Direct liquid cooling becomes essential specification for data centers running high-power AI accelerators like NVIDIA Blackwell.

The reality of AI infrastructure has arrived: power-hungry chips such as the NVIDIA Blackwell have rendered traditional air cooling obsolete. With rack densities climbing from 20 kW to over 140 kW and the 1 MW milestone approaching, the industry is transitioning from speculative planning into an urgent, high-stakes deployment phase. Direct liquid cooling (DLC) has become the preferred cooling method for extreme chip power densities across the IT industry.

Specifying a DLC system requires balancing liquid-cooled IT requirements against specific data center conditions. The combination of liquid-cooled IT with traditional heat rejection systems presents distinct challenges. Successful DLC implementation begins with rigorous specification—here are five common obstacles operators face.

**Material incompatibility and galvanic corrosion.** When two metals with different electrochemical potentials connect in the presence of a fluid, galvanic corrosion can occur, degrading materials and creating debris that clogs cold plates and damages servers. This risk intensifies when materials vary between the CDU and connected components.

To minimize risk, keep the anodic index difference small across all wetted surfaces. Follow IT manufacturer guidelines strictly and consult acceptable materials lists from organizations like the Open Compute Project (OCP) or the American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE). Maintain a comprehensive material registry, avoid aluminum unless using proper inhibitors, and consider separate cooling loops if cold plate materials differ across IT vendors.

**Competing liquid and air-cooling economics.** Liquid cooling permits higher chilled water temperatures than air cooling, delivering substantial efficiency gains: for every 0.6°C (1°F) increase in chilled water temperature, chiller efficiency rises one to two percent—two to four percent for variable speed chillers. Additional savings come from economizer hours when the chiller operates at partial load or cycles off, a benefit particularly significant in cooler climates.

Maximizing these savings typically requires investing in a separate high-temperature chiller plant dedicated to liquid-cooled loads, while maintaining an existing low-temperature chiller for air-cooled equipment. Alternatively, operators can accept reduced energy savings by sharing a single low-temperature chiller across both loads. The optimal choice depends on a 10-year total cost of ownership (TCO) analysis informed by local climate: hotter regions generally justify a dedicated high-temperature chiller due to increased economizer availability, while colder climates favor a shared low-temperature chiller.

**Direct coupling between servers and cooling infrastructure.** Unlike air cooling—where ducts may direct air to a single rack but remain loosely coupled to individual servers—DLC systems physically connect coolant distribution piping directly to IT equipment. This tight coupling introduces new variables around space, leak risks, and specific coolant flow requirements.

Distribution piping must connect the CDU to each server's cold plates. Work closely with IT and infrastructure vendors to specify compatible quick-connect couplings and dedicate a single cooling loop to each server model to prevent uneven cooling. While liquid cooling is more effective than air, it carries the risk of catastrophic leaks. Mitigate this through direct and indirect leak detection systems—such as cable detectors and turbidity sensors—monitored by centralized control systems.

Liquid-cooled servers depend on centralized CDU pumps and controls, requiring TCS designs calibrated precisely to each server's specifications. Manage flow resistance using constant differential pressure control with energy valves to stabilize pressure across racks.

**Absence of CDU efficiency standards.** Without standardized testing protocols, vendors set their own variables for fluid type, flow rates, fluid temperature, and ambient temperature when rating CDU efficiency or capacity. This allows vendors to claim optimal performance under favorable conditions while making meaningful comparison of different CDUs nearly impossible.

Air cooling benefits from established standards; liquid cooling does not. Until ASHRAE finalizes official testing standards for liquid cooling, request vendor performance ratings based on draft test points. From a design perspective, limit CDU capacity to no more than 10–20 percent of your liquid-cooled servers per loop, minimizing the blast radius if a failure occurs.

**Provisioning for unknown liquid-cooled IT.** Data centers often must provision IT space for liquid-cooled equipment without knowing the specific hardware that will eventually arrive—a common scenario for colocation providers preparing for new tenants. Unlike air cooling, DLC directly couples coolant supply to each server, meaning every CDU and its TCS loop must support both maximum racks at the lowest density and minimum racks at the highest density. Wide density ranges risk oversizing piping or stranding cooling capacity.

Mitigate this by sizing main loop piping diameter for bulk CDU capacity and making it long enough to accommodate the maximum number of racks at the lowest density. However, size branch piping and manifolds for the highest potential rack density. Alternatively, designate distinct data center areas for different power densities or develop reference designs for two or three DLC "pods" at varied densities that can be finalized once specific IT equipment is confirmed.

Moving from conceptual pilots to robust, production-ready liquid cooling environments demands careful attention to specification challenges before proceeding to installation and operations. With disciplined specification practices and the right technical partner, operators can mitigate the inherent risks of high-density cooling and establish the foundation for sustainable, high-performance AI growth.