Nvidia said on 21 October 2026, at its GTC Berlin conference, that its DSX reference designs for AI data centres can cut facility cooling water use to near zero. It put the potential saving at 54 million liters a day for a 2-gigawatt deployment, which it said equals the daily water use of 180,000 households. The company also said its DSX MaxLPS software helps operators generate the same number of tokens with up to 30% less power.
The figures are Nvidia’s own, from its release and from trials it reported with partners. As we read it, they land in a week when local opposition and power limits have cost data centre projects in Australia and the US, as our earlier coverage reported.
Business Pill · GRID COST ALLOCATION
A short explainer on grid cost allocation, who pays when large new loads connect to the power grid. It teaches the general idea only and says nothing about any company in this story.
The Keynote
NVIDIA GTC Berlin 2026 keynote, 21 October (stream posted by Yahoo Finance). Quotes in this piece come from NVIDIA’s release and live blog, not from the video.
The key insight: As we read it, Nvidia is now quoting water and power savings, not only chip speed, as the selling point for the AI data centres it supplies.
Near-zero cooling water. Nvidia said its DSX reference designs use closed-loop liquid cooling and dry coolers to enable near-zero facility cooling water consumption, saving a potential 54 million liters a day at 2 gigawatts compared with conventional cooling-tower systems.
Up to 30% less power for the same output. DSX MaxLPS dynamically allocates power within an AI factory’s provisioned budget, Nvidia said, and CoreWeave, Firmus, Google Cloud, IREN, Lambda, Nebius, Nscale and Sharon AI are exploring the software.
Grid-flexible data centres. DSX Flex lets operators adapt workload power use to utility needs, demand-response events and energy prices, according to the release.
Simulation and operations tools. Nvidia launched SimReady Central on Hugging Face with power and cooling equipment models, and added three DSX OS open source components in developer preview.
Nscale’s trial of Nvidia’s DSX MaxLPS software in Iceland, as Nvidia reported it: 49% higher inference throughput within the same 264-kilowatt power budget, shown as an index from 100 to 149 (our arithmetic). Nvidia also says its DSX designs can save up to 54 million liters of water a day at 2-gigawatt scale.
What Was Posted on X
AI factories need to respond when the power grid needs relief.
At #OCPSummit26, NVIDIAโs Ian Buck explains how adjusting power use and shifting computing workloads can help, and why collaboration across the industry matters.
NVIDIA’s own account, two days before the GTC Berlin keynote, says AI factories need to respond when the power grid needs relief, with a clip of NVIDIA’s Ian Buck at the OCP Summit explaining how adjusting power use and shifting computing workloads can help. It is the same idea the release now sells as DSX Flex.
Jensen Huang just highlighted NVIDIA’s growing AI presence in ๐ช๐บ Europe, with over 12,000 companies in its ecosystem, more than $10 billion invested over the past year, and 14 new startup investments this week.
Wall St Engine says Jensen Huang highlighted more than 12,000 companies in NVIDIA’s European ecosystem, more than $10 billion invested over the past year and 14 new startup investments this week. Those figures are the poster’s summary; NVIDIA’s live blog says the company committed nearly $10 billion to Europe’s technology ecosystem.
How the Cooling Works, in Plain Terms
Conventional cooling towers remove heat by evaporating water, so they need a steady supply of replacement water, Nvidia said in the release. Its DSX designs instead pipe a cooling liquid past the computer chips in a closed loop and send the heat to outdoor dry coolers, which Nvidia compared to large radiators that release heat into the air without evaporating water.
Nvidia said its Vera Rubin systems support 100% liquid cooling with coolant temperatures of up to 45 degrees Celsius, and that with dry-cooler-based designs this architecture can reduce facility cooling water consumption to near zero. It added a limit: local climate determines whether supplemental mechanical cooling is needed during the hottest periods.
The Trials Nvidia Reported
Nscale, Iceland. Using DSX MaxLPS across four GB300 NVL72 racks, Nscale demonstrated 49% higher inference throughput and 37% more managed GPU capacity within the same 264-kilowatt provisioned power budget, according to Nvidia.
Lambda. In a proof of concept on HGX B200 systems, Lambda demonstrated 24% more cluster token throughput and 19% more GPU capacity within a fixed power budget, Nvidia said.
UK grid trial. Emerald AI, EPRI, Nebius, the UK National Grid and Nvidia demonstrated reductions of up to 40% in a Blackwell Ultra cluster’s power draw in response to simulated grid events, while critical workloads kept running, the release said.
“Europe’s AI ambitions demand that we make every megawatt count,” Nscale founder and CEO Josh Payne said in the release. Warren Barkley, Nvidia’s vice president of DSX productmanagement, said: “Europe’s technology leadership depends on how efficiently we turn power and infrastructure into intelligence.”
Who Else Is in the Ecosystem
Nvidia named AVEVA, Dassault Systèmes and Siemens among its technology partners, and ABB, Eaton, GE Vernova, Hitachi Energy, Schneider Electric, Siemens Energy, Trane Technologies and Vertiv among its energy and cooling providers. It said Trane Technologies has published seven thermal management reference designs, including three 250-megawatt DSX-based cooling designs, and that Hitachi Energy’s battery storage system earned DSX Ready qualification.
In his keynote, Jensen Huang said, according to Nvidia’s live blog: “To be part of this AI economy, you don’t want to just be consuming the AI; you want to be producing the AI.”
Connecting the Coverage
Our levy piece reported on 21 October that Australian senator David Pocock proposed a 5% levy on data centres, citing Microsoft’s A$2.3 billion of local data centre revenue, as SBS News and Switzer Daily reported. Our NextDC piece the same day covered the operator removing about A$140 million from its fiscal 2027 guidance after the distributor Jemena suspended power works for its M3 Melbourne site, citing community opposition.
Our Loudoun piece reported, citing the Financial Times and RegLayer data, that at least 47 large US data centre projects were delayed, withdrawn or blocked in 2026, and our Goldman piece reported the bank’s 19 October forecast that global data centre power demand will rise 170% between 2025 and 2030. Our Germany pact piece reported that the federal government planned to announce an AI partnership with Nvidia at the Invest in Germany Summit in Berlin, as deutschland.de reported. What is new here is Nvidia putting numbers on how much water and power its own designs can save. Our companion piece on DSX POD covers the other release Nvidia issued the same morning, which says factory-built Vera Rubin AI factories in modular containers can cut deployment at a prepared site “from six months to as little as 14 days.”
What It Means for Operators and Communities
As we read it, Nvidia is selling efficiency as a way to get permits and grid connections, not only as a cost saving. The release itself says DSX Flex can help enable faster grid connection.
As we read it, the power claim cuts two ways: Nvidia says operators can use the savings to run more GPUs within the same power budget, which means the same grid connection could carry more computing rather than a smaller load.
The Structural Read
As we read it, the water figure speaks to the people who decide permits, while the power figure speaks to operators: Nvidia says MaxLPS can produce the same tokens with up to 30% less power, or run more GPUs inside the same budget.
As we read it, the second use matters more for the grid. Nvidia’s own trial results describe more throughput and more GPU capacity in the same power envelope, not a smaller load.
As we read it, DSX Flex is the piece aimed at utilities: Nvidia says it lets AI factories cut or shift power use during grid events, and the release links it to faster grid connection.
Josh Payne, founder and CEO of Nscale, in NVIDIA’s release, 21 October 2026
“Europe’s AI ambitions demand that we make every megawatt count.”
Three Implications
FOR DATA CENTRE OPERATORS Nvidia says CoreWeave, Firmus, Google Cloud, IREN, Lambda, Nebius, Nscale and Sharon AI are exploring DSX MaxLPS, which it says delivers the same tokens with up to 30% less power.
FOR UTILITIES AND GRID OPERATORS In a trial Nvidia reported with Emerald AI, EPRI, Nebius and the UK National Grid, a Blackwell Ultra cluster cut its power draw by up to 40% in simulated grid events.
FOR LOCAL AUTHORITIES Nvidia says dry-cooler designs can save 54 million liters of water a day at 2 gigawatts, but adds that local climate decides whether extra mechanical cooling is needed.
The framework states its thesis in one line: “The structural bottleneck isn’t computing, it’s electrons.”
As we read it, Nvidia’s DSX release applies that thesis from the supplier’s side: its headline claims are about power and cooling water, not chip performance. The framework was published in December 2025 and does not discuss DSX.
What Is Not Established
All figures are Nvidia’s or its partners’, from Nvidia’s release; we did not see independent measurements. The 54 million liters is a potential saving at 2-gigawatt scale against cooling-tower systems, and the release says hot climates may still need supplemental mechanical cooling. The 30% and the 40% are up-to figures, and the Lambda and Nscale results come from specific test setups. The release does not give prices, the number of European sites using DSX or deployment dates for the partners exploring MaxLPS.
Business Engineer Framework
The State of AI Data Centers
A Business Engineer analysis of the power, grid and capital constraints behind the AI data-centre build-out.
Nvidia said on 21 October 2026 that its DSX designs can bring facility cooling water use to near zero, saving a potential 54 million liters a day at 2 gigawatts, and that its MaxLPS software can produce the same tokens with up to 30% less power. As we read it, the chip supplier is now pitching its designs as an answer to the local resistance that has slowed data centre projects.
A note on sourcing. This piece is built on NVIDIA’s GTC Berlin press release of 21 October 2026, read in full, and the opening of NVIDIA’s live blog of Jensen Huang’s keynote. All efficiency figures and trial results are NVIDIA’s or its partners’ as NVIDIA reported them; we did not see independent measurements. The chart index is our arithmetic. We did not contact NVIDIA or the partners named. Nothing here is legal, financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
