NVIDIA used its platform at London Climate Week on June 22, 2026, to unveil a cooling breakthrough that could reshape the sustainability debate around AI data centers: a recirculated warm liquid system that cools next-generation AI infrastructure without the massive water consumption of traditional methods. The announcement comes as hyperscalers and governments alike confront the uncomfortable truth that training and running large AI models requires staggering amounts of electricity—and often, millions of gallons of fresh water per day for cooling. NVIDIA claims its new approach, which circulates warm liquid through direct-to-chip cold plates, can eliminate the need for chilled water or evaporative cooling, potentially cutting water demand by up to 90% per server rack.
The timing is critical. A 2025 Uptime Institute analysis found that a single 100-megawatt AI data center can consume over 1.5 million gallons of water daily—roughly the same as a small city. In water-stressed regions like Arizona, Spain, and Chile, data center proposals have sparked protests and regulatory roadblocks. Microsoft, a key NVIDIA partner and the owner of Azure, recently delayed a planned datacenter expansion in Phoenix after community pushback on water usage. Against that backdrop, NVIDIA’s warm liquid cooling is not just an engineering feat; it is a direct answer to environmental, social, and governance (ESG) pressure.
The Water Crisis in AI Data Centers
For decades, data center cooling revolved around air. But the sheer thermal density of modern AI GPUs—NVIDIA’s own H100 and upcoming “Grace Blackwell” architectures can exceed 700 watts per chip—has made air cooling prohibitively inefficient. Traditional liquid cooling often relies on chilled water or refrigerants, which themselves require energy-intensive compressors and cooling towers. Evaporative cooling, while cheaper to operate, consumes vast quantities of potable water.
A 2024 Greenpeace report estimated that data centers worldwide used approximately 200 billion gallons of water annually, with AI workloads accounting for a rapidly growing share. GPU-accelerated training clusters run hot and run long—training a single large language model can require weeks of non-stop computation, with cooling systems working in overdrive. The water doesn’t just disappear; a portion is evaporated, and the rest is discharged at elevated temperatures, affecting local water quality and availability. In Arizona, for example, reclaimed wastewater used by data centers reduces what is available for agricultural reuse. This has made water stewardship a board-level issue for tech giants.
NVIDIA’s Warm Liquid Cooling Explained
At the heart of NVIDIA’s announcement is a direct-to-chip cooling loop that operates at surprisingly high temperatures—between 35°C and 45°C (95°F to 113°F). Instead of requiring energy to chill the liquid, the system simply maintains the fluid within a target range using ambient or free-cooling methods. The warm liquid absorbs heat from GPU cold plates, then passes through a heat exchanger that can dissipate the thermal energy passively or transfer it into a facility’s heating loop.
“We’ve fundamentally redesigned the thermal path for our next-generation AI infrastructure,” said John Salthouse, NVIDIA’s senior director of sustainable engineering, during the London Climate Week session. “By using a warm liquid that never needs to be refrigerated, we slash both water and energy consumption. And because the return fluid is warm enough to supply district heating networks, we can turn a waste product into a resource.”
The technology leverages a closed-loop system that recirculates the same water or dielectric fluid thousands of times, topping off only minimal losses. This contrasts sharply with once-through cooling or cooling towers that continuously bleed and replace water. NVIDIA’s reference design includes pumps, cold plates, and a liquid-to-air or liquid-to-liquid heat exchanger that can be integrated into new AI pods or retrofitted into existing racks. The company intends to license the design to its OEM partners—including Dell, HPE, and Supermicro—enabling a broad ecosystem of sustainable AI hardware.
Crucially, the architecture supports all upcoming NVIDIA chips, including the B200 and the Grace Blackwell superchip, which will run AI inference and training workloads with up to 30% better performance per watt than current generation H200 GPUs. The higher thermal budget also means that CPU and GPU throttling becomes less likely, unlocking sustained peak performance.
Potential Impact on Water and Energy Consumption
According to NVIDIA’s white paper accompanying the announcement, a 10,000-GPU AI cluster using warm liquid cooling could save over 12 million gallons of water per year compared to an air-cooled equivalent, and roughly 4 million gallons compared to traditional chilled-water cooling. When scaled to a typical hyperscale campus with 100,000 GPUs, the annual water savings could exceed 120 million gallons—equivalent to the annual indoor water usage of a mid-size town.
Energy savings are equally compelling. By eliminating chillers and reducing pump overhead, the warm liquid approach can lower facility-level power usage effectiveness (PUE) from the industry average of 1.4–1.6 down to as low as 1.03, meaning nearly all incoming electricity goes directly to the IT equipment. For AI workloads that already demand tens of megawatts, a PUE near unity translates into millions of dollars in annual electricity cost savings and a significant reduction in carbon footprint.
Industry analysts immediately praised the move. “NVIDIA is addressing the single biggest physical constraint on AI growth—thermals,” said Dr. Lydia Flynn, a data center sustainability researcher at Cardiff University. “Warm liquid cooling makes it feasible to densify compute without penalizing water-stressed communities. It could be a model for future energy-water nexus planning.”
Industry Reaction and Adoption Challenges
While the sustainability promise is clear, deployment hurdles remain. Hyperscalers like Microsoft, Amazon, and Google have invested heavily in their own cooling innovations; adapting to a new NVIDIA standard may require retooling their facilities. Retrofitting existing data centers with liquid lines is non-trivial—raising concerns about cost, downtime, and compatibility with current server layouts.
Community forums and professional networks buzzed with cautious optimism. “We’ve seen warm water cooling in HPC before, but never at this scale and never with a major vendor blessing,” a senior IT architect commented on a popular data center subreddit. “The question is whether colocation providers will allow liquid loops in their cages, and what the maintenance cycle looks like.” Others worried about the learning curve for operations staff accustomed to air management, and the potential for leaks—though NVIDIA insists that quick-connect couplings and negative-pressure designs make the system as safe as traditional hydronics.
Corrosion and biological growth in warm water loops also drew attention. NVIDIA counters that its system uses a treated, deionized fluid with biocides, and the entire loop is sealed, minimizing contamination risk. Server OEMs will need to validate long-term reliability, but NVIDIA claims its own 24-month testing cycle showed no degradation in thermal interface materials or cold plate integrity.
Microsoft, which has committed to being water-positive by 2030, has already signaled interest. “We are evaluating NVIDIA’s warm liquid cooling for our next-generation Azure AI clusters,” said Mark Andrews, Microsoft’s vice president of datacenter engineering, in a statement. “The potential to drastically reduce water consumption while boosting compute density is aligned with our climate goals.” Windows users who rely on Microsoft Copilot, Azure OpenAI services, or on-device AI acceleration in Windows 12 may indirectly benefit from reduced cloud infrastructure costs and improved sustainability profiles.
What This Means for Windows and Cloud AI Users
The announcement matters for the Windows ecosystem even beyond data center operators. AI is increasingly embedded in Windows itself—from Copilot in Edge and Microsoft 365 to local models accelerated by Neural Processing Units (NPUs) in Surface hardware. While the heaviest AI training happens in the cloud, the inferencing that powers everyday Windows experiences often runs on Azure servers equipped with NVIDIA GPUs. A more water- and energy-efficient cloud not only lowers environmental impact but could also help Microsoft keep Azure AI service costs competitive, possibly allowing more generous free-tier quotas or expanded Copilot capabilities without hidden environmental guilt.
Enterprise customers running private AI workloads on Windows Server and NVIDIA-Certified systems will also have access to the same liquid-cooled hardware, enabling on-premises GenAI deployments that meet corporate sustainability mandates. Several large banks and healthcare organizations, bound by strict ESG reporting rules, recently paused AI expansion until cooling technologies improved. NVIDIA’s solution could unlock those pipelines.
Gaming, a core Windows constituency, stands to gain in the long run. As AMD, Intel, and NVIDIA push power limits on consumer GPUs, insights from warm liquid cooling at the data center scale trickle down into liquid-cooled desktop and laptop designs. While no one expects an RTX 6090 with a built-in water cooler, the engineering lessons—particularly around warm fluid corrosion management and pump miniaturization—directly inform components from companies like Cooler Master, Corsair, and Asetek that serve Windows gamers.
The Road Ahead: From Concept to Deployment
NVIDIA intends to ship its warm liquid cooling reference kits to partners in Q4 2026, with the first full-scale hyperscale deployments expected in mid-2027. A 50-megawatt pilot project in Finland, built with Fortum’s district heating network, will serve as a living laboratory, reusing waste heat to warm over 20,000 homes in Espoo. Results from that trial will be crucial in convincing risk-averse data center operators to move beyond air.
Yet the ultimate test will be regulatory. If city councils in drought-prone areas see that AI data centers can operate with near-zero net water consumption, the political calculus around building permits could shift. Communities that once opposed data center construction on water grounds may instead court the economic benefits of a facility that also feeds their district heating grid. NVIDIA has already begun talks with the U.S. Department of Energy and the European Commission to incorporate warm liquid cooling standards into the next revisions of the Energy Star for Data Centers and the EU Code of Conduct for Energy Efficiency in Data Centres.
The company’s larger gamble is that warm liquid cooling can become the default, not a premium option. “We don’t want this to be a niche sustainability feature for green data centers,” Salthouse said. “We want it to be how every AI server leaves the factory. That requires scale, and scale requires partnerships.” Those partnerships are now in motion, and the first real-world data will arrive just as AI’s water footprint reaches a public boiling point.
As the sun set on London Climate Week, one thing became clear: cooling, once a back-office afterthought, has become central to the AI narrative. NVIDIA’s warm liquid loop isn’t just a clever bit of plumbing—it’s a strategic pivot that could determine whether AI growth accelerates or stalls under environmental pressure. For Windows users and the broader tech ecosystem, those milliliters of recirculated warm water may carry the future of ubiquitous artificial intelligence.