In large-scale data centers, electrical architecture has always played a defining role in performance, resilience, and long-term operational efficiency. For decades, alternating current (AC) has been the dominant model, supported by well-established standards and operational practices.

Today, however, the rapid expansion of artificial intelligence workloads is challenging some of these long-standing assumptions. As rack densities are moving beyond 200kW and approach the megawatt range, power delivery is becoming a central design constraint.

No longer is the question solely how to ensure uptime, but also how to deliver power efficiently, at scale, and within increasingly constrained physical and economic environments.

In this context, direct current (DC) is emerging as part of a broader evaluation of data center power architectures.

When power density reshapes design principles

The impact of AI is not incremental – it is structural. High-density environments introduce concentrated and dynamic loads that differ significantly from traditional IT profiles. Electrical infrastructures, once sized for predictable growth, must now support extreme power densities within limited space. In some cases, the footprint of electrical systems can approach that of the IT equipment itself, raising questions around layout efficiency and infrastructure scaling.

At the same time, higher current levels lead to larger conductors, increased material usage, and additional thermal constraints. As one Socomec innovation expert explains: “AI is not just increasing power demand – it is changing how that power is used within the infrastructure. This forces a reassessment of how electrical systems are designed and integrated.”

DC revisited: An evolving option

Direct current is not new to data centers. It has long been used in telecommunications and explored in various IT applications. What is different today, is the context in which it is being reconsidered.

Historically, IT and telecom racks operated at power levels of only a few tens of kilowatts, relying on internal 50V or 240VDC distribution supplied by front-end AC/DC converters. With the rise of AI workloads, rack power density is increasing dramatically – reaching several hundred kilowatts, with some projects targeting up to 1MW per rack.

At 50VDC, this would result in currents of several thousand amperes, leading not only to excessive thermal losses but also to an unsustainable consumption of copper and a significant increase in cabling size, weight, and physical space within the rack.

This paradigm shift therefore requires a move to higher DC voltages, such as 800VDC – leveraging the existing electric vehicle ecosystem – or up to 1,500VDC, which corresponds to the upper limit of the low voltage standard.

The renewed interest in DC is largely driven by IT requirements. High-performance computing systems increasingly operate internally on DC, prompting consideration of whether power could be delivered closer to its end-use format. High-voltage DC distribution – often discussed around 800VDC – is one of the approaches being explored, with potential benefits in efficiency, compactness and scalability.

Its role, however, remains under evaluation. As a Socomec direct current specialist notes: “The question is not whether DC can work – it already does in certain contexts. The real challenge is understanding how it behaves at scale, within complex and highly interconnected data centre systems.”

Multiple configurations are currently being considered, including decentralised conversion, sidecar-based systems, and hybrid AC/DC architectures. At this stage, no single model has emerged as a clear industry standard.

Efficiency versus system complexity

DC is often associated with improved efficiency, particularly through the reduction of conversion stages. While this can be valid, the picture is more nuanced.

Introducing DC into a data center does not necessarily eliminate complexity – and it can redistribute it. Conversion equipment may still be required elsewhere in the system, and protection strategies must be adapted to different electrical behaviours. Operational models must also evolve to accommodate new constraints.

As the Socomec expert explains: “Efficiency is an important driver, but it cannot be considered in isolation. Protection, maintainability, selectivity, uptime and operational robustness are equally critical when evaluating any architectural shift. They need to be considered at system level to allow efficient redundancy.”

A hybrid transition

The transition towards new power architectures is unlikely to be binary. For the foreseeable future, AC and DC are expected to coexist within the same facilities. Hybrid approaches – where AC remains dominant at site level for traditional workloads, and DC is introduced for specific high-density loads – are emerging as a pragmatic response.

This reflects the diversity of modern data center workloads. Not all applications require extreme density, and not all systems are designed to operate on DC. Flexibility, therefore, becomes a key design principle.

Beyond DC: A broader transformation

It is important to view DC as part of a wider evolution. The first step will be a local DC distribution within the IT room, thanks to AC/DC SideCar, before broader expansion of the DC perimeter.

This transition will be shaped by the growing use of medium-voltage distribution (Solid State Transformers – SST), new protection technologies (Semi-Conductor Circuit Breakers – SCCB), and the integration of energy storage solutions and new associated power conversion and source transfer solutions. Together, these trends are expanding the range of architectural options available.

As the expert adds: “The industry is not converging towards a single solution. It is expanding its toolbox. The challenge is to ensure that these technologies can be integrated into coherent and scalable architectures.”

Designing for flexibility

The history of data center design suggests that no single architecture remains dominant indefinitely. Instead, systems evolve in response to changing technologies, economic pressures, and operational feedback. The current transition appears to follow a similar pattern.

Rather than converging rapidly towards a new standard, the industry is entering a phase of experimentation. AC, DC, and hybrid architectures will coexist, each addressing different use cases and constraints. For designers and operators, this implies a shift in mindset: the objective is no longer to select a definitive architecture, but to design infrastructures that remain adaptable over time.

Technical standards for DC power distribution within data centres are currently being defined, notably through industry-driven initiatives such as the Open Compute Project (OCP). This standardisation phase is shaping future architectures and accelerating market structuring.

Hyperscalers are expected to be the first adopters, driven by the need to support the extremely high-power racks required for intensive AI training workloads. These use cases are accelerating the move toward higher DC distribution voltages at rack level, around 800VDC, a direction notably supported by AI chips manufacturer and associated roadmap for native 800V rack architectures.

Several hyperscale technology providers are actively evaluating DC architectures based on voltage levels in the 800V range, with a longer-term perspective extending toward 1,500VDC (or ±750 V), which corresponds to the upper limits of low-voltage DC standards. This evolution reinforces both the market and technological direction and the importance of alignment with emerging industry standards.

In that sense, the question is not simply whether DC represents the future of data centres. It is whether current architectures are sufficiently flexible to accommodate whatever the future may bring.