For years, cooling a data hall was a settled question: raised floor, CRAC units, contained aisles, done. That still works, right up until a rack asks for more power than air can carry away. A single high-end accelerator can pull a thousand watts on its own, and eight of them in one chassis push a cabinet past what a well-built air-cooled room handles gracefully. Cooling is no longer a facilities detail bolted on at the end. It decides which servers you can buy, how much of your floor you can actually use, and who gets paged when something fails at 2 a.m.

Rack density is the filter that eliminates options fastest

Start with the number your racks will need, not the number they need today. As a working rule of thumb, conventional air cooling handles roughly 10–15 kW per rack comfortably. With aggressive containment, blanking and rear-door heat exchangers, 20–25 kW is reachable, but the last few kilowatts cost disproportionately more air volume and fan power.

Direct-to-chip liquid cooling picks up where that curve steepens. Cold plates on processors and memory let a rack run at 30 kW and well beyond, with the practical ceiling set by your coolant distribution units and the facility water loop behind them.

Immersion sits at the top. Because the fluid surrounds every component, a single tank can absorb tens or even hundreds of kilowatts across the servers inside it. The limit becomes the tank design, the power feed, and how much weight your floor can take — not the airflow.

Air cooling: familiar, cheap to staff, hard to stretch

Air's weakness is physics. Water carries roughly 3,500 times more heat per unit of volume than air does, so an air-cooled hall must move enormous quantities of it. Fan energy climbs steeply as you push more airflow, and fans can consume a meaningful slice of a server's total draw.

Where air still wins is in mixed, moderate-density environments and existing buildings. It needs no water supply, no chemistry program, no leak detection. Any mechanical technician can service a CRAC unit or swap a filter. If your racks will stay under about 15 kW and you have room to contain aisles and raise supply temperatures into the ASHRAE allowable range, spending on liquid is money you do not have to spend. Rear-door heat exchangers are the useful compromise: they add a water loop but leave the servers themselves air-cooled and standard.

Direct-to-chip liquid: the practical middle ground

Cold plates sit directly on the hottest silicon, and a technology cooling loop carries heat to a CDU that rejects it to the facility water. Most major server vendors now offer liquid-ready SKUs, which matters enormously for warranty and support.

What you still need is residual air cooling for drives, network cards and power supplies, unless you specify otherwise. The operational model stays familiar: technicians pull a server out of the rack as they always have. The new work is on the loop — fluid quality, filtration, biocide or inhibitor programs, quick-disconnect inspection, and leak detection you trust.

Immersion: full submersion, full commitment

Servers go into a tank of dielectric fluid, either single-phase where the fluid stays liquid and is pumped to a heat exchanger, or two-phase where it boils on hot components and condenses on a coil above. Two-phase moves heat beautifully but adds fluid handling complexity and typically pricier fluids, some with environmental profiles worth checking before you commit.

Immersion is not a drop-in. Fans usually come out, some heatsinks are removed, and gaskets, cables, connectors and labels must be compatible with the fluid. Service changes character: reaching a component means lifting a lid, draining or wiping, and accepting that a five-minute swap becomes a planned event. In exchange you get extraordinary density, near-silent halls, and warm fluid that can feed dry coolers or heat-recovery systems far more easily than lukewarm air.

Retrofit complexity: what the building has to give up

Air is the least invasive upgrade. Containment panels, blanking, sealing cable openings and a controls tune-up can often be done in occupied halls over a few weekends.

Liquid cooling asks the building harder questions:

• Water: Is there a supply, and can quality be maintained without corrosion or scaling?
• Weight: Full tanks and CDUs are heavy. Slab loading may need a structural review.
• Height and access: Immersion lids need vertical clearance; tanks need aisle space for service.
• Pipe routing: Manifolds and secondary loops have to reach the racks without blocking airflow or service paths.
• Containment and drainage: Leak detection, drip trays and a plan for containing spills.
• Power distribution: Busway ratings, whip lengths and redundancy for both the IT load and the CDU pumps.
• Commissioning: Flushing, pressure testing, verifying flow and temperature at every rack, and documenting maintenance bypass procedures.

Maintenance and the people who do it

Air maintenance is well understood: filters, belts, coil cleaning, humidity control. Liquid adds fluid sampling, pump service, CDU filter changes, and periodic checks of every quick disconnect and hose. Immersion adds fluid testing, top-ups, eventual replacement and disposal, plus pump and coil upkeep on the tank.

The quiet risk is staffing. Far fewer technicians have hands-on immersion experience, and vendor documentation varies in quality. Decide early who owns the cold plate loop — the facilities team or the IT team — and write it into the service contract. Confirm the exact server SKU is supported in your chosen cooling mode before purchase, not after.

Total cost: count the things that don't show up on a quote

The energy savings from liquid are real — mostly from removing server fans and running chillers less — but they are rarely the whole story. Model these together:

• Capex: CDUs, tanks, manifolds, cold plates, piping, leak detection, heavier racks and power distribution.
• Opex: pump and fan energy, chiller hours, water treatment, and any fluid consumed or replaced.
• Floor space: higher density means fewer racks, fewer aisles, and sometimes a smaller building footprint.
• Risk: one leak in the wrong place can cost more than years of saved electricity. Price the mitigation you actually need.
• Deferred capacity: if liquid lets you use a hall you already own, that is often the largest line item of all.

A sensible way to decide

Measure before you commit. Log rack inlet temperatures and actual power draw per cabinet for a couple of weeks — nameplate numbers mislead in both directions. Then set a density target for the next three years, not just the next deployment, and ask what your building can physically bear; floor loading, water chemistry and electrical capacity deserve a qualified mechanical or structural engineer's eye rather than a guess.

Next, confirm vendor support for the exact server model in the exact cooling mode you plan to run. Then pilot it: one liquid-cooled rack or one immersion tank, instrumented, run long enough to include a component failure and a firmware update. Compare service time, PUE, fan energy and technician comfort against your air baseline.

Finally, be honest about your operational appetite. Air remains the right answer for plenty of halls. Direct-to-chip is usually the least disruptive way to reach high density in an existing building. Immersion rewards sites with genuinely extreme density, cheap floor space relative to compute, and a team willing to relearn how service works. Pick the one your staff can run on their worst day, not their best.

Photo: Akela999 / Pixabay