This series has spent thirteen essays documenting the water, power, land, environmental, financial, and legal dimensions of the data center boom. This final essay covers something more hopeful: the cooling technologies that could dramatically reduce the water and power footprint described in Essay #2 and Essay #3 already exist, are commercially available today, and are not being deployed at anywhere near the scale they could be. Understanding what’s technically possible is essential ammunition for any community asking a developer why their proposed facility isn’t using it.
As introduced in Essay #1, data centers cool their servers through air cooling, water cooling, or liquid immersion. This essay goes deeper on the real-world performance differences between these approaches, measured using the industry’s standard efficiency metric: Power Usage Effectiveness, or PUE — the ratio of total facility energy consumption to the energy actually used by computing equipment. A PUE of 1.0 would be perfectly efficient, with zero overhead for cooling and infrastructure; in practice, every facility falls somewhere above that.
The shift toward liquid cooling isn’t primarily driven by environmental concern — it’s being forced by simple physics. NVIDIA’s GPU roadmap shows power consumption doubling roughly every two years, reaching 1,500 watts per chip by 2026 — a density that traditional air cooling simply cannot keep pace with. Industry data confirms how fast this has moved: according to the Uptime Institute, average data center rack power density increased 38 percent between 2022 and 2024 alone, with the steepest growth in AI and hyperscale deployments — power densities that once maxed out around 15 kilowatts per rack are now pushing 80 to 120 kilowatts in AI clusters.
This matters directly for the community fights described throughout this series: liquid cooling adoption is accelerating not because operators have become more environmentally conscious, but because air cooling has hit a hard physical ceiling for the densities modern AI hardware requires. Liquid cooling penetration was estimated at only around 3 percent of deployments in 2021, and is projected to reach approximately 37 percent by 2026 — meaning a developer proposing a new AI-focused facility today has far less excuse than they did even three years ago for defaulting to the most water- and power-intensive cooling method available.
Liquid immersion cooling cuts cooling-related energy costs by roughly 40 percent and uses approximately 90 percent less water than traditional evaporative cooling approaches — a figure that should be central to every conversation about the water footprint issues covered in Essay #2. If a developer’s facility is proposed for a drought-prone region, that 90 percent reduction is not a marginal improvement; it can be the difference between a facility that strains a community’s water supply and one that doesn’t.
If liquid cooling is this much more efficient, the obvious question is why the industry hasn’t already converted entirely. The honest answer involves real, non-trivial barriers, not corporate indifference alone. Liquid cooling requires fundamentally different technical expertise than air cooling — most technicians are trained on fans and air handlers, while liquid systems demand specialized knowledge of hydraulic systems and fluid dynamics, and immersion cooling specifically requires technicians to physically lift servers out of dielectric fluid using special handling procedures. This workforce transition is genuinely difficult and creates real operational risk during the changeover period, including dependency on outside vendors that can create knowledge gaps.
Cost is the other major barrier, particularly for the most efficient option. Two-phase immersion delivers the best efficiency of any current approach, but the specialized dielectric fluid required is expensive enough that the economics remain difficult to justify outside of the highest-density AI deployments — meaning the very best available technology is currently reserved mostly for cutting-edge AI training clusters, not standard cloud infrastructure, even though the environmental case for broader adoption is strong.
It’s worth being direct about a complication that doesn’t get enough attention: switching to liquid cooling generally reduces water consumption dramatically, but the relationship to overall environmental impact is more nuanced than a simple win. Research shows cold plates and immersion cooling cut greenhouse gas emissions by 15 to 21 percent compared to air cooling — but Microsoft’s own researchers found that switching to 100 percent renewable energy could cut emissions by 85 to 90 percent regardless of which cooling technology is used. In other words, the choice of cooling technology matters most directly for water consumption; the choice of power source matters most for carbon emissions. A facility could adopt the most water-efficient cooling available and still run on a fossil-fuel-heavy grid, exactly the dynamic discussed in Essay #5.
“Organizations face an inflection point where incremental improvements to air cooling cannot match exponential growth in heat density. The decision made today locks in operational costs for the next decade.” — Industry analysis of AI-driven cooling requirements, March 2026
Given that this technology exists, is commercially mature, and is already used in roughly a third of new deployments industry-wide, a community facing a new proposal has every right to ask specific, technical questions rather than accepting a generic sustainability pledge.
The technology to build a dramatically less water-intensive, more power-efficient data center exists right now, is being adopted industry-wide at real scale, and is not some speculative future innovation. When a developer proposes a facility using older, more water-intensive cooling methods, that is a choice, not an inevitability — and it’s a choice communities have standing to question, especially given everything covered in the rest of this series: the water scarcity detailed in Essay #2, the grid strain in Essay #3, the tax incentives in Essay #9, and the legal and organizing tools in Essays #7, #8, and #11 that exist precisely to make sure that choice gets made with real community input, not despite it. The data center boom driven by AI, covered in Essay #10, is not slowing down. The question this entire series has tried to answer is not whether data centers will keep being built, but whether they’ll be built well, with full disclosure, fair terms, and genuine accountability to the people who will live next to them for decades to come.