In sum, what you need to know:
- Shorter deployment cycles: Data center timelines have gone from an average of 5 years to 18-24 months.
- Rapid refresh cycles: Rapid Nvidia GPU releases are significant enough that it’s not profitable to stick with “old” hardware.
- Prefabricated components: Prefabricated, integrated blocks streamline the design phase and ensure quick coupling.
Industry research suggests scaling global data center infrastructure may require well over $7 trillion in investment by 2030, with AI-related demand driving what McKinsey & Company says could be 156 GW of capacity in that timeframe. That incredible explosion of demand is triggering a much higher interest in prefabricated modular data centers (PMDCs).
Just yesterday, Schneider Electric hosted a panel session, “Is the future of data centres modular” at the 2026 Datacloud Global Congress in Cannes. Leading up to that discussion and during the event, experts such as Arcadis’ Christian Goldsmith (data center Global Solutions Lead) and Hitachi’s Kasper Berggren (Business Development Manager for Data Centers in Europe), joined speakers from Refine, RITTAL and FTI Consulting to talk about what’s driving the rapid growth of the modular data center market. No longer niche, modular data centers are a strategic infrastructure model, and the market for them may surge from $42.24 billion in 2026 to over $167 billion by 2034.
With deployment cycles shrinking from an average of 5 years to just 18 – 24 months, and data center operators increasingly building multiple facilities, concurrently, it’s becoming very difficult to keep up with AI infrastructure refresh cycles. For example, Nvidia releases new GPU architectures every 18 to 24 months, each generation offering 50% greater performance or 30% efficiency savings.
For these reasons, modular design is gaining attention as data center owners and operators look for ways to “drive speed, certainty, and scalability,” as explained by Arcadis Global’s Goldsmith, who spoke in a pre-session video at Cannes about balancing standardization and flexibility in data center design. While standardization can greatly reduce design and assembly timelines, it does come at a cost in terms of rigidity of component dimensions and fixed configurations that limit how much a facility can be tailored. Despite the risk of less flexibility, many data center operators prefer modular approaches that streamline the design phase in the race to meet capacity demands. These approaches usually include some or all of the below top-10 components:
- Electrical power blocks –prefabricated systems integrating medium-voltage switchgear, step-down transformers, and uninterruptible power supplies (UPS) on steel skids or in containers;
- Power Distribution Units (PDUs) –input/output voltages, breaker configurations, and busway connections eliminate custom electrical engineering;
- Uninterruptible Power Supplies (UPS) –battery packs and hot-swappable power modules;
- Mechanical cooling skids— chillers, heat exchangers, pumping packages, valves, manifolds, expansion tanks, and controls-and-implementation. For high-density situations, Coolant Distribution Units (CDUs) for liquid-to-liquid heat exchange in direct-to-chip or immersion cooling;
- Cooling Infrastructure: Direct-to-chip liquid cooling manifolds, rear-door heat exchangers, and CRAH units use uniform pipe connections and flow rates;
- Server Racks and Frames: Standard 19-inch or Open Compute Project (OCP) 21-inch racks feature uniform rail spacing, cable management, and mounting points;
- Airflow Containment: Hot or cold aisle containment structures that physically separate exhaust from intake air to maximize cooling efficiency.
- Cable Management Systems: Fiber trunk cables, color-coded patch panels, and overhead raceways for uniform network topology;
- Environmental Controls – fire suppression nozzles, gas detectors, and automated door locks follow uniform placement templates;
- Management Software (DCIM) – unified monitoring through sensors that use standardized communication protocols like SNMP or Modbus to track power and cooling.
Of all the components, the thermal management block is garnering the most attention whether talking about legacy enterprise systems or next-gen AI clusters. In fact, the average for rack power density has surged to 27 kW per rack, according to the 10th anniversary edition of the AFCOM State of the Data Center report. Consider that a fully configured NVIDIA DGX H100 system has a rated maximum power draw of 10.2 kW, with some system configurations peaking at 11.3. As another example, the NVIDIA Blackwell B200 platform (server node) can draw as much as 120 kW to 135 kW per single rack (for full rack scale).
This incredible YoY change is rapidly changing thermal profiles, especially as systems are stacked. In modular designs, there is an infrastructure “wall,” in which standardized air-cooling infrastructure no longer works. At around 49 kW, ib modular designs, there’s a need to pivot to options like direct-to-chip liquid cooling or immersion cooling.
This is a topic RCRTech will be covering in upcoming RCRTV AI TechTalk episodes later in June. Stay tuned!
