The race to 800V DC is often framed as a greenfield opportunity.
New AI campuses can be designed around a direct-current backbone from day one: rectification at the facility edge, DC distribution through the white space, and rack-level conversion only where it is needed. The efficiency gains are compelling. Fewer conversion stages. Lower copper intensity. Faster response to volatile AI loads. A power architecture that aligns more naturally with batteries, solar, and modern compute.
But most AI infrastructure will not be built on empty land.
It will be deployed in existing data centers: facilities built around 480V or 415V AC distribution, centralized UPS systems, PDUs, busways, and racks that convert power to DC internally. These sites are already connected to the grid. They are commissioned, staffed, and operational. In many markets, they also have something increasingly difficult to obtain: available power capacity.
The question for most operators is not whether DC is the future. It is how to introduce DC into the AC data center they already have.
A full conversion of a live facility is rarely practical. Switchgear, transformers, UPS systems, and distribution paths were designed, permitted, and commissioned as an integrated architecture. Replacing them wholesale creates downtime risk, strands capital, and triggers a new cycle of engineering approvals, all while tenants and workloads still need to run.
At the same time, the racks arriving in these facilities are changing faster than the buildings around them.
GPU clusters are moving from tens of kilowatts per rack to hundreds. Next-generation platforms are increasingly designed around 800V DC input. The legacy AC distribution path can still serve conventional IT loads, but it becomes progressively more constrained as AI workloads increase density, volatility, and cooling requirements.
This is where hybrid AC/DC architecture becomes important.
Rather than rebuilding the entire facility, operators can preserve the upstream AC plant and introduce DC where it delivers the most value: high-density AI pods, dedicated GPU clusters, or new halls built within an existing campus.
The transition is deliberate:
Keep the utility service, medium-voltage switchgear, transformers, and viable AC infrastructure already in place.
Introduce 800V DC distribution for the highest-density workloads.
Expand the DC footprint as rack densities rise, equipment matures, and each new deployment justifies moving the conversion boundary upstream.
The model is familiar. Liquid cooling entered air-cooled data centers in much the same way. It began with rack-adjacent cooling units that did not require a complete redesign of the building. It then expanded into dedicated liquid loops and, eventually, into facility-wide cooling strategies for AI-scale compute.
DC is following a similar path. The key architectural decision in a hybrid facility is where the AC-to-DC boundary sits.
At first, that boundary may exist near the rack. AC feeds a localized power shelf or sidecar, which converts and distributes 800V DC to a high-density cluster. As DC capacity grows, conversion can move upstream into pod-level power centers, row-level distribution, or dedicated DC zones supplied from centralized rectifiers.
The building remains predominantly AC. The AI load is DC-native.
This approach reduces the barrier to adoption, but it does not reduce the engineering challenge. Hybrid systems introduce design problems that pure AC and pure DC facilities do not face independently.
Protection is the first. AC breakers benefit from current zero crossings that occur naturally during every waveform cycle. DC faults have no such interruption point. A DC arc can persist unless it is actively isolated. At every AC/DC boundary, protection must be coordinated across both systems so that a DC fault is cleared in microseconds without unnecessarily tripping upstream AC infrastructure or cascading across a larger load domain.
Grounding and bonding also require careful design. An 800V unipolar DC bus behaves differently from a ±400V bipolar architecture, and both must coexist safely with the facility’s existing AC grounding system. Isolation strategy, fault detection, touch-voltage protection, and maintenance procedures must be defined at the system level rather than treated as equipment-level details.
Then there is load volatility. Large AI training clusters can move from partial load to peak demand in milliseconds. In a pure DC environment, localized storage, fast converters, and supervisory control can be designed around those dynamics from the beginning. In a hybrid environment, those ramp events pass through the AC/DC interface and into infrastructure that may never have been sized for synchronized GPU behavior.
Energy storage therefore becomes more than a backup system. Capacitors near the rack, batteries at the pod level, and facility-scale storage can act as dynamic buffers, absorbing short-duration transients and protecting both the DC bus and the upstream AC plant. The architecture shifts from static delivery to managed energy flow.
Thermal design matters as well. Early DC sidecars and localized conversion equipment are often air-cooled. A single unit can reject more than 20 kW of heat, effectively placing a dense air-cooled load beside liquid-cooled racks. That heat cannot be treated as incidental. Cooling plans must account for it in airflow modeling, containment strategy, and white-space capacity planning.
People and procedures must evolve with the architecture. Technicians trained on 480V AC will increasingly work alongside 800V DC systems. Lockout and tagout procedures, arc-flash studies, equipment labeling, commissioning processes, and workforce training all need to account for both power domains. The transition to DC is not only a change in equipment. It is a change in operational discipline.
Finally, there is visibility. AC and DC equipment often come from different vendors, expose different telemetry, and report into different operational tools. Without a unified view, the interface between them becomes a blind spot: precisely where operators need the most insight into power quality, transient behavior, fault response, and available capacity.
This is why hybrid architecture cannot be approached as a collection of converters and busways. It requires a coordinated control layer that can observe both sides of the boundary, model system behavior, and manage power as a dynamic resource.
The opportunity is significant. Hybrid AC/DC allows operators to deploy higher-density AI infrastructure without betting an entire facility on a single cutover. It preserves the value of existing AC infrastructure while creating a practical route to 800V DC where the economics are strongest.
The AC plant continues to do what it does well. DC goes where density demands it. And the boundary between them moves upstream only when engineering, operational readiness, and economics support the next phase.
The future is DC. For most operating data centers, the path there runs through AC.
Contact us to learn how Ennovria can help you model, validate and phase your move from AC to hybrid AC/DC.
