HN Debrief

45°C cooling design cuts data center water use to near zero

  • AI
  • Infrastructure
  • Climate
  • Hardware

Nvidia’s post argues that AI racks can be redesigned so the entire server, not just the hottest chips, rides a warm liquid loop. The coolant comes in around 45°C, leaves hotter, and in favorable climates can dump heat straight to outdoor dry coolers instead of relying on evaporative towers that consume fresh water. That is how they get to the headline claim of near-zero water use. It is not zero water in an absolute sense. It means little or no ongoing cooling water consumption after the loop is filled.

If you run or finance data center infrastructure, treat this as a siting and systems-design story, not a miracle breakthrough. The prize is lower water use and simpler cooling in the right climates, but only if the whole facility stack, local grid, and nearby heat reuse options line up.

Discussion mood

Mixed but mostly skeptical of the marketing. People liked the direction of cutting evaporative cooling and saw real value in warm-water loops, but they pushed back hard on Nvidia presenting established cooling ideas as novel and on the article overstating both the breakthrough and the universality of the water savings.

Key insights

  1. 01

    Nordic heat reuse is already practical

    Existing projects in Finland and nearby regions show that data center waste heat is not a futuristic add-on. It already heats towns, pools, and district networks. The key trick is usually a heat pump on the production side that boosts data center output into the 60 to 90°C range district heating expects, which means the limiting factor is less the server room and more whether a heat network is nearby.

    If you are evaluating a new facility in a cold region, map district heating and heat pump integration before treating waste heat as mere exhaust. The commercial upside comes from co-locating with existing heat demand, not from the cooling design alone.

      Attribution:
    • lrasinen #1
    • hiAndrewQuinn #1
    • helsinkiandrew #1
    • kukkamario #1
  2. 02

    Climate still decides whether dry cooling works

    Warm-liquid loops only get you out of evaporative cooling when outdoor air stays far enough below your return-water temperature. Real operators noted that once ambient temperatures climb into the high 30s Celsius, you are back to supplemental cooling, which means you still need towers or chillers in the design. Even moderate places now see heat spikes often enough that backup infrastructure is not optional.

    Do not read 'near zero water' as a global template. Ask for site-specific weather bins, design-day assumptions, and how many hours per year the facility expects to fall back to wet or compressor-based cooling.

      Attribution:
    • matt-p #1 #2
    • pjdesno #1
  3. 03

    The real change is full-rack liquid cooling

    The strongest technical clarification was that Nvidia is not inventing liquid cooling. The meaningful step is pushing liquid to the parts that used to remain air cooled, so you can stop carrying a parallel cooling regime for RAM, storage, networking, power gear, and other rack components. That simplification is what makes higher loop temperatures and drier heat rejection more plausible at scale.

    When vendors pitch 'liquid cooling,' ask what percentage of the rack still depends on air. Hybrid systems keep dragging along the same building-level constraints that full liquid designs are trying to escape.

      Attribution:
    • RachelF #1
    • frollogaston #1
    • XorNot #1
    • Toutouxc #1
  4. 04

    Cooling power savings are smaller than the headline implies

    Operators called out the article’s use of historical cooling overhead figures. Many modern facilities already run at a Power Usage Effectiveness around 1.2 or better, so cooling is not 40% of total draw anymore. That means the absolute electricity savings from this design may be modest even if the water savings and operational flexibility are real.

    Model this as a water, permitting, and deployment-speed improvement first. If your business case depends on huge total-energy savings, you are probably leaning on outdated assumptions.

      Attribution:
    • pjdesno #1
    • dgacmu #1
  5. 05

    Forty-five to fifty-five degrees is useful heat

    Several technically grounded comments corrected a common mistake about the temperature numbers. A 55°C outlet is not too low to matter. It can directly serve low-temperature heating systems and is an excellent source temperature for a heat pump, even if it is marginal for legacy domestic hot water systems that need higher temperatures for safety and compatibility.

    If you own buildings, campuses, or industrial loads near a data center, check whether you can use low-grade heat before dismissing it. A heat pump can turn mediocre exhaust into a valuable utility stream.

      Attribution:
    • amluto #1 #2
    • lrasinen #1
    • krab #1

Against the grain

  1. 01

    The AI water panic is overstated

    A minority view argued that data center water use has become a meme detached from scale. Those comments claimed the most viral figures often count indirect water use from power generation or present local incidents in misleading ways, which makes the political salience much bigger than the actual share of water demand in most regions.

    If water risk matters to your strategy, get local operational numbers instead of relying on viral per-query claims. The decision hinge is usually basin-level scarcity and cooling method, not generic 'AI drinks water' rhetoric.

      Attribution:
    • protoster #1
    • harry19023 #1
    • rcpt #1
  2. 02

    Evaporative cooling is common because it works

    One operator pushed back on the idea that water use is fictitious or trivial in practice. In a real facility, evaporative cooling can consume around 10,000 gallons per day per megawatt because latent heat is such an efficient way to reject heat outdoors. That does not make it automatically unacceptable, but it does mean the water tradeoff is concrete, not imaginary.

    Treat cooling-water consumption as a measurable engineering choice, not a culture-war proxy. Request gallons per megawatt-day and seasonal operating modes from providers instead of broad assurances.

      Attribution:
    • pjdesno #1
  3. 03

    Waste heat can still be a local environmental issue

    Some comments rejected the idea that swapping water towers for dry coolers makes the externality disappear. Even if the heat does not meaningfully affect global climate, concentrated thermal discharge can matter locally, especially when heat ends up in waterways or intensifies urban heat island effects around large industrial sites.

    When reviewing sites, separate global carbon arguments from local thermal impacts. Ask where the rejected heat goes, into air, water, or a reuse network, and what limits apply in that environment.

      Attribution:
    • FridgeSeal #1
    • newsclues #1
    • thinkingtoilet #1
    • stonogo #1

In plain english

AI ↩
Artificial intelligence, here mainly referring to generative models used to create or assist with images and animation.
district heating ↩
A system that produces heat centrally and distributes hot water or steam through pipes to many buildings.
evaporative cooling ↩
A cooling method that removes heat by evaporating water, which is energy efficient but consumes water.
heat pump ↩
A machine that moves heat from one place to another and can raise low-temperature heat to a more useful temperature for heating.
HPC ↩
High-performance computing, large-scale computing systems used for scientific, engineering, or other very demanding workloads.
warm-water cooling ↩
A liquid cooling approach that uses relatively high coolant temperatures, making it easier to reject heat without chillers.

Reference links

District heating and heat reuse examples

Alternative cooling deployments and references

Water use and environmental impact debate