Ask an operator with a few hundred machines across two or three generations where a particular unit is, and the answer is usually some version of “it is in with the rest of mine.” That is the intuitive way to organize a floor: each customer’s hardware together, in one place, easy to point at.

It is also the arrangement that produces the strangest fault reports in hosting, because a row containing three different machine models is not one thermal environment. It is three, sharing air.

Machines in the same row are not in the same conditions

A row of identical units behaves predictably. Each one draws a similar volume of air at a similar temperature, heats it by a similar amount, and pushes it out at a similar rate. The hot aisle gets hot evenly. The pattern repeats down the row, which is exactly what makes it possible to design for.

Put a mix in the same row and that stops being true. Generations differ in how much air they move and how much heat they put into it. A machine that moves a lot of air is not just cooling itself, it is influencing the pressure on both sides of the rack around it. A machine that moves less air sitting beside it does not get the same inlet conditions it would get among its own kind.

The effects are not dramatic enough to look like a cooling failure. They look like a specific machine that throttles more than its neighbors, or a unit that runs a few degrees warmer than the identical unit two slots along, or one position in a rack that keeps having problems regardless of which machine is put in it. Those reports get filed as hardware faults. Some of them are placement.

Our guide to rack design and airflow layout covers how to size airflow for a hall. The point here is narrower: a hall correctly sized in aggregate can still have rows that do not work, if the machines in them are not alike.

Recirculation is the mechanism

The specific failure is short-cycling. Hot exhaust finds a path back to an intake instead of leaving the aisle, and a machine ends up breathing air it already heated.

Mixed rows create those paths, because machines of different generations are not the same physical size and do not present the same face to the aisle. Gaps between units of different depths, blanking that fitted the old model and not the new one, and differences in where the intake actually sits all give hot air somewhere to go that it should not have.

A machine that recirculates does not fail. It derates, quietly, and keeps hashing at reduced output. Nothing alarms. The fleet dashboard shows the unit online. The loss shows up only as an efficiency number that is slightly worse than it should be, which is the kind of problem that persists for months because nothing ever asks for attention.

Containment helps, which is why it exists, and our note on hot and cold aisle containment is worth reading alongside this. But containment works by making the aisle behave as one sealed volume, and that assumption is weakest exactly where the row is irregular.

The electrical side is less forgiving than the thermal side

Thermal problems degrade. Electrical problems trip.

Different generations do not draw the same power, and they do not all accept the same input voltage. A circuit sized and balanced for one model is not automatically correct for another, and a rack populated by swapping a few units for a newer, hungrier generation can quietly end up closer to its limit than anyone intended.

Nothing is wrong while the fleet runs steadily. The problem appears at startup. Machines coming up together draw more than they draw running, and a rack that sits comfortably within its circuit in normal operation can exceed it when everything in it starts at once after an outage. That is the classic post-outage fault: the site comes back, a breaker trips, and the obvious suspect is the machine that happened to be starting when it went.

Mixed racks make this harder to reason about, because the total depends on exactly which models ended up in which positions, and that arrangement changes every time a unit is swapped. A rack of identical machines has one answer. A mixed rack has an answer that drifts as the fleet changes.

Power quality compounds it. Power supplies of different vintages behave differently on the line, and our article on harmonic distortion and power quality covers what that does upstream.

Why grouping by model wins

Organizing the floor by machine rather than by owner fixes most of this by removing the irregularity rather than managing it.

• Each zone has one thermal profile. Set points, fan curves and containment can be tuned for the hardware actually in that space instead of compromised across three kinds at once.
• Circuits become predictable. A rack of one model has a known draw running and a known draw starting. Staged restart after an outage can be planned per zone instead of guessed.
• Faults become comparable. When a hundred identical machines sit in identical conditions, an outlier is genuinely an outlier. In a mixed row you cannot tell a failing unit from a badly placed one, which is the diagnostic cost nobody prices.
• Maintenance batches. Spares, firmware and cleaning intervals are per model. Grouping them means a technician works one procedure through a zone rather than switching between three.
• Curtailment gets cleaner. Powering down the least efficient hardware first is simple when that hardware is physically together and on known circuits.

The problem arrives with every new generation

This is not a one-time layout decision. It recurs, because fleets do not refresh all at once. Machines are bought in batches as capital allows and as hardware becomes available, which means a site that was uniform eighteen months ago is mixed today whether anyone planned it or not.

The temptation when new units arrive is to put them where the old ones were. The slot is free, the power is already run, and the rack is already there. It is the fastest thing to do and it is how mixed rows happen without anyone deciding to create one.

A newer generation in an old position usually wants more power and moves air differently, so the position that suited the outgoing machine is not automatically right for the incoming one. The honest version of a fleet refresh is a re-layout: decide where the new zone is, move what needs moving, and accept that this costs labor. Sites that skip it accumulate irregularity one batch at a time until nobody can say why a particular aisle runs warm.

The same applies to partial failures. When a unit dies and is replaced from stock with whatever model is available, that single swap puts a different machine into an otherwise uniform row. One swap is nothing. Two years of swaps is a mixed row that nobody intended and no record explains.

Mixing cooling types is the harder version

Everything above concerns air-cooled machines of different generations. Mixing cooling methods in one building is the same problem with less margin.

Air, immersion and hydro-cooled hardware do not just differ in how heat leaves the machine. They differ in where the heat ends up, what infrastructure has to be present, and what the room is expected to do. An immersion deployment moves the thermal load into a fluid loop and a heat exchanger; an air-cooled row expects the room to carry it away. Those are not two configurations of one facility so much as two facilities that happen to share a roof.

Our overview of air, immersion and hydro cooling covers the methods themselves. The layout consequence is that these belong in separate, purpose-built areas with their own thermal design, not interleaved to keep one customer’s hardware together. A facility that offers more than one cooling method and cannot tell you where the boundary between them sits is worth more questions.

What this means for a customer

The immediate implication is that your machines will probably not be in one place, and that is a sign the facility knows what it is doing rather than a sign it has lost track of your fleet.

What matters is not adjacency. It is that the facility can tell you, per machine, where a unit is and what circuit it is on, and that your serial manifest is maintained independently of where things physically sit. A host that groups by model and keeps good records gives you better conditions than one that keeps your fleet together and tunes the room to an average.

There is one case where it genuinely matters to be together: if you need physical access to your own hardware, or you are running something the facility does not otherwise host. Worth raising before signing rather than discovering at install.

The question worth asking a prospective host is simply how they decide where a machine goes. An answer about thermal zones and circuit loading is a good sign. An answer about keeping each customer’s units together is a reasonable thing for a small site to do and a warning sign at a large one.

If you are placing a fleet that spans more than one generation and want to know how it would be zoned, talk to us. It is a more useful conversation before the machines ship than after they are racked.