Is Your Data Center Thermally Ready for AI? What to Validate Before the Dense Racks Arrive
Keep It Cool — Practical insights on data center airflow and cooling performance.
A rack that ran fine at 8kW will not run fine at 30kW, and the room around it changed the day you added the load, whether or not anyone measured it.
That's the part most of the AI cooling conversation skips. The headlines are all about liquid cooling and 130kW racks, which is useful if you're building a hyperscale AI factory from scratch, and a lot less useful if you're an operator with a working air-cooled room and a mandate to support a few high-density racks by next quarter. Most operators are in that second group, and this is written for them.
What's happening
Rack density has climbed from around 6kW a decade ago to roughly 16kW today, and AI workloads want 30 to 40kW or more, with some 2026 surveys putting the average rack already at 27kW. The top end lives in another world entirely, where a fully loaded NVIDIA GB200 NVL72 rack pulls 120 to 130kW.
Air cooling was built for 8 to 12kW racks, and its practical ceiling sits around 20 to 25kW under good conditions, with direct-to-chip liquid cooling becoming the requirement closer to 35kW. So there's a band, roughly 15 to 35kW, where operators are pushing air past its comfort zone but haven't moved to liquid yet, and that band is where the surprises live.
The liquid-cooling headlines miss something, though. Most operators are not going fully liquid, they're going hybrid, adding a few dense racks to a room that's still mostly air-cooled and running it that way for years. The AFCOM 2026 report found more than one in four racks now use liquid cooling, with another 28% of operators planning to adopt it within two years, and that's a long transition where the air side still has to work the whole time. Even inside a liquid-cooled rack, the cold plates only handle the processors, and the rest of the server, 20 to 30% of the heat load, still comes off into the air.
Why it matters
Air cooling fails at high density in a specific way. As the load climbs, the air can't carry heat away fast enough at the intake, so you get recirculation, bypass, and hot spots, even when the room has plenty of cooling capacity on paper.
That last part is the trap, because the room average can look completely normal while one new dense rack starves for air at the top. We've seen rooms where the BMS reported uniform conditions and a single rack row had a 15-degree spread between bottom and top, and the averages smoothed over the exact problem the operator needed to see. Add a 30kW rack to that room and the math gets worse, not better, with more heat concentrated in one spot and drawn through an air path that was already marginal.
The assumptions you're working from are probably stale too. Meta has run 120kW racks in buildings designed around 20kW air-cooled assumptions, and if your CFD model was built at commissioning, every server swap, tile move, and containment tweak since then has pulled the room further from the model. The model describes the room you designed, not the room you have.
What to look for
Before you add density, check the things that air cooling gets wrong first. None of this requires touching equipment or taking anything offline.
Rack inlet temperature at three heights. A rack can pass at the middle and cook at the top, and one reading per rack hides that spread.
Delta-T across the racks you're about to load up. If the temperature difference is already off before you add the dense gear, it won't get better afterward.
Bypass and recirculation near the planned dense racks. Cold air that returns to the unit without doing work, and hot exhaust pulled back into the intake, both get worse as density climbs.
Blanking panels and containment gaps in that row. These are the cheap fixes that matter most exactly when the load goes up.
The resiliency window. If a cooling unit near the dense racks fails, you want to know how long you have before a critical threshold, and at high density that window shrinks fast, so it's better to have the number before the failure than during it.
What to do next
Baseline the room before you add the load, not after the hot spot shows up. A before-and-after check gives you two things, telling you whether the room can take the density at all, and giving you a clean comparison once the racks are in, so that if something changes you'll know what and where instead of guessing.
This isn't a CFD study or a six-month engagement. It's a portable scan that runs in a live room with no installation, no shutdown, and nothing taken offline, and you can have rack-level data and a report to show leadership inside a day.
The operators who get caught out are the ones who treat the air side as solved because they're planning for liquid, and the cutover runs for years, with the air-cooled racks, the supporting spaces, and the transition period being exactly where thermal surprises happen. Checking the air before you add the load costs a lot less than finding out after.
Take the Cooling Risk Quiz to see where your room stands, or book a call to talk through a thermal assessment before your next density increase.
About Purkay Labs
Purkay Labs helps data center operators see what their cooling is doing at the rack. Our portable thermal assessments collect rack-level temperature and airflow data, then turn it into heat maps and a practical report that shows where cooling is reaching the equipment, where it is missing the load, and what to check next. It is a fast way to benchmark the room as it is today, especially before a layout change, capacity discussion, or leadership review.