Written by: Rita

Trend Guide

Bernstein released an in-depth introductory report on liquid-cooled cold plates on July 5th. The core judgment is clear: cold plates are the most unobtrusive but widely used component in liquid cooling systems for AI data centers, benefitting in the short term from strong demand for increased liquid cooling penetration. However, in the medium to long term, they face the risk of commoditization or even elimination, with the business model being far less attractive than that of CDU (Cooling Distribution Unit) in the same system.

A GB200 NVL72 cabinet requires 108 cold plates, which is a massive quantity, but the manufacturing threshold for cold plates is not high; once the design matures, it is easy to become a bulk commodity, and the service revenue is almost zero, requiring direct replacement rather than repair. Looking further down the line, new technologies like silicon micro-channel etching could potentially eliminate the cold plate category altogether.

Cold Plates are the Only Component in the Liquid Cooling System that Contacts Chips

The two core components of the liquid cooling system are the CDU and cold plates. The CDU is responsible for pumping the cooling liquid out, bringing it back, cooling it down, and pumping it out again; it is the heart of the system. The cold plates are responsible for adhering to the surfaces of GPUs and CPUs, using the cooling liquid in internal micro-channels to carry heat away, making them the only components that physically contact the chips.

A NVIDIA GB200 NVL72 cabinet has 18 compute trays, each containing 4 Blackwell GPUs and 2 CPUs, totaling 72 GPUs and 36 CPUs, each paired with a cold plate, resulting in 108 cold plates per cabinet. As the Rubin architecture fully transitions to liquid cooling, the demand for cold plates will continue to grow.

The key performance indicators for cold plates include five metrics: thermal capacity, heat flow density, pressure drop, thermal resistance, and internal channel geometry. Currently, mainstream cold plates have a single-chip heat dissipation capability of about 2 to 5 kilowatts, with leading manufacturers having released products capable of 15 kilowatts. The target for heat flow density is over 100 to 130 watts per square centimeter, with high-end designs aiming for over 150 watts.

The Short-Term Demand for Cold Plates is Certain, but the Business Model is Weak

The cold plate market is currently about 2 to 3 billion dollars, with Bernstein estimating growth to 6 to 7 billion dollars by 2030 under a baseline scenario, implying an annualized growth rate of about 20% to 30%.

The incremental growth comes from three areas: the penetration rate of liquid cooling in new installations transitioning from the current 30% to 40% towards an absolute dominance by 2030; the power density of single cabinets continuously climbing with GPU iteration; and NVIDIA having made it clear that the Rubin architecture will be 100% liquid cooled, without fans.

However, the commercial model of cold plates has structural weaknesses. If cold plates break, they are replaced rather than repaired, typically in bulk. Opening a liquid cooling loop contaminates the thermal interface material on all cold plates, making it more cost-effective to replace them all rather than just one. Maintenance is the responsibility of the server OEMs, with cold plate manufacturers only earning hardware gross margins and having almost no service revenue. This is in stark contrast to the CDU, which has multiple repairable components such as pumps, heat exchangers, and control logic, with a longer lifecycle and a higher service revenue proportion.

Bernstein describes the cold plate business as “design-driven, manufacturing outsourced,” with several cold plate manufacturers being design companies themselves, outsourcing production to third parties. This means that it is difficult to maintain high hardware gross margins in the long term.

Commoditization and Technological Replacement are Long-Term Hurdles

There are two long-term risks for cold plates.

The first is commoditization. The OCP (Open Computing Project) has defined operational performance standards for cold plates, including allowable pressure drops, working temperature ranges, leakage rates, and stress limits for chip packaging but has not specified a particular design path. Manufacturers can freely design internal channel structures as long as they meet the standards, meaning that design itself is currently the source of value. However, once performance standards stabilize and GPU thermal design power no longer rises rapidly, the designs of cold plates from different manufacturers will converge, shifting value from design to manufacturing, compressing hardware gross margins. Bernstein believes this time window spans from 2028 to 2030.

The second risk is technological replacement. Directly etched silicon cooling technology (such as Microsoft’s investment in the Corinth project) could potentially replace cold plates by integrating micro-channels at the chip packaging level, eliminating the need for cold plates altogether. This technology is still in its early stages, but once commercialized, the cold plate category could be uprooted. Bernstein assesses the risk of this development as “moderate,” with timelines extending beyond 2030.

Bernstein's Position: Optimistic About Equipment Stocks, Cold Plates are Just a Transition

Bernstein's ratings for cold plate-related stocks are clear: Vertiv (VRT) target price is $416, nVent (NVT) target price is $218, Eaton (ETN) target price is $534, Schneider (SBGSY) target price is $310, Trane (TT) target price is $550, and Johnson Controls (JCI) target price is $176, all expected to outperform the market. Carrier (CARR) has a target price of $75, in line with the market.

These seven stocks are primarily giants in electrical equipment and HVAC, covering multiple aspects such as CDU, power distribution, and thermal management, with cold plates being just a small part of their exposure to AI infrastructure. Bernstein is not truly optimistic about cold plates themselves, but rather about the segments within the entire liquid cooling system that have service revenues and technological barriers.

Trend Perspective

The most valuable point of Bernstein’s report is its dissection of a pattern through the cold plate component: the part most lacking in AI infrastructure is not necessarily the best business.

Both CDU and cold plates belong to liquid cooling systems, but their business models are worlds apart. CDU is complex, has a high unit price, includes service revenue, and enjoys strong customer loyalty. Cold plates are simple, have a low unit price, lack service revenue, and are simply replaced when broken. The segments with the largest demand are often not the ones with the highest profit margins. For investors, when understanding AI hardware, it is essential to distinguish which segments earn money from scarcity and which earn from volume. Volume-based revenue comes quickly but can also vanish quickly.

Disclaimer

This article is a compilation and interpretation of third-party brokerage research reports by Trend Research. The ratings, target prices, profit forecasts, and related judgments quoted in the text are solely the views of the analysts from that brokerage firm and do not represent the views of Trend Research, nor do they constitute any investment advice.

The market carries risks, and decisions should be made independently. This article should not serve as the basis for buying or selling any securities.