AI GPU Cooling Revolution: Deionized Water, Ethylene Glycol & Propylene Glycol – The Ultimate Liquid Cooling Guide
Table of Contents
What you will learn
📋 What You'll Learn
This guide walks you through ai gpu cooling revolution: deionized water, ethylene glycol & propylene glycol – the ultimate liquid cooling guide with detailed instructions.
The chemistry keeping AI alive
Every ChatGPT query, every Midjourney image, every autonomous vehicle model runs on GPUs that would melt without glycol-based liquid cooling. Here is exactly how it works.
Data centers have always needed cooling. But the AI boom changed the math. A traditional server rack draws about 12 kW of power. An AI training rack packed with NVIDIA H100 or B200 GPUs draws 85 kW or more — and every watt becomes heat that must be removed. Air conditioning alone cannot keep up. The industry is turning to liquid cooling, and at the heart of every liquid cooling loop is a familiar chemical: glycol.
Why AI Datacenters Need Liquid Cooling
Photo by Đào Hiếu on Unsplash
A single NVIDIA H100 GPU dissipates roughly 700 watts of thermal energy. Pack eight of them into a server, add networking and storage, and you have a rack generating heat equivalent to dozens of space heaters running simultaneously in a space the size of a refrigerator.
Air cooling hits a physics wall around 30-40 kW per rack. Beyond that, you need liquid. And with AI rack densities projected to reach 200-250 kW per rack in the next few years, liquid cooling is not optional — it is the only viable path.
Cooling consumes 25-40% of total datacenter energy. The U.S. Department of Energy allocated $40 million in 2025 for innovative cooling technologies. Liquid cooling adoption is growing at 25-40% annually — roughly double the overall datacenter growth rate. Currently, 19% of datacenters use liquid cooling, with another 36% planning to adopt it within 24 months.
How Liquid Cooling Works in Datacenters
There are several approaches, and most of them involve glycol at some point in the loop.
Direct-to-Chip (D2C)
Cold plates mounted directly on GPUs/CPUs. Glycol-water mix flows through micro-channels in the plate, absorbing heat at the source. The dominant method for AI racks. AMD and NVIDIA both recommend D2C for their current AI accelerators.
Rear-Door Heat Exchangers
A liquid-cooled coil replaces the rear door of a server rack. Hot exhaust air passes through the coil, transferring heat to a glycol loop. Good for retrofitting existing air-cooled facilities without re-engineering the server hardware.
CRAH Units + Glycol Chillers
Computer Room Air Handlers push chilled air through the data hall. The chilled water or glycol loop connects to roof-mounted dry coolers or central chillers. The traditional approach, still dominant for moderate-density facilities.
Coolant Distribution Units (CDUs)
The bridge between server-level liquid cooling and facility-level heat rejection. CDUs manage the glycol secondary loop, controlling temperature, pressure, and flow rate. They reject heat to cooling towers, chillers, or ambient air.
Where Glycol Fits in the Cooling Loop
Photo by Arno Senoner on Unsplash
A typical datacenter liquid cooling system has two loops:
- Primary (facility) loop — Connects the building's chillers or cooling towers to the CDUs. Typically runs a glycol-water solution to prevent freezing in outdoor piping and rooftop equipment.
- Secondary (server) loop — Circulates coolant from the CDUs through cold plates on the GPUs. This loop typically uses a propylene glycol-water mix or deionized water with corrosion inhibitors, depending on facility requirements.
Glycol serves three purposes in these loops:
- Freeze protection — Outdoor piping, rooftop dry coolers, and any exposed runs need protection to sub-zero temperatures
- Corrosion inhibition — Inhibited glycol protects copper, aluminum, and steel components from electrochemical degradation over years of continuous operation
- Biocide function — Glycol solutions resist microbial growth that can foul heat exchangers and clog micro-channels in cold plates
Propylene Glycol vs Ethylene Glycol for Datacenters
This is the central decision for any datacenter cooling engineer. Both work. The trade-offs are real.
| Property | Propylene Glycol (PG) | Ethylene Glycol (EG) |
|---|---|---|
| Toxicity | Low (FDA GRAS) | Toxic if ingested |
| Heat Transfer | Good (slightly lower) | Superior (10-15% better) |
| Viscosity | Higher (more pump energy) | Lower (easier to pump) |
| Freeze Protect @ 50% | -28 °F (-33 °C) | -34 °F (-37 °C) |
| Environmental Spill Risk | Low — biodegradable | Higher — regulated disposal |
| Datacenter Usage | Secondary (server) loops, colocation | Primary (facility) loops, outdoor piping |
| Regulatory | No special handling required | Requires spill containment in many jurisdictions |
Propylene glycol is gaining share in datacenter cooling. Colocation providers and hyperscalers increasingly prefer PG for server-level loops because a leak inside a million-dollar GPU rack is far less problematic when the coolant is non-toxic. EG remains dominant for outdoor facility loops where thermal performance matters most and human contact risk is minimal.
Concentration Selection for Datacenter Systems
Most datacenter glycol loops run at 30-50% concentration. The right number depends on the coldest temperature any part of the loop will see.
| Scenario | Recommended | Why |
|---|---|---|
| Indoor-only loops (no outdoor exposure) | 30% PG or DI water | Minimal freeze risk; maximize heat transfer |
| Rooftop dry coolers, mild climate | 30% PG or EG | Protection to +5 °F (PG) or -5 °F (EG) |
| Rooftop coolers, cold climate | 50% PG or 50% EG | Protection to -28 °F (PG) or -34 °F (EG) |
| Outdoor CDU in extreme cold | 60/40 EG | Max freeze protection to -52 °F |
| Semiconductor fab cleanroom | EG Semiconductor Grade | Ultra-pure; no trace metal contamination |
Going above 60% glycol provides almost no additional freeze protection and significantly degrades heat transfer performance. In a datacenter where every degree of cooling efficiency affects PUE and energy costs, this matters. Match the concentration to your actual minimum temperature exposure — no more.
Inhibited vs Uninhibited: Why It Matters for Datacenter Loops
Datacenter cooling loops run 24/7/365, often for 10-15 years before a major system overhaul. Over that time, uninhibited glycol degrades into glycolic acid, which attacks copper heat exchangers, aluminum cold plates, and steel piping.
Inhibited glycol contains a package of corrosion inhibitors — typically a blend of organic acid technology (OAT) compounds — that neutralize this degradation. The inhibitors form a protective film on metal surfaces, extending fluid life from 2-3 years to 8-10+ years.
For datacenters, inhibited glycol is non-negotiable. The cost of premature heat exchanger failure or cold plate fouling in a rack of AI GPUs vastly exceeds the price difference between inhibited and uninhibited glycol.
Glycol Maintenance and Monitoring
Installing glycol is not a set-and-forget operation. Datacenter facility teams should:
- Test glycol annually — Check freeze point (refractometer), pH (should stay 7.5-9.0), and inhibitor reserve. Degrading inhibitors mean accelerating corrosion.
- Monitor for dilution — Makeup water additions reduce glycol concentration over time. Track concentration and top up as needed.
- Watch for contamination — Particulates, biological growth, or cross-contamination from other loop chemicals can foul micro-channel cold plates.
- Replace on schedule — Even inhibited glycol has a finite life. Plan for complete fluid replacement every 5-10 years depending on test results.
Datacenter Glycol Products from Alliance Chemical
We supply glycol in every grade, concentration, and volume datacenter operators need — from 1-gallon jugs for testing to 55-gallon drums and totes for full system fills.
Propylene Glycol — For Server and Secondary Loops
Inhibited
Inhibited
Inhibited
Inhibited
USP
ACS
Ethylene Glycol — For Facility and Outdoor Loops
Pre-mixed
60/40
30/70
Inhibited
ACS
Semiconductor
ACS + Inh
Premium
Pre-mixed
Cooling an AI datacenter?
Our team supplies glycol to datacenter operators across the country. Every order ships with full SDS documentation and Certificate of Analysis. Bulk pricing available for drum and tote quantities.
Contact SalesFrequently Asked Questions
What type of glycol do AI datacenters use for cooling?
Most AI datacenters use a combination of propylene glycol (PG) and ethylene glycol (EG). PG is preferred for server-level secondary loops because it is non-toxic, making leak scenarios less damaging to equipment. EG is used in facility-level primary loops and outdoor piping where superior heat transfer performance matters most.
Why do AI GPUs need liquid cooling instead of air cooling?
A single AI GPU like the NVIDIA H100 generates about 700 watts of heat. A full AI training rack can draw 85 kW or more. Air cooling becomes physically insufficient above 30-40 kW per rack. Liquid cooling using glycol-water solutions can handle these heat loads because liquids transfer heat approximately 25 times more efficiently than air.
What glycol concentration is used in datacenter cooling?
Most datacenter glycol loops run at 30-50% concentration. Indoor-only loops may use 30% or even deionized water. Systems with outdoor-exposed piping in cold climates typically use 50% glycol for freeze protection to approximately minus 30 degrees Fahrenheit. Going above 60% is not recommended as it degrades heat transfer with minimal additional freeze protection.
Why use inhibited glycol in datacenters?
Datacenter cooling loops run continuously for 10-15 years. Without corrosion inhibitors, glycol degrades into glycolic acid that attacks copper heat exchangers, aluminum cold plates, and steel piping. Inhibited glycol contains organic acid technology compounds that protect metal surfaces, extending fluid life from 2-3 years to 8-10 years.
Can you use propylene glycol in a datacenter chiller loop?
Yes. Propylene glycol works in datacenter chiller loops, though it has slightly lower heat transfer efficiency (10-15% less) and higher viscosity than ethylene glycol. Many colocation providers and hyperscalers prefer PG for server-level loops because a coolant leak on a million-dollar GPU rack is far less damaging when the fluid is non-toxic.
