Direct-to-Chip Liquid Cooling for Data Centers: The Developer's Decision Guide

Single-phase cold-plate liquid cooling has become the default specification for AI-era data centers, and developers who understand it will make better decisions on infrastructure, site, and budget.

Air cooling had a long run. For rack densities below 10kW, it remained cost-effective and operationally familiar. The combination of fans, CRAC units, hot-aisle containment, and chilled water coils handled what the market required.

AI inference and training workloads effectively ended that era. NVIDIA H-series GPU racks push above 80kW. Next-generation AI accelerator clusters are being spec'd at 100-150kW per rack, with 500kW racks under development. No CRAC configuration removes that heat reliably. The question is not whether to deploy liquid cooling for AI-scale data centers. The question is which type, when, and what it requires of the site.

The Two Main Options

Direct-to-chip (DTC) liquid cooling circulates coolant through cold plates mounted directly on CPUs and GPUs. The heat transfers from the chip into the liquid, which carries it to a coolant distribution unit (CDU) and then to a heat rejection system (dry coolers or cooling towers). The server chassis still exists. The server still has a cabinet. The physical data center looks broadly similar to an air-cooled facility from the outside.

Immersion cooling submerges entire servers in a dielectric fluid inside purpose-built tanks. There are no traditional server cabinets. The heat transfers directly from components into the fluid, which circulates to a heat exchanger and secondary loop. Achievable power usage effectiveness (pPUE) approaches 1.01 -- effectively all electrical energy goes to compute rather than cooling.

In 2026, single-phase DTC cold plates account for approximately 55% of the liquid cooling market. Immersion remains under 10% of installations by unit count. The market preference is not accidental.

Why Single-Phase DTC Wins for Most Developers

Schneider Electric's current guidance describes single-phase DTC as "the default choice likely to dominate the next 5-10 years." The reasons are practical:

Retrofit compatibility. DTC cold plates can be introduced row by row into existing data centers, tied into existing chilled-water plants through CDUs. Immersion requires dedicated tanks, reinforced slabs for fluid weight, drainage systems, lifting equipment for server maintenance, and often separate building areas. DTC integrates with the building you already have. Immersion requires a different building.

Vendor and hardware ecosystem. Major OEMs -- Dell, HPE, Lenovo, and OCP platform partners -- offer factory-integrated DTC configurations that carry standard warranties and serviceability paths. Immersion cooling requires hardware validated for specific fluids, and warranty coverage varies significantly by OEM. For developers who do not control the tenant's hardware selection, DTC is the safer infrastructure specification.

Operational familiarity. Data center operators trained on air-cooled infrastructure can transition to DTC with manageable retraining. The fundamental data center structure (cabinets, hot/cold aisles, structured cabling) remains recognizable. Immersion changes the operational model significantly: servers are lifted from tanks, handled wet, and returned to fluid. Staff training, PPE requirements, maintenance procedures, and spare-parts strategy all change.

Warm-water operation. Modern single-phase DTC systems operate with supply water temperatures around 40-45C. That temperature range enables free cooling via dry coolers across a wide range of climate conditions, eliminating or dramatically reducing chiller use. A well-designed DTC facility can achieve overall PUE of 1.05-1.20, versus 1.5 or higher for legacy air-cooled designs and comparable to immersion in many climates.

When Immersion Makes Sense

Immersion cooling is appropriate for specific scenarios, not general deployment:

Ultra-high density greenfield builds. When targeting rack densities above 150kW, or when space constraints limit the number of racks that can achieve the required compute density with DTC, immersion's thermal performance advantage justifies the operational complexity.

Heat reuse integration. Immersion systems can maintain loop temperatures suitable for district heating, greenhouse heating, or industrial process heat recovery. In cold climates with nearby heat offtakers, the economics of heat reuse can offset the higher infrastructure cost.

Highly standardized hardware environments. Operators who control the hardware specification and can lock in immersion-validated server SKUs across a large fleet reduce the compatibility risk. Hyperscalers running uniform, purpose-built hardware are better positioned than colocation operators who must support diverse tenant configurations.

For most developers in 2026, DTC is the specification and immersion is reserved for specific halls or campus zones where its advantages are clearly justified by the workload and site context.

Site Criteria Changes

Specifying liquid cooling changes what you need from a site.

Water access becomes a design variable. Air-cooled data centers have a water story (cooling tower makeup) but not a water supply story. DTC with dry coolers in a temperate climate can minimize water consumption significantly. DTC with evaporative cooling or hybrid towers still requires water. The site's water rights, municipal supply capacity, and drought exposure need to be evaluated as part of the infrastructure design decision, not separately.

Heat rejection infrastructure replaces cooling aisle. DTC facilities need exterior CDUs, dry coolers, or cooling towers sized for the total liquid loop load. The mechanical yard footprint, setback requirements for dry coolers, and structural support for roof-mounted equipment are site planning inputs from the earliest design stages.

Power topology stays substantially the same. One advantage of DTC over immersion is that the electrical distribution design -- switchgear, transformers, power paths to racks -- remains similar to an air-cooled equivalent at the same IT load. Rack-level power delivery changes (480V feeds and higher-current circuits for 100kW racks), but the macro power design framework does not require rearchitecting.

Climate matters more. DTC with warm-water operation (supply at 40-45C) allows free cooling via dry coolers across most of the year in temperate climates. Northern markets with more cooling hours below ambient gain a measurable PUE advantage. Hot, arid markets may require supplemental cooling during peak summer periods. The climate analysis belongs in feasibility, not in operations.

Infrastructure Requirements

The two-loop architecture. Standard DTC design uses a facility loop (water or glycol from heat rejection to CDUs) and an IT loop (higher-purity fluid from CDUs to rack manifolds and cold plates). Variable-flow pumps and variable-speed drives optimize efficiency at partial loads. N+1 redundancy on pumps and CDUs is minimum for Tier III equivalent.

Rack and manifold design. DTC racks require supply and return manifolds with drip-less quick-connect fittings at each cold plate connection. Row-level distribution manifolds feed individual racks. The manifold design, pressure ratings, and material choices (typically copper or stainless steel) affect serviceability and longevity. Blanking panels and supplemental air management remain relevant for residual heat loads from components not covered by cold plates (switches, storage, ancillary equipment).

Leak detection. Liquid in a data center creates risk that air-cooled environments do not have. Continuous monitoring of supply and return temperatures, pressure, and flow rate at manifold and rack level is standard. Leak detection sensors in subfloor, rack, and manifold locations should be integrated into the DCIM/BMS system with automated isolation and alarm logic.

Heat rejection sizing. Dry coolers sized for the IT loop capacity plus a 20-25% margin for peak ambient conditions, future expansion, and degradation over time. In markets where summer ambient temperatures approach the supply water temperature, redundant cooling capacity or supplemental chilling becomes necessary. The mechanical engineer's climate analysis should drive this, not the equipment vendor's standard spec sheet.

Common Specification Mistakes

Sizing cold plates to steady-state load. GPU clusters do not draw constant power. Peak transient loads during model inference bursts or training checkpoint writes can exceed average load by 30-50%. Cold plate selection and loop sizing should account for peak, not average, heat load.

Ignoring component coverage. Cold plates on CPUs and GPUs address the majority of rack heat load, but VRMs, memory, NICs, and storage can represent 15-25% of total heat in dense AI configurations. Advanced DTC designs include auxiliary cold plates or conductive paths for these components. The specification should define coverage requirements explicitly, not leave it to the server OEM.

Treating DTC as a complete cooling solution. Even fully DTC-cooled racks have residual air-side loads from uncooled components. Hot-aisle containment, basic airflow management, and CRAC/CRAH units sized for the residual load remain necessary. Removing air-side infrastructure entirely in a DTC deployment creates hot spot risks around non-liquid-cooled equipment.

Skipping warm-water design. Many early DTC deployments connected to chilled-water plants at 7-12C supply temperatures because it was the easy integration path. Warm-water operation at 40-45C supply eliminates chiller runtime in most climates and enables heat reuse. Retrofitting a cold-water DTC installation to warm-water operation requires CDU and heat exchanger changes. Design for warm water from the start.

The Bottom Line for Developers

DTC liquid cooling should be the default specification for any data center designed for AI workloads in 2026. The technology is proven, the vendor ecosystem is mature, the operational model is familiar, and the site and infrastructure requirements are manageable within standard development workflows.

Immersion cooling is a specific tool for specific conditions. Committing a general-purpose data center campus to immersion without a clear operational plan, hardware strategy, and heat reuse rationale adds cost and complexity without proportionate benefit.

The developers and owners who get ahead in AI data center delivery are those who treat liquid cooling not as a specialty topic but as a baseline infrastructure competency, one that belongs in feasibility analysis, site selection criteria, and construction specifications from the first page of the program.