Bernstein Research Report Analysis: AI Liquid Cooling Market to Triple in 4 Years, Cold Plates Face Commoditization Risk
TechFlow Selected TechFlow Selected
Bernstein Research Report Analysis: AI Liquid Cooling Market to Triple in 4 Years, Cold Plates Face Commoditization Risk
The components most lacking in AI infrastructure are not necessarily the best business.
Navigating Web3 tides with focused insightsWritten by: Rita
TechFlow Guide
On July 5, Bernstein released an in-depth primer report on liquid cold plates. The core judgment is straightforward: cold plates are the most unassuming but most heavily used component in AI data center liquid cooling systems. While demand is strong in the short term due to increased liquid cooling penetration, they face risks of commoditization or even obsolescence in the medium to long term, with a business model far less attractive than the CDU (Coolant Distribution Unit) within the same system.
A single GB200 NVL72 cabinet requires 108 cold plates, representing huge volume, but the manufacturing threshold for cold plates themselves is not high. Once the design matures, they easily become commodities, with almost zero service revenue; if broken, they are replaced directly, not repaired. Looking further long-term, new technologies such as silicon-based microchannel etching could eliminate the cold plate category entirely.
Cold Plates Are the Only Components in Liquid Cooling Systems That Contact Chips
The two core components in a liquid cooling system are the CDU and the cold plate. The CDU is responsible for pumping the coolant out, collecting it, cooling it down, and pumping it out again; it is the heart of the system. The cold plate is attached to the surface of GPUs and CPUs, using coolant in internal microchannels to carry away heat, and is the only component physically contacting the chips.
An NVIDIA GB200 NVL72 cabinet has 18 compute trays, each with 4 Blackwell GPUs plus 2 CPUs, totaling 72 GPUs plus 36 CPUs. Each is paired with one cold plate, totaling 108 per cabinet. As the Rubin architecture fully shifts to liquid cooling, cold plate usage will continue to expand.
There are five key performance indicators for cold plates: thermal capacity, heat flux density, pressure drop, thermal resistance, and internal flow channel geometry. Currently, mainstream cold plates have a single-chip heat dissipation capacity of about 2 to 5 kilowatts, with leading manufacturers having released 15-kilowatt products. The heat flux density target is above 100 to 130 watts per square centimeter, with high-end designs moving towards above 150 watts.
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. Under Bernstein's base case, it will grow to $6 to $7 billion by 2030, implying an annualized growth rate of about 20% to 30%.
The incremental growth comes from three levels: liquid cooling penetration in new installations is moving from the current 30% to 40% towards absolute dominance by 2030; single-cabinet power density continues to climb with GPU iterations; NVIDIA has confirmed the Rubin architecture will be 100% liquid cooling, without fans.
However, the cold plate business model has structural shortcomings. When cold plates break, they are not repaired but replaced directly, usually in batches. Opening a liquid cooling loop once will contaminate the thermal interface material of all cold plates; replacing one individually is less cost-effective than replacing all of them. Maintenance is handled by server OEMs; cold plate manufacturers only earn hardware gross margins with almost no service revenue. This is completely different from CDUs, which have multiple repairable components such as pumps, heat exchangers, and control logic, have longer lifecycles, and a higher proportion of service revenue.
Bernstein describes the cold plate business as "design-driven, manufacturing outsourced." Many cold plate manufacturers are design companies themselves, outsourcing production to third parties. This means hardware gross margins are difficult to maintain at high levels in the long term.
Commoditization and Technological Substitution Are Two Long-Term Hurdles
There are two long-term risks for cold plates.
The first is commoditization. The OCP (Open Compute Project) has defined operational performance standards for cold plates, including allowable pressure drop, operating temperature range, leakage rate, and chip package stress limits, but has not specified specific design paths. Manufacturers can freely design internal flow channel structures while meeting standards, meaning design itself is the current source of value. However, once performance standards stabilize and GPU thermal design power stops climbing rapidly, cold plate designs from different manufacturers will converge, value will shift from the design end to the manufacturing end, and hardware gross margins will compress accordingly. Bernstein believes this time window is between 2028 and 2030.
The second is technological substitution. Direct etched silicon cooling technology (such as the Corinth project invested by Microsoft) could directly replace cold plates, integrating microchannels at the chip packaging level, eliminating the need for the cold plate component entirely. This technology is currently in early stages, but once commercialized, the cold plate category will be uprooted. Bernstein's judgment on this risk is "moderate," with a time dimension after 2030.
Bernstein's Stance: Bullish on Equipment Stocks, Cold Plates Themselves Are Just Transitional
Bernstein's ratings on cold plate-related targets are clear: Vertiv (VRT) target price $416, nVent (NVT) target price $218, Eaton (ETN) target price $534, Schneider (SBGSY) target price $310, Trane (TT) target price $550, Johnson Controls (JCI) target price $176, all Outperform. Carrier (CARR) target price $75, in line with the market.
These seven targets are basically electrical equipment and HVAC giants, with businesses spanning CDU, power distribution, thermal management, and other links; cold plates are just a small part of their AI infrastructure exposure. What Bernstein is truly bullish on is not the cold plates themselves, but those links within the entire liquid cooling system that have service revenue and technical barriers.
TechFlow Perspective
The most valuable point of this Bernstein report is using the cold plate component to dissect a rule: the most scarce component in AI infrastructure is not necessarily the best business.
CDUs and cold plates belong to the same liquid cooling system, but their business models are worlds apart. CDUs are complex, have high unit prices, generate service revenue, and have strong customer stickiness. Cold plates are simple, have low unit prices, generate no service revenue, and are replaced when broken. The link with the highest demand is often not the link with the highest profit margin. For investors, when understanding AI hardware, it is necessary to distinguish clearly which link earns money from scarcity and which link earns money from volume. Volume money comes fast and goes fast.
Disclaimer
This article is a compilation and interpretation by TechFlow Research of third-party brokerage research reports. The ratings, target prices, earnings forecasts, and related judgments cited in the text are the views of the brokerage analysts, represent only the position of their affiliated institutions, do not represent the views of TechFlow Research, and do not constitute any investment advice.
The market involves risks; decisions must be independent. This article should not be used as a basis for buying or selling any securities.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@chaoxiangooo
