When is a datacenter an AI datacenter?

Why "traditional" and "AI" are soft terms, and why power per rack is the number almost everything else hangs on.

Robert Lukkenaer, Head of Datacenters at Project Enki | September 2026

I am often asked what exactly separates a traditional datacenter from an AI datacenter. Most people are looking for a hard definition: a specific number of kilowatts per rack, a type of chip, or a line you can draw somewhere.

That line does not exist. Traditional, high density, HPC and AI are terms that point in a direction. Underneath them sits one number that determines almost the entire design: power per rack.

To understand why that single number carries so much weight, it helps to look first at where datacenters came from.

From a patch cable at Science Park to corporate IT

Datacenters started small and fairly by accident. In places where a lot of fiber connections happened to come together, such as Science Park in Amsterdam, someone put up a building with racks in it. You rented a full rack or a single position in one, ran a cable, and your web server was online.

Then came the companies. They had their own server rooms, including cooling and emergency power, and found that managing all of it themselves was expensive and impractical. Corporate IT moved to external datacenters: mail servers, ERP systems, document management and other standard applications, usually spread across a limited number of racks.

A rack like that used roughly 2 to 5 kW on average. That is a rough rule of thumb, since a rack full of storage draws something quite different from a rack full of CPUs. As a starting point for designing a building, though, that range worked well for years.

Until processors started drawing more and more power.

Compute that has to work together

A processor that computes more draws more power and produces more heat. Follow that through and you arrive at high-performance computing: workloads in which large numbers of processors work on a single problem together.

This combines CPUs and GPUs, because each is good at a different kind of work. A GPU is exceptionally fast at certain parallel calculations, but it is not suited to every task. A CPU is far more general purpose. A CPU can in principle do the work of a GPU, but it takes much longer and uses considerably more energy doing so. In the other direction, a GPU cannot simply take over every CPU task. The system only works well when both types of processor are properly matched to each other.

To work together efficiently, those chips have to sit close to each other. Processors wait for data from other processors, and the shorter the connection, the smaller the delay. Shorter connections mean a more compact layout, and that puts a lot of power into a small footprint: a concentrated load.

About ten years ago I was already working with an air-cooled GPU rack of around 30 kW. Standing in front of it felt like looking into a hairdryer. You can put a rack like that into a datacenter designed for 5 kW per rack, but that single rack then consumes the power and cooling capacity of five or six surrounding rack positions as well. You end up renting floor space that has to stay largely empty.

Why air eventually stops working

Air is an inefficient medium for removing large amounts of heat. Take a chip of roughly 10 by 10 centimeters that has to shed one kilowatt of heat. With heat sinks and enough airflow that is still achievable, but the volume of air required quickly becomes impractical. Water absorbs far more heat per liter. At higher power levels, liquid cooling is therefore the obvious route.

In our design, cooling water enters at around 40 °C and leaves at around 50 °C. Our current design calculations assume that roughly 70 percent of the heat is carried away directly by the liquid. The remainder still ends up in the room. If you do nothing about that, the room temperature rises to the point where maintenance is barely safe or workable.

So even a liquid-cooled datacenter still needs air cooling for the residual heat.

The number that actually matters

Now to today's power levels. There are racks that draw around 150 kW, and the industry is working on concepts heading toward one megawatt per rack. That last figure is an extreme amount of power. I honestly cannot yet picture in full how you would make the electrical connection, the cabling and the heat removal work at rack level in a way that is practical and maintainable.

This is where a misunderstanding creeps in easily. An AI datacenter is not filled exclusively with 150 kW GPU racks. Around those racks sit storage, network equipment, CPU capacity, patching and cooling infrastructure. Averaged across a complete scale unit, you therefore land much lower.

For our current design calculations we use roughly 70 kW per rack as the average across a 2.8 MW scale unit.

Do the math in floor tiles

The simplest way to show what this means is to count 600 by 600 millimeter floor tiles.

Take an IT load of 2.8 MW. At 5 kW per rack you need 560 racks. A rack covers about two tiles, and you need roughly one tile at the front and one at the back to reach the equipment. That is four tiles per rack, or 2,240 tiles: close to 800 square meters.

At 70 kW per rack, the same 2.8 MW needs around 40 racks. Applying the same simple rule gives 160 tiles, slightly over 57 square meters.

The electrical power and the amount of heat are identical. The floor area is roughly one fourteenth. That, in essence, is what high density means.

Where the savings are and where they are not

On the powertrain there is little to save in principle. A megawatt stays a megawatt. It has to be transformed, distributed and protected, and depending on the availability requirements it also has to be buffered or backed up by emergency power. The electrical installation of an AI datacenter is therefore not automatically smaller than that of a traditional datacenter with the same total power.

There is a saving in the electrical distribution, because the high-draw AI racks are moving to a different voltage: from 400 V AC to 800 V DC. Higher voltage means lower current, and therefore thinner conductors.

The savings sit mainly in the whitespace, and possibly in part of the cooling plant.

The latter follows from the higher cooling water temperature. A traditional air-cooled datacenter runs on relatively cold water, around 20 °C, and depending on the design uses free cooling, aquifer thermal energy storage, cooling or hybrid towers, and chillers. At a supply temperature of 40 °C, that balance shifts. In the Dutch climate you can run on free cooling for a large part of the year. Because supplementary cooling is only needed during a limited number of warm hours, the central cooling plant can be much smaller. Hybrid cooling towers, and with them the evaporation of drinking water, may be avoidable altogether.

You still need cooling for the residual heat in the room, but that is a much smaller share of total cooling demand.

High density is not the same as AI

This is where the terms often get mixed up.

High density says something about power density: how many kilowatts you want to install per rack or per square meter.

HPC and AI say something about how the computing system and the workload are put together. The parts work on a task jointly and should wait for each other as little as possible.

Those are different axes. Bitcoin mining illustrates it well. A mining installation can have very high power density without being HPC. Its processors largely run the same deep calculation independently and barely need to exchange data with each other. When training an AI model, many processes run in parallel and intermediate results have to be exchanged and brought together continuously. That makes the links between the machines almost as decisive as the chips themselves.

The distinction between HPC and AI is much smaller in structural and building services terms. The difference lies mainly in the mix of CPUs, GPUs, memory and network, and in the workload running on it. As far as I can see, the industry has no generally accepted hard boundary.

What you can say is this: a modern AI datacenter is almost always high density. A high-density datacenter is not automatically an AI datacenter.

Training and inference

Another distinction that often gets skipped is the one between training and inference. On the energy side, that is probably more relevant than the AI label itself.

Inference is generally online work. The model is running, users or agents ask questions and expect an answer within a short time. That calls for high availability and for capacity that can follow user demand.

Training is batch work. A training run can take days, and it can technically be interrupted and resumed later from a checkpoint. How much interruption is economically acceptable depends on the workload, on the contracts and above all on GPU utilization. Being technically interruptible does not automatically mean that a lot of downtime still makes a good business case.

Training and inference can also be combined on the same infrastructure. If inference demand drops overnight, the freed-up capacity can be used for training. Whether that is efficient and commercially attractive in practice depends on how the workloads are scheduled and on the value of the available capacity.

It seems plausible that a large-scale training cluster is on average denser and more network-intensive than many inference environments. Training requires heavy interconnects, such as InfiniBand and NVLink-type networks. Inference can reach very high densities as well, though, certainly with large models and high user volumes. I would therefore not draw a hard density boundary between training and inference.

Why you cannot simply convert an existing datacenter

The logical follow-up question is whether an existing hall along the highway, designed for low-density corporate IT, can be converted for AI.

It can be done. It is usually not simple.

The first problem is the distribution of cooling capacity. Air-cooled datacenters often use water to exchange heat as well. The whitespace holds CRAH units, for example, which are essentially heat exchangers with fans, fed with cold water. The pipes and pumps are sized for the original load. If you want to remove far more heat in the same place, you need a higher water flow rate, and the pipes, valves and pumps often turn out to be too small.

The second problem is temperature. You cannot simply raise the water temperature while existing air-cooled equipment keeps running. At the same time, the datacenter usually has to stay operational throughout the rebuild.

On top of that comes a structural inefficiency. If new equipment can be cooled with water at 40 °C while the existing plant was designed for 20 °C, you keep spending energy to make the water colder than it needs to be.

Waste heat at 50 °C is also far easier to put to use in a district heating network than waste heat at 30 °C.

And then there is floor area. I saw this myself at our academic computing center. A supercomputer of roughly one megawatt stood in a hall designed for 5 kW per rack. Its forty racks averaged around 25 kW, which meant a large part of the original floor was no longer needed as computer floor. We eventually turned it into a boardroom and storage space. That is a workable solution, and it is still expensive datacenter space no longer serving its original purpose.

At the same time, the electrical architecture of new AI datacenters is still in motion. The industry is working on the shift from 400 V AC to 800 V DC and on more flexible conversion systems, including solid-state transformers. Those may eventually make the powertrain more compact and more flexible, but the technology is still new, so it warrants caution.

Why low density was built anyway

That is largely a matter of timing.

AI had of course been around far longer, but it only reached a wide audience with the introduction of ChatGPT at the end of November 2022. Development moved quickly after that. Designing, permitting, financing and building a datacenter takes years, though. Anyone who built a business case in 2021 around 5 kW per rack then went on to build a building around 5 kW per rack. At the time of the investment decision, that was often a defensible starting point.

The awkward part is that the central powertrain of such a building is sometimes still perfectly usable at a higher density. The megawatts are there. It is the distribution from the central plant to the racks, the cooling pipework and the spatial layout that were designed for a different load profile. That last part is exactly the part that is invasive to change.

In closing

So the answer to the original question is that there is a real difference between a traditional datacenter and an AI datacenter, and that no hard boundary runs between them.

A datacenter remains a datacenter. The difference lies in power density, cooling method, technical build-up and the type of workload it was designed for. Traditional, high density, HPC and AI are useful terms, as long as we do not treat them as hard classifications.

The hard numbers sit underneath: power per rack, the supply temperature of the cooling water, the volume of data machines have to exchange with each other, and the degree to which a workload can be interrupted.

One development is clear: the number of square meters per megawatt is falling, at least in the whitespace, and in the grey space the much more compact solid-state transformer may start to make a difference. Datacenters do not necessarily get smaller as a result. They get more compact per unit of power. That makes some new locations usable, and it makes halls designed for 5 kW per rack increasingly hard to use for the newest generations of computing equipment. Refurbishment while the datacenter stays operational is technically complex and expensive. For some older datacenters, new construction will therefore become more attractive than a major rebuild.

At Project Enki we design for high power density and warm-water liquid cooling from the start. That avoids the constraints of an existing low-density design. In return we get a set of other technical and operational questions. I will write about those next time.

Robert Lukkenaer is Head of Datacenters at Project Enki, a Dutch company developing offshore infrastructure for high-density AI and HPC datacenters, located directly at offshore wind farms and using seawater as the cooling source. Project Enki is working on the first European offshore AI infrastructure.

projectenki.com | contact@projectenki.com