The short
- Density shift: Below ~20–25 kW per rack, air can hold; above ~30 kW, liquid options take the lead on stability and energy.
 - Cost lens: Fans and floor space push air OPEX higher at scale; direct-to-chip cuts blower load and raises heat-recovery options.
 - Rollout tell: Sites adding rear-door exchangers today often pave the path to direct-to-chip later without uprooting rows.
 
Rack density × cooling type × annual OPEX
| Rack kW | Cooling type | Annual energy for cooling (kWh/rack) | Typical notes | 
|---|---|---|---|
| 10–15 | Air (CRAC/CRAH) | 18,000–28,000 | Hot/cold aisle with blanking panels works | 
| 20–25 | Air + containment | 28,000–40,000 | Strong containment, higher fan power, watch noise | 
| 25–35 | Rear-door heat exchanger | 22,000–30,000 | Good retrofit path; water loop at row ends | 
| 35–60 | Direct-to-chip (cold plates) | 16,000–24,000 | Most heat to liquid; fans throttle back | 
| 60–100+ | Immersion (single/dual-phase) | 14,000–22,000 | Top efficiency; layout shifts and service training | 
Energy ranges assume standard setpoints and steady utilization. Your site’s PUE, climate, and water temps will swing results.
Break-even — a quick way to think about it
Rule: If fan energy + raised inlet temps force throttling at target density, your next dollar should go to liquid, not bigger blowers.
- Air path adds: More fan power, tighter containment, and floor space for plenum work.
 - Liquid path adds: CDUs, distribution lines, quick-disconnects; but lowers fan energy and boosts heat reuse.
 
Kit choices (and what they imply)
Rear-door heat exchangers
Retrofit friendly, keeps racks; good for mixed rows where only a few racks are hot. Great first step if you plan direct-to-chip later.
Direct-to-chip
Catches most heat right at CPUs/GPUs; reduces air load; preserves standard racks; ideal for 35–60 kW lanes.
Immersion
Highest thermal headroom; different service rhythm; great for ultra-dense AI clusters with predictable load.
Space, power, water — the three checks
- Space: If adding air paths cuts sellable rack count, liquid often wins even if kit CAPEX is higher.
 - Power: Fan horsepower adds up; direct-to-chip slashes it and frees watts for compute.
 - Water: Warm-water loops (higher supply temps) reduce chiller load and enable heat reuse.
 
What to watch
- Sites crossing 30 kW+ with rear-door rollouts; how quickly they layer in direct-to-chip.
 - Cold-plate designs that avoid board swaps — quicker adoption with standard racks.
 - Heat-reuse pilots: district loops that turn waste heat into a revenue credit.
 
FAQ
- Is liquid always cheaper? Not at low densities. The win shows up as you push density and fan energy climbs.
 - Do I need new racks? Rear-door and direct-to-chip can ride on standard racks; immersion changes layout.