Direct-to-chip liquid cooling vs immersion cooling for AI
Direct-to-chip liquid cooling (DLC) and immersion cooling are the two primary architectures for AI-density thermal management above 30-40 kW per rack. DLC uses cold plates + cooling distribution units (CDUs); immersion submerges hardware in dielectric fluid (single-phase oil or two-phase fluorinated). Both persist because they solve different constraints. DLC dominates GPU-training clusters and is easier to retrofit; immersion offers higher volumetric density and simpler thermal management for extreme-density edge cases.
01Bottom-line verdict
02Comparison table
| Direct-to-chip (DLC) | Immersion (1-phase + 2-phase) | |
|---|---|---|
| Architecture | Cold plates on GPU/CPU + CDU + secondary loop | Servers submerged in dielectric fluid |
| Density supported | Up to ~600 kW per rack demonstrated | Up to ~1000 kW per rack in dense configurations |
| Retrofit into existing rack forms | Straightforward with cold plate + CDU | Requires immersion tank; not rack-form-compatible |
| Fluid choice | Water-glycol or dielectric coolant | Single-phase: mineral or synthetic oil. Two-phase: fluorinated (3M Novec, Solvay Galden) |
| Vendor concentration | CoolIT (Ecolab) + Motivair + Asetek + Boyd | Submer + LiquidStack + GRC + Iceotope |
| Hyperscaler adoption (2026) | Meta, MSFT, GOOG, AWS AI clusters using DLC | Growing but smaller absolute footprint |
| Fluid supply chain | Water + coolants: broad, low-risk | Two-phase fluorinated: concentrated (3M, Solvay), PFAS-adjacent regulatory pressure |
| Maintenance access | Standard rack access, hot-swap possible | Tank access; harder to service individual components |
| OCP standards support | Advanced Cooling Solutions workstream | Advanced Cooling Solutions workstream |
| Capex per MW cooling | ~$1.2-1.5M/MW for greenfield DLC | ~$1.5-2M/MW single-phase; higher for two-phase |
03Decision framework
Choose Direct-to-chip (DLC) when
- Density 40-500 kW/rack. DLC covers the mainstream AI-density band
- Retrofit of existing air-cooled facility. Cold plates + CDU integrate into existing rack forms
- GPU-training cluster (Meta, MSFT, GOOG, AWS pattern). Reference deployments are DLC
- Serviceability priority. Standard rack access preserved
- Fluorinated fluid supply concerns. Water + glycol coolants are unconstrained
Choose Immersion (1-phase + 2-phase) when
- Ultra-density new-build (>500 kW/rack). Immersion volumetric density can exceed DLC at extremes
- Extreme-ambient siting (desert, tropics). Immersion thermal envelope more permissive
- Specific accelerator packages designed for immersion. Some ASICs + custom silicon optimised for immersion
- Sound-sensitive locations. Immersion runs quieter (no rack fans)
- Willing to underwrite fluid supply-chain risk. Two-phase fluorinated requires supply-chain diligence
04Deep-dive research
05Primary sources
06Frequently asked
When is air cooling still viable for AI?
Below ~30-40 kW/rack sustained. Above that, air-cooling infrastructure scales sub-linearly in cost + energy compared to liquid alternatives.
Why is two-phase fluorinated fluid a supply concern?
Manufacturing capacity for specific fluorinated fluids (3M Novec, Solvay Galden) is concentrated among a small number of chemical suppliers, some facing regulatory pressure on PFAS/fluorochemistry. Growth in immersion demand could outpace fluid supply expansion.
Can I mix DLC + immersion in one facility?
Yes, common in newer facilities to allocate DLC for GPU-training and immersion for specific edge-case density zones.
What is the biggest vendor-side risk in DLC?
CDU + high-current disconnect connector reliability, specifically for hyperscaler-scale deployments. Fewer failure points than immersion but each failure is more consequential.
| Dimension | Direct-to-chip liquid cooling (DLC) | Single-phase immersion | Two-phase immersion |
|---|---|---|---|
| Thermal capacity per die | Up to 1,500W per die | Up to 400W per die | Up to 2,000W+ per die |
| Coolant type | Water-glycol mix (facility loop) + dielectric secondary | Synthetic hydrocarbon dielectric (mineral oil, PAO) | Fluorocarbon dielectric (boiling at ~50°C) |
| Server modification required | Cold plate + hose fittings only | Sealed / dielectric-safe server | Sealed / dielectric-safe server |
| Hyperscaler deployment status (2026) | Broad: Meta, Microsoft, Google, Oracle | Limited: bespoke deployments | Emerging: OCP Sustainability Project |
| Top vendors | CoolIT (Ecolab), Motivair (Schneider), Vertiv, Boyd (Eaton) | Submer, GRC, LiquidStack | LiquidStack, ZutaCore, Iceotope |
| Chemistry supply chain risk | Low (water-glycol widely available) | Medium (single-source dielectrics) | High (PFAS regulatory pressure) |
| CDU + secondary loop | Required for chip-package coolant separation | Not required | Not required |
| Best fit | Rack retrofit + hyperscale greenfield 50-500 kW/rack | Edge + specialist AI workloads | 1,500W+ per die (Rubin Ultra, custom silicon) |