AI Rack Cooling: Applying Directed Energy Thermal Strategies to High-Density Data Centers (50kW-1MW+)
Devin Pellicone, Head of Data Center Solutions, ACT
AI computing is redefining data center energy density, with 100 kW racks fast becoming the standard and megawatt-class deployments approaching reality. This shift creates new challenges: traditional air cooling systems can’t handle sudden, concentrated heat spikes, and slow control loops risk throttling vital workloads. By leveraging lessons learned from directed-energy projects—where immense pulsed loads demand immediate, stable, and uniform cooling—data center teams are now pioneering two-phase direct-to-chip solutions that enable performance, reliability, and scale from 50 kW to 1MW+.
ACT has spent over a decade engineering two-phase cooling systems for directed-energy programs where heat loads are on the same order of magnitude and stability is non-negotiable. These payloads come online suddenly at immense energy levels and require immediate and efficient cooling to keep from completely melting down under the load. Sound familiar?
We like to think of AI thermal management as a stability and reliability problem first and a capacity problem second. Over the past decade, we have proven the effectiveness of two-phase direct-to-chip cooling to handle the highest heat fluxes and most demanding thermal budgets across many tactical applications and environments. As far as cooling solutions are concerned, it doesn’t get better than two-phase cooling. If we take some lessons from the directed energy sector in managing massive, pulsed energy loads, repeatably and reliably, two-phase cooling represents one of very few solutions to making the 1MW rack a reality in the very near future.
Quick Take
- 25–50 kW/rack → Use rear-door heat exchangers (RDHx) as a practical bridge. Start planning your liquid transition.
- 50-200 kW/rack → The first major step towards liquid cooling. Single and two-phase direct to chip cooling (DTC) solutions offer more efficient heat transfer for the higher concentration of heat.
- 200 kW to 1 MW+/rack → Two-phase DTC is the most attractive solution and the only technology proven to scale to 1MW+ rack densities and cover the entire advanced GPU roadmap. Offers >300W/cm2 capacity with less than 0.7 LPM/kW of flow rate.
Which Cooling, When?
| Situation in Your Racks | Good Fit | How Hard is it to Add? | Improvements? | Why Pick This? |
|---|---|---|---|---|
| Up to ~ 40 kW per rack (lighter loads) | Air + aisle containment | Rear-door heat exchanger (RDHx) | Easy | Basic control | Lowest disruption. Practical bridge while you plan for liquid cooling. |
| 40-200 kW per rack (high-power AI) | Direct-to-Chip (water/glycol or two-phase) | Moderate | Much steadier temperatures; directed cooling to main heat dissipating components | The necessary first major liquid step. Pulls heat off CPUs/GPUs at the source, curbing localized spikes. |
| 200 kW to 1MW+ per rack (future AI/HPC) | Two-Phase Direct-to-Chip (refrigerant) | Moderate-Higher | Fast responding; Stable and uniform component temperatures even in dynamic events; 1/3 to 1/4 the flow requirement of single-phase DTC | The definitive scaling solution. Proven to cover the entire advanced GPU roadmap up to 1MW+ densities. Delivers ±2°C uniformity and industry-leading 300W/cm2+ cold plate capacity. |
Final selection depends on server SKUs, duty cycles, and facility-water specs. A well-engineered CDU decouples the secondary loop from dynamics in the facility water feed, easing adoption and protecting the balance of plant. Many teams bridge with RDHx while prepping DTC, then step to two-phase as density/flux and duty cycles rise — especially when uniformity and transients become the constraint.
How Two-Phase Cooling Loops Work in AI Racks
A reliable two-phase deployment is a system, not a part. We design and build the complete chain.
- Custom-Engineered Cold Plates: Optimize for low thermal resistance (0.060°C-cm²/W) and uniform heat extraction, capable of handling 7.5 kW and heat flux exceeding 300 W/cm².
- Accumulator Design: Maintains stability across varying load profiles and accommodates future rack installations and modularity.
- Intelligent N+1 Pumps: Uses proprietary control loops to ensure stable operation with turndown ratios in load all the way to zero from peak load.
- Engineered Manifolds: Precisely balance fluid flow across multiple parallel servers and racks
- Comprehensive Telemetry: Monitors temperatures, pressures and flow, and reports back the health of the system to operators
- Peripheral Heat Management. Modular solutions for traditionally air-cooled components to achieve a fan-less rack solution
Key Performance Factors When Evaluating Two-Phase Cooling Vendors
Don’t trust claims—trust the numbers. These are the critical questions to ask any vendor:
- CDU Capacity & Derating Curves: What is the actual heat rejection capacity at higher 30°C or 40°C facility-water temperatures?
- Uniformity & Transient Response: What is the time-to-setpoint after a 30% load step? (This is the ultimate measure of stability and throttle avoidance.)
- Single-Source Accountability: Who owns the entire thermal chain from the cold plate and manifold to the CDU and facility heat rejection system?
- Maintenance & Openness: Does the solution include necessary features like quick-disconnects, advanced leak detection, spares, and telemetry for integration?
AI Cooling FAQs: Air vs Liquid vs Two-Phase
Q: When do we move from air to liquid?
A: Continue the use of air cooling if your site reliably supports the load; shift to liquid when sustained density pushes past ~40–50 kW per rack. If you experience thermal throttling after verifying containment, airflow, and RDHx capacity, treat that as your trigger. For Blackwell-class deployments, liquid is the default choice for higher densities (≈50–60 kW+ or when packing >2 high-end nodes per rack).
Q: Do we jump straight to two-phase?
A: Single-phase direct-to-chip comfortably spans ~50–150 kW/rack and, in purpose-built designs, can stretch to ~200 kW/rack. Consider two-phase when you foresee loads exceeding~200 kW/rack in the future and implement it now. Requirements for warmer facility supply water, or if you’re constrained by per-socket heat flux are also good use cases for two-phase. The exact cutover depends on CDU capacity, facility water temps, and server mix; two-phase primarily delivers higher rack density, reduced fluid flow rates and velocities, better heat-flux handling, and the benefits of non-conductive (dielectric) coolants.
Q; Can our existing facility water infrastructure support liquid cooling?
A: Yes. The CDU is key to isolating chip-side loops and easing adoption while protecting your legacy facility water.
Closing insight
What separates a stable, high-performance computing system from a throttled or unstable one is often the effectiveness of the cooling solution that has been implemented. That’s solvable when you treat the rack like a directed-energy problem and design the whole system accordingly.
Bottom-line: two-phase DTC isn’t just next-gen cooling — it’s the only proven path to stable megawatt-class AI compute.
See Me at SC25
If you’re wrestling with thermal control or planning a density jump, book a 30-minute technical review with me at SC25. I’ll walk you through our two-phase demo, show the 200kW two-phase CDU, and preview how we’re scaling to megawatt-class. Schedule time. Or email me anytime at devin.pellicone@1-act.com with questions.
