Cold plates are a key component in AI data center liquid-cooling conversion, but long-term moats are limited
AI summary card
Cold plates are a key component in AI data center liquid-cooling conversion, but long-term moats are limited
The report expects the cold plate market to grow from about $2-3 billion today to about $6-7 billion by 2030, with near-term benefits from increasing DTC liquid-cooling adoption, but commoditization and replacement risk rising as specs standardize, manufacturing scales, and alternative cooling technologies mature.
- Cold plates directly touch GPUs/CPUs and transfer heat to coolant, forming the core heat transfer component in DTC liquid-cooling architectures.
- NVIDIA Rubin’s push toward 100% liquid cooling, combined with hyperscalers introducing liquid cooling, is expected to support strong cold-plate demand from 2026 to 2028.
- Each NVIDIA GB200 NVL72 rack needs about 108 cold plates; cold plates are high-usage items, but the structure is relatively simple and service revenues are limited.
- The base case in the report estimates the cold-plate market at about $6-7 billion by 2030, versus about $2-3 billion currently.
- Longer-term risks include two-phase DTC standardization across the supply chain after maturity, falling $/kW pricing of cold plates, and direct silicon-etched microchannel cooling potentially reducing or eliminating cold-plate demand.
Report interpretation
Overview
This report is a deep-dive primer by Bernstein on cold plate technology for liquid cooling in the U.S. multi-industry and electrical equipment space. Starting from rapidly rising thermal density in AI data center racks, it explains why traditional air cooling is no longer sufficient for modern GPU/CPU clusters, and why direct-to-chip liquid cooling is becoming a key architecture, with cold plates serving as the core component that transfers chip surface heat into coolant. The report also presents market size, competitive landscape, technical metrics, and investment implications: near-term cold-plate demand has relatively high visibility, but long-term commoditization, low service attach rates, and substitution technologies need to be monitored.
Core views
The core view is that cold plates have strong incremental potential from 2026 to 2028 as liquid-cooling penetration rises, with platforms such as NVIDIA Rubin driving a full-liquid-cooling ecosystem, and because each high-density AI rack requires a large number of cold plates. However, unlike CDUs, cold plates have lower system complexity and service attachment, with most value concentrated in design and manufacturing. By 2028–2030, as two-phase DTC matures, supply chains expand, and hyperscalers progressively standardize specs, innovation and scarcity premiums may decline; after 2030, if solutions like direct silicon-etched microchannel cooling commercialize, cold plates may face partial substitution or even long-term structural obsolescence risk.
Analysis framework
The report combines technical teardown, value-chain comparison, and a top-down market-sizing methodology. On the technical side it breaks down key metrics such as thermal capacity, heat flux, pressure drop, thermal resistance, and channel geometry; on the industry side it compares cold plate suppliers Boyd, CoolIT, CoolerMaster, and Accelsius, along with HVAC/electrical players JCI, Carrier, Eaton, and Schneider; for market sizing, it uses incremental data center GW, per-GPU power draw, cold plate $/kW assumptions, and replacement/repair demand as key drivers.
Methodology notes
Estimate market size from incremental data center GW and per-kW cooling value of cold plates
The report starts from incremental compute and rack power density, combines GPU power draw, liquid-cooling penetration, cold-plate heat-removal capability, and $/kW assumptions to derive the cold-plate market size for 2030.
Thermal capacity, heat flux, pressure drop, thermal resistance, and channel geometry
Cold-plate performance depends on how much heat it can remove, whether it can manage local hot spots, whether pumping resistance is manageable, heat-transfer efficiency from chip to coolant, and internal channel design.
Difference between hardware margins and service income
The report compares cold plates with CDUs, arguing that cold plates are structurally simpler and are typically replaced rather than repaired over lifecycle, resulting in limited service attach and making long-term margins more vulnerable to commoditization pressure.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- VRTCompany exposed to data center liquid cooling and electrical thermal management; the report gives it an outperformance rating.
- Strengths
- Target price $416, benefiting from expansion of AI data center infrastructure and liquid-cooling ecosystem development.
- Weaknesses
- If hardware segments such as cold plates become commoditized, profitability of single components may come under pressure.
- Comparison
- Compared with pure hardware cold-plate suppliers, Vertiv is more exposed to system-level data center infrastructure.
- Risks
- AI capex volatility, slower-than-expected liquid-cooling rollout pace, and intensified competition.
- NVTCompany related to electrical interconnects and rack infrastructure; the report gives it an outperformance rating.
- Strengths
- Target price $218, potentially benefiting from higher-density data center electrical and thermal-management demand.
- Weaknesses
- The direct exposure to cold plates should be assessed based on its specific product mix.
- Comparison
- More focused on electrical equipment and rack ecosystem exposure; cold plates are not the only driver.
- Risks
- Data-center order cycles, valuation pullback, and supply-chain competition.
- CARRHVAC and thermal-management related company; the report gives it an in-line-with-index rating.
- Strengths
- Target price $75, with experience in cooling and HVAC systems.
- Weaknesses
- The rating is relatively neutral, with limited upside versus current price in the report.
- Comparison
- Compared with outperformance names, the investment appeal is weaker in this report.
- Risks
- Competition in liquid-cooling markets, HVAC business cycle effects, and margin pressure.
- TTTrane has thermal-management and HVAC exposure; the report gives it an outperformance rating.
- Strengths
- Target price $550, potentially benefiting from long-term growth in data center cooling demand.
- Weaknesses
- The stock-to-target gap in the table indicates limited upside or cyclical valuation premium.
- Comparison
- A large, mature industrial company; cold-plate opportunities are likely only part of broader business.
- Risks
- Valuation, execution of orders, and changes in liquid-cooling technology pathways.
- JCIJohnson Controls participates in the liquid-cooling ecosystem and invested in Accelsius; the report gives it an outperformance rating.
- Strengths
- Target price $176, with capabilities in large building systems, HVAC, and data center cooling.
- Weaknesses
- Limited service attachment and manufacturing economics in cold plates may constrain standalone profitability.
- Comparison
- It participates via investments and ecosystem positioning rather than as a pure cold-plate component company.
- Risks
- M&A integration, technology-path changes, and customer concentration.
- ETNEaton acquired Boyd, strengthening cold-plate and liquid-cooling capabilities; the report gives it an outperformance rating.
- Strengths
- Target price $534, with strong exposure to electrical equipment and data center infrastructure.
- Weaknesses
- Commoditization of cold plates could affect long-term margins of the acquired assets.
- Comparison
- Compared with pure cold-plate suppliers, Eaton has a broader electrical equipment platform.
- Risks
- Post-acquisition integration, manufacturing expansion in the supply chain, and a slowdown in AI data center capex.
- SU.FPSchneider Electric has data center electrical and thermal-management ecosystem exposure; the report gives it an outperformance rating.
- Strengths
- Target price €310, benefiting from global data center buildout and electrification demand.
- Weaknesses
- Cold-plate contribution is unclear, with most earnings likely from the broader data center ecosystem.
- Comparison
- Alongside Eaton and Vertiv as beneficiaries in data center electrical equipment.
- Risks
- European industrial demand, exchange rates, competition, and pace of liquid-cooling adoption.
Key data
- Base cold-plate market size in 2030about $6-7BThe report estimates the current market at about $2-3B, rising to about $6-7B in the 2030 base case.
- Liquid-cooling penetration in new data centerscurrently about 30-40%, moving toward majority adoption by around 2030Rising liquid-cooling penetration is one of the main drivers of cold-plate shipment growth.
- Number of cold plates per NVIDIA GB200 NVL72 rackabout 108 units18 compute trays, 72 Blackwell GPUs, and 36 CPUs; each GPU/CPU requires an individual cold plate.
- Current typical cold-plate thermal capacityabout 2-5 kW/chipNext-generation designs can reach above 5-10 kW; CoolIT has previously announced a 15 kW cold plate.
- Current hyperscaler heat-flux targetabout 100-130+ W/cm²Leading designs are moving toward 150+ W/cm² for handling local hot spots on AI chips.
- Current pressure-drop rangeabout 15-25 psi, over 30+ psi for high-performance designsLower pressure drop is more favorable for lowering pump power, but there is a trade-off with high heat-flux designs.
- High-performance cold-plate thermal resistance<0.04°C/WLower thermal resistance means lower chip temperature for the same heat output.
- DOE COOLERCHIPS grant$5 millionNVIDIA and partners received funding to develop advanced liquid-cooling systems.
Impact & implications
The investment implication is to focus in the near term on industrial and electrical equipment companies with exposure to cold plates, liquid cooling, or data center thermal management, as rising AI rack power density makes cold plates a high-usage essential component. But valuation and long-term margins should be discounted: cold plates are not high service-attach businesses, and as customer specs standardize, contract manufacturing ecosystems form, and two-phase DTC matures, hardware margins may compress. Comparatively, companies with system-level capability, strong customer relationships, scale manufacturing, or a more complete liquid-cooling ecosystem are more likely to sustain competitiveness.
Risks
- Cold-plate structures are relatively simple, and as customer specs converge and manufacturing scales, hardware prices and margins may be compressed.
- Cold-plate service attachment is low; under faults, replacements are typically used rather than repair, making it harder for manufacturers to monetize ongoing maintenance in the way CDUs can.
- If two-phase DTC matures in 2028-2030, it may alter current single-phase cold-plate designs and supply-chain dynamics.
- If direct silicon-etched microchannel cooling commercializes, it could reduce or even eliminate some cold-plate demand.
- Immersion cooling is not viewed by the report as a near-term serious mainstream substitute, but it remains an uncertainty in the data center cooling technology path.
- The market-size estimate is highly sensitive to incremental GW, GPU power consumption, liquid-cooling penetration, and $/kW assumptions, so forecasts carry significant sensitivity.
What to watch
- Whether NVIDIA Rubin and subsequent platforms advance a 100% liquid-cooling ecosystem as planned.
- The extent to which hyperscalers standardize cold-plate specifications, supplier lists, and contract-manufacturing ecosystems.
- Commercialization progress, reliability, and supply-chain maturity of two-phase DTC cold plates.
- Whether cold-plate $/kW pricing declines as assumed in the report, and the impact of price compression on supplier gross margins.
- Capacity expansion, customer qualification, and M&A activity at cold-plate players such as Boyd, CoolIT, CoolerMaster, and Accelsius.
- R&D and commercialization progress by Microsoft, TSMC, Corintis, and others in direct silicon-etched microchannel cooling.