Send us your configuration
A modern office door standing open onto a 1970s machine room of tape drives, consoles and a green-screen terminal

8-GPU AI servers · Air vs liquid

Air is reaching its limit.

The AI server on your quote is cooled the old way.

You've specced the GPUs, memory and storage for your AI workloads. The cooling is still the old answer. An air-cooled 8-GPU server sits close to the edge of what air can cool, and you don't yet know what you'll ask of it tomorrow.

LM TEK specialises in liquid cooling. The LM TEK 4U8GPU10KW is the alternative: eight GPUs in one 4U server, with no facility plumbing and no external CDU.

01 // It has happened before

Nobody asked for liquid-cooled engines.

Nobody looked forward to the pumps, the hoses and the radiator, or to changing how engines had always been built. And yet no modern car has an air-cooled engine. Not even the small, boxy city car that never drives further than the local supermarket.

The air-cooled flat-six in the engine bay of a classic Porsche 911, its large cooling fan and shroud at the front
Porsche's air-cooled flat-six: a cooling system with an engine attached. Retired nearly 30 years ago. Photo: Bill Abbott, CC BY-SA 2.0, via Wikimedia Commons. Resized.

02 // What you will ask of it tomorrow

Cars did not go liquid-cooled because they could not hold 2,500 rpm on a cold day.

They switched because ordinary driving includes the hard days too. Your server's hard days are just as ordinary, and you have not planned them yet.

An air-cooled 8-GPU server · what you ask of its cooling

Close to the edge. Still running. Nothing left for the next thing you ask of it.

Illustrative. Not a measurement of any server. Tick what tomorrow brings.

Re: GPU server slowing down under load
"Yes, but you added a rack unit and ran it for twelve hours."
You will not want to hear this as the reason your installed infrastructure is now the problem.

Buy the cooling for the days you have not planned.

The LM TEK 4U8GPU10KW takes the heat off each GPU with liquid, at the source, and sizes chassis, loop and power for eight GPUs on day one.

Before you sign, ask your vendor

"What have you tested with all eight GPUs at full load, at what room temperature, and for how long?"

A temperature class on the spec sheet is not an answer. It says the server will not fail. It does not say it will not slow down.

Put the same question to us.

100%

GPU servers are where car engines were in the 1990s. NVIDIA's next generation of rack systems is already 100% liquid-cooled.

"The Rubin generation of NVIDIA AI infrastructure is the world's first to achieve 100% liquid cooling"

Source: NVIDIA, "Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI's Biggest Machines", June 2026.

Loading the LM TEK 4U8GPU10KW Loading the 3D model
Illustrative visualisation. Flow paths and temperatures are a simplified model, not CFD or measurement.

03 // Making the move

You don't need a new data centre to make it.

The LM TEK 4U8GPU10KW brings liquid cooling to the PCIe GPUs you buy today: a closed loop inside the chassis, no facility plumbing, no external CDU.

Its cooling is designed by the EK engineering team, 25 years in high-performance liquid cooling. The chassis was drawn around the loop, not the other way round.

LM TEK 4U8GPU10KW

GPU slots
8 × full-height PCIe Gen 5
Form factor
4U · 19-inch rack
Power
5 × 2 kW with redundancy
Ambient
5 – 30 °C
Coolant
5 – 60 °C
Facility
No facility plumbing · no external CDU

What it does not do

It does not take the heat out of the room.

The heat leaves the chassis as warm air. Your room cooling still has to remove it.

It is not rated above 30 °C.

Rated for 5 – 30 °C ambient. Operation above 30 °C ambient is not supported.

Two pumps, no pump redundancy.

Both pumps are needed. If either fails, the system shuts down to protect the GPUs.

Switching this late

Three steps from the quote on your desk.

  1. 01 Send the configuration. The one you are about to order.
  2. 02 We map it. We show how it maps to the published compatibility list.
  3. 03 A partner quotes it. We connect you with a regional partner, who quotes and builds the server.

The LM TEK 4U8GPU10KW is a barebone platform. Motherboard, CPUs, GPUs, memory, storage and network adapters are sourced separately, so you choose them. A GPU outside the published compatibility list is a bespoke project with LM TEK engineering.

Before you sign

About to sign? Send us the configuration first.

We show how it maps to the published compatibility list and connect you with a regional partner to quote it. Our partners install servers every day.

Rather talk first? Leave the configuration field empty and we'll arrange a short call.

We use your details only to map your configuration and arrange a partner introduction. Privacy notice

Choose which categories LM TEK may set. Strictly necessary cookies keep the site working and cannot be switched off here. Your choice is stored for 12 months and you can change it at any time from the footer.

Cookie Policy