Whitepaper: Thermal Boundaries in AI Infrastructure - Matching cooling technologies to compute through the power inflection, Version 1.00 (October 2026)

Overview

The paper argues that cooling, long a supporting utility, now sets the pace at which compute can be deployed. It defines three boundaries bounding every method: the efficiency boundary (the knee of the curve of cooling share against device power), the economic boundary (where total cost of ownership, with energy only one term, turns against the method) and the physics boundary (where no achievable flow, surface or temperature can hold the junction in range, so the method becomes impossible). The gap between the economic and physics boundaries is the avoidance zone, whose width decides whether a transition is abrupt or gradual. Version 1.00 was published on 12 October 2026.

The central claim is the least-disruption law: a new cooling method is adopted only when the incumbent can no longer do the job, because each new method brings dependencies the ecosystem has good reason to defer. Air was forced past its physics boundary in a single generation, cold plate became the default overnight, and a preparation debt followed. In the physics chapter the paper compares specific heat for air at 1,005 J/(kg K), water 4,184, PG25 3,950, fluorocarbon dielectrics 1,050 and oil-based dielectrics 1,850, and notes that die-level heat flux stayed roughly flat at about 50-90 W/cm2 until 2024 while package power tripled through larger die area.

The technology chapters assess air, single- and two-phase cold plate, single- and two-phase immersion, and microfluidics. For air, fan power rises roughly with the cube of flow, so doubling airflow can demand close to eight times the fan power; the practical air-cooled rack ceiling is taken as 45 kW, extendable to about 60-80 kW with rear-door heat exchangers. For a single-phase cold plate at 5 LPM of water and a 10 C allowable coolant rise, sensible heat transport is limited to about 3.46 kW, making roughly 3-4 kW per plate a practical boundary; this is an architectural boundary, not an intrinsic thermodynamic ceiling. The conclusion projects the cold plate physics boundary near 4,000 W for a reticle-class die and notes frontier package power doubling roughly every two years.

The reliability analysis uses typical annual failure rates of 2.5% for an air-cooled GPU, 1.5% for a liquid-cooled GPU, 0.8% for a central CDU, 0.15% per quick disconnect and 0.1% per cold plate per GPU. A shared CDU raises modelled rack-level outage probability from about 0.01% per year (air) to about 0.25% per year (DLC), roughly 25 times, even with N+1 pumps. Rack power anchors are 120-132 kW for GB200 NVL72, about 140 kW for GB300 class, about 230-250 kW for VR200 NVL144 and about 600 kW for the announced Kyber-class Rubin Ultra. The economics use a 200 MW reference case with facility-water delta-T of 8 C for liquid versus 20 C for air, 4% UPS loss and 88% average PSU efficiency.

The conclusion is a toolbox: no method wins on every axis. Air is cheapest and most limited, cold plate is moderate at first and grows expensive when stretched, immersion costs more to stand up but earns it back in reliability and efficiency, two-phase methods reach furthest but pay in fluids, containment and regulation, and microfluidics amplifies the others. The facility is a second forcing function through power, water and land. An epilogue argues commissioning is becoming the bottleneck and must verify not only steady state but load-step response, leak detection and automated isolation, and handoffs between methods. It is an argument-driven paper; many values are modelled or vendor-reported rather than measured standards.

Key points

1. Three boundaries are defined: efficiency (knee of the cooling-share curve), economic (total cost of ownership) and physics (impossibility); the gap between the economic and physics boundaries is the avoidance zone.
2. Least-disruption law: a new cooling method is adopted only when the incumbent can no longer do the job; air was forced past its physics boundary in one generation and cold plate became the default abruptly.
3. Specific heat: air 1,005, water 4,184, PG25 3,950, fluorocarbon dielectrics 1,050, oil-based dielectrics 1,850 J/(kg K).
4. Die-level heat flux stayed roughly flat at about 50-90 W/cm2 until 2024 while package power tripled through larger die area.
5. Typical annual failure rates used: air-cooled GPU 2.5%, liquid-cooled GPU 1.5%, central CDU 0.8%, quick disconnect 0.15% per fitting, cold plate 0.1% per GPU.
6. A shared CDU raises modelled rack-level outage probability from about 0.01% per year (air) to about 0.25% per year (DLC), roughly 25 times, even with N+1 pumps.
7. Practical air-cooled rack ceiling is taken as 45 kW, extendable to about 60-80 kW with rear-door heat exchangers.
8. Rack power anchors: GB200 NVL72 120-132 kW, GB300 class about 140 kW, VR200 NVL144 about 230-250 kW, Kyber-class Rubin Ultra about 600 kW (announced).
9. For a cold plate at 5 LPM water and a 10 C allowable coolant rise, sensible heat transport is limited to about 3.46 kW, so roughly 3-4 kW per cold plate is a practical boundary.
10. The economic reference case assumes facility-water delta-T of 8 C for liquid versus 20 C for air, 4% UPS loss and 88% average PSU efficiency; no method wins on every axis.

Using it in design

• Choose cooling architecture against target rack power and each method's usable envelope: a 45 kW air ceiling (60-80 kW with rear doors), 120-250 kW class cold-plate racks, and the paper's anchor values for announced 600 kW-class racks.
• In cold plate design treat 3-4 kW per package as the practical limit and use the 3.46 kW result for 5 LPM and a 10 C rise directly when setting coolant flow and allowable delta-T.
• Note that a shared CDU as a single point of failure raises rack-level outage probability by about 25 times; test redundancy, isolation valves and CDU partitioning against that effect.
• For each method write down its efficiency, economic and physics boundaries and its avoidance zone, and base the decision on total cost of ownership including reliability, capital, labour and floor space rather than energy alone.
• For hybrid architectures (a surface method on the highest-flux accelerator with immersion for the distributed remainder), add the handoffs between methods and load-step response to commissioning acceptance tests.
• Treat the site's power, water and land limits as a second forcing function and compare cooling strategies under those limits for density and heat reuse.

Who should read it

Engineers selecting AI data centre cooling architectures, cooling and IT equipment suppliers, and executives making strategic investment decisions.

Document contents

1. Version History and Acknowledgments
2. Abstract
3. Introduction
4. Cooling becomes the constraint
5. Runaway power curves are not new
6. Why there is no rescue this time
7. A different kind of transition
8. Open Compute Project Tenets
9. The physics that sets the boundaries
10. The governing equation
11. Why liquids outperform air
12. The shrinking temperature budget
13. Flux and total power
14. The three boundaries
15. The efficiency boundary
16. The economic boundary
17. The physics boundary
18. The avoidance zone
19. The adoption law
20. The shape of the heat
21. The thermal path (acquisition, transport and rejection)
22. The cooling technologies (air, single-phase cold plate, two-phase cold plate, single-phase immersion, two-phase immersion, microfluidics)
23. The economics of cooling
24. The toolbox, a multi-technology roadmap
25. Epilogue: Commissioning the AI datacenter
26. Conclusion
27. Appendix A Detailed source material for graphs
28. Appendix B Thermohydraulic Limits of Single Phase Cold Plate
29. Appendix C Reader's reference
30. Appendix D License

Read the document

Publication details

Authors

Rolf Brink; Mohan J. Kumar; Steve Mills; Kelley Mullick

Publisher

Open Compute Project Foundation

Year

2026

Document language

English

Source type

Industry body

Access

Open licence — downloadable

Licence

CC BY-SA 4.0 — share with attribution under the same licence

Licence evidence

www.opencompute.org

Citation