Liquid-to-Chip vs DLC vs D2C: the 3 Liquid Cooling Variants That Confuse 90% of Operators

CDU vendors, hyperscalers, and NVIDIA’s brochures mix three terms as if they were synonyms: liquid-to-chip, direct liquid cooling, and direct-to-chip. The technical difference between the three is on the order of 30 to 80 kW of capacity per rack, and choosing the wrong one for your cluster can mean your site ends up with 2 MW of capacity when you actually need 3.5 MW, or vice versa. This article breaks down the 3 architectures, when each one makes sense, and the objective criteria to avoid falling for vendor marketing.

The 3 Architectures (Without the Marketing Jargon)

The three variants differ in where the liquid touches the silicon and which server components are being actively cooled. The detail is subtler than it looks.

Liquid-to-Chip (L2C): a Cold Plate on the CPU, Everything Else on Air

Liquid-to-chip is the most conservative variant. A cold plate is installed on the CPU (sometimes also on the chipset or the VRM), and the rest of the server components (memory, disks, NICs, motherboard VRMs) keep being air-cooled. The liquid enters the cold plate and removes the CPU’s heat, and the rest of the server operates inside the enclosure like a traditional server.

Capacity per rack: typically between 25 kW and 40 kW. Above that, the air-cooled components (memory, NVMe) start to throttle because the hot air coming out of the recirculated cold plate is not enough to cool them. Practical limitation: if your rack is going to 50+ kW, L2C on its own is not enough.

Direct Liquid Cooling (DLC): the Cold Plate Covers CPU + Memory + VRM

DLC is a superset of L2C. Instead of a cold plate only on the CPU, there is an extended cold plate that covers the CPU, the memory modules (DIMM), and the voltage regulator modules (VRM). The liquid removes the heat of the entire compute block, not just of the CPU.

Capacity per rack: typically between 45 kW and 80 kW. Memory and VRMs, which are the main heat generators after the CPU, are now actively cooled. The components that stay outside the cold plate (disks, motherboard, NICs) are typically low-consumption, so the enclosure air handles them without any problem.

Direct-to-Chip (D2C): Extended Cold Plate, GPU and HBM Included

D2C is the densest variant, and it is what NVIDIA calls “direct-to-chip” in its DGX H100/H200 architectures. The cold plate extends over the CPU, the GPU (or multiple GPUs), and the HBM memory that is co-packaged with the GPU. In a DGX server with 8 H100 GPUs, there are 9 cold plates (8 for GPUs + 1 for the CPU).

Capacity per rack: typically between 80 kW and 130 kW for high-density servers. Servers with B200 or next generations can reach 150+ kW per rack. The practical limitation is no longer thermal but electrical: at 130 kW per rack, 60A breakers are not enough, and you need high-capacity PDUs (generally 415V three-phase direct to the rack).

The 5 Operational Differences That Matter

1. Cost of the CDU and the Distribution Loop

An L2C system needs less liquid flow (lower flow rate) and can run with smaller CDUs. A D2C system needs 3 to 5 times more flow, which requires larger pumps, larger-diameter piping, and higher-capacity CDUs. The CAPEX difference of the cooling system is typically 1.5x to 2.5x between L2C and D2C for the same number of racks.

2. Leak Risk

More cold plates = more potential leak points. A D2C system in a DGX rack has 9 quick-disconnects per server. If each one has a failure rate of 0.1% per year, in a cluster of 100 servers you have 90 failure points. The correct practice is to install humidity sensors on the rack floor, and secondary containment (drip tray) on each server. But the operational risk is real and has to be budgeted.

3. Compatibility With Existing Servers

L2C works with standard servers that have a cold plate on the CPU socket, which is most of the market. DLC and D2C require servers designed for an extended cold plate, which is a subset of the market (typically high-density OEM servers such as Dell PowerEdge XE, HPE Cray, Supermicro, Lenovo ThinkSystem, or NVIDIA DGX). If your cluster is heterogeneous with general-purpose servers, L2C is the only realistic option; DLC and D2C require a hardware upgrade.

4. Energy Efficiency of the Cooling

In a well-designed system, D2C is marginally more efficient in PUE than L2C because the heat is extracted at a higher liquid temperature (CPU+GPU+HBM vs CPU only), which allows free cooling for longer. The difference is 0.01 to 0.03 in PUE, which looks marginal but on a 10 MW site it is 100 to 300 kW of continuous savings.

5. Maintenance Complexity

L2C is the simplest: swapping a server is disconnecting 2 hoses and disconnecting power. D2C requires a procedure to drain the cold plate before removing the server (or having a bypass system that isolates the server from the loop without draining the entire cluster). DLC is in between. Operational complexity impacts the MTTR (Mean Time To Repair) of planned changes.

When Each One Makes Sense (Decision Rule)

1. L2C: heterogeneous clusters with general-purpose servers, density between 20-40 kW per rack, tight cooling budget. It is the most conservative option and works with almost any modern server.
2. DLC: dense general-purpose compute clusters (HPC, enterprise virtualization, VDI), 40-70 kW per rack, servers designed for an extended cold plate. A balance between density and compatibility.
3. D2C: GPU clusters for AI (training or inference at scale), 80-130 kW per rack, DGX servers or equivalents. It is the only option that supports that density. The investment is high, but there is no alternative.

The 3 Most Common Mistakes When Choosing

4. Buying D2C when you only need L2C. The most expensive mistake. If your target density is 35 kW per rack, L2C is enough. Buying D2C for 35 kW leaves you with inflated CAPEX in CDU, pumping, piping, and containment, with no operational benefit.
5. Assuming that a “liquid-cooled” server is D2C. Many servers on the market are sold as “liquid cooling ready” but come with a cold plate only on the CPU. That is L2C, not D2C. Check the datasheet: if the cold plate only covers the CPU, it is L2C. If it covers CPU+memory or CPU+GPU, it is DLC or D2C respectively.
6. Undersizing the CDU for D2C. The most expensive technical mistake. A DGX H100 server consumes 10-11 kW of cooling per unit. A rack with 4 of those servers consumes 40-44 kW. If you size the CDU at 80% of that “to have margin”, you come up short when the cluster reaches 80% load. Size for the real sustained peak, not for the nominal value plus margin.

The Verdict

The three architectures (L2C, DLC, D2C) are not interchangeable. Each one is optimized for a different density range and a different server type. The right decision depends on three variables: your target density per rack, the type of servers you operate or will operate, and the CAPEX available for the cooling system. If your workload is heterogeneous with a mix of CPU and GPU, L2C + DLC in the GPU racks is the most common combination in production. If your workload is 100% GPU for AI, D2C is the only answer. And if your density is below 25 kW per rack, none of the three is right for you; well-designed air is still cheaper and simpler.

Sources

ASHRAE TC 9.9 — Thermal Guidelines for Data Processing Environments (5th Edition, 2021, liquid cooling classifications) —.https://www.ashrae.org/technical-resources/bookstore/thermal-guidelines-for-data-processing-environments.

NVIDIA — DGX H100/H200 System Design Guide (D2C architecture, cold plate specifications) —.https://docs.nvidia.com/dgx/.

Open Compute Project — Liquid Cooling Specifications (open standard for DLC/D2C) —.https://www.opencompute.org/.

Uptime Institute — Data Center Site Infrastructure Tier Standard (liquid cooling considerations) —.https://uptimeinstitute.com/resources/asset/tier-classification-system.

TIA-942-B-2023 — Telecommunications Infrastructure Standard for Data Centers (liquid cooling in Section 6) —.https://www.tia.org/standards.