprise rack-scale platforms
- same engineering team, same materials discipline, same published pressure envelope
A barebone you spec, not a SKU you buy.
RM-4U8G ships as a barebone - chassis, integrated cooling loop, power distribution, and platform support. The motherboard, the CPUs, the GPUs, the RAM, the storage are sourced separately and chosen per deployment. Same chassis, same cooling loop, same power topology - across three orders of magnitude of compute capacity.
| Tier | Configuration | Use case |
|---|---|---|
| Modest start | 2× RTX 6000 Ada · 1× AMD EPYC | Workstation crossover · departmental AI |
| Mid-tier | 4× RTX A6000 · 2× AMD EPYC | Lab / research group |
| Production maximum | 8× H200 NVL · 2× Xeon | Enterprise inference / fine-tuning |
Understanding scales - how to size your first deployment without throwing early spend away.
The three tiers above map directly onto Stages 1–4 of the staged-investment model. A starter deployment with two RTX 6000 Ada cards is the same chassis as a maxed-out 8× H200 system; you upgrade by replacing GPUs and motherboard, not the server.
Reference configurations · examples drawn from real customer builds. Final BOM is specified per deployment - LM TEK supplies the platform; the GPUs, CPUs, memory, and storage are sourced separately and can vary.
Chassis
4U · standard rack
GPU support
Up to 8 · full-height
Cooling
Integrated liquid loop
Specifying a deployment?
Tell us what you are trying to do, what data is in scope, and what scale you are starting at. We will come back with a configuration recommendation and - if helpful - an introduction to a system integrator partner who can take it from spec to deployed system.