4 Percent in the Data Center, 54 at the Power Plant: Where Nvidia’s Water Promise Really Ends
Nvidia cools AI data centers with warm water in a closed loop, saving nearly everything on-site.
6 min read
Nvidia claims it has largely solved the water problem for data centers. A closed-loop system using warm water cools the next generation of chips without any new water evaporating. The manufacturer says this cuts consumption by up to 100 percent-but only up to the data center wall. The bigger thirst arises where the power comes from.
Key Takeaways
- The loop is real: Nvidia’s 45-degree cooling runs without active chillers and saves almost all on-site cooling water in favorable climates.
- The system boundary is misleading: On-site cooling is only a small slice of total water demand; the bulk is generated where electricity is produced.
- Operators must look at the full ledger: When reviewing water commitments, ask exactly where the line is drawn.
Related:200 kW per rack: firstcolo’s new build in Rosbach / Smart cooling: AI slashes energy use by half
What Nvidia has announced
At the London Climate Week in late June 2026, Nvidia unveiled a cooling architecture that uses warm instead of chilled water. The coolant is a mix of water and propylene glycol-similar to automotive antifreeze-that circulates at about 45 °C. Because the fluid stays in a closed loop, no water evaporates to cool the chips.
The savings are real. In mild climates the system runs without active chillers, relying solely on dry coolers. Nvidia estimates the reduction versus classic cooling towers at roughly 9.8 million liters per megawatt per year, dropping to almost zero. First reference kits ship to partners in Q4 2026, with large-scale deployments from mid-2027. A 50 MW pilot in Finland already feeds its waste heat into a district-heating network, warming more than 20,000 homes in Espoo.
| On-site cooling | Classic cooling tower | Nvidia’s warm loop |
|---|---|---|
| On-site water use | ≈ 9.8 million liters per MW per year | almost zero |
| Active chiller | required | unnecessary in mild climates |
| Waste heat | mostly unused | usable for district heating |
Why “solved on-site” only works locally
The figure of almost zero refers to the water that evaporates inside the data centre itself. Yet this line item is the smaller one. An analysis by Xylem and Global Water Intelligence from 2026 places the additional water demand triggered by AI by 2050 in very clear perspective: on-site cooling accounts for only a fraction. The lion’s share arises in power generation, with another large slice coming from chip fabrication.
What is the system boundary? The imaginary line that defines where a balance sheet stops counting. Draw it tightly around the server room and water use looks small. Extend it to power generation and chip fabrication and the full demand becomes visible.
This shifts the problem rather than erasing it. Warmer cooling allows denser racks. Denser racks need more electricity. More electricity means more water at the power plant. If the boundary is drawn tightly around the server room, an impressive balance can be presented. If it is drawn all the way to the power station and the chip fab, AI’s thirst remains.
Still real progress
This is not an indictment of Nvidia. The closed loop solves a genuine headache that many sites in water-scarce regions face. Where cooling towers once evaporated millions of litres, the new architecture saves precisely where the conflict with local communities is most acute-and the usable waste heat is more than a side benefit, as the Espoo example shows.
The mistake is not in the technology but in the interpretation. When a single, well-engineered component is sold as the solution to the entire puzzle, a false sense of security is created. For the debate on AI and resources, the honest statement is the better one: an important step, not the end of the road.
Questions operators should ask now
For your own planning, the headline matters less than the point where a promise ends. Four questions separate a verifiable figure from a marketing number.
- Where does the balance stop? Does the water pledge cover only on-site cooling or the entire chain back to the power plant?
- Where does the power come from? A water-saving site next to a thirsty power station simply moves the problem elsewhere.
- Is the waste heat used? A district-heating connection turns the load’s waste heat into a benefit.
- Does the balance hold in summer? Ditching the chiller depends on the climate. In hot regions the math changes.
Frequently Asked Questions
How does Nvidia’s warm-water cooling work?
A closed loop circulates a mix of water and propylene glycol at around 45 °C past the chips. Because the liquid circulates and does not evaporate, no new water is consumed for cooling. In mild climates, the active chiller can even be omitted.
Is the claim of 100 % less water accurate?
For the water that evaporates on-site to dissipate heat, the savings are indeed very high. However, this applies only to that specific line item; the far larger water demand-stemming from power generation and chip fabrication-remains unchanged.
When will the technology be available?
Nvidia plans to ship reference kits to partners in Q4 2026. First large installations are expected from mid-2027. A 50-megawatt pilot in Finland is already heating more than 20,000 homes via its district-heating network.
What does this mean for your own data-centre planning?
Warm-liquid cooling is becoming the standard for high rack densities. It is essential to account for the entire water balance, factor in electricity sourcing, and plan heat-reuse strategies from day one.
Editor’s Reading List
- Sustainable IT: greening the last mile in logistics
- KRITIS in the Cloud: what secures migration
- Europe’s first exascale system: who may compute?
Source of title image: AI-generated (July 2026)
Image source: AI-generated (July 2026)
