INSIGHTThermal engineering · Analysis
There is no winner across all six axes, and any page that names one is selling something. Air cooling does not fail because air cannot carry the heat — it fails because a rack cannot be given enough floor area to receive the air.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
The short answer
First principles
The usual argument against air is that it is a poor coolant. True, but it sends operators to the wrong test. Air will carry any heat load you like if you move enough of it; what it cannot do is arrive. A rack occupies roughly one floor tile of plan area, and ASHRAE's TC 9.9 committee puts best-of-breed raised-floor delivery at about 1,900 cfm per tile.[1] That single number turns the cooling question into a floor-plan question: how many tiles of air does one tile of rack demand?
Assumptions
Cross-check. ASHRAE independently states that a 40–50 kW rack can demand up to 5,000 cfm against that 1,900 cfm tile.[1] The model above returns 5,265 cfm at 45 kW — within about 5%, which is the confidence every other row inherits. The committee states the same physics as a rack-fill limit: servers drawing 100 cfm or more per U would fill only 19U if one tile were the whole supply.[1]
A second bill lands on the critical power budget rather than the mechanical one. ASHRAE reports that fan power of 10% to 20% of server power is not uncommon on denser servers, and that moving from 2% to 10% is equivalent to cutting usable UPS capacity by 8%.[1] On a 1 MW critical block that swing is roughly 80 kW — capacity bought, cooled, and protected as if it were compute, then spent on moving air. That number, not any coolant property, tends to end the argument.
Figure 1 · Original analysis
Every bar is one rack drawing its air from best-in-class raised-floor delivery at a 15 K rise. The dashed line is the floor plan itself: one rack, one tile. Past it, a rack is asking for air the plan area in front of it cannot supply.
Air will carry any heat load you like if you move enough of it; what it cannot do is arrive.
1,900
cfm per tile — best-in-class raised-floor delivery
Table 1 · The numbers behind Figure 1
Read the tile column as a plan-area ratio, not a physical law: it is what best-in-class raised-floor delivery would have to put in front of one rack to carry that load at a 15 K rise.
| Rack load | Airflow (cfm) | Airflow (m³/s) | Tiles of delivery | Reading |
|---|---|---|---|---|
| 10 kW | 1,170 | 0.55 | 0.6 | Conventional. Nothing is broken. |
| 20 kW | 2,340 | 1.10 | 1.2 | Containment becomes mandatory. |
| 30 kW | 3,510 | 1.66 | 1.8 | Needs close-coupled or overhead supplement. |
| 45 kW | 5,265 | 2.48 | 2.8 | Matches ASHRAE's published figure — see cross-check. |
| 60 kW | 7,020 | 3.31 | 3.7 | Air delivery, not heat capacity, is now the limit. |
| 120 kW | 14,040 | 6.63 | 7.4 | No plausible floor plan delivers this to one rack. |
Table 2 · Original analysis
Six axes, three approaches. Read column-down for a method, row-across for where each one genuinely wins — because each one does.
| Axis | Air | Direct-to-chip | Immersion |
|---|---|---|---|
| AX-01Density ceiling | Federal-lab guidance places traditional air-cooled racks at 1–5 kW and calls liquid necessary across the 5–80 kW range.[3] | High enough that vendors ship it as the default: the GB200 NVL72 is described as a “rack-scale, liquid-cooled design” of 72 GPUs and 36 CPUs.[4] | ASHRAE credits immersion with broad temperature support, high heat capture, and high density — but single-phase natural convection has its own ceiling, past which forced convection or two-phase is required.[1] |
| AX-02Retrofit difficulty | None. It is the incumbent, and its capital is already sunk. | Moderate. Where facility water has not been brought to the rack, ASHRAE offers liquid-to-air heat exchangers — in the rack or at the row — as an interim bridge, at the cost of rejecting that heat back into the room rather than to facility water. Rack weights, aisle pitch, and rolling-load paths still bind.[1] | Highest. Horizontal tanks break the rack-and-aisle floor plan, and the load path — rolling loads, ramps, elevators — must carry a filled tank.[1] |
| AX-03Serviceability | Pull the server, swap it, no fluid involved. This is air’s strongest axis. | Dripless quick disconnects keep the rack orientation and the service motion, but cold plates add airflow impedance and hybrid air/liquid boards keep fan power high.[1] | ASHRAE: “a crane or two-man lift is often required to remove IT equipment hardware for service.” Yet immersion may cool high-power DIMMs without the serviceability problems cold plates create on the same parts.[1] |
| AX-04Fluid handling | No working fluid at the rack. The fluid problem is displaced to the chiller plant. | Treated water or glycol in a closed technology loop. Compatible metals and water chemistry are called crucial, alongside filtration with bypass and hydronic redundancy.[3] | The whole board is wetted. ASHRAE recommends a materials-compatibility assessment and a warranty-impact evaluation before deployment;[1] OCP publishes a dedicated component immersion compatibility test guideline.[8] |
| AX-05Ecosystem maturity | Total. Every vendor, every integrator, every technician already speaks it. | Converging fast: OCP maintains vendor-neutral cold-plate requirements[6] inside a broader cooling-environments programme.[5] | Active but younger. OCP immersion requirements sit at Rev 2.10[7] with a separate 2026 compatibility guideline;[8] ASHRAE describes a field of many startup and established players.[1] |
| AX-06Heat capture fraction | By definition none to liquid — every watt leaves through the room. | Depends entirely on what the plates touch. Federal-lab practice specifies at least 95% of rack heat captured directly to liquid;[3] a CPU-only design that leaves memory on air misses that and erodes the economic case.[1] | Essentially complete by construction — the fluid touches everything on the board. |
Reading the matrix
AX-03 is the row worth re-reading. Received wisdom makes immersion the serviceability loser, and for whole-chassis work it is — ASHRAE is explicit about the crane. But the same paper observes that immersion may cool high-power memory adequately without the serviceability problems cold plates introduce on those same parts.[1] Cold plates are excellent where they touch and irrelevant where they do not, and memory is exactly that boundary.
Run these in order and stop at the first that is true for your site. The method is a consequence of the answer, not a preference expressed before it.
Practice
Compute your own airflow number before evaluating a single product. Rack kW and your real server delta-T give the cfm; compare it against what your delivery method can put in front of that rack. If the ratio exceeds one, the shortlist is already decided.
Ask vendors for heat capture fraction, not cooling type. Federal-lab practice specifies at least 95% of rack heat captured directly to liquid;[3]a quotation that says “liquid-cooled” without a capture fraction is describing a hybrid whose residual air plant you still have to build.
Move fan power into the critical-power line of the model. It runs on the same UPS as the GPUs, and ASHRAE puts the 2%-to-10% move at 8% of UPS capacity.[1]
Name an owner for fluid before the first fill. Chemistry, filtration, material compatibility, and warranty impact are operating disciplines with a schedule, not a commissioning task that closes.[1]
Re-ask the maturity question annually. Uptime Institute's 2025 operator survey — its fifteenth — frames the year as rising costs, worsening power constraints, and challenges in meeting the demands for AI.[9] The inputs that decide this matrix are the ones actively moving.
Honest limits
The calculation above is a heuristic built on published figures, and it fails in several identifiable ways.
In the product
PODOS answers the fourth constraint — the one where the building binds. Each PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs, with a closed direct-to-chip liquid cooling loop specified as part of the enclosure rather than added to a room. That is not a claim that cold plates beat immersion on every axis — the matrix says plainly that they do not. It is a decision about which constraint the product removes: the site's. Because the loop, the thermal enclosure, and the rejection interfaces are integrated and tested in the factory, the cooling system arrives commissioned — one reason PODOS targets a 90-day window from order to commissioning for a standard unit.
Review schedule: 90 days. ASHRAE facility water classes[2] and the OCP requirement revisions cited here are the fastest-moving inputs on this page.
Bring the rack load, the delivery method, and the floor plan. Engineering will tell you which row of the matrix you are actually standing on.