The race to power the AI data centre
The rapid growth of artificial intelligence is driving a surge in data centre energy demand, pushing grid capacity and traditional electrical infrastructure to their limits. Here, Meir Adest, Co-Founder and VP Core Technologies at SolarEdge, explains why a shift from AC infrastructure to DC-native energy architecture is now critical for meeting rising compute demands and improving overall data centre efficiency and profitability.
Artificial intelligence (AI) is reshaping the data centre landscape at extraordinary speed. Hyperscale operators are racing to deploy infrastructure capable of supporting increasingly power-hungry workloads, with more than 100GW of new AI data centre capacity expected globally between 2026 and 2030. At the heart of this shift is a fundamental change in compute density.
Traditional server racks typically operate at around 10kW. Today, AI racks are approaching 1MW, representing a two orders of magnitude increase in power demand within the same physical footprint. This escalation is exposing a critical constraint that has, until now, remained largely invisible: the limitations of the underlying energy architecture in sustaining future demand.
The limitations of AC infrastructure
Despite the digital nature of modern computing, the energy systems powering data centres remain rooted in a model developed more than a century ago. Alternating current (AC) became the standard because it enabled efficient long distance transmission at a time when direct current (DC) could not. However, every core component inside a data centre, from GPUs to batteries, fundamentally operates on DC power. As a result, today’s infrastructure relies on a continuous cycle of conversions: AC from the grid is stepped down, converted to DC for storage, inverted back to AC for distribution and finally converted again to DC at the chip level.
This process introduces a ‘conversion tax’ – between 10% and 30% of incoming energy can be lost before it ever reaches the compute load. Crucially, these losses create a double penalty – this valuable energy is lost as heat, meaning this heat must then be actively removed with cooling systems, requiring additional energy consumption. In an era where power availability is becoming a limiting factor for data centre expansion, every percentage point of efficiency matters. Each incremental loss represents not only reduced compute capacity from a fixed grid connection, but also substantial financial impact at scale. Across gigawatt-class facilities, even a single percentage point of inefficiency can translate into millions in lost value, reinforcing the need to maximise the proportion of energy that reaches the processors themselves.
Why AI Is forcing a move to DC
Historically, these inefficiencies were tolerable. Rack densities were low enough that energy losses could be absorbed without fundamentally constraining performance, but AI changes this. With racks approaching megawatt scale, the physical and thermal limits of AC infrastructure are being reached. NVIDIA and other technology leaders are now advocating for 800V DC architectures as the only viable path to support megawatt-scale racks.
The rationale is straightforward. Higher voltage DC systems enable lower current for the same power delivery, reducing resistive losses and allowing significantly more power to be transmitted through existing conductors. In practical terms, 800V DC can deliver substantially more power while reducing copper usage, installation complexity and cost. However, implementing this requires a fundamental redesign of how energy is delivered within the data centre.
A staged evolution
The transition to DC-native architecture is not a single step, but a staged evolution. Today, most data centres operate in what can be described as Stage 0, where power is distributed entirely in AC, with DC conversion only occurring at the point of use. Stage 1 introduces targeted improvements through retrofits such as sidecar solutions, enabling the introduction of 800V DC racks, and resulting in 85-93% system efficiency. While valuable, these approaches are widely viewed as interim measures. A more significant shift occurs with Stage 2, where elements of the infrastructure, mainly the UPS systems, begin to operate on DC.
This hybrid approach removes unnecessary conversions, so system efficiency can increase to 91-96%. The real shift begins at Stage 3, where data centres become fundamentally DC-native. Here, traditional transformers are replaced by solid-state transformers that convert medium voltage AC directly to DC. This simplifies the power chain and unlocks efficiency of 94-97%. Stage 4 builds on this with fully integrated DC architecture, combining advanced ultra-efficient solid-state transformer connected directly to 34.5V grids, DC based UPS systems and intelligent energy management. At this stage, total system efficiencies approach 98%, dramatically increasing the proportion of incoming power that can be used for compute.
While some operators will progress incrementally, leapfrogging intermediate stages entirely is ultimately more cost effective and optimises operations. For data centres planned today, designing for 800V DC rather than preparing to retrofit this architecture will likely prove the more effective approach.
Integrating conversion and control
At the core of this transition is the solid-state transformer (SST). Unlike conventional transformers, which are typically large, site-specific and limited in functionality, solid-state transformers will integrate conversion and control into a single, modular system. They are designed to connect directly to medium voltage grid infrastructure and deliver high voltage DC in a single step, eliminating multiple layers of conversion.
Advancements in semiconductor materials, particularly silicon carbide, are enabling SSTs to operate at higher voltages with lower switching losses, improving efficiency and performance. Beyond efficiency gains, SSTs can offer additional advantages. Their modular design allows for faster manufacturing and deployment compared to traditional bespoke transformers, reducing lead times and enabling quicker time to operation. Built-in redundancy at the module level also enhances system resilience while ensuring no single point of failure.
However, technical challenges remain. Although 800V DC architecture already exists in utility scale solar and electric vehicles, implementing this for SSTs is no easy feat. Directly interfacing with grid-level voltages, such as Europe’s 33kV grid or the 34.5KV grid in the US, without intermediate conversion stages is complex and requires much expertise in DC energy architecture.
Implications for the grid
The benefits of DC architecture extend beyond the data centre itself. As power demand from AI infrastructure grows, concerns around grid capacity and stability are intensifying. In many regions, access to sufficient power is already limiting data centre expansion. By improving energy efficiency, DC-native systems enable operators to extract more compute from the same grid connection, reducing the need for additional capacity. Lower losses also translate into reduced cooling requirements, further easing overall energy demand.
In addition, integrating DC based energy storage and intelligent control systems should allow data centres to interact with the grid more dynamically. DC UPS systems, for example, can act as a buffer, smoothing demand fluctuations and mitigating the impact of sudden load changes.
Overcoming barriers to adoption
Despite the clear benefits, the transition to DC is not without challenges. First, there is the question of infrastructure change. Moving beyond early stage retrofits requires replacing key components within the power chain, which can be complex. Second, regulatory frameworks for DC systems are not yet fully mature. Standards for safety, certification and operation are still being developed, particularly for higher voltage levels. Finally, there is a skills gap – designing and deploying DC-native systems requires specialist expertise that is not yet widespread across the industry. However, these challenges are typical of any major technological shift and with the groundwork already being laid, solutions will soon appear.
The transition timeline
This transition is already taking shape. 2026 is expected to see a few prototyping and early pilot deployments of DC architectures and SSTs. By 2027, SSTs are expected to be commercially available and initial commercial projects should emerge. From 2028 onwards, as megawatt-scale AI racks become mainstream, DC infrastructure will need to scale rapidly to meet demand. Looking further ahead, the industry is already considering a future transition to 1,500V DC systems. While this offers additional efficiency gains, it introduces new technical and safety challenges that will need to be addressed through further innovation and standardisation.
The inevitability of evolution
As AI continues to push the boundaries of power density, the limitations of traditional infrastructure are being reached. DC architecture offers a path to higher efficiency, greater scalability and improved alignment with the needs of next-generation workloads. The question is no longer whether this transition will happen, but how quickly the industry can execute it.
Data centres are the first to face this challenge at scale, but they are unlikely to be the last. As the technology matures, the shift towards DC-native energy systems may prove to be far more widespread.
