Nvidia Has Nearly Eliminated Cooling-Water Use in AI Data Centers

📋 Today’s Story in Three Lines

• Nvidia has designed Rubin data centers to use 45°C warm-water cooling.
• The design can nearly eliminate on-site cooling-water use, but it addresses only part of the water footprint once power generation is included.
• Check cooling methods and water use by energy source separately in your infrastructure contracts.

Today, we are unpacking exactly how much of a data center is covered by the phrase “uses almost no water.”

📌 Today’s Deep Dive — Rubin Saves Water by Running Servers Hotter

What happened

Nvidia has unveiled a reference design for next-generation Rubin data centers built around 100% liquid cooling. The point is not to make servers colder. Instead, coolant enters at temperatures as high as 45°C and leaves at 55°C after passing over the chips. That is hot water to a person, but still cool enough to carry heat away from silicon. The Verge

The coolant circulates continuously through a closed loop. Once filled, the same water can be reused throughout the facility’s lifetime, removing the need to keep evaporating fresh water to cool the chips. Nvidia says the design can reduce on-site water use by as much as 100% where climate conditions allow. TechCrunch

Why this approach now

Traditional cooling towers reject heat by evaporating water. Nvidia’s comparison figure is roughly 2.6 million gallons per megawatt each year. Warmer coolant, however, creates a larger temperature difference with the outside air. In many regions, that makes it possible to reject heat through outdoor dry coolers and reduce reliance on evaporative cooling and chillers. The Verge

Put simply, the design carries heat away with hot water instead of evaporating water to keep chips cold. Facilities that need fewer fans and chillers may also cut electricity use and noise. Nvidia says cloud providers and data-center operators building Rubin facilities are choosing this transition. It has not disclosed how much more these facilities cost to build than conventional air-cooled sites. The Verge

Where the numbers stop

This is where the boundary matters. “Nearly all water” means the water used for cooling inside the data-center facility. It excludes water consumed by power plants and semiconductor manufacturing. An analysis cited by TechCrunch says including those external sources can make a facility’s total water footprint two to three times larger. By that estimate, Nvidia’s design addresses roughly one-quarter to one-third of the total. TechCrunch

The energy source also makes a large difference. The figures presented are 1.17 liters of water per kWh for natural gas and 2.2 liters for coal. Wind uses 0.01 liters, while solar uses 0.03 liters including manufacturing and panel cleaning. Fossil fuels currently supply about half of data-center electricity, and gas and coal are projected to provide more than 40% of the additional power data centers will consume through 2030. TechCrunch

Why it matters

This distinction changes how AI products should describe their environmental impact. When a cloud provider claims a “100% reduction in water use,” the conclusion depends on whether that figure covers only facility cooling or the entire electricity supply chain. Efficient cooling and a low total water footprint are not the same thing.

That does not make the design trivial. Using the benchmark of 2.6 million gallons per MW per year, a facility could save roughly 27,000 liters of on-site cooling water per day. That is a concrete change when seeking data-center permits or negotiating with communities in water-stressed regions. But if off-site power generation remains unchanged, the total benefit should not be overstated.

What to watch next

First, watch operating data from real Rubin facilities. We need to see how consistently the reference design’s “up to 100%” figure holds across different climates. Second, watch construction and maintenance costs. Savings on water, fans, and chillers may not drive rapid adoption if the initial infrastructure cost is too high.

Third, watch power procurement. Even when coolant circulates in a closed loop, water use remains outside the facility if its electricity comes from water-intensive power plants. When evaluating cloud sustainability metrics, on-site cooling and the energy source should appear in the same comparison.

⚡ Quick News

• OpenAI is supporting shared standards for advanced AI — Through the Appia Foundation, it says it will support evaluation systems, safety practices, and international cooperation. Source
• GPT-5 Pro helped solve a three-year-old immunology problem — It surfaced clues about T-cell behavior that could expand research into cancer and autoimmune disease. Source
• Samsung Electronics is deploying ChatGPT and Codex to employees worldwide — OpenAI presented it as a large-scale enterprise AI adoption case. Source
• OpenAI has launched an open-source vulnerability repair program — Patch the Planet uses AI and expert review to find, validate, and fix security vulnerabilities. Source
• Oracle’s AI investment is colliding with 21,000 layoffs — The company is reportedly cutting labor costs while investing billions of dollars in data centers. Source
• Claude Tag learns organizational context from Slack conversations — The always-on AI coworker can also be read as a strategy for acquiring workflow and internal-company knowledge. Source

🇺🇸 What You Should Do Now

1. Ask your cloud provider where its water-use metric stops. Add one line to your contract review checklist stating whether WUE covers only on-site cooling or also includes power generation.
2. Model your AI workload’s water use by energy source. The figures provided here are 1.17 liters per kWh for gas, 2.2 liters for coal, 0.01 liters for wind, and 0.03 liters for solar. Ask the supplier for its disclosed energy mix and multiply it by your monthly consumption to reveal the gap between claims and reality.
3. Do not repeat “100% less water” in customer-facing copy without qualification. Report on-site cooling water and the total water footprint separately, and state that unverified construction and chip-manufacturing stages are outside the scope.

Today’s takeaway: AI data centers can shrink their water problem inside the cooling loop, but they cannot make it disappear from the power grid.

—
AI Daily · A morning brief for people building with AI