TechnicalNews

The Less Water Data Centers Use, The More Power They Consume. Here’s Why

CNBC

AI data centers face a fundamental trade-off between water and power consumption for cooling. While liquid cooling systems use minimal water inside servers, the heat rejection process requires either significant water usage through evaporative cooling or substantial additional power for dry cooling, creating infrastructure challenges for communities nationwide.

Summary

AI data centers, particularly those running GPU-intensive workloads, generate extreme heat requiring sophisticated cooling solutions. Modern GPU racks like Nvidia's Grace Blackwell and AMD's Helios operate at temperatures around 80°C and would fail without continuous cooling. The transcript explores how data centers manage this challenge through liquid cooling systems, which are roughly 3,000 times more effective than air cooling at removing heat from chips.

Liquid cooling comes in two distinct phases: the closed-loop system inside servers that recirculates coolant without water loss, and the external heat rejection process that must expel absorbed heat to the atmosphere. Hyperscalers often claim zero water usage by only counting the internal loop, but this masks the significant water or power required for the second phase. The transcript reveals that reducing water consumption for heat rejection necessarily increases power demands, creating a critical trade-off.

Three primary cooling methods exist for heat rejection: dry cooling (using ambient air via fans), evaporative cooling (most water-intensive), and hybrid approaches. Dry cooling requires substantial power during warm months when air conditioning-style units must compensate for higher ambient temperatures. Evaporative cooling towers consume massive quantities of water through evaporation. The largest data centers can use up to 5-8 million gallons per day—equivalent to a small town of 50,000 people—straining local infrastructure.

The broader context shows over 2,000 new data centers planned across the U.S., with 71% of survey respondents opposing local data center construction. Communities face water rate increases and power shortages as data centers compete with residential needs. Solutions emerging include lower-heat chip designs from companies like DeepX and SK Hynix's IHBM with integrated cooling elements, which aim to reduce the fundamental heat generation problem before it requires aggressive cooling interventions.

Key Insights

  • Water is approximately 3,000 times more effective than air at removing heat from chips, enabling much higher server density in the same physical footprint
  • Hyperscalers label their closed-loop cooling systems as using zero water for most of the year, but this accounting ignores the water and power required to reject heat from the building to the ambient environment
  • Saving water in data center cooling necessarily means using more power, as dry cooling systems require power-hungry air conditioning units during warm months when ambient air temperatures are high
  • Equinix's global water footprint across all operations in 2025 was equivalent to approximately 16-18 holes golf courses in the southwestern United States, consuming 1.2 billion gallons with a 10% increase from 2024
  • New data centers are projected to account for most of the increase in electricity demand over the next decade, prompting major tech companies to build their own power plants to meet cooling and operational needs

Topics

Liquid cooling systems in AI data centersWater vs. power trade-offs in heat rejectionHeat exchanger and external cooling methodsCommunity backlash and infrastructure strainLow-power chip design innovations

Transcript

[0:02] So here we can see all of the liquid cooling infrastructure for the Grace Blackwell rack inside this Equinix data center in San Jose, California. There are hundreds of server racks, some executing massive AI compute tasks for the companies renting space here to run thousands of graphics and central processing units. There are eight GPU racks in here, and each rack has 72 GPUs inside of it, and all those chips get hot under heavy workloads. Server can boil your eggs. They are running at 80°C. [0:34] The chips themselves would melt down if they weren't continuously cooled. And keeping these chips cool requires a lot of water and power, which is a major reason why there's a…

Full transcript available for MurmurCast members

Sign Up to Access

More from CNBC

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.