The power math comes first

A typical legacy hall was built around 5 to 15 kW per rack, raised floor, perimeter CRAH units, and a utility service, UPS, and generator plant sized to match.  When ownership says "we need 80 kW racks," the conversation drifts to chillers and cold plates.  That is the second conversation.

Run the arithmetic on a 12 MW critical IT hall.  At 12 kW per rack it holds about 1,000 racks.  At 100 kW per rack the same 12 MW feeds 120 racks.  Your electrical capacity did not go stranded.  Your floor space did.  Roughly 880 rack positions now sit in a room that has no power left to give them.

That changes the business case.  You are not filling a hall with AI racks.  You are carving out dense pods inside a hall and deciding what happens to the rest of the floor.  Options are to keep legacy air-cooled colocation or enterprise load in the remainder, decommission it, or reuse the space for the mechanical and electrical equipment the dense pods will need.  The owners who skip this step buy cooling for 12 MW of density and discover they can only power a fifth of it.

Audit every asset before you reallocate a dollar

Existing asset Usually reusable? What to check
Utility service, MV switchgear Yes Available headroom, protection settings, step-load behavior
Generators Usually AI training loads swing hard and fast.  Verify governor response and step-load acceptance with the OEM
UPS and batteries Often Topology, ride-through time, and whether the load will accept a lower redundancy tier
PDUs, RPPs, branch circuits Rarely Built for 30 to 60 A at 208 V.  A 100 kW rack needs higher-amperage busway and different rack power
CRAH and CRAC units Partly Keep for the residual air load, retire the rest
Chillers Yes, at higher setpoints Many liquid systems accept much warmer facility water, which opens up economizer hours
Chilled water piping Seldom for the rack loop Wrong materials, wrong size, wrong cleanliness for a technology cooling loop
Raised floor and slab Check structure first Liquid-filled racks are heavy.  Many pods need slab-on-grade or reinforced zones

Anything marked "check" needs an engineer's signature before it enters a budget.

The 70/30 problem

Direct-to-chip cold plates typically pull somewhere around 70 to 80 percent of a rack's heat into liquid.  The rest still leaves as hot air.  On a 100 kW rack, that is 20 to 30 kW of air-side load, which is more than an entire legacy rack ever produced.

So the air system does not go away.  It gets resized for a different job: fewer, hotter pockets of residual load.  Rear-door heat exchangers are the usual answer for that remainder, and they tie back into the same facility water.  Owners who retire all air handling to save money end up with GPUs that run warm and memory and power supplies that cook.

Mechanical loop redesign

Think in three loops.

  1. Facility water system. Chillers, dry coolers, or cooling towers produce water for the building.  This loop is the part you most likely already own.
  2. CDU heat exchange. The coolant distribution unit separates the facility loop from the rack loop.  Row-level CDUs run in the hundreds of kilowatts.  Facility-scale units reach into the megawatt range.  Pick per pod, not per building.
  3. Technology cooling system. The sealed secondary loop that carries coolant, usually a propylene glycol mix, to manifolds and cold plates.  This is a new build.  Fine filtration, compatible wetted materials, and a disciplined flush-and-fill procedure matter more here than anywhere else in the plant.

Supply temperature is where the money hides.  Many direct-to-chip designs accept facility water in the 30 to 40 degree Celsius range and above, ASHRAE W32 through W45 classes.  At those temperatures you can run dry coolers for much of the year and cut chiller hours sharply.  Confirm the server vendor's required inlet temperature before you size anything.

Redundancy needs a harder look than it usually gets.  An air-cooled hall coasts through a pump failure on thermal mass.  A cold-plate rack does not.  Accelerators throttle or trip within seconds of losing flow.  So CDU pumps, controls, and the leak-detection and valve logic belong on UPS power, with N+1 CDUs per pod and dual pumps in each.  If your design review does not include a flow-loss scenario with a stopwatch, it is incomplete.

Immersion belongs in the conversation, though rarely as the default.  Single-phase and two-phase tanks handle extreme density, and they strip out fans.  The costs show up elsewhere: slab loading, dielectric fluid cost and handling, server warranty and OEM support constraints, and a maintenance workflow your technicians have not done before.  Most current large-scale accelerator platforms ship with direct-to-chip as the reference design.  Use immersion where a specific customer or workload justifies it, and prove OEM support in writing first.

Structure and live-hall risk

A fully loaded liquid-cooled rack can exceed 3,000 pounds.  Legacy slabs and raised floors were often designed for something between 150 and 250 pounds per square foot.  Get a structural engineer's calculation on the real floor, not the drawings, and budget for reinforcement or slab-on-grade pods.

Then there is the hall you are working next to.  Pipe in a live white space is a leak waiting for a calendar date.  Hold these rules on every conversion:

  • No wet work above or adjacent to energized equipment without isolation and a method of procedure approved by operations.
  • Prefabricate and pressure-test skids and manifolds off-line.  Flush and fill the loop before it gets anywhere near a server.
  • Automatic leak detection with zone isolation valves, tested before first energization.
  • Dust and vibration controls, because the adjacent customers' hardware is still running.

Reallocating the capital

Legacy facilities carry a refresh budget: CRAH replacements, battery cycles, raised-floor repairs, PDU upgrades.  A conversion program redirects that budget.  Stop spending on air-side refresh inside the pods being converted.  Start spending on CDUs and secondary loops, high-amperage busway and rack power, structural work, leak detection and controls integration with the EPMS and BMS, and a longer commissioning window.

Test the conversion against a new build.  Published estimates for new critical capacity have clustered around roughly $10 million to $15 million per MW of IT load excluding the IT equipment, and AI-oriented builds sit toward the high end.  Use your own bids rather than those numbers, but the logic holds.  A retrofit pod only wins if its cost per usable MW beats new construction, or if it reaches revenue years sooner.  In markets where a new interconnection takes three to seven years, speed alone usually carries the case.

Sequence the money by pod.  Fund a pilot pod of 1 to 2 MW first, with its own commissioning and a load bank test at full heat output.  Release the next tranche only after the pilot clears thermal acceptance, flow-loss testing, and a leak response drill.  Liquid-cooled load banks exist for exactly this purpose.  Use them.

Decision gates

  1. Power and structure sign-off. Deliverable MW to the pod, floor loading, and cooling plant capacity verified by the engineer of record.
  2. Customer and OEM requirements frozen. Rack power, inlet temperature, flow rate per rack, and warranty conditions in writing.
  3. Pilot pod complete. Flushed, filled, load-banked, and signed off by commissioning.
  4. Operations readiness. Staff trained on CDU operation, fluid chemistry checks, and leak response before the first customer rack lands.
  5. Scale decision. Remaining pods released only against signed demand.

Related briefs: Behind-the-Meter Power and Advanced Nuclear | OEM Capacity Reservations and Supply Chain Capital Protection