# AI liquid cooling is only as good as the building water loop By MarketScale Newsroom · Published 2026-11-14 · Building Management on MarketScale Canonical: https://www.marketscale.com/industries/building-management/dow-scientist-ai-liquid-cooling-is-only-as-good-as-the-building-water-loop > Dipti Singh says the facility water behind each CDU sets AI servers’ cooling margin, making chemistry and flushing specs a commissioning item. ## Key points - A CDU walls facility water off from the servers, yet every degree of warmer supply water and every film of deposit on the heat exchanger comes out of the chip's cooling margin. - Flushing velocity, turnover volume, passivation criteria and turbidity endpoints are cheapest to enforce when they are written into the construction spec, before the loop carries IT load. - A closed loop that saves water on site can push demand upstream to power generation, so a low WUE figure is only meaningful when read next to PUE and the regional grid mix. A cold plate on a GPU never touches the building's water. It still depends on it. Dipti Singh makes that argument in a sponsored DataCenterDynamics piece dated November 13, 2026. She writes that in liquid-cooled AI halls, the primary facility water system now serves as the heat-rejection backbone for the coolant distribution units (CDUs) that protect the racks. Her prescription follows from that. Operators should stop treating the system as a building utility inspected on a schedule and should manage it continuously as an engineered thermal system. For the facilities engineer or commissioning lead specifying the next liquid-cooled hall, this matters most where the drawings meet the water. The useful part of Singh's piece is less the warning than the mechanics. She explains how a problem on the facility side reaches the chip, and what goes into the spec to stop it. ## Two loops sharing one heat budget A liquid-cooled AI deployment runs two circuits. The technology cooling system (TCS) carries coolant through server manifolds and cold plates. The facility water system (FWS) carries heat away to final rejection. A heat exchanger inside the CDU sits between them, so facility water never enters the servers. The loops are separate hydraulically. Thermally they are still one system: the cold plates can only shed heat as fast as the facility side carries it off. Singh puts this in terms of the basic heat-transfer relation, Q = UAΔT. Heat moved equals how readily heat crosses the exchanger surface, times the surface area, times the temperature difference driving the flow. Hold the cold plate, the coolant and the chip temperature constant, and warmer supply water shrinks that temperature difference. Primary-side fouling, low flow, unstable temperature, biological growth and corrosion debris each eat into the CDU's thermal margin, she writes. The loss can show up as higher secondary-loop temperatures and, at the extreme, shutdowns. A CDU walls facility water off from the servers, yet every degree of warmer supply water and every film of deposit on the heat exchanger comes out of the chip's cooling margin. Singh draws a practical conclusion from that chain. The old model of periodic HVAC water treatment should give way to continuous, risk-based loop management aimed at keeping heat transfer stable as AI loads change. For a site whose water treatment program was scoped around comfort cooling, that is a change in scope worth checking against the existing service contract. ## What goes into the primary-side spec Singh says the facility and technology loops need separate chemistry specifications. The parameters monitored on the facility side, though, should look much like those on the technology loop. Her list: - pH and conductivity - Chloride, hardness, alkalinity and silica - Dissolved oxygen, where relevant - Inhibitor residual and glycol concentration - Suspended solids and corrosion metals - Microbial indicators One assumption she tells engineers to drop is that deionized water is safe by default. Without buffering and inhibitors, it can aggressively pick up ions from the metals and other materials it touches. The purest fill water can be the most corrosive. Her harder point is about timing. The work of protecting the loop starts while the system is still being fabricated and built. In her view, specifications should spell out capped storage for pipe, clean methods of assembly and the use of temporary strainers. They should also set the flushing velocity and the turnover volume, the criteria for passivation, acceptance limits for differential pressure across filters, endpoints for particles or turbidity, and rules for preserving branches that stay idle until a later phase. On the facility loop, permanent side-stream filtration then restricts what can reach the plate heat exchanger's narrow channels. On the technology loop, final filtration protects the manifolds and cold plates. Nearly all of this is a construction-phase decision. A flushing endpoint or a capped-storage rule is easy to enforce while the contractor is still on site. Once the loop is filled and carrying IT load, it is a remediation project. That makes the mechanical spec and the commissioning plan the place where Singh's advice either sticks or doesn't. ## Sizing the pipe for the second phase Singh also flags a design limit that tends to show up only after expansion. As liquid-cooled deployments grow, pipe sizing can cap coolant delivery, so she recommends checking headers, branches and CDU connections against both the initial and the expanded loads. Design reviews should map which racks lose cooling in each failure, confirm how much cooling remains during maintenance, and verify that local CDU controls keep coolant moving if the link to the building management system drops. Commissioning should record clean-system performance at representative loads and test how controls respond to failures. ASHRAE's integrated design guidance, part of its AI Data Center Energy Performance Framework, says rack densities have climbed from roughly 120 kW to several hundred kilowatts, with megawatt-class racks expected in the near term. When integrated design is not applied effectively, it says, facilities risk stranded capacity, operational instability and costly retrofits. Singh raises a related concern. Pipe sizing can become a constraint on coolant delivery as liquid-cooled deployments grow, so she says headers, branches and CDU connections should be verified against the initial load and against the expanded load as well. Block Field ## Water saved on site, power spent elsewhere Singh's piece also covers the tradeoff that sits behind every choice of heat rejection. Evaporative systems use more water but less power. A closed loop recirculates its coolant and uses little water itself, but it isn't zero-water, because total site use depends on how heat leaves the building. Dry cooling avoids evaporation but runs more compressor and fan power, which pushes power usage effectiveness (PUE) up. Her rule of thumb: water-stressed sites with plenty of power may accept that penalty, while power-constrained sites with plenty of water may favor evaporative cooling. She also notes that water usage effectiveness (WUE) counts water per kWh of IT energy and leaves out the water used to generate the electricity. Steven Lucente, a business development manager for Xylem's Applied Water business who leads its data center business development, looked at water and energy as a pair in an October 5, 2026 article in Data Center Knowledge sponsored by Xylem's Bell & Gossett. He writes that the power a data center draws can bring a water footprint of its own, and the size of that footprint depends on the regional generation mix. So if a cooling approach cuts on-site water use but needs more electricity, it may simply move part of that water demand upstream instead of eliminating it. He suggests project teams begin with the full water cycle: the source of the water, the quality it needs, how it is treated and moved, how much can be reused, how heat is rejected and what happens before discharge. Block Field Manfreid Chua, ZutaCore's VP of Business Development, AI and Sustainability, cites the IEA figures in an Enterprise Viewpoint piece. He argues that the challenge of supporting much higher computational density while using energy and resources responsibly starts at the chip. Together the figures explain why Singh wants primary-loop management judged alongside site water and energy performance rather than as a maintenance line item. At that scale, a few degrees of lost exchanger margin can mean real power, real water or fewer usable racks. ASHRAE's AI data center framework says warm-water, chiller-less cooling enables near-zero water use in many climates. Read on its own, that makes primary-loop water look like a shrinking concern. The heat-transfer math points the other way on reliability. Warmer supply water is the very condition that narrows the temperature difference, which suggests a warm-water design has less margin left to lose to fouling or low flow, and gains the most from the chemistry and filtration discipline Singh describes. Both vendors' pieces end in the same place: real-time visibility. Singh calls for monitoring water quality continuously, and Lucente points to digital monitoring of water quality, flow and equipment condition. The test of whether operators take this up will show in the next round of liquid-cooled hall specifications. Specifications that set the flushing, filter differential-pressure and particle or turbidity endpoints Singh recommends would suggest operators are treating the loop as engineered infrastructure. ## Sources - [Why AI infrastructure must rethink facility water loop management](https://www.datacenterdynamics.com/en/opinions/why-ai-infrastructure-must-rethink-facility-water-loop-management/) (DataCenterDynamics) - [Rethinking WUE: Data Center Water Strategy Beyond The ...](https://www.datacenterknowledge.com/cooling/rethinking-wue-data-center-water-strategy-beyond-the-facitily) (Data Center Knowledge) - [AI infrastructure needs a new approach to sustainability](https://enterpriseviewpoint.com/ai-infrastructure-needs-a-new-approach-to-sustainability/) - [Integrated Design Principles | AI Data Center Energy ...](https://www.ashrae.org/technical-resources/topics-and-initiatives/ai-data-center/integrated-design-principles) (ASHRAE) Tags: Dow, Data center cooling, Liquid cooling, Coolant distribution unit, Water treatment, AI infrastructure, Facilities management, Commissioning, WUE, PUE, ASHRAE, Xylem, Mechanical engineering, Data center operations --- Source: MarketScale, https://www.marketscale.com/industries/building-management/dow-scientist-ai-liquid-cooling-is-only-as-good-as-the-building-water-loop. Published for AI indexing and citation; cite the canonical URL. Site guide for agents: https://www.marketscale.com/llms.txt