Data center cooling typically uses 3–5 million gallons of water per MW annually. Blowdown recovery can cut make-up demand by 40%, supporting 12–15 cooling-tower cycles at TDS 3,000–5,000 mg/L, with RO flux at 15–25 LMH and ATP below 100 RLU.
Why do AI heat loads break conventional data-center water treatment?
AI and ML workloads raise rack heat density from 10–20 kW to 50–100 kW. That shift drives liquid cooling—direct-to-chip or immersion—with stricter water limits such as conductivity <0.1 μS/cm and TSS <10 μm. A 2025 Uptime Institute survey found that 70% of existing facilities lack treatment capacity for these loads without risking equipment failure.
Moving from air-cooled designs (PUE 1.6–2.0) to liquid-cooled architectures (PUE 1.1–1.3) often ignores Water Usage Effectiveness (WUE). A 10 MW site in Phoenix cut water use 35% after adding blowdown recovery, yet poor pretreatment caused $200K in downtime from scaling (Saltworks 2024 report). The three primary failure modes in 2026 are: (1) scaling in microchannels that chokes flow, (2) microbial fouling that insulates heat exchangers, and (3) accelerated corrosion in copper and nickel loops from poor pH and chloride control.
High-density GPU clusters, including NVIDIA Blackwell-class racks, can produce localized heat fluxes above 500 W/cm². A calcium carbonate film thinner than 0.1 mm can cut thermal conductivity by about 10% and trigger thermal throttling. Water-positive targets by 2030 are pushing reclamation that IT campuses previously deferred as too complex.
Permits in Northern Virginia and Singapore increasingly set WUE below 0.4 L/kWh. Meeting that limit needs higher cycles of concentration (CoC), real-time chemistry monitoring, automated dosing, and high-recovery filtration to manage minerals and biology in recirculating loops. Bleed-and-feed alone is no longer viable for these duty cycles.
What water-quality limits should open, closed, and direct-to-chip loops meet?
Heat transfer in liquid-cooled halls follows the Dittus-Boelter relation (Nu = 0.023 Re^0.8 Pr^0.4). Fouling can cut the Nusselt number by 20–40% and force a 15–30% rise in cooling energy to hold setpoints (ASHRAE 2026). Open-loop towers typically keep TDS below 3,000 mg/L; closed secondary circuits need TDS <50 mg/L; primary cold-plate loops need much tighter purity.
The following table outlines the 2026 ASHRAE TC 9.9 water quality targets for mission-critical cooling loops:
| Parameter | Open Loop (Cooling Tower) | Closed Loop (Secondary) | Direct-to-Chip (Primary) |
|---|---|---|---|
| TDS (mg/L) | <3,000 | <50 | <5 |
| pH Range | 7.5 – 8.5 | 8.0 – 9.0 | 7.0 – 8.0 |
| Hardness (as CaCO₃) | <100 mg/L | <10 mg/L | <1 mg/L |
| Chloride (mg/L) | <500 | <20 | <5 |
| Silica (mg/L) | <50 | <10 | <1 |
| Conductivity (μS/cm) | <4,500 | <100 | <0.1 |
For multi-metal systems, hold molybdate at 2–5 ppm for mild steel and zinc at 1–3 ppm for galvanized parts. Keep microbial load below 100 RLU by ATP testing. WUE = (make-up water + blowdown) / IT energy (L/kWh). A 10 MW hall at 1.5 L/kWh uses 131,400 m³/year; poor chemistry can raise that by 20% through excess blowdown.
Target LSI of +0.2 to +0.5 in open loops for a thin protective film. Direct-to-chip microchannels cannot tolerate that same scaling potential, so primary loops move to DI or EDI water at <0.1 μS/cm to limit stray-current electrolysis and copper pinhole leaks. Keep dissolved oxygen below 10 ppb with vacuum degassing or hydrazine alternatives to limit magnetite sludge. Rotate non-oxidizing biocides (DBNPA or isothiazolinone) with chlorine dioxide. Side-stream filtration of 5–10% of circulating flow removes solids that seed scale and biofilm—especially where towers scrub dusty ambient air.
How should you choose RO, DAF, or evaporation for 40% blowdown recovery?
RO remains the main path to about 40% make-up reduction because it rejects 95–99% of TDS at moderate energy. Typical blowdown RO runs at 15–25 LMH and recovers 70–85% of the blowdown stream. Pretreatment controls fouling: where TSS or organics are high, DAF systems for cooling tower blowdown pretreatment remove 90–95% of TSS and 95% of FOG before membranes.
Where discharge permits force Zero Liquid Discharge, evaporation crystallization for zero liquid discharge (ZLD) in data centers can treat TDS up to 100,000 mg/L and recover 90–95% of the waste stream. CapEx/OPEX and blowdown chemistry decide the split between RO and thermal concentration.
Standard towers often run 3–5 cycles; advanced recovery targets 10–15 cycles. Silica then becomes the limiter and can form hard, glass-like scale. Warm lime softening or magnesium oxide dosing before RO precipitates silica so membranes can hold higher recovery. Disc-tube RO (DTRO) helps on high-fouling or poor make-up sources, including reclaimed municipal water: open-channel modules tolerate higher solids and osmotic pressure. Pairing DAF solids removal with DTRO salt rejection cuts brine volume and lowers WUE.
| Technology | Recovery Rate | TDS Limit (mg/L) | CapEx ($/m³/day) | OPEX ($/m³) | Ideal Use Case |
|---|---|---|---|---|---|
| Reverse Osmosis (RO) | 70–85% | <5,000 | $800 – $1,200 | $0.40 – $0.65 | Standard recovery in low-silica areas |
| Dissolved Air Flotation (DAF) | Pre-treatment | N/A (TSS focus) | $300 – $500 | $0.10 – $0.20 | High organic/solids loading removal |
| Mechanical Vapor Recompression (MVR) | 95–98% | <250,000 | $4,000 – $7,000 | $2.50 – $4.50 | Strict ZLD requirements/Water scarcity |
| Electrodialysis Reversal (EDR) |
Who this is for / Who should look elsewhere / Next step
This guide is for plant owners and process engineers sizing treatment for liquid-cooled GPU halls, cooling-tower blowdown recovery, and WUE-constrained permits. Air-cooled edge sites with negligible evaporative load, or campuses that only need once-through makeup without recirculation chemistry control, should look elsewhere. Next step: lock make-up analysis, target CoC, silica and chloride ceilings, and discharge limits, then select DAF pretreatment, RO/DTRO, or MVR against those constraints.
