Why this exists
The $700B Class-A AI-tenant capex cycle is producing a new question from owners we have not seen at this volume before: my prospective tenant just told me they want 60–120 kW racks — what do I do about cooling, and on what timeline?
The honest answer is rarely "rip and replace." It is a 90-day branching decision that depends on rack density, current cooling profile, plenum and column geometry, jurisdiction code regime, and how much of the capex envelope an AI-tenant lease will actually absorb. This piece is the decision tree we walk owners through.
ANALYSIS_GRADE: BENCHMARK. Numbers below are benchmark-grade, not simulation. Treat them as a triage frame to scope a Stage-2 engineering study, not as a procurement specification. Sources cited inline.
Step 1 — Input parameters (collect before opening the decision tree)
The decision tree only works if you can fill in five inputs. If you cannot, the first deliverable is closing those gaps — typically a two-week tenant-intake + site walk.
- Tenant rack density (kW/rack): The single most-load-bearing input. Below 30 kW you have time; 30–70 kW you have choices; 70–120 kW the choice has been made for you.
- Floor area committed to the AI envelope (sf): Distinguishes a partial-floor pilot from a full-floor build-out. Drives whether the cooling decision affects the rest of the stack.
- Jurisdiction: New York City, Singapore, Hong Kong, Austin, and Frankfurt each have different code-triggered upgrade obligations, different water-use restrictions, and different grid-interconnection lead times.
- Current cooling profile: Air-cooled with raised-floor plenum height, water-cooled with chilled-water loop temperature, or air-cooled with hot-aisle containment retrofit. Determines the cheapest pivot path.
- Lease economics: Tenant capex contribution, term length, and rent uplift envelope. Decides which retrofit class actually pencils.
Step 2 — Air-cooled feasibility check (the 70 kW ceiling)
Conventional perimeter-cooled or raised-floor air systems hit a thermodynamic ceiling between 25 and 35 kW per rack. With aggressive hot-aisle containment, rear-door heat exchangers, and tight floor management, well-engineered air can reach 60–70 kW per rack. Beyond that, air physics stops cooperating.
| Rack density | Air-cooled feasibility | Typical envelope |
|---|---|---|
| ≤ 25 kW/rack | Conventional air works | Plenum-tuning, containment, controls upgrade |
| 25–45 kW/rack | Air works with containment + rear-door HX | Hot-aisle containment + RDHx units, ~$1,200–2,400/rack |
| 45–70 kW/rack | Air is at its physical ceiling | Full containment + RDHx + chilled-water loop temperature drop, ~$3,500–6,500/rack |
| 70–120 kW/rack | Air is no longer feasible | Direct-to-chip liquid loop required |
| > 120 kW/rack | Immersion territory | Single-phase or two-phase immersion |
A second gating check: plenum height and column spacing. If your slab-to-slab is under 13 feet, or your columns are tighter than 30 feet on-center, retrofitting in-row chilled-water cabinets eats usable floor area faster than the rent uplift justifies. Retrofit Compliance Scan flags this geometry constraint inside the 90-day window before it surfaces in a leasing committee.
Step 3 — Liquid-cooled pivot triggers
The owner triggers a liquid-cooled pivot when any one of these is true: rack density crosses 70 kW; the tenant's AI workload roadmap shows a 24-month path to 100+ kW racks; the jurisdiction tightens code on chiller-plant electrical load (NYC LL97, EU EPBD, Singapore Green Mark 2026 are the live ones); or the air-cooled CapEx-per-rack number passes parity with rear-door heat-exchanger CapEx, which historically happens around the 60 kW threshold.
Direct-to-chip (DTC) liquid cooling is the default first pivot. It keeps the rest of the building on chilled water, contains the high-density envelope to specific halls, and avoids the floor-load and waterproofing scope of full immersion. Operational hand-off cost ranges $8,000–22,000/rack depending on whether you are sharing CDU (coolant distribution units) across racks or running per-rack CDUs.
Step 4 — Immersion-cooled edge cases
Immersion (single-phase mineral oil or two-phase fluorochemical) shows up in three situations: rack densities over 120 kW that DTC cannot service; tenants who want PUE below 1.05 and are willing to pay for it; or owners who have a structural advantage (existing pool deck, basement with floor-load capacity over 250 psf) that makes immersion economical. For 95% of Class-A office-to-AI conversions, immersion is the wrong answer — the floor-load, fluid-handling, and code-trigger overhead overwhelm the efficiency gain.
The other place immersion shows up is in greenfield BTS (build-to-suit) campuses where the tenant has pre-committed to a specific chip platform. That is a different transaction class and not in scope for this piece — it belongs in the greenfield decision tree, which a later post will cover.
Step 5 — Cost envelope by jurisdiction
Capex-per-rack varies more by jurisdiction than by cooling class. The dominant cost drivers are not the cooling equipment — they are electrical interconnect upgrades, water permits, fire-suppression code triggers, and structural review. The benchmark band below covers the cooling-system retrofit alone; full project cost typically runs 2.4×–3.6× this number once electrical, structural, and code-triggered scope is added.
| Jurisdiction | Air-cooled retrofit ($/rack) | DTC liquid ($/rack) | Immersion ($/rack) | Dominant cost driver |
|---|---|---|---|---|
| New York City | $4,200–7,800 | $14,000–24,000 | $28,000–48,000 | LL97 carbon penalty + DOB structural review |
| Singapore | $3,800–6,400 | $11,000–19,000 | $23,000–38,000 | Green Mark 2026 chiller efficiency floor + water permit |
| Hong Kong | $4,500–7,200 | $13,500–22,000 | $26,000–42,000 | BD electrical interconnect + cooling-tower siting |
| Austin | $2,800–5,000 | $9,500–16,500 | $19,000–31,000 | ERCOT interconnect queue + city building code |
These are 2026 working bands sourced from owner-reported retrofit pricing on conversion projects in the last 18 months, anchored against our $700B AI-tenant capex frame. They are not RSMeans-grade estimates and they will be wrong for any specific project — that is what Stage-2 engineering exists for. Use them to decide whether to commission Stage 2, not to underwrite a lease.
What this connects to
The cooling decision does not sit alone. Three other AISB analyses pair with this one:
- Data-Center Cooling for the Class-A AI Tenant — the broader $700B capex frame and why the 60–120 kW band is the strategic question for 2026 Class-A office conversion.
- Retrofit Compliance Scan — the 90-day audit that catches code-triggered upgrade obligations before they surface in committee.
- Open Protocol Moat — why owners with multi-vendor BMS and IoT infrastructure have more cooling-decision optionality than single-OEM-locked buildings.
What to do next
If you are inside the 90-day window with an AI-tenant inquiry, the right first move is not to commission an engineering study. It is to run the inputs in Step 1 against the matrix in Steps 2–5, then ask three operator-questions: what is the tenant's 24-month density roadmap, what does my code regime trigger at the new cooling capacity, and what is my lease economics envelope? If two of the three answers are uncertain, your first deliverable is a two-week intake — not a $200K engineering report. Bring that intake to /ask/ and we will walk it with you.
This post is a sequel to Data-Center Cooling for the Class-A AI Tenant and pairs with the broader Retrofit Compliance Scan framework.