Liquid Cooling's Hidden Supplier Network Is Becoming the Real AI Cooling Trade
September 16, 2026
Altsets
Research by Altsets Research
As AI racks shift from room-level air cooling toward direct liquid cooling, more of the thermal-management opportunity is moving into cold plates, CDUs, pumps, manifolds, connectors, heat exchangers, and the suppliers that make those systems work.
Data used:Altsets Supply Chain Intelligence: 90k+ entities, 400k+ relationships, 20+ years of history.
Key findings
- NVIDIA's Vera Rubin reference architecture continues using direct-to-chip liquid cooling even as Jensen Huang said water chillers would not be required for the generation.
- NVIDIA and nVent describe a core liquid-cooling architecture built around CDUs, TCS manifolds, and standardized rack-drop interfaces that can scale from Grace Blackwell into Vera Rubin.
- Bank of America estimated that liquid cooling could rise from roughly 30% of new AI data center installations today to 70% by 2030.
- The supplied article identifies listed exposure across nVent, Delta Electronics, Asia Vital Components, Auras Technology, and Shenzhen Envicool, while avoiding product-level revenue attribution that the evidence does not support.
The most important change in AI cooling is not simply that data centers need more cooling. It is that cooling spend is moving closer to the chip. As NVIDIA pushes increasingly dense rack-scale systems, the critical thermal path is shifting from large room-level chillers and air handlers toward cold plates, pumps, coolant distribution units, manifolds, quick-disconnect connectors, heat exchangers, and fluid-control components. That creates a different group of potential beneficiaries, including nVent, Delta Electronics, Asia Vital Components, Auras Technology, and Shenzhen Envicool.
The distinction became clearer in January when NVIDIA CEO Jensen Huang said the company's Vera Rubin generation would not require water chillers for data centers. Cooling-related HVAC stocks fell after the remarks, while analysts pointed to nVent as a company that could benefit because its exposure is concentrated further inside data center power and liquid-cooling infrastructure rather than conventional space cooling.[1]
That apparent contradiction is the key to the supply-chain story. Eliminating a traditional chiller does not eliminate cooling. It changes where the thermal-management equipment sits.
Cooling is moving from the room into the rack
Direct-to-chip liquid cooling creates a physical chain that looks very different from traditional air cooling. A cold plate sits against a GPU, CPU, or other high-heat component. Coolant carries the heat through hoses and connectors into a rack manifold. Pumps maintain flow and pressure. A coolant distribution unit, or CDU, controls the secondary loop and transfers the heat through a heat exchanger into the facility-side cooling system.
Each additional liquid-cooled rack therefore creates demand for a collection of components that barely existed in conventional server racks at comparable scale.
The thermal path is moving closer to the chip
Direct-to-chip cooling creates a component chain inside the rack and row rather than relying only on room-level HVAC.
- 01GPU, CPU, acceleratorThe heat source sits inside the compute tray.
- 02Cold plateLiquid collects heat directly from the high-power device.
- 03Hoses and quick disconnectsFluid moves between trays, racks, and serviceable interfaces.
- 04Rack manifoldFlow is distributed across the rack.
- 05Pumps and CDUFlow, pressure, temperature, and the secondary loop are controlled.
- 06Heat exchangerHeat transfers into the facility-side cooling system.
Removing a traditional chiller does not remove the cooling stack. It changes which parts of the stack carry more of the thermal-management function.
NVIDIA's own 2026 infrastructure guidance makes this transition explicit. Its Vera Rubin reference architecture continues using direct-to-chip cooling and is designed around CDUs, technology-cooling-system manifolds, and standardized rack-drop interfaces. NVIDIA and nVent have also emphasized that the same basic liquid-cooling architecture can scale from Grace Blackwell into Vera Rubin even as the compute platform changes.[2]
That continuity matters for investors. GPUs have short product cycles. A qualified manifold, connector, pump, cold plate, or CDU architecture that survives several accelerator generations can participate in the buildout without depending on one specific GPU model.
Liquid cooling is a stack of different component businesses
Different parts scale with different physical units, from devices to racks to facility loops.
| Layer | Primary function | What demand scales with |
|---|---|---|
| Cold plates | Move heat from high-power silicon into the liquid loop | Number and thermal density of cooled devices |
| Hoses and connectors | Provide serviceable fluid paths between components | Connections across trays and racks |
| Rack manifolds | Distribute coolant across the rack | Number and configuration of liquid-cooled racks |
| Pumps and valves | Maintain flow, pressure, and control | Loop design, redundancy, and rack deployment |
| CDUs | Control the secondary loop and isolate facility-side water | Rack groups, thermal load, and deployment architecture |
| Heat exchangers | Transfer heat into the facility-side system | Thermal load and facility integration |
The component economics also differ by layer. Cold plates are tied closely to the number of high-power devices being cooled. Manifolds and connectors multiply across trays and racks. Pumps, valves, sensors, and heat exchangers sit higher in the loop and may support entire racks or groups of racks. Coolant itself is technically essential, but in many single-phase systems it is unlikely to represent the same hardware value per installation as the engineered components controlling flow, heat transfer, sealing, and reliability.
This means "liquid cooling" should not be analyzed as one market. It is a stack of separate supply chains.
The less obvious suppliers are already gaining importance
The strongest public evidence is appearing in Asia, where much of the AI server supply chain is concentrated. Reuters reported in September that Delta Electronics, Asia Vital Components and Auras Technology are among the thermal-management suppliers participating in NVIDIA's ecosystem, alongside China's Shenzhen Envicool. Bank of America estimated that liquid cooling could rise from roughly 30% of new AI data center installations today to 70% by 2030.[3]
Bank of America expects liquid cooling to become the majority architecture
Estimated share of new AI data center installations using liquid cooling.
These companies occupy different positions.
Delta Electronics is broader than liquid cooling, with exposure to data center power and thermal infrastructure. Asia Vital Components and Auras sit closer to the thermal hardware around servers, including the cold-plate and related cooling-component layer. Envicool has expanded around CDUs and other data center thermal-management systems. In the United States, nVent provides another listed route into the rack and row-level cooling architecture rather than traditional space cooling.
The investable supplier set extends beyond conventional HVAC
The same AI rack can pull demand through several thermal-management layers.
| Company | Position described in the article | Interpretation boundary |
|---|---|---|
| nVent | Rack and row-level data center liquid-cooling infrastructure | Not presented as a pure-play liquid-cooling company or as an exclusive NVIDIA supplier |
| Delta Electronics | Data center power and thermal infrastructure | Broader exposure than liquid cooling alone |
| Asia Vital Components | Thermal hardware around servers, including the cold-plate layer | No product-level revenue attribution is asserted |
| Auras Technology | Thermal-management hardware in the NVIDIA ecosystem | No product-level revenue attribution is asserted |
| Shenzhen Envicool | CDUs and other data center thermal-management systems | No exclusive supplier relationship is asserted |
Scope and limitations
This article is a presentation pass on the supplied draft. It distinguishes room-level cooling from rack and chip-level liquid-cooling infrastructure and preserves the draft's focus on cold plates, CDUs, pumps, manifolds, connectors, heat exchangers, and fluid-control components.
The listed suppliers should not be read as a complete AI liquid-cooling supply chain. The article does not provide enough evidence to assign product-level revenue exposure, market share, exclusive sourcing, customer concentration, or profit sensitivity to the companies named.
The Bank of America adoption figures are external estimates reported by Reuters. They are not Altsets forecasts. The NVIDIA and nVent architecture discussion describes reference-architecture continuity and does not establish that every Vera Rubin deployment will use an identical bill of materials or supplier set.
The supplied draft ended before its source list and later sections. This version restores only the three source references already implied by the draft and does not invent additional article claims.
For evidence limits and relationship methodology, see the Altsets methodology.
Sources
-
Reuters, "Data center cooling-related stocks drop after Nvidia CEO Huang's comments," January 6, 2026. https://www.reuters.com/business/data-center-cooling-related-stocks-drop-after-nvidia-ceo-huangs-comments-2026-01-06/
-
NVIDIA On-Demand, "Liquid Cooling Readiness for Vera-Rubin Reference Architectures (Presented by nVent)," GTC San Jose 2026. https://www.nvidia.com/en-us/on-demand/session/gtc26-ex82328/
-
Reuters, "Not just Nvidia: these power and cooling firms are riding the trillion-dollar data centre boom," September 1, 2026. https://www.reuters.com/business/energy/not-just-nvidia-these-power-cooling-firms-are-riding-trillion-dollar-data-centre-2026-09-01/
How to Cite This
According to Altsets Supply Chain Intelligence (altsets.com), the AI cooling opportunity is shifting closer to the chip as direct-to-chip liquid cooling moves thermal-management demand toward cold plates, CDUs, pumps, manifolds, connectors, heat exchangers, and fluid-control hardware.
For research inquiries or data access: press@altsets.com
Go Deeper
Browse additional casefiles on AI infrastructure, industrial bottlenecks, and second-order supplier exposure.
Use Altsets guided analysis to compare companies, inspect supplier and customer exposure, and trace dependency paths.
Review relationship definitions, evidence limits, missing-data treatment, and historical methodology.
Get new Altsets research when we publish.
Original supply-chain research, market casefiles, and new data findings.