If your CRAC units are redlining and your GPU racks are throttling, you’re not alone. Average rack density jumped roughly 69% year-over-year to around 27 kW in 2026, and over 70% of existing U.S. data center floor space was originally designed for air-cooled densities below 15 kW per rack. That gap is exactly why cooling retrofits have become one of the most consequential — and expensive if you get it wrong — decisions facility operators are making this year.
This guide breaks down what CRAC, CRAH, and liquid cooling retrofits actually cost, when each makes sense, the vendors actually shipping this equipment, and the real payback math behind switching.
Why This Decision Got Urgent
Traditional CRAC (Computer Room Air Conditioner) and CRAH (Computer Room Air Handler) systems hit a hard physical limit as density climbs. Moving enough chilled air through a raised floor plenum becomes cost-prohibitive above roughly 15 kW per rack — you simply run out of practical airflow volume before you run out of heat to remove. A 6.4 MW hall at advanced air-cooling density can require on the order of 1,000,000 CFM of total airflow management. At that point, the infrastructure needed to keep pushing air gets more expensive than switching cooling methods entirely.
Modern GPU hardware has made this worse fast. High-end chips now peak over 2,000 watts each, and the NVIDIA B200 effectively demands liquid cooling at its 1,200W+ TDP. If you’re deploying current-generation AI hardware, air cooling isn’t really a design choice anymore above a certain density — it’s a constraint you’re forced to solve around. The market reflects this: liquid cooling systems, including CDUs, manifolds, and immersion, are now the fastest-growing segment of the broader cooling market, driven directly by rack densities climbing past 50–100 kW.
What Each Option Actually Costs
CRAC/CRAH (air-based, with hot/cold aisle containment). Mature, lowest upfront cost, widely supported by every facility technician. Works well for uniform loads under roughly 12–15 kW/rack. Diminishing returns kick in fast above that — you’re paying more for airflow management with less thermal benefit each step up. This remains the right call for a large share of standard enterprise deployments; don’t over-engineer a facility that doesn’t need it.
Rear-Door Heat Exchangers (RDHx). A practical middle-ground retrofit: a chilled-water coil mounted on the rack’s rear door absorbs 60–80% of the rack’s heat load right at the point of generation, before that air even enters the room. It connects to existing chilled water infrastructure and doesn’t require touching the servers themselves — one of the more popular retrofit paths precisely because of that. Vendors like Motivair and Vertiv have built substantial product lines around exactly this use case, since it’s often the fastest way to buy density headroom without a full facility redesign.
Liquid-to-Chip (Cold Plate). Direct thermal contact between coolant and the CPU/GPU. Delivers 3–5x higher heat transfer than air and has become the de facto standard for deployments at 30 kW/rack and above. Expect a CAPEX premium of roughly $2,500–$4,500 per kW compared to traditional air cooling — real money, but it’s what’s actually required to run modern GPU clusters without throttling. Dell, HPE, Lenovo, and Supermicro all now offer factory-integrated direct-to-chip servers in standard 19-inch racks, which has meaningfully lowered the integration risk compared to a few years ago.
Two-Phase Immersion. Servers submerged directly in dielectric fluid. Highest density support (100+ kW/rack) and the lowest achievable PUE, often in the 1.03–1.05 range. Requires full hardware requalification and specialized maintenance — this is the option for genuinely extreme density, not a default choice. Vendors including Submer, GRC, and Asperitas have moved this from a niche technology to a real commercial category, though it still represents a much smaller share of new deployments than direct-to-chip.
Real Retrofit Cost Ranges (2026 Vendor Data)
| Approach | Typical Cost | Notes |
|---|---|---|
| Air-cooled modular unit (100 kW) | $120,000 – $180,000 | 5–7 year ROI via energy savings |
| Liquid cold plate (per 42U rack, 40 kW) | $85,000 – $110,000 | 3–4 year ROI |
| RDHx retrofit | Lower than full liquid conversion | Uses existing chilled water loop |
| Full liquid retrofit vs. new-build immersion | ~20% of new-build cost | Delivers ~70% of new-build performance |
One documented real-world case: a pharmaceutical company retrofitted a 2008-era facility to support 800 NVIDIA H100 GPUs for $4.2 million, versus a projected $35 million for a comparable new build — in four months instead of eighteen. That gap is the core argument for retrofitting over rebuilding wherever the existing structure can support it.
The Payback Math
The efficiency gain is where retrofits actually pay for themselves. Air-cooled facilities typically run a PUE of 1.5 or higher; well-executed liquid cooling environments often land under 1.2. That difference compounds: a 10 MW facility switching from air to cold-plate liquid cooling has been documented saving the energy-cost equivalent of roughly 24 million RMB per year in one Chinese industry case study — with payback landing well under 24 months even at the high end of capital cost estimates.
Need Expert Guidance?
Talk to a Data Center Expert
21+ years of hands-on experience in data center design, operations & infrastructure. Book a quick discovery call to discuss your project.
📞 Book a Discovery CallImmersion cooling adds a second savings lever beyond energy: footprint. A 60–75% smaller physical footprint can save real money on facility square footage — one estimate puts this at roughly $1.5M per year for a 10,000 sq ft facility condensed to 2,500 sq ft at $200/sq ft annually. That’s before counting the energy savings at all. Immersion also eliminates several recurring maintenance costs entirely — no CRAC/CRAH filter changes, no cooling tower biological growth control, no compressor servicing — though it introduces its own upkeep in the form of periodic fluid testing and reconditioning, typically every 5–8 years at $4–50 per liter depending on treatment approach.
What to Check Before You Retrofit
Before ordering cold plates or committing to any liquid cooling conversion, verify three things:
- Electrical infrastructure — can your existing power distribution actually support the new density, or does this retrofit trigger a separate electrical upgrade project?
- Water loop capacity — for RDHx and cold-plate approaches, does your existing chilled water infrastructure have the headroom, or does it need expansion first?
- Floor loading — liquid cooling hardware adds real weight. Cold plate retrofits typically add 15–25 kg per rack; confirm your raised floor and structural specs can handle it before you commit capital.
Retrofit costs generally run 15–30% higher than an equivalent greenfield deployment, due to integration complexity and the facility modifications above — budget for that premium rather than being surprised by it mid-project.
Also worth planning for early: fire suppression compatibility. Cold plate retrofits often require switching from older systems like FM-200 to something like NOVEC 1230 — an easy detail to miss until it holds up a project close to completion. Similarly, don’t underestimate the operational learning curve: hybrid deployments that mix air and liquid cooling during a phased transition allow gradual staff training and capital spreading, but retrofit costs in this scenario typically run 15–30% higher than a comparable greenfield immersion deployment, specifically due to that integration complexity.
CRAC vs. CRAH vs. Liquid Cooling: Quick Decision Guide
- Under 12 kW/rack, uniform load: Stick with well-executed CRAC/CRAH and proper containment. No need to over-engineer this.
- 12–30 kW/rack, mixed or growing density: RDHx is often the pragmatic middle step — meaningful heat removal without a full liquid conversion project.
- 30 kW/rack and above, especially GPU/AI workloads: Cold plate liquid cooling is effectively the current standard, not an exotic upgrade.
- 60–70 kW/rack and above, or extreme footprint/PUE requirements: Immersion cooling becomes cost-justified — below that threshold it’s usually more complexity than the workload requires.
Who to Talk to: Vendor Landscape
For RDHx and hybrid chilled-water retrofits, Vertiv and Motivair are among the most established names, with strong existing service networks that reduce the risk of a retrofit going unsupported. For direct-to-chip cold plate deployments, most enterprises are now buying factory-integrated hardware directly from Dell, HPE, Lenovo, or Supermicro rather than retrofitting third-party cold plates onto existing servers — this significantly de-risks compatibility questions. For immersion specifically, Submer, GRC (Green Revolution Cooling), and Asperitas are the names with the most commercial deployment history, though given the maintenance and hardware requalification requirements, this is a category worth bringing in a specialist consultant for rather than self-directing.
Frequently Asked Questions
Can I retrofit liquid cooling into an existing air-cooled facility?
Yes, but with real caveats — compatible server chassis, structural reinforcement for coolant piping, and updated fire suppression are all typically required. Most successful retrofits plan for these upfront rather than discovering them mid-project.
Do I need immersion cooling for GPU servers?
Not necessarily. Cold plates handle most AI workloads efficiently up to roughly 45 kW/rack. Immersion becomes cost-justified mainly above 60–70 kW/rack, or when footprint, acoustic, or PUE requirements rule out the alternatives.
How long does a typical cooling retrofit take?
Documented real-world retrofits have run as fast as four months for a full liquid cooling conversion — significantly faster than an 18-month-plus new build, which is a major part of the retrofit case beyond just cost.
What’s the difference between a CRAC and a CRAH unit?
A CRAC has its own integrated refrigeration/compressor and cools air directly. A CRAH instead relies on chilled water supplied from a separate central chiller plant, which is generally more energy-efficient at scale but adds a dependency on that chilled water infrastructure being sized correctly.
Is a hybrid air-and-liquid deployment a reasonable long-term approach, or just a transition step?
Both, depending on your facility. Many operators run hybrid deployments indefinitely — air cooling for standard enterprise racks, liquid cooling reserved for AI/HPC racks — rather than treating it purely as a stepping stone to full liquid conversion.
Modeling your own facility’s cooling upgrade path? Our Data Center Efficiency Suite includes PUE and TCO calculators to help you compare options before committing capital — or book a consulting session to work through your specific retrofit constraints directly.
Written by
Raajeev Ratra
Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management
Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.