For four decades, air cooling was the data center’s unsung workhorse. Racks drew 2 to 3 kilowatts in the 1980s, rising to 5 to 8 as servers became denser through the 2000s. Better fans, airflow containment and hot/cold aisle architecture extended its run. It was cheap, familiar and sufficient.
GPUs changed the equation. Nvidia’s A100, released in 2020, drew 400 watts per chip. The H100 pushed that to 700 watts. The B200 hits 1,000 watts, and the GB200 NVL72 rack pulls 120 to 130 kW total. Air cooling, optimized for 8 to 12 kW racks, has no viable answer for that kind of density.
Liquid cooling does. A January 2026 Dell’Oro Group report found the market nearly doubled in 2025, approaching $3 billion, and is forecast to reach $7 billion by 2029. A November 2025 S&P Global Market Intelligence 451 Research survey found only 45% of data centers now run purely on air cooling, down from 48% in 2024, with 59% planning to implement liquid cooling within five years.
The question is no longer whether to adopt liquid cooling, but which form and how fast.
The physics of the problem
The failure of air cooling at AI rack densities comes down to a basic material property.
Air resists heat transfer. That is not a design flaw; it is the same property that makes double-pane windows insulate a room. Water conducts heat roughly 25 times better than air at rest. In motion, the gap is still larger. A fan pushing air across a hot chip surface pulls heat away at around 50 watts per square meter per degree. Coolant flowing through the copper microchannels of a cold plate pulls heat away at around 15,000 watts per square meter per degree, roughly 300 times faster. Think of the difference between waving a magazine over a hot pan and dropping the magazine into cold water. Both move heat. Only one keeps up.
Average rack power density has more than doubled in two years, from 8 kW to 17 kW, and is projected to reach 30 kW by 2027, according to an October 2024 McKinsey report, with AI training racks already well ahead of that average.
Those limits show up in GPU clock speed. H100 GPUs under inadequate air cooling can throttle to a fraction of their rated clock speed within seconds of a sustained training run. In distributed jobs across thousands of GPUs, one throttled chip can stall the entire run. The DOE estimates cooling accounts for up to 40% of data center energy use.
JLL research establishes three density thresholds:
- Up to ~20 kW per rack: air cooling is adequate
- Up to ~100 kW: rear-door heat exchangers extend viability
- Above ~175 kW: immersion cooling is required
Direct-to-chip cooling fills the middle band, handling densities between ~100 and ~175 kW where rear-door exchangers fall short and immersion is not yet warranted.
Hot water changes the economics
Mechanical chillers are one of the biggest energy draws in any liquid-cooled data center, and until recently there were an unavoidable cost of liquid cooling. Nvidia’s Vera Rubin processor is changing that.
At CES in January 2026, Jensen Huang announced that Vera Rubin supports liquid cooling at 45 degrees Celsius, high enough for data centers to reject heat through dry coolers using ambient air rather than mechanical chillers. Nvidia’s CES press release confirmed Rubin is in full production, with customer availability in the second half of 2026. According to Nvida’s product specifications, the Vera Rubin NVL72 uses warm-water, single-phase direct liquid cooling at a 45°C supply temperature, allowing data centers to reject heat through dry coolers using ambient air rather than energy-intensive chiller systems.
Lenovo’s Neptune platform predates the current market interest in high-temperature liquid cooling by more than a decade. Lenovo’s technical documentation confirms the company has delivered direct water-cooling solutions since 2012 under the Neptune brand, routing heat to roof-mounted radiators that dissipate it passively with no evaporative cooling required.
Lenovo’s product datasheets note that in most climates, water-side economizers can supply water below 45°C for most of the year, bypassing chillers, which Lenovo identifies as the most significant energy consumer in the data center.
Running hotter carries operational tradeoffs. At supply temperatures approaching 45 degrees Celsius, ambient air in the data hall may require conditioning for staff, and not all ancillary equipment is rated for the same thermal range.
3 technologies, 3 sets of tradeoffs
Liquid cooling is not a single technology. Three approaches dominate, each suited to different deployment contexts and risk profiles.
Direct-to-chip cold plate cooling
Of the three, direct-to-chip has the clearest near-term path. Direct-to-chip (DTC) cold plate cooling commands roughly 47% of the AI data center liquid cooling segment, according to Future Market Insights’ AI Datacenter Liquid Cooling Market report, and is the technology Nvidia specifies for its GB200 compute nodes.
Cold plates with copper or aluminum microchannels mount directly to GPUs and CPUs. Nvidia’s GB200 multi-node tuning guide confirms the architecture: compute processors and network switch ASICs are liquid-cooled, while remaining components (storage, power distribution, ancillary networking) are still air-cooled. Real-world GB200 deployments are hybrid by design, not fully liquid.
Unlike immersion, DTC does not require a facility rebuild. It works with existing chilled water infrastructure, deploys rack by rack and needs no purpose-built tanks or modified server form factors.
Single-phase immersion cooling
Where DTC is partial, single-phase immersion is total. Entire servers are submerged in tanks of dielectric fluid that remain liquid throughout, capturing close to 100% of IT heat with no fans. PUE ranges from 1.02 to 1.10.
Chip manufacturer certification had been one of the missing pieces for enterprise immersion adoption. In May 2025, Shell became the first immersion fluid provider to receive official certification from a major chip manufacturer.
Shell’s press release confirmed Intel’s endorsement of its fluids for 4th and 5th generation Xeon processors, with Intel providing a warranty rider for immersion-cooled chips. An Intel customer spotlight documenting the joint validation puts the electricity consumption reduction at up to 48%.
The certification does not change the deployment profile. Immersion requires purpose-built tanks, modified fanless servers and dedicated floor space, making it a fit for greenfield builds rather than retrofits.
2-phase immersion cooling
Two-phase immersion leads on efficiency, but the cost and regulatory picture is harder. Two-phase delivers a PUE of 1.01 to 1.03, the best of any approach, and supports rack densities of 150 to 250-plus kW. LiquidStack offers tanks rated to 252 kW with a passive circulation loop that requires no pumps.
The efficiency gains are real. The fluid costs are prohibitive. Two-phase systems rely on fluorocarbon fluids containing PFAS (per- and polyfluoroalkyl substances, a class of synthetic chemicals that persist indefinitely in the environment and are facing tightening restrictions in both the EU and US). 3M has exited production of Novec (its branded line of fluorinated immersion fluids) and all PFAS-based products entirely.
Fluorocarbon fluids cost multiples of the hydrocarbon fluids used in single-phase systems, and the gap compounds at rack scale. A May 2025 Microsoft lifecycle assessment published in Nature found that optimized cold plates or single-phase immersion can match two-phase efficiency without the PFAS exposure.
Foundry
The infrastructure decisions are already made
The largest AI infrastructure operators are not evaluating liquid cooling. They have deployed it, standardized it, and structured future builds around it.
Google has run liquid cooling across more than 2,000 TPU pod deployments at gigawatt scale for seven years, achieving twice the chip density of equivalent air-cooled configurations, according to Google’s Cloud Blog.
In August 2024, Microsoft moved all new data center designs to closed-loop, zero-water-evaporation liquid cooling, saving more than 125 million liters of water per facility per year, per the Microsoft Cloud Blog.
Meta committed $800 million to a liquid-cooled AI data center in Indiana and debuted a 140 kW liquid-cooled rack at OCP Global Summit 2024. Air cooling served the industry well for four decades. The operators building the next four decades of AI infrastructure have already moved on.