AI-Ready Data Center Engineering: Liquid Cooling Retrofits for Existing Facilities

The data center industry was built around a thermal envelope that AI workloads have already left behind. A typical enterprise colocation facility designed in 2018 was sized for 5–8 kW per rack, with a handful of "high-density" cabinets pushed to 15–20 kW. Modern AI training racks routinely consume 60 kW, 100 kW, sometimes 150+ kW per rack. The math doesn't work. Air can't carry the heat.

Most owners don't have the option of building greenfield. The question is: how do you engineer the conversion of an existing, operating facility into something that can support AI workloads — without taking the building down, without losing existing tenants, and without spending the cost of a new build?

This piece walks through the engineering decisions in the order they actually have to be made.

1. Decide the target density profile, not "AI-ready"

"AI-ready" is a marketing term. The engineering question is: which AI workloads, at what density, in what fraction of the hall?

• 30–40 kW per rack. Modern dense GPU inference, high-end virtualization. Achievable with rear-door heat exchangers on existing IT and contained hot aisles.
• 60–80 kW per rack. AI training with current- generation accelerators. Requires direct-to-chip cold plates on the IT or aggressive in-row cooling, plus serious upgrades to the chilled-water plant.
• 100–150+ kW per rack. Frontier-class training. Almost always direct-to-chip; sometimes single-phase or two-phase immersion. Requires rethinking the hall, not retrofitting it.

Most retrofit programs target a fraction of the hall — say, 15–25 percent — not the whole facility. Pick the zone, model the mixed-mode load on the rest, and engineer for both halves of the hall.

2. Run the chilled-water plant capacity math first

Every heavy-density retrofit begins and ends with the chilled-water plant. If the plant can't carry the marginal load, none of the downstream choices matter.

• Total ton-hours required at the new density vs. installed plant capacity at design summer conditions.
• Approach temperature delta — direct-to-chip and RDHX prefer warmer chilled water (often 60–75°F supply) than legacy CRAH systems. Warmer supply often allows free cooling for more of the year, which changes operating economics.
• Pipe sizing and flow capacity from plant to hall. Many legacy halls have piping sized for 10 kW per rack. Bring the same flow to a 60 kW rack and the differential pressure equation changes everything.
• Pump redundancy and standby capacity. N+1 at 5 kW per rack may not remain N+1 at 60 kW per rack.

We've seen retrofit programs proceed for months under the assumption that the chilled-water plant has headroom, only to discover at commissioning that the existing distribution can't deliver flow at the new approach temperature. Run the plant math first.

3. Choose the heat-removal pathway: RDHX vs. direct-to-chip vs. immersion

Three paths cover most retrofit work. Each has a different engineering envelope and a different operational footprint.

Rear-door heat exchanger (RDHX)

RDHX is the lowest-friction conversion path. The IT stays air-cooled. Hot exhaust passes through a chilled-water heat exchanger before re-entering the room. Most rack form factors accept an RDHX without modification.

• Practical density ceiling: 30–60 kW per rack.
• Minimally disruptive to existing tenants and existing IT — the conversion can happen rack-by-rack.
• Adds chilled-water plumbing to every retrofitted rack. Leak detection and engineered drip pans become operational disciplines.
• Hot-aisle containment integrity matters even more — bypass air past the RDHX defeats the purpose.

Direct-to-chip (DTC) liquid cooling

DTC moves cold plates directly onto CPUs and GPUs. The cooling fluid goes inside the chassis. Air still cools low-power components like memory and drives.

• Practical density ceiling: 100–150+ kW per rack.
• Requires IT vendor support — most current-generation AI accelerators ship DTC-ready, but some legacy IT does not.
• Adds a coolant distribution unit (CDU) per row or per rack to isolate the IT-side fluid loop from the facility-side chilled-water loop.
• Operational discipline shifts toward fluid chemistry, manifold health, and CDU monitoring. Skill-mapped staffing becomes non-optional.

Immersion cooling

Immersion submerges the IT in dielectric fluid. Single-phase circulates the fluid; two-phase relies on phase change at the chip surface.

• Practical density ceiling: 200+ kW per rack.
• Requires structurally and operationally different halls. Floor loading, fluid handling, and IT installation all change.
• Best suited to greenfield zones or fully isolated retrofit areas rather than rack-by-rack conversion.
• Highest engineering and operational lift. Highest density payoff.

4. Engineer the electrical for the new load curve

Electrical engineering for heavy-density retrofits has its own set of decisions:

• Branch circuit and busway sizing. Existing busway rated for a 5 kW per rack design point will not support 60 kW per rack without modification or replacement.
• Rack PDU choice. 415V 3-phase rack PDUs at higher amperage are typical for AI racks. Single-phase 208V at 30A is not enough.
• UPS topology. Some AI training workloads tolerate short power excursions and run on raw utility for cost reasons. Decide deliberately whether the AI zone is on full UPS protection or a different topology.
• Generator and fuel. A 30 percent density increase across a hall often translates to a meaningful generator capacity question. Don't discover that at commissioning either.
• Coordination and arc-flash. A heavy-density retrofit changes available fault current and protective device coordination. New short-circuit study, new coordination study, new arc-flash labels.

5. Structural and floor-loading reality

High-density racks weigh more — some immersion tanks can exceed 4,000 lb fully fluid-loaded. Some legacy raised-access floors can't carry them. Concrete-slab installations are usually fine but can have local loading issues at the rack base.

• Existing floor live and dead load capacity vs. proposed equipment weight.
• Localized loading at rack base, lift truck pathways, and crane paths during installation.
• Vibration isolation for high-pump cooling systems near sensitive IT.

6. Network and structured cabling implications

AI fabric is denser than the typical leaf-spine deployment. Plan for:

• Higher fiber count per rack — 200/400/800 GbE links proliferating.
• Top-of-rack vs. middle-of-row vs. end-of-row choices in the heavy- density zone.
• Pathway capacity. Tray fill, conduit, and overhead routing all need a re-look.
• DWDM or coherent optics support for cross-facility AI clusters when the build spans multiple halls.

7. Operating the heavy-density zone

Engineering the hall is half the work. Operating it is the other half. Specifically:

• Leak detection at the rack, manifold, and CDU level. Operational response must be sub-minute.
• Fluid chemistry monitoring (DTC and immersion). Particulate, pH, conductivity all matter.
• Skill mapping for DCT and CIE roles in the heavy-density zone — most operators have not run liquid cooling at scale and need formal cross-training.
• Runbooks for fluid spill, manifold failure, CDU failure, and chilled- water plant fault under the new load curve.

See Operations Management for the operating discipline that runs alongside this engineering work.

8. Sequence the program around live tenants

The hardest part of an AI-ready retrofit is staging the work without disrupting the customers paying today's bills. Practical sequencing:

• Identify a containable retrofit zone — physically separable, ideally with its own electrical and cooling tap.
• Move existing tenants out of the zone first, into the rest of the hall, before any heavy work begins.
• Sequence chilled-water plant upgrades during shoulder seasons (fall / spring) when the load can absorb temporary capacity reductions.
• Commission and load-test in stages — bring the heavy-density zone up at 30 percent of design, then 60, then 100, with operational observation between steps.

How CR Technology runs this work

We engineer AI-ready retrofits as integrated programs across power, cooling, network, structural, and operations — not as a stack of single-discipline drawings. Specialty licensed work (civil, structural, MEP, fire protection) is coordinated and integrated under one accountable team. We warrant the combined deliverable.

See Engineering Services for the full pillar, or reach out at info@castlerocktechnology.com to scope a retrofit assessment for your facility.