As AI and high-powered computing push rack loads to 30 kW or even 100 kW in advanced clusters, old-fashioned air cooling hits its wall at about 20 kW. Out-of-control temperatures don’t just waste energy and spike PUE; they also shorten chip lifespan, force slowdowns, and drain both performance and budgets. By using fluids with superior heat absorption and transfer, liquid cooling brings PUE down to 1.1–1.2, cuts energy usage by 20–30%, and is quickly becoming a must-have for modern data centers.
But what kinds of liquid cooling are out there? And why do they beat air solutions in power, cost and reliability? Let’s answer all that—plus practical roll-out advice and market outlook—in the following six sections for a complete guide to the era of liquid cooling.
“Liquid cooling” isn’t one thing—it’s a spectrum, from warm water doors to fully immersed “bathing” machines.
There are four major types: cold plate, single-phase immersion, two-phase immersion, and direct chip cooling. Cold plates lead, thanks to being tried-and-true—they’re used in 90% of projects.
Cold plates are the mainstay, yet immersion and jet cooling are key for extreme high-power scenarios.
Cold plate cooling has become a data center go-to for its crisis-proof stability, affordable price, and easy retrofitting.
Looking at a 30 kW rack, cold plate capex runs about 21,000 RMB/IT kW, and monthly opex is just 685 RMB/IT kW—lower than either air cooling or immersion.
Copper or aluminum cold plates, shaped with grooves or microchannels, deliver about 25 W/cm²·K. Since the coolant never actually touches the electronics, retrofitting is painless—just add a CDU and some piping. The ASHRAE Datacom Series, EU Code of Conduct, and US DOE all spell out water purity and temperature specs for cold plates. And in the real world? When Equinix switched its A100 GPUs from air to cold plate cooling, PUE fell from 1.6 to 1.15, saving 30% on power, and doubled per-rack density to 30 kW.
When GPUs start pumping out 800 W or even 1 kW apiece, even cold plates begin to sweat.
Single-phase immersion easily cools 1 kW GPUs. Two-phase immersion and jet cooling raise heat-handling limits even further (over 300 W/cm²·K).
Immersion cooling fully submerges servers in dielectric fluid, eliminating the “air gap” of traditional designs. Hypertec’s bottom-flow GPU racks paired with custom radiators efficiently cool 1000 W GPUs non-stop. Two-phase immersion takes advantage of phase change for big gains but faces hurdles over refrigerant costs and safety. Jet cooling, by actually spraying chips with a mist of coolant, offers the highest possible theoretical heat removal, but is only used in research owing to complexity. As it stands, these techs see limited use in AI R&D and defense, but are held back from wider adoption by price and lack of mature standards.
Every project lives and dies by three questions: What will it cost? How quick is payback? Will it really save?
For the typical 30 kW rack, cold plate TCO is just 1,057 RMB/IT kW·month—beating air cooling by 14%. Single-phase immersion, with pricier upfront cost, is about even with air in total cost.
Cold plate costs are mostly down to the plate, pipes, and CDU—about 60% of the total. Ongoing costs are from pumps and water treatment. Single-phase immersion gets expensive (31,000 RMB/IT kW) from its fluids and custom tanks. Air seems cheap, but running it with all the extra AC means liquid cooling pays for itself in two years. According to NVIDIA, switching every CPU server on the planet to GPUs and liquid cooling could save over a trillion kWh per year—that’s worth $100 billion.
Liquid cooling saves energy, but a single leak can undo it all.
Follow these seven iron rules—ASHRAE water quality, dual-loop CDN redundancy, quarterly pressure checks, stable coolants, standard racks, plan for 30–100 kW growth, and AI thermal control—for a risk-free system
ASHRAE advises water circuits under 100 µS/cm; Europe calls for 20–45°C coolant. Use EPDM/FKM seals for cold plates and run seasonal (1 bar) air tests; keep an eye on fluid levels and chemical stability in immersion systems. US DOE recommends heat exchange between facility and technical cooling loops. Hypertec’s smart plumbing cut required checkups from monthly to quarterly. Best case: link to a DCIM platform and let machine learning set pump speeds and valves, trimming PUE even further.
Sometimes, markets speak louder than specs.
By 2028, projections see cold plate cooling in 25% of CPU servers (135 billion RMB) and 55% of GPU servers (>500 billion RMB).
Expect over 36 million CPU servers (average 290 W) and 15 million GPU servers (average 1800 W) shipped worldwide. Cold plate liquid cooling costs just 3,400 RMB/kW, compared to 9,600 RMB for immersion. In China, major companies like Yingwei Tech and Inspur cover the entire value chain; overseas, CoolIT and Asetek dominate high-end deployments. Looking ahead: aluminum cold plates and plastic pipes cut costs 15%; nanofluids could push heat transfer higher, but still need to prove stability and cost. Microchannel cooling for chips could break 300 W/cm², but isn’t on the market yet. AI-powered controls and edge centers mean even more growth.
If a tech can’t deliver in live use, it’s just a slide deck.
NVIDIA’s liquid-cooled A100 and H100 GPUs have cut PUEs from 1.6 to 1.15, doubled compute density, and are rolling out en masse in Equinix data centers.
The A100 uses copper cold plates with microchannels to cool 70% of hotspots; the H100 brings closed-loop direct cooling for lower water refills. They support PCIe, SXM, and OAM, and work in ASUS and Foxconn racks—no need for special cabinets. At 30% energy savings, NVIDIA claims a global switch to liquid-cooled GPUs could save over a trillion kWh a year: as much as France uses in a year. Equinix pilots prove rack density can jump from 15 kW to 30 kW, with enough cooling headroom for 100 kW/rack AI setups.
In today’s ultra-dense, low PUE world, liquid cooling isn’t a luxury—it’s required. Cold plates are in 90% of racks for good reason; immersion and jet cooling are now breaking through for the most power-hungry apps. With robust safety, maintenance, and AI-powered controls, your facility will be ready for the next leap. By 2028, the liquid cooling market will top 600 billion RMB, with microchannels and nanofluids poised as the next big things. Know your seven implementation rules, four global standards, and the industry players—and you’ll lead the race for the coolest (and hottest!) compute.
Here are some essential liquid cooling terms to help make sense of the technology:
ES 
EN 
DE 
KO 