800VDC and Power Architecture Transitions: Designing for a Future That Arrives Unevenly

Is DC Voltage for Real this time?

800VDC is real, and it’s almost here—but do you need it in your facility? Industry projections estimate 800VDC adoption will reach 15–25% of data centers by the end of the decade, concentrated almost exclusively in megawatt-scale AI training clusters. Rather than requiring ground-up facility overhauls, initial deployments are taking shape as "side-car" architectures—bolted directly onto conventional AC infrastructure to power high-density pods.

This design consideration is one of the most important in data center power today. The question shoudn't be "When do we switch to DC?" but rather "How do we serve AC loads profitably today while preserving a low-cost path to DC for the pods that need it?" Get that balance wrong and you strand capital—overbuild full-DC too early and you pay for immature gear serving non-existent loads; ignore DC entirely and your facility gets disqualified from next-gen AI deployments.

Why 800VDC Exists at All

The physics case is straightforward, and we cover the full voltage-architecture picture in our medium voltage design guide. The short version: power scales with voltage times current, but copper, connector count, and resistive losses scale with current. At 120 kW per rack the conventional 415/480V AC chain strains; at 500 kW it fails on copper mass alone; at 1 MW a single rack's LV busbar would need roughly 200 kg of copper — and NVIDIA calculates that a gigawatt-scale facility built on today's architecture would need up to 200,000 kg of copper in rack busbars alone. Moving distribution to 800V DC transmits 85% more power through the same conductor size, eliminates AC skin-effect and reactive power losses, and recovers up to 5% end-to-end efficiency at exactly the moment every megawatt of loss has to be re-removed by the cooling plant.

And the strain isn't only in the distribution path — it's inside the rack. Today's AI racks convert AC to 54V DC in power shelves that ride in the rack itself: a GB200 NVL72 already needs up to eight of them, and NVIDIA projects that at megawatt scale, 54V power shelves would consume up to 64U of rack space — leaving no room for compute. The alternative, dedicated power-only racks beside every compute rack, just duplicates infrastructure. Higher-voltage DC to the rack is how the in-rack real estate problem gets solved.

Inside NVIDIA's 800V HVDC Reference Architecture

NVIDIA's published 800V HVDC architecture formalized the target, and its details matter for anyone planning a facility. Power converts once, at the facility perimeter: 13.8 kV utility AC is rectified directly to 800V DC in the power room, and everything downstream stays DC. Overhead 800V busways replace today's 415V AC busways in the white space, racks accept simple two-conductor 800V feeds, and the remaining DC/DC conversion (800V down to 54V/12V and core voltage) happens inside the rack, close to the GPUs.

The claimed benefits, from NVIDIA's own analysis:

Two details in NVIDIA's announcement deserve special attention from facility planners. First, the timeline: full-scale production begins in 2027, timed to the Kyber rack generation — consistent with the adoption curve below. Second, the form of the first real deployment: at GTC 2025, NVIDIA demonstrated an 800V sidecar powering 576 Rubin Ultra GPUs in a single Kyber rack. The flagship vendor's own proof-of-concept is a side-car, not a facility redesign — which is exactly the migration path this guide recommends. The ecosystem behind it spans silicon , power components, and systems integrators (Like Legrand).

NVIDIA is candid that this is "a holistic redesign," not a component swap — with open work remaining on safety standards for facility-level DC, workforce training, fault detection and serviceability, the transformer-vs-SST conversion question, and energy storage to buffer the subsecond power swings GPU training loads produce. Those open items are the substance of the pros-and-cons that follow. Its not a question of if, only when and how.

What a Side-Car Design Actually Looks Like

The phrase "side-car" describes the architecture that will carry most real-world 800VDC deployments through this decade: a dedicated DC conversion plant added alongside the facility's AC infrastructure, sharing the same medium-voltage backbone, serving only the AI pods that need it. The rest of the facility keeps running proven AC distribution.

The appeal is obvious once it's drawn: the side-car isolates all of the new-technology risk — converter maturity, DC protection, workforce unfamiliarity — inside a bounded scope, while the facility's existing AC business keeps running on equipment and practices that are fully understood. It also phases capital: DC conversion capacity gets added pod by pod, tracking actual AI load commitments instead of a speculative full-facility bet.

The Case For and Against 800VDC

800VDC represents a major design shift—and like any early-adopter transition, weighing the move requires balancing real operational trade-offs against future-proofing. Here is the engineering reality on both sides:

Who Adopts, Who Waits — and Why the Curve Flattens

The segmentation is rational, not sluggish. A colocation operator running 20–60 kW racks on paid-for AC infrastructure gains nothing from a DC conversion that would interrupt revenue, retrain staff, and adopt immature protection technology. Meanwhile a hyperscaler standing up a 300 MW training campus with 600 kW racks has no real alternative — the copper math alone forces the decision. Both are making the right call. The design challenge is for everyone in between: facilities that serve AC loads today and want to win megawatt-class AI tenants tomorrow.

New Build vs. Retrofit: The Actual Decision Framework

If You're Designing a New Build

• Provision, don't commit. Reserve side-car space, structural capacity, and conduit/bus pathways adjacent to AI-capable halls. Concrete and clearances are cheap at design time and brutally expensive later.
• Size the MV backbone for both futures. The shared MV switchgear in Fig. 1 is the one element both architectures need — spec it (and order it early, given 2–3 year transformer lead times) with capacity for a future DC plant.
• Make the white-space distribution layer transition-ready. This is the busway decision, covered below — the overhead infrastructure you install for AC should be the same physical architecture a DC pod will want.
• Keep protection studies modular. Document today's AC coordination as its own bounded study so a future DC side-car gets its own analysis instead of unwinding assumptions.
• Plan for energy storage in the DC pod. GPU training loads swing on subsecond timescales, and NVIDIA's architecture leans on energy storage to buffer those spikes — reserve space and structural capacity for battery systems alongside the conversion plant, not as an afterthought.

If You're Evaluating a Retrofit

• Side-car only. Full-facility DC conversions of operating AC plants almost never pencil — the disruption to revenue-bearing load outweighs the efficiency gains for the remaining life of the equipment.
• Anchor the side-car to a committed tenant or workload. DC conversion capacity should follow signed AI load, not speculation.
• Audit the MV headroom first. The most common retrofit blocker isn't floor space — it's an MV backbone with no spare capacity and a multi-year queue for more.
• Treat the DC pod as a greenfield inside your building: its own protection philosophy, its own commissioning plan, its own operating procedures and training.

Architecture Comparison at a Glance

The Future Proofed Solution: Track Busway

Here is the reassuring news amidst the uncertainty: one key piece of your infrastructure design already has a clear answer and doesn't need to change either way. Whether a facility stays with AC or transitions to 800VDC, overhead track busway remains the most effective white-space distribution architecture.

Look at what 800VDC actually changes and what it doesn't. It changes the voltage, the conversion plant, and the protection scheme. It does not change the physical problem in the white space: megawatts of power have to travel overhead down a row of racks and drop to each rack through a connection that can be added, moved, upgraded, and metered without an outage. Starline track busway was engineered to solve, these issues regardless of the voltage.

"Nvidia's roadmap states 800V HVDC reference architecture doesn't replace busway; it specifies that "800V busways replace 415VAC busways" in the white space. The form factor survives the transition intact. Overhead busway is the distribution layer on both sides of the architecture diagram — the voltage on the bus changes, the architecture doesn't.

This is why we tell every client planning an AI-capable facility the same thing: standardizing on Starline track busway for white-space distribution is the safest decision you can make today to be well positioned for future designs.

• If 800VDC arrives on schedule for your loads — your overhead infrastructure, supports, routing, and operating practices are already built around the track-and-tap-off model. The transition is an equipment conversation, not an architecture conversation.
• If it arrives late, or your loads stay AC longer than projected — you're running the best AC distribution system available at high density, with continuous-slot access, hot-swappable metered tap-offs, and A+B redundancy a proven design trusted by hyperscale end users.

Along side Starline sits the broader Legrand critical-power ecosystem — Kratos MV switchgear on the gray-space side, rack PDU and containment in the white space — which means the vendor developing your busway is the same organization engineering the DC-era portfolio around it. As Starline's authorized representative, Arkhon Power Group has direct factory access to that roadmap: what's shipping, what's listed, and what's coming for DC applications.

The Bottom Line

800VDC is the right architecture for megawatt-class AI, although it may still be a minority architecture in 2030. The winning strategy is neither betting the facility on it nor ignoring it: build conventional where conventional works, provision side-car capacity where AI loads are plausible, and standardize on the one piece of infrastructure that serves both futures — overhead track busway. That's a strategy you can commit capital to today without knowing exactly how fast the transition runs.