Thinking inside the box: Why modular data centers are no longer optional for AI deployments
Four pressures are pushing data center construction into the factory. Two are about timescales that won’t sit still, two are about what a “traditional” build simply can’t absorb.
A data center used to be built the way a building is: shell and core first, systems inside decided later...sometimes long after ground had broken. That worked when a facility’s return didn’t depend on which week it opened. It stopped working once speed itself became the thing driving revenue.
As Corenix CEO Martin Renkis puts it, 10 megawatts of data center capacity can generate close to a 100 million dollars a year gross revenue once it’s live. In this market, weeks matter. A modular build finished exactly on schedule is still faster than a traditional build finished exactly on schedule. An investor who has already committed capital to GPUs wants the quickest route to a return on that investment, and they’re going to choose modular every time because they can start to generate revenue faster. That’s not even factoring in delays, which in the construction world often compound once multiple suppliers and contractors are on site in the final stages of a build. The point where the finishing line is in sight is the point where operators are most sensitive to delays. Therefore, the benefits of modular therefore don’t start and end with speed; it’s also about risk reduction at the critical point in a build where IT, power, and cooling come together on site.
Time is certainly money, but Renkis is keen to push back on the industry narrative around obsolescence. A chip doesn’t become worthless the moment a newer one is released. Older hardware holds its value and stays in productive use far longer than that narrative suggests. But being on the latest GPU platform makes real financial sense in terms of price per token, and a build finished before the next generation of hardware is released means the chip specified at the start of the project is still the current one when the site goes live.
A timeline that won’t sit still
Speed also has to run the other way: the ability to adapt as technology and power availability change, without waiting on a fresh construction cycle. New chip generations, new cooling requirements, new power coming online in stages…all of it moves faster than a conventional build can absorb, and a facility that takes years to deliver locks in decisions made at the start of the build, not the ones that make sense by the time it opens.
That gap between how fast the market moves and how fast a conventional build can respond is what breaks the traditional “building first” model.
Demand behaves the same way. The requirements of accelerated computing, power, cooling, serviceability, and the ability to support successive generations of high density compute, have grown faster than a traditional construction model can keep up with. What the market needs is a faster, more predictable path from an approved technical architecture to operational capacity, wherever that capacity needs to sit.
That demand doesn’t arrive on a predictable schedule either. It shows up in bursts, such as a pilot project that suddenly needs to scale, or a customer commitment that lands faster than planned. Committing to a full, fixed shell before that demand is proven means either overbuilding capacity that sits unused or underbuilding and running into a ceiling within months. NVIDIA’s own read on the market points the same way: in August 2026, its CEO named land, power, and shell, alongside chips, packaging, memory, and networking, as the critical resources an AI factory now has to secure. A repeatable delivery model, where capacity gets added in step with what’s needed, answers that shift.
What a traditional build can’t absorb
Two further pressures have nothing to do with how fast the target moves, and everything to do with what a conventional concrete build simply cannot absorb.
Labor is one of them. Electricians and construction personnel are in short supply, and a conventional build is capped by how many qualified trades can be on one site at once. Electricians account for roughly 30 to 40% of the total construction hours on a data center project, and industry analysis forecasts a shortage of that trade emerging from 2027, concentrated in the US where the buildout is heaviest. Standardized work on a production line depends on factory capacity instead, which can be planned, staffed, and repeated in a way a single construction site never can.
Sustainability is the other. A conventional shell uses much more concrete than a modular one, and concrete carries a real cost to the environment. That carbon is locked in before a single server is powered on, and there’s no later point at which it gets corrected in the way operating efficiency can be fine tuned over time. Utilizing less concrete, by moving structural components into factory-controlled production instead, is the only point in the process where this gets addressed.
Factory tested, site proven
Everything so far explains why factory building exists. It doesn’t explain why it can be trusted.
On a conventional build, power, cooling, IT, and network systems are typically brought together sequentially on site, one contractor’s work waiting on the last, with the whole system tested for the first time only once everyone is finished. A factory-built platform inverts that. Those systems are built and proven in parallel, each tested on its own before they ever meet, so the platform that reaches site arrives already proven as one working system, not separate parts hoping to fit together once they’re on site. The result is something closer to plug and play, and a process that’s more predictable from start to finish.
Independent modeling of a fully modular build puts the construction window at around 36% shorter than a conventional equivalent, a saving of several months.
That two-stage pattern is standard across modular delivery. A platform is designed, built, integrated and factory tested as a complete system before it ever leaves the production facility. Installation, connection to site infrastructure, site acceptance testing, commissioning and full operational readiness then happen at the client’s own location. Both stages have to work, and work together, for a modular build to deliver the time it saved in the factory.
Why Corenix
Corenix was built to answer all of this at once as a single delivery model. It designs, builds, integrates, and factory tests complete data center platforms, engineered to NVIDIA reference designs and customized to each client’s specific requirements, then carries that same accountability through installation, connection, site acceptance testing, commissioning and operational readiness at the client’s location. One team stays responsible for the full sequence, from an approved technical architecture through to operational capacity in the place it is needed.
That approach draws on Submer Group’s background in high density thermal engineering and liquid cooling, built up over more than a decade of solving exactly the problems described above: speed, unpredictable demand, labor, and the environmental cost of getting a build wrong. Corenix is where that experience becomes a dedicated way of delivering AI infrastructure at the scale the market is now asking for.
The articles that follow will go further into what that looks like in practice: how the platform is engineered, what it delivers in the field and the people building it.
.svg)
