Liquid Cooling Revolution: Why AI Data Centers Can No Longer Afford Air Cooling

May 2026
Archive: May 2026
AI accelerator power has breached 1000W per chip, rendering traditional air cooling physically inadequate. AINews reports that liquid cooling is accelerating from an optional upgrade to a mandatory infrastructure for next-generation intelligent computing centers, fundamentally reshaping server architecture, cost models, and even data center geography.

The thermal demands of modern AI hardware have reached a tipping point. NVIDIA's B200 GPU, for instance, has a thermal design power (TDP) exceeding 1000W, and racks packed with these chips can easily surpass 100kW. Air cooling, even with high-speed fans, cannot effectively dissipate this heat density without consuming enormous energy and creating hotspots that throttle performance. This is not a marginal improvement cycle; it is a physics-driven infrastructure revolution.

Liquid cooling, particularly direct-to-chip cold plate technology, is now the de facto standard for high-density AI clusters. The transition is not merely about better cooling. It forces a complete rethinking of server motherboard layouts to accommodate coolant distribution, drives data center operators to prioritize water availability alongside power, and fundamentally alters operational expenditure by slashing fan energy consumption. For companies racing to deploy large-scale AI agents, world models, and video generation systems, the choice of cooling technology is no longer a technical preference—it is a binary question of whether their compute infrastructure can survive its own waste heat.

This article provides an in-depth analysis of the liquid cooling revolution, covering the technical underpinnings, key industry players, market dynamics, and the strategic implications for every organization building AI infrastructure today.

Technical Deep Dive

The physics of heat transfer is unforgiving. Air has a specific heat capacity of roughly 1.005 kJ/(kg·K) and a thermal conductivity of about 0.026 W/(m·K). Water, by contrast, has a specific heat capacity of 4.18 kJ/(kg·K) and a thermal conductivity of 0.6 W/(m·K). This means water can carry away heat approximately 4 times more efficiently per unit mass and conduct it over 20 times better than air. When a single GPU like the NVIDIA B200 generates 1000W of heat, the volume of air required to cool it at a reasonable temperature gradient becomes impractical—requiring high-velocity fans that consume significant power and generate noise.

Two primary liquid cooling architectures have emerged:

1. Direct-to-Chip (Cold Plate) Cooling: This is the most widely adopted approach for high-density AI clusters. A cold plate, typically made of copper or aluminum, is mounted directly on the GPU and other high-heat components. Coolant (usually a water-glycol mixture) flows through microchannels inside the plate, absorbing heat directly from the chip. The heated coolant is then pumped to a heat exchanger (CDU - Coolant Distribution Unit) where it transfers heat to a facility water loop, which ultimately rejects it to the outside environment via cooling towers or dry coolers. This method is highly efficient, can handle heat fluxes exceeding 1000 W/cm², and requires minimal modification to existing server form factors.

2. Immersion Cooling: Servers are fully submerged in a dielectric (non-conductive) fluid, such as a fluorocarbon or engineered hydrocarbon oil. The fluid boils at a low temperature (single-phase or two-phase), absorbing heat directly from all components. Two-phase immersion cooling is particularly effective because the phase change from liquid to gas absorbs a massive amount of latent heat. While immersion cooling offers the highest cooling density and eliminates fans entirely, it presents challenges for serviceability, component compatibility (some plastics degrade), and the sheer volume of fluid required.

Engineering Challenges: The transition to liquid cooling demands a redesign of the server motherboard. Coolant distribution manifolds must be integrated, requiring precise alignment of quick-disconnect fittings. Leak detection systems are mandatory—a single leak can destroy millions of dollars of hardware. The coolant itself must be chemically treated to prevent corrosion, biological growth, and scaling. Furthermore, the entire data center plumbing—pumps, valves, pipes, and heat exchangers—must be designed for high reliability, as a pump failure can lead to rapid thermal runaway.

Data Table: Cooling Technology Comparison

| Technology | Typical PUE | Max Rack Density (kW) | Capital Cost (per kW) | Maintenance Complexity | GPU Temp Stability |
|---|---|---|---|---|---|
| Air Cooling (CRAC/CRAH) | 1.4 - 1.8 | 15 - 30 | $8 - $12 | Low | Moderate (fluctuates) |
| Direct-to-Chip Liquid Cooling | 1.05 - 1.15 | 50 - 150+ | $10 - $15 | Medium | Excellent (steady) |
| Single-Phase Immersion | 1.02 - 1.10 | 100 - 200+ | $12 - $18 | High | Excellent (steady) |
| Two-Phase Immersion | 1.01 - 1.05 | 150 - 300+ | $15 - $25 | Very High | Best (isothermal) |

Data Takeaway: Direct-to-chip cooling offers the best balance of density, cost, and serviceability for most current AI workloads, while immersion cooling is reserved for the highest-density deployments where PUE (Power Usage Effectiveness) optimization is paramount.

Relevant Open-Source Project: The Open Compute Project (OCP) has published several open specifications for liquid cooling, including the "Open Rack V3" standard which defines coolant distribution architectures. The GitHub repository (github.com/opencomputeproject) contains detailed mechanical drawings, thermal models, and best practices that have been adopted by major hyperscalers. As of May 2026, the OCP liquid cooling sub-project has over 1,200 stars and active contributions from engineers at Meta, Google, and Microsoft.

Key Players & Case Studies

The liquid cooling ecosystem is a mix of established infrastructure giants and specialized startups.

CoolIT Systems: A dominant player in direct-to-chip cooling, CoolIT provides CDUs and cold plates to major OEMs like Dell, HPE, and Lenovo. Their Rack DLC (Direct Liquid Cooling) solution is deployed in some of the world's largest AI clusters. They have shipped over 1 million cooling units globally, with a focus on high-reliability, low-leak designs.

Asetek: A pioneer in data center liquid cooling, Asetek's technology is used by many hyperscalers. Their patented technology focuses on server-level liquid cooling loops. They have a strong track record in the HPC (High-Performance Computing) market, which has directly translated into AI deployments.

LiquidStack: A leader in immersion cooling, LiquidStack's two-phase immersion technology has been adopted by companies like Bitcoin miner Hut 8 for high-density GPU clusters. Their systems can handle up to 200kW per rack, making them ideal for next-generation AI hardware.

NVIDIA: The chipmaker itself is a key driver. NVIDIA's DGX systems are increasingly designed with liquid cooling in mind. The DGX B200, for instance, is offered with a liquid-cooled option that reduces its power consumption by up to 30% compared to air-cooled versions. NVIDIA has also published reference architectures for liquid-cooled data centers.

Case Study: Microsoft's Project Natick While not directly about AI, Microsoft's underwater data center experiment (Project Natick) demonstrated the viability of liquid cooling in extreme environments. The lessons learned about corrosion resistance, pump reliability, and remote monitoring are directly applicable to terrestrial AI data centers.

Data Table: Key Liquid Cooling Solution Providers

| Company | Core Technology | Key Product | Target Market | Notable Customer/Deployment |
|---|---|---|---|---|
| CoolIT Systems | Direct-to-Chip | Rack DLC CDU | Hyperscalers, OEMs | Dell, HPE, Lenovo |
| Asetek | Direct-to-Chip | ServerLynx | HPC, AI | NASA, various universities |
| LiquidStack | Two-Phase Immersion | LiquidStack T-Engine | High-density AI, Crypto | Hut 8, CoreWeave |
| Submer | Single-Phase Immersion | SmartPod | Edge, AI, HPC | Various European data centers |
| GRC (Green Revolution Cooling) | Single-Phase Immersion | ICEraQ | Enterprise, AI | US Department of Defense |

Data Takeaway: The market is bifurcating: direct-to-chip solutions dominate for mainstream AI deployments where serviceability is critical, while immersion cooling is carving a niche for ultra-dense, purpose-built AI clusters.

Industry Impact & Market Dynamics

The liquid cooling market is experiencing explosive growth. According to industry estimates, the global data center liquid cooling market was valued at approximately $3.5 billion in 2025 and is projected to reach $12.8 billion by 2030, a compound annual growth rate (CAGR) of 29.6%. This growth is directly correlated with the TDP of AI accelerators.

Impact on Server Architecture: Server motherboards are being redesigned to accommodate coolant distribution. The traditional layout, with fans at the front and power supplies at the back, is giving way to designs where coolant manifolds run along the sides or through the center. This requires closer collaboration between chip designers, server OEMs, and cooling vendors.

Impact on Data Center Location: Water availability is becoming a critical factor in site selection. A 100MW liquid-cooled data center can consume up to 1 million gallons of water per day for evaporative cooling towers. This has led to a push for closed-loop, dry cooling systems that reject heat to the air without water evaporation, but these are less efficient in hot climates. Data center operators are now evaluating locations based on water risk indices, not just power availability and latency.

Impact on Operational Costs: The elimination of fans reduces electricity consumption by 10-20% of the total IT load. However, this is partially offset by the power required for pumps and chillers. The net effect is a reduction in PUE from ~1.4 (air) to ~1.1 (liquid), translating to significant OpEx savings over the life of the facility. For a 100MW facility, a 0.3 improvement in PUE can save over $10 million annually in electricity costs.

Data Table: Market Growth Projections

| Year | Market Size ($B) | CAGR (%) | Key Driver |
|---|---|---|---|
| 2023 | 2.1 | — | Early AI adoption |
| 2025 | 3.5 | 29.6 | B200/GB200 deployment |
| 2027 | 6.0 | 30.0 | Widespread AI agent infrastructure |
| 2030 | 12.8 | 28.5 | World model and video generation scaling |

Data Takeaway: The market is doubling every 2-3 years, driven entirely by the thermal demands of AI hardware. This is not a cyclical trend but a structural shift.

Risks, Limitations & Open Questions

Despite its necessity, liquid cooling is not without risks.

1. Leak Risk: The single greatest fear. While modern connectors and leak detection systems are highly reliable, the consequence of a leak in a high-density GPU cluster is catastrophic. A single drip can short-circuit a $300,000 server. The industry is moving toward dielectric coolants that are non-conductive, but these are more expensive and have lower thermal performance than water.

2. Maintenance Complexity: Liquid cooling systems require specialized technicians. Pumps, valves, and CDUs need regular maintenance. The coolant chemistry must be monitored and adjusted. This increases operational complexity and requires retraining of data center staff.

3. Vendor Lock-in: Proprietary cooling solutions can lock operators into a single vendor for spare parts and servicing. The industry is pushing for standardization (e.g., OCP standards) to mitigate this, but it remains a concern.

4. Environmental Impact of Coolants: Some dielectric fluids used in immersion cooling have high global warming potential (GWP). Fluorocarbon-based fluids, for example, can have a GWP thousands of times higher than CO2. The industry is moving toward more environmentally friendly alternatives, but the transition is slow.

5. Retrofitting Existing Data Centers: Converting an air-cooled data center to liquid cooling is expensive and disruptive. It often requires shutting down sections of the facility, installing new piping, and reinforcing floors. Many operators are choosing to build new, liquid-cooled facilities rather than retrofit.

AINews Verdict & Predictions

Verdict: Liquid cooling is no longer a choice; it is a prerequisite for any organization serious about deploying next-generation AI infrastructure. The physics are immutable. Air cooling has hit its ceiling, and the AI industry is accelerating past it.

Predictions:

1. By 2027, over 60% of new hyperscale data center capacity will be liquid-cooled. The cost premium for liquid cooling will shrink as volume increases, making it the default option for any new build exceeding 10MW.

2. Direct-to-chip cooling will dominate the next 5 years, but immersion cooling will capture a growing share (from ~10% today to ~25% by 2030) as chip TDPs exceed 2000W and rack densities surpass 200kW.

3. Water availability will become as important as power availability in data center site selection. This will drive investment in water-efficient cooling technologies and shift new builds toward regions with abundant water or cooler climates.

4. The server OEM market will consolidate around liquid-cooled designs. Dell, HPE, and Lenovo will phase out air-cooled high-density server SKUs by 2028, forcing enterprises to adopt liquid cooling even for non-AI workloads.

5. A major leak incident at a hyperscaler will occur within the next 18 months, causing significant downtime and hardware loss. This will accelerate the adoption of dielectric coolants and redundant leak detection systems.

What to Watch Next: The development of on-chip microfluidic cooling, where coolant flows directly through microchannels etched into the silicon substrate. This could eliminate the need for cold plates entirely and push cooling efficiency to its theoretical limit. Several research groups, including those at MIT and Stanford, have demonstrated prototypes. If commercialized, this could be the next revolution.

Archive

May 20261707 published articles

Further Reading

Beyond Electricity: The Hidden Cost Structure of AI Compute RevealedThe public narrative around AI's insatiable appetite focuses on electricity consumption. However, a comprehensive industAI's Infrastructure War: OpenAI-Plaid, Cerebras IPO, Intel Apple Chip ShiftOpenAI partners with fintech Plaid to embed ChatGPT into personal finance management. Cerebras goes public at a $67 billNovo Nordisk and OpenAI: Can AI Forge the Next Weight Loss Blockbuster?Novo Nordisk has entered a landmark partnership with OpenAI, deploying generative AI and large language models to acceleFrom OpenAI's Core to Challenger: The Architect Rewriting AI's Emotional BlueprintA former OpenAI technical leader is quietly building a new AI system that rejects the 'bigger is better' dogma. Instead

常见问题

这篇关于“Liquid Cooling Revolution: Why AI Data Centers Can No Longer Afford Air Cooling”的文章讲了什么?

The thermal demands of modern AI hardware have reached a tipping point. NVIDIA's B200 GPU, for instance, has a thermal design power (TDP) exceeding 1000W, and racks packed with the…

从“liquid cooling vs air cooling for AI data center”看,这件事为什么值得关注?

The physics of heat transfer is unforgiving. Air has a specific heat capacity of roughly 1.005 kJ/(kg·K) and a thermal conductivity of about 0.026 W/(m·K). Water, by contrast, has a specific heat capacity of 4.18 kJ/(kg·…

如果想继续追踪“direct to chip cooling vs immersion cooling cost comparison”,应该重点看什么?

可以继续查看本文整理的原文链接、相关文章和 AI 分析部分,快速了解事件背景、影响与后续进展。