How Reliable Data Centers are Built: A Practical Engineer’s Perspective

Author: Sergey Voronin, Head of Engineering Systems at “GIGANT Computer Systems”

Today, discussions about data centers almost always start with the “hardware”: which UPSs to choose, how many air conditioners are needed, which chiller is better. But if this is the starting point for a project, it is beginning from the wrong place. A data center is not just a set of racks and engineering systems; it is a tool to provide specific business services and manage the level of risks a company is willing to tolerate. The logic is simple. First, the business has a goal: launch a new service, ensure continuity of critical systems, meet regulatory requirements, or move infrastructure “on-premises” from a leased data center. Based on this, the IT landscape is formed—servers, storage systems, software platforms. Only then is this IT infrastructure “landed” onto the engineering systems.

What actually runs on the servers—an internal corporate system, a high-load application, or even mining—is secondary for engineers. From an infrastructure perspective, there are only a few tasks: properly supply and reserve power, ensure cooling and permissible environmental parameters, set up basic security, and provide predictable site operation.

That’s why a competent data center project always begins not with equipment model comparisons, but with analyzing business requirements and the customer's real needs: which services cannot be stopped, acceptable downtime, risks the company is willing to take, and the load density it can support in the future. Only after this is it possible to build out the engineering architecture. If this stage is skipped or done formally, subsequent mistakes are almost guaranteed. The most common is a misjudgment of one’s actual needs: excessive power requirements, unnecessary racks for “future growth,” incorrect calculations of heat and consumption. All this seems trivial until it becomes the foundation of an engineering scheme that cannot be changed without fighting the already constructed reality.

Startup Mistakes That Become Costly Later

Building “for reserve” is one of the most widespread reasons for unnecessary expenses in data center creation. Almost every second project starts with the client oversizing their future infrastructure. The scenario is typical: today, the company has two or three racks, but it assumes in a couple of years there will be a dozen, and the entire technical layout is immediately designed for this hypothetical load.

The problem is that engineering systems are always built for the full design capacity. If the real load is half that, the overpayment becomes tangible. Moreover, an “empty” data center operates unstably: cooling systems sized for a large thermal load cycle on and off at low loads and enter abnormal operating modes. Therefore, it is crucial at the start to honestly define growth horizons and design the infrastructure so it can be expanded in phases, without overpaying for capacity that may never be needed.

Another error leading to budget inflation is incorrect power consumption calculation. Clients mechanically add up the specification kilowatts of servers, storage, and auxiliary equipment, getting a figure that has nothing to do with real operation. Most servers are equipped with multiple power supplies, but actually use only some of them—the rest operate as reserves. As a result, a power supply system is designed for a load that the equipment will never actually draw, and the data center’s cost increases for no technical reason.

Even the best technical specification can be easily undermined by the wrong choice of premises. On paper it may look perfect, but during the survey it turns out the building cannot physically withstand the weight of the intended infrastructure. Office and administrative buildings often have floor loads rated at 500–600 kg per square meter, whereas racks and engineering nodes typically require about twice as much. In such conditions, the premises physically cannot host a data center.

Even if load-bearing capacity is sufficient, another limiting factor may arise—the absence of adequate electrical power. The area might be ideal, but the building lacks available power. Securing additional technical conditions and bringing new kilowatts to an existing building is often too expensive or even technically impossible.

Hidden engineering communications also play a role. Main heating, water supply, or sewage pipes may run through the premises—systems that have no place in a data center. If it is impossible to reroute such utilities, the premises must be ruled out.

Sometimes the problem is not the premises themselves, but what is above them. If “wet” areas such as kitchens, restrooms, or technical rooms are located above the intended server hall, the risk of leaks becomes critical. Any leak from above is a direct path to outages and downtime, so such premises are ruled out at the early analysis stage.

To avoid these mistakes, a project needs an integrator from the start—not just as an executor, but as a strategic engineering consultant. The integrator helps correctly assess future load, consider site limitations, and temper inflated expectations before they turn into unnecessary costs. Audits, consultations, and discussions about development options at the beginning make the project predictable and help avoid costly reworks.

For example, an integrator may suggest that instead of building a data center “for future growth,” it’s better to determine the real expansion horizon and break it into convenient modules. If 100 kW is needed, there is no reason to install two large air conditioning units “for the whole volume”—they will not work optimally at low load. It’s more reasonable to split the system into 25 kW modules and connect them as the load increases. The same applies to power: modular UPSs allow capacity increases incrementally by adding power modules into the same platform. This avoids overpaying for excess capacity and enables gradual scaling of the data center.

Data Center Power Supply: Design Principles, Redundancy Levels, and Common Mistakes

From an engineering infrastructure standpoint, there are two systems without which a data center simply isn't a data center: power supply and cooling. A site can theoretically operate for a while without video surveillance or even a full fire safety system, but never without power and cold. Therefore, power supply inevitably becomes the most expensive, complex, and critical part of the project. The overall engineering layout depends on how well it’s designed. The starting point is always the same: choosing a reliability level.

Engineering practice is based on Tier I–IV standards, and Tier III is most commonly chosen as the optimal balance of cost and reliability. This level assumes redundancy of all key components by the n+1 scheme, ensuring system operation even if one element fails. Tier IV requires two fully independent power supply circuits, completely duplicating each other, so it is rare and mostly found on commercial sites where downtime is critical for revenue. In some regions, such facilities do exist: annual scheduled shutdowns there are literally only a few hours.

But even an ideal redundancy scheme does not help if mistakes are made at the design or operation stage. A typical story: the client neglects scheduled maintenance—UPSs, switchgear, air conditioners run “until something breaks.” This results in accidents and downtimes that could have been avoided. Layout mistakes are common too. If equipment is installed without regard for service zones, it becomes physically impossible to access: changing a filter in an air conditioner, servicing a UPS battery block, or dismantling a unit. It may seem minor, but such details lead to forced outages or expensive redesigns. More telling is the case of a data center inspection where the cooling system, which worked about ten years, failed much earlier than expected. The reason: the design simply did not provide for oil traps in the lines, leading to accelerated equipment wear. This is a classic example of how a single design error lays the groundwork for future operational problems.

Cooling as the Foundation for Data Center Stability: Which Technologies Work Best

The cooling system is, along with power supply, the data center’s second critical pillar. Without it, the equipment will last only a few minutes before the server hall temperature starts rising rapidly. It is cooling that ultimately determines whether the site operates stably or turns into a “sauna” during any emergency. In engineering practice, there are two classic approaches depending on the scale of the future data center.

For small server rooms and compact data centers, direct expansion precision systems (DX) are commonly used. They reliably maintain conditions in a narrow range and are well-suited for capacities where compactness and contour simplicity are critical. The key limitation is the refrigerant line length: outdoor units must be placed directly adjacent to the server hall.

When capacity grows—starting approximately from 200–300 kW and above—it is more rational to switch to a water-based Chiller–Fan Coil scheme. It provides lower power consumption, enables phased capacity buildup, and delivers the greatest benefit when free cooling is used. In cool periods, the chiller runs on outside air almost without compressors, significantly reducing operating costs and improving the PUE of large sites.

In a cold climate, free cooling becomes a key element in reducing power consumption. Two approaches are used: direct, where outside air is filtered and supplied directly into the server hall, and indirect, where air passes only through a heat exchanger, never contacting the internal circuit. The direct version is simpler in architecture but demands more from filtration and operations control. The indirect method offers higher equipment safety while preserving the energy savings of utilizing low outside temperatures.

What Determines the Reliability of a Future Data Center

Climatic conditions are among the first factors considered when designing a data center’s cooling and power systems. Engineers always start by analyzing the minimum and maximum temperatures in the region, as equipment performance depends directly on these extremes—whether under the highest load or, conversely, in the harshest conditions of the year. The air conditioning system operates differently at +25°C and +35°C, and its capacity is specified to handle the thermal load in the most stressful conditions, not the “average.”

At the same time, the common belief that “building data centers where it's cold is cheaper” is only partly true. Yes, a mild and cool climate allows for more frequent free cooling and reduces the average annual PUE. That’s why such solutions are especially effective in regions with colder climates. However, at low temperatures (minus thirty and below), other challenges appear. Equipment needs not just to cool the server hall, but also to remain operational year-round under such conditions. Low-temperature kits are used for this, and not all manufacturers can guarantee correct system operation at temperatures as low as minus sixty. Nevertheless, such projects are feasible and implemented where this is a genuine operational necessity.

As for peak loads, the approach is pragmatic. Engineering systems are designed for the full, one-hundred percent load of the equipment—without “discounts” for averaging or forecasted demand factors. Yes, both the power and cooling systems have temporary overload capabilities and can operate briefly above nominal, but relying on this as normal is unacceptable. In exceptional cases, design may use lower factors when the client is certain equipment will never run at 100%—it will stay within 70–80% of nominal, for example. But this is the exception, not the rule: reliable data centers are always designed for full thermal and electrical load profiles to remain stable during any consumption peak or external condition swings. This brings up an important topic—the choice of the data center format. Once it’s clear how the site will handle maximum loads and the harshest environmental conditions, it’s clear what type of facility is worth building. For some projects, building traditionally makes sense; for others, time-to-launch and site constraints are critical, making modular and container data centers particularly sought after.

Modular and Container-Based Formats: Practical Necessity, Not a Compromise

Modular and container data centers are consistently in demand—not because they’re trendy or a “new format.” They are chosen when conventional construction is impossible, unprofitable in terms of timeline, or unjustified by scale. The main factor here is speed: such solutions allow you to deploy a working data center many times faster than a traditional building, and for many projects, time to launch defines success. When business needs new capacity in 3–4 months, not 18, there is simply no alternative.

The second reason is predictability. The modular approach eliminates most construction risks: engineering systems are assembled and tested at the factory, and onsite, it’s just installation and commissioning. The client gets a facility with a predetermined cost, timeline, and parameters, without surprises during construction or loss of quality.

Another argument is the ability to scale without unnecessary capital tie-up. In the modular format, there’s no need to build infrastructure for future growth and hope it’s eventually filled. You can bring the first capacity online, monitor real usage dynamics, and later add more blocks as needed. This is critical because many companies overestimate future growth, and excess space and capacity then sit idle for years. Modular architecture mitigates this risk.

Container solutions are in demand for specific and local needs. When a facility is temporary, remote, industrial, or distributed across multiple sites—rapid deployment, operation, and removal are key advantages. There have been situations where a container data center was optimal even within city limits because the building had no area that met the basic requirements for loads and power supply. In such cases, a container solution is the only technically feasible alternative, and no building renovations will solve the problem.

Modular complexes are also popular in other scenarios—especially where there is land available, but traditional construction would lengthen timelines or lead to excessive investment. This approach allows deploying a full-fledged data center without going through the entire capital construction cycle. That’s why such solutions are particularly popular for companies needing to quickly deploy infrastructure on their own sites without extended permitting processes, while retaining the flexibility to adjust construction volume to actual needs.

The demand is not created only by remote regions. In cities, the problem is the same: a shortage of suitable sites and electrical capacity. Most inquiries are not due to a desire to “try modular format” but because there are no real alternatives—buildings are old, have weak floor loads, capacities are taken, and there’s simply nowhere to build a traditional data center.

Looking at the market overall, the picture is logical. Large projects still stick with traditional construction—for megawatt loads and thousands of square meters, it’s the only rational way. But projects from several racks up to a few dozen—up to around a hundred—are increasingly executed modularly. It’s not because this is necessarily cheaper, but because it’s faster, more predictable, and more accurately matches real task volumes, without redesigning buildings and unnecessary costs. That’s why modular and container solutions maintain a steady niche: they fulfill tasks that traditional data centers cannot—whether due to time, site, or economics.

Key Technological Directions Shaping the Data Center Market

Looking at the coming years, the drivers of engineering infrastructure development for data centers are quite clear. Data volume growth remains a strong trend, directly influencing the need for new computational capacity. Localization also continues to play a significant role: organizations need reliable facilities within their jurisdiction, and this requirement will only intensify.

But, probably, the most significant factor will be everything related to the development and training of artificial intelligence systems. Global experience shows that AI projects require massive energy and computational resources, and the infrastructure for such loads must be designed at a different level of density and resiliency. As soon as such initiatives start to scale, they will inevitably impact the data center market.

In other words, the need for new sites and high-quality engineering infrastructure will not decrease—on the contrary, it will grow. This means only one thing: there is plenty of work ahead, and the challenges will become ever more complex and interesting.