Liquid cooling for AI: from edge case to essential infrastructure

Updated December 11, 2025

December 2025 Update: Liquid cooling market surging from $2.8B (2025) to $21B+ by 2032 (30%+ CAGR). Current NVIDIA racks at 132kW; next-gen requiring 240kW. GB200 NVL72 enabling 25x cost savings (>$4M annually for 50MW facility). Direct-to-chip now handles up to 1,600W per component. Accelsius NeuCool cooling 4,500W per GPU socket with 40°C warm facility water.

The global liquid cooling market will surge from $2.8 billion in 2025 to over $21 billion by 2032, with compound annual growth exceeding 30%.¹ At the halfway mark of 2025, the shift from air to liquid cooling moved from experimental to operational.² When fully loaded, the latest NVIDIA-based GPU servers require 132 kilowatts per rack. The next generation, expected within a year, will require 240 kilowatts.³ Traditional air cooling cannot dissipate heat at these densities. Liquid cooling transformed from a hyperscaler luxury to a requirement for any organization deploying current-generation AI infrastructure.

The economics reinforce the shift. Data centers spend an estimated $1.9 to $2.8 million per megawatt annually on cooling.⁴ Deploying liquid-cooled GB200 NVL72 systems enables hyperscale data centers to achieve up to 25x cost savings, translating to over $4 million in annual savings for a 50-megawatt facility.⁵ Organizations that resist the transition will find themselves unable to deploy the GPU generations that define AI capability.

The physics driving the transition

AI-optimized servers and GPU-dense clusters push power densities beyond 50 kilowatts per rack, reaching levels where traditional air cooling cannot ensure stable or efficient heat dissipation.⁶ According to the Uptime Institute, average data center rack power density increased 38% from 2022 to 2024, with the steepest growth in AI and hyperscale deployments.⁷ Power densities that once maxed at 15 kilowatts now push 80 to 120 kilowatts in AI clusters.⁸

The fundamental advantage of liquid cooling lies in thermodynamics. At nearly 1,000 times the density of air, liquid excels at carrying heat away thanks to superior heat capacitance and thermal conductivity.⁹ By efficiently transferring heat from high-performance GPUs, liquid cooling reduces reliance on energy-intensive cooling fans. The result: 11% average reduction in server energy consumption while eliminating 80% of traditional cooling infrastructure space requirements.¹⁰

Air cooling systems struggle to handle power densities above 10 to 15 kilowatts per rack.¹¹ Many AI workloads require racks running at 30 to 60 kilowatts or more.¹² The gap between what air cooling delivers and what AI infrastructure demands grows with each GPU generation.

Direct-to-chip cooling dominates production

Direct-to-chip cooling rapidly became the most common liquid cooling form deployed in production environments.¹³ Cold plates mount directly onto CPUs, GPUs, memory modules, and voltage regulators. A closed-loop system circulates coolant through these plates, removing heat at the source.¹⁴

NVIDIA's GB200 NVL72 and GB300 NVL72 systems use direct-to-chip liquid cooling as the standard configuration.¹⁵ Unlike evaporative or immersion cooling, the NVL72's liquid cooling operates as a closed-loop system where coolant does not evaporate or require replacement, saving water.¹⁶ The architecture delivers 40x higher revenue potential, 30x higher throughput, 25x more energy efficiency, and 300x more water efficiency than traditional air-cooled systems.¹⁷

Direct-to-chip solutions now handle up to 1,600 watts per component, enabling 58% higher server density compared to air cooling while reducing infrastructure energy consumption by 40%.¹⁸ Supermicro's DLC-2 enabled NVIDIA HGX B200 system captures up to 98% of system heat by liquid-cooling CPUs, GPUs, DIMMs, PCIe switches, voltage regulators, and power supplies, enabling quiet data center operation at noise levels as low as 50 decibels.¹⁹

Accelsius achieved two thermal milestones with its NeuCool technology: successfully cooling 4,500 watts per GPU socket and maintaining safe GPU temperatures in a fully loaded 250-kilowatt AI rack using warm 40°C facility water.²⁰ The ability to use warm water rather than chilled water reduces cooling infrastructure requirements and operating costs.

Immersion cooling scales for extreme density

Immersion cooling submerges servers in dielectric fluid, achieving over 100 kilowatts per rack and, in some designs, scaling to 250 kilowatts.²¹ Systems like GRC's ICEraQ achieve cooling capacity up to 368 kilowatts per system while maintaining power usage effectiveness below 1.03.²² The approach eliminates fans entirely and enables operators to pack 10 to 15 times more compute into the same footprint.²³

The data center immersion cooling market reached $4.87 billion in 2025 and will grow to $11.10 billion by 2030 at a 17.91% compound annual growth rate.²⁴ Single-phase systems retain the largest market share due to installation familiarity, yet two-phase designs win pilots where extreme density and pump-free architectures prove essential.²⁵

Compared to traditional air cooling, single-phase immersion cooling reduces electricity demand by up to nearly half, contributes to CO2 emissions reductions of up to 30%, and supports up to 99% less water consumption.²⁶ The efficiency gains translate directly into faster time-to-revenue for AI services. The ability to drive higher utilization from every square foot remains the strongest economic lever motivating hyperscale adoption.²⁷

In May 2025, Intel partnered with Shell Global Solutions to launch the first Intel-certified immersion cooling solution for 4th and 5th generation Xeon processors, enabling high-performance thermal management at production scale.²⁸ The partnership signals that immersion cooling reached the certification and support levels enterprise deployments require.

Hyperscaler deployments set the standard

Microsoft's Azure AI clusters, Google's TPU deployments, and Meta's LLaMA model training nodes all shifted to liquid cooling.²⁹ Microsoft's advanced AI supercomputer, unveiled in 2025, features exclusively liquid-cooled racks supporting GPT-Next training workloads.³⁰ The hyperscaler commitments validate liquid cooling as production-ready infrastructure rather than experimental technology.

HPE shipped its first NVIDIA Blackwell family solution, the GB200 NVL72, in February 2025.³¹ HPE built seven of the top ten fastest supercomputers in the world, establishing deep expertise in direct liquid cooling.³² The company's reference architectures provide blueprints for enterprise deployments.

Vertiv's reference architecture for NVIDIA GB200 NVL72 servers reduces annual energy consumption by 25%, cuts rack space requirements by 75%, and shrinks power footprint by 30%.³³ Schneider Electric's liquid cooling infrastructure supports up to 132 kilowatts per rack for GB200 NVL72 AI data centers.³⁴ The vendor ecosystem now offers turnkey solutions rather than requiring custom engineering.

Meta developed Air-Assisted Liquid Cooling with Microsoft as a hybrid, retrofittable solution.³⁵ The approach enabled Meta to begin integrating liquid cooling without overhauling its entire air-cooled infrastructure, demonstrating pragmatic transition paths for organizations with existing facilities.

Retrofitting challenges persist

Retrofitting an operating data center to accommodate more powerful processors presents significant technical and logistical challenges.³⁶ Some operators conclude that building new facilities proves easier than upgrading existing ones.³⁷ The decision depends on facility age, remaining useful life, and the scale of planned AI deployments.

Liquid cooling requires specialized infrastructure including liquid distribution units, cold plates, immersion tanks, and coolant pumps.³⁸ Retrofitting involves modifying server racks, adding leak prevention systems, and ensuring regulatory compliance.³⁹ Brownfield sites face architectural and infrastructure limitations that greenfield projects avoid.

Lower adoption rates for infrastructure-intensive solutions like immersion cooling, at 20.4% among brownfield sites, reflect practical constraints.⁴⁰ These constraints include extensive retrofitting to accommodate tanks, limited floor space, and challenges integrating with existing power and cooling infrastructure.⁴¹ Brownfield sites appear more likely to adopt incremental solutions like liquid-to-air cooling that avoid complete infrastructure overhaul.⁴²

Schneider Electric partnered with NVIDIA on three retrofit reference designs for data center operators seeking performance improvements without redesigning facilities from scratch.⁴³ The designs acknowledge that most organizations cannot build greenfield AI data centers and must work within existing constraints.

Operational complexity increases

Because liquid systems cool only chips, supplemental air cooling still handles 20% to 30% of the total thermal load.⁴⁴ Hybrid cooling architectures require expertise that many organizations lack internally.⁴⁵ The operational shift proves as significant as the mechanical upgrade itself.

Liquid cooling introduces new operational requirements: leak detection, hydraulic redundancy, coolant quality control, and technician upskilling.⁴⁶ Traditional data center operations teams may not have experience with plumbing, pumps, and heat exchangers at the scale AI infrastructure demands. The skill gap affects deployment timelines and ongoing operations.

ZutaCore developed direct-to-chip liquid cooling systems supporting the GB200 superchip, which combines NVIDIA Grace ARM processors with Blackwell GPUs.⁴⁷ Third-party solutions expand options but also complicate vendor management and support arrangements.

Supply chain issues could complicate hybrid cooling plans, potentially worsened by trade policy changes.⁴⁸ The rapid increase in computing power means data centers on the bleeding edge today may rapidly fall behind.⁴⁹ Designing facilities with capacity for future power densities proves challenging when the target continues moving.

Regional adoption patterns

North America leads market adoption through production-scale rollouts by hyperscale cloud providers.⁵⁰ The US market will grow from $1.09 billion in 2024 to $6.39 billion by 2034.⁵¹ Hyperscaler investments from AWS, Google, and Microsoft drive adoption as enterprises follow their lead.

Asia-Pacific exhibits the steepest growth as Japan, China, and South Korea champion liquid-cooled AI clusters.⁵² Conventional air cooling proves cost-prohibitive in hot, humid climates.⁵³ Immersion cooling offers sustainable, space-efficient solutions particularly suited to regional conditions. Asia-Pacific leads the global immersion cooling market throughout the forecast period.⁵⁴

The geographic distribution reflects both climate considerations and the concentration of AI infrastructure investment. Regions with aggressive AI development programs drive cooling innovation out of necessity.

Strategic planning considerations

Organizations planning AI infrastructure must factor liquid cooling into facility and budget decisions. The choice between direct-to-chip and immersion cooling depends on deployment scale, retrofit constraints, and operational capabilities.

For new deployments, liquid cooling should be the default specification for any rack exceeding 30 kilowatts. Planning for 100-kilowatt-plus densities anticipates GPU roadmaps through 2027. Facilities designed today without liquid cooling infrastructure will face expensive retrofits or replacement within years.

For existing facilities, evaluate retrofit feasibility honestly. Schneider Electric's reference designs provide starting points, but significant engineering work remains required. Hybrid approaches that layer liquid cooling onto air-cooled infrastructure offer incremental paths forward.

The cold plate segment accounts for over 43% market share in 2025, efficiently managing CPU and GPU thermal loads.⁵⁵ Organizations should standardize on proven approaches rather than waiting for emerging technologies to mature. Immersion cooling continues growing fastest, delivering up to 80% higher energy efficiency with power usage effectiveness scores as low as 1.02 to 1.03.⁵⁶

Liquid cooling represents infrastructure that enables AI capability. Organizations that delay the transition will find themselves unable to deploy the GPU generations their AI initiatives require. The technology reached production maturity; the remaining question is how quickly organizations can implement it.

References

1. Data Center Frontier. "Liquid Cooling Comes to a Boil: Tracking Data Center Investment, Innovation, and Infrastructure at the 2025 Midpoint." 2025. https://www.datacenterfrontier.com/cooling/article/55292167/liquid-cooling-comes-to-a-boil-tracking-data-center-investment-innovation-and-infrastructure-at-the-2025-midpoint

2. Data Center Frontier. "Liquid Cooling Comes to a Boil."

3. Datacenters.com. "Why Liquid Cooling Is the New Standard for Data Centers in 2025." 2025. https://www.datacenters.com/news/why-liquid-cooling-is-becoming-the-data-center-standard

4. NVIDIA Blog. "Blackwell Platform Boosts Water Efficiency by Over 300x." 2025. https://blogs.nvidia.com/blog/blackwell-platform-water-efficiency-liquid-cooling-data-centers-ai-factories/

5. NVIDIA Blog. "Blackwell Platform Boosts Water Efficiency by Over 300x."

6. ARC Compute. "Liquid Cooling & Green AI Infrastructure: Designing Sustainable GPU Data Centers." 2025. https://www.arccompute.io/arc-blog/liquid-cooling-green-ai-infrastructure-designing-sustainable-gpu-data-centers

7. Data Center Frontier. "Liquid Cooling Comes to a Boil."

8. Data Center Frontier. "Liquid Cooling Comes to a Boil."

9. Applied Digital. "Why Modern AI Data Centers Are Harnessing the Power of Liquid Cooling." 2025. https://www.applieddigital.com/insights/why-modern-ai-data-centers-are-harnessing-the-power-of-liquid-cooling

10. Applied Digital. "Why Modern AI Data Centers Are Harnessing the Power of Liquid Cooling."

11. Datacenters.com. "Liquid Cooling vs. Air Cooling: Which One Will Dominate Data Centers in 2025?" 2025. https://www.datacenters.com/news/liquid-cooling-vs-air-cooling-which-one-will-dominate-data-centers-in-2025

12. Datacenters.com. "Liquid Cooling vs. Air Cooling."

13. Data Center Frontier. "Liquid Cooling Comes to a Boil."

14. ARC Compute. "Liquid Cooling & Green AI Infrastructure."

15. Data Centre Magazine. "How Nvidia is Boosting Water Efficiency with Blackwell." 2025. https://datacentremagazine.com/hyperscale/how-nvidia-is-boosting-water-efficiency-with-blackwell

16. NVIDIA Blog. "Blackwell Platform Boosts Water Efficiency by Over 300x."

17. NVIDIA Blog. "Blackwell Platform Boosts Water Efficiency by Over 300x."

18. ARC Compute. "Liquid Cooling & Green AI Infrastructure."

19. Supermicro. "Supermicro Expands Its NVIDIA Blackwell System Portfolio with New Direct Liquid-Cooled Systems." 2025. https://www.supermicro.com/en/pressreleases/supermicro-expands-its-nvidia-blackwell-system-portfolio-new-direct-liquid-cooled-dlc

20. SiliconANGLE. "AI data center cooling becomes the boiling point." November 2025. https://siliconangle.com/2025/11/20/ai-data-center-cooling-becomes-boiling-point-sc25/

21. ARC Compute. "Liquid Cooling & Green AI Infrastructure."

22. Data Center Frontier. "Liquid Cooling Comes to a Boil."

23. Data Center Frontier. "Liquid Cooling Comes to a Boil."

24. Mordor Intelligence. "Data Center Immersion Cooling Market Size, Growth, Competition 2025-2030." 2025. https://www.mordorintelligence.com/industry-reports/immersion-cooling-market-in-data-centers

25. Data Center Frontier. "Liquid Cooling Comes to a Boil."

26. Intel Newsroom. "Intel and Shell Advance Immersion Cooling in Xeon-Based Data Centers." May 2025. https://newsroom.intel.com/data-center/intel-shell-advance-immersion-cooling-xeon-based-data-centers

27. Data Center Frontier. "Liquid Cooling Comes to a Boil."

28. Intel Newsroom. "Intel and Shell Advance Immersion Cooling."

29. Datacenters.com. "Why Liquid Cooling Is the New Standard for Data Centers in 2025."

30. Datacenters.com. "Why Liquid Cooling Is the New Standard for Data Centers in 2025."

31. HPE Newsroom. "HPE announces shipment of its first NVIDIA Grace Blackwell system." February 2025. https://www.hpe.com/us/en/newsroom/press-release/2025/02/hpe-announces-shipment-of-its-first-nvidia-grace-blackwell-system.html

32. HPE Newsroom. "HPE announces shipment of its first NVIDIA Grace Blackwell system."

33. SiliconANGLE. "AI data center cooling becomes the boiling point."

34. Schneider Electric Blog. "Why liquid cooling for AI data centers is harder than it looks." August 2025. https://blog.se.com/datacenter/2025/08/12/why-liquid-cooling-for-ai-data-centers-is-harder-than-it-looks/

35. EnkiAI. "Meta's Liquid Cooling 2025: Inside the $65B AI Overhaul." 2025. https://enkiai.com/data-center/metas-liquid-cooling-2025-inside-the-65b-ai-overhaul

36. Utility Dive. "The 2025 outlook for data center cooling." 2025. https://www.utilitydive.com/news/2025-outlook-data-center-cooling-electricity-demand-ai-dual-phase-direct-to-chip-energy-efficiency/738120/

37. Utility Dive. "The 2025 outlook for data center cooling."

38. Datacenters.com. "Why Liquid Cooling Is the New Standard for Data Centers in 2025."

39. Datacenters.com. "Why Liquid Cooling Is the New Standard for Data Centers in 2025."

40. Hyperion Research. "Insights into Liquid Cooling in Brownfield HPC and AI Data Centers." March 2025. https://hyperionresearch.com/wp-content/uploads/2025/04/Hyperion-Research-Quick-Take-Liquid-Cooling-March-2025-1.pdf

41. Hyperion Research. "Insights into Liquid Cooling in Brownfield HPC and AI Data Centers."

42. Hyperion Research. "Insights into Liquid Cooling in Brownfield HPC and AI Data Centers."

43. Utility Dive. "The 2025 outlook for data center cooling."

44. Schneider Electric Blog. "Why liquid cooling for AI data centers is harder than it looks."

45. Schneider Electric Blog. "Why liquid cooling for AI data centers is harder than it looks."

46. Data Center Dynamics. "Liquid cooling: the future of data center architecture and operations." 2025. https://www.datacenterdynamics.com/en/opinions/liquid-cooling-the-future-of-data-center-architecture-and-operations/

47. ZutaCore Blog. "First Direct-to-Chip liquid cooling cold plates for Nvidia Grace Blackwell Superchip." 2025. https://blog.zutacore.com/news/first-direct-to-chip-liquid-cooling-cold-plates-for-nvidia-grace-blackwell-superchip

48. Utility Dive. "The 2025 outlook for data center cooling."

49. Utility Dive. "The 2025 outlook for data center cooling."

50. Data Center Frontier. "Liquid Cooling Comes to a Boil."

51. Persistence Market Research. "Data Center Liquid Cooling Market Size & Forecast, 2032." 2025. https://www.persistencemarketresearch.com/market-research/data-center-liquid-cooling-market.asp

52. Mordor Intelligence. "Data Center Immersion Cooling Market."

53. Mordor Intelligence. "Data Center Immersion Cooling Market."

54. Mordor Intelligence. "Data Center Immersion Cooling Market."

55. Persistence Market Research. "Data Center Liquid Cooling Market."

56. Data Center Frontier. "Liquid Cooling Comes to a Boil."

SEO Elements

Squarespace Excerpt (159 characters): Liquid cooling market surges from $2.8B to $21B by 2032. GPU racks hit 132kW today, 240kW next year. Why liquid cooling became essential AI infrastructure.

SEO Title (54 characters): Liquid Cooling for AI: Essential Data Center Infrastructure

SEO Description (155 characters): GPU racks require 132kW today, 240kW by 2026. Liquid cooling market grows 30% CAGR to $21B. Analysis of direct-to-chip and immersion cooling for AI infrastructure.

URL Slugs: - Primary: liquid-cooling-ai-data-center-infrastructure-essential - Alt 1: direct-to-chip-cooling-nvidia-blackwell-deployment - Alt 2: immersion-cooling-gpu-data-center-2025 - Alt 3: ai-data-center-cooling-retrofit-challenges

Key takeaways

For infrastructure architects: - Liquid cooling market: $2.8B (2025) → $21B (2032) at 30%+ CAGR - Current GPU racks: 132kW; next-gen (within 1 year): 240kW—air cooling cannot dissipate at these densities - Air cooling struggles above 10-15kW per rack; AI workloads require 30-60kW or more

For finance teams: - Cooling costs: $1.9-2.8M per MW annually; liquid-cooled GB200 NVL72 saves $4M+ annually for 50MW facility - Direct-to-chip enables 58% higher server density; reduces infrastructure energy consumption by 40% - GRC ICEraQ immersion: up to 368kW per system with PUE below 1.03

For operations teams: - Supermicro DLC-2 captures 98% of system heat; enables 50dB quiet operation - Liquid cooling requires 20-30% supplemental air cooling for remaining thermal load - Brownfield immersion adoption only 20.4% due to retrofitting complexity

For technology planning: - NVIDIA GB200 NVL72/GB300 NVL72: direct-to-chip as standard configuration (closed-loop, no water loss) - Accelsius NeuCool: 4,500W per GPU socket using 40°C warm facility water - Cold plate segment: 43%+ market share in 2025; immersion delivers up to 80% higher energy efficiency