Medium Voltage Is Moving Onto the Data Center Floor: What Engineers Need to Plan For Now

The Power Density Challenge

The power chain that has served data centers for thirty years — utility medium voltage stepped down to 480V, distributed as low-voltage AC, conditioned by a UPS, and routed to the rack through a PDU — was built for 5–15 kW racks. AI training racks are now running 120–140 kW, next-generation Blackwell Ultra and Rubin platforms are being specified at 250–900 kW, and NVIDIA's own roadmap points toward single-rack systems approaching 1 MW before the end of the decade.

That's not an incremental jump. It's an order-of-magnitude change happening inside the same four walls, on the same floor plan, with the same code books. Every engineer specifying a new AI-ready facility today is running into the same question: how much closer can medium voltage get to the rack, and what has to change to make that safe, code-compliant, and maintainable?

Why the Traditional Chain Is Breaking Down

Walk the conventional path and the strain shows up at every stage. Utility MV — typically 13.8 kV or 34.5 kV — comes into the building and passes through a line-frequency transformer down to 480V. That 480V AC goes to LV switchgear, then to a UPS for conditioning and ride-through, then to a PDU, which finally routes power to the rack. Inside the rack, it gets converted yet again to the DC bus the GPUs actually run on.

Each of those conversion stages is a place where copper, iron, and heat pile up. At 1 MW per rack under a traditional low-voltage architecture, a single rack's busbar alone would require somewhere in the neighborhood of 200 kg of copper — and scaled across a gigawatt-class campus, that copper burden becomes genuinely unworkable from both a cost and a physical-space standpoint. Every added conversion stage also adds loss, and at megawatt density that loss becomes heat that has to be removed by cooling systems already working at their limit.

There's a second, quieter problem: floor space. As rack density climbs, the proportion of the technical space consumed by low-voltage gear — transfer switches, switchboards, UPS systems, distribution boards, batteries — grows disproportionately relative to the compute space it's supporting. Every square foot given to LV distribution is a square foot not holding GPUs. That math is what's forcing the industry to ask whether MV distribution can move closer to the load — not just as an efficiency play, but as a floor-plan necessity.

Two Paths Engineers Are Choosing Between

Broadly, the industry is converging on two design responses, and most real projects will end up somewhere in a hybrid of the two.

Path One: Keep the Transformer, Shorten Everything Else

This is the conservative path — conventional MV switchgear and transformers, but sited and sized to push 480V (or higher) distribution as close to the row or pod as the building allows, minimizing conductor runs and consolidating LV gear into fewer, larger, better-coordinated blocks. This is where busway-based distribution earns its keep: it scales in place, supports higher ampacities per linear foot than parallel cable runs, and can be reconfigured as load pods are added — which matters when a facility that starts at 30 kW/rack is likely to be retrofitted for 80–100+ kW/rack within its operating life.

Path Two: Replace the Transformer

This is the more disruptive approach, and it's the one generating the most engineering attention right now — the solid-state transformer (SST). Rather than a passive iron-core transformer feeding a separate AC/DC conversion stage, an SST integrates a medium-voltage front end, high-frequency isolation, and downstream DC regulation into a single controlled device. Instead of stepping MV down to 480V AC and converting to DC later near the rack, an SST can take 13.8 kV or 34.5 kV utility MV and convert it directly to 800V DC (or ±400V DC) in one stage, right at the point of use.

The efficiency case is real: eliminating the AC floor and the extra conversion stage can cut distribution losses meaningfully at megawatt scale, reduce copper mass, and free up floor space that would otherwise hold LV switchgear. NVIDIA's own 800V HVDC reference architecture — introduced for the next generation of rack-scale systems — is built around exactly this idea, and it's why nearly every major power electronics supplier (Eaton, ABB, Hitachi Energy, Delta, Wolfspeed, along with a wave of well-funded startups) now has an MV-SST program aimed at data centers.

What Changes for the Engineer of Record

This isn't a drop-in swap, and it changes real engineering deliverables — not just equipment selection.

Protection Philosophy Shifts from Passive to Active

A conventional transformer-fed system relies on high available fault current and time-current coordination to clear faults. An SST is converter-limited — the semiconductor devices doing the conversion can't tolerate the fault currents a magnetic transformer shrugs off, so protection has to be handled through sub-cycle current limiting and coordinated shutdown logic embedded in the SST's control system. That turns your protection coordination study from a breaker time-current curve exercise into something closer to control-system validation. If you're used to specifying selective coordination the traditional way, budget real time to understand how a given SST vendor's protection scheme actually behaves under low-fault-current conditions, because it will not look like a conventional study.

Isolation and Insulation Coordination Get Harder, Not Easier

Bridging medium voltage to a DC bus at the rack means the isolation job that a conventional line-frequency transformer does for free now has to happen inside a high-frequency transformer stage — and insulation engineers report needing roughly three times the insulation margin at MV voltage classes to do it safely. This is a genuine design constraint, not a vendor talking point, and it affects physical clearances and enclosure design.

Redundancy No Longer Looks Like N+1 UPS Modules

Modular SST platforms are being built around dozens to hundreds of small power modules operating in coordinated series/parallel arrays, with the system designed to keep running even if a meaningful percentage of modules fail. That's a fundamentally different redundancy conversation with your client than "how many UPS modules do we need" — it's closer to specifying a fault-tolerant distributed system.

Lifecycle Mismatch Is a Real Specification Risk

GPU generations turn over in 18–24 months. Power distribution equipment gets specified for a 15–20 year service life. A switchgear lineup or SST platform installed to support this year's rack density needs a realistic path to supporting two or three future hardware generations without a full rebuild — which argues for modular, field-expandable architectures over fixed-capacity gear, even at a cost premium today.

Architecture Comparison at a Glance

The Next Challenge: Lead Times

Here's the practical constraint that should be driving today's decisions more than any efficiency number: medium-voltage transformer lead times are currently stretching out to as long as three years, and that's for conventional equipment — SST platforms from most vendors are still moving through demonstration and early pilot phases, with volume shipments generally not expected before 2027–2028.

If you're specifying a facility that needs to be energized in the next 18–24 months, you are very likely specifying conventional MV/LV architecture regardless of where the industry ultimately lands. The engineering decision today isn't "SST or transformer" — it's "how do I design a conventional system today that doesn't paint the owner into a corner when SST-based, higher-voltage-DC architectures mature."

That means:

• Reserve physical space and structural capacity for future MV-adjacent power rooms, even if the initial build uses conventional switchgear. Retrofitting SST capability into a facility not designed for it is far more disruptive than reserving the footprint up front.
• Specify busway and distribution infrastructure with headroom and reconfigurability, not just capacity for day-one load. A busway system sized and laid out to support 40 kW/rack today should have a documented, engineered path to 100+ kW/rack without a full re-pull.
• Engage transformer and switchgear procurement early — given multi-year lead times, the ordering decision is often the actual critical path on the schedule, ahead of construction.
• Build the protection and coordination study around the equipment you're actually installing now, and treat any future SST integration as its own protection study rather than assuming today's coordination scheme will carry forward.

Where to Source MV Equipment: Providers We Work With

Lead time and engineering flexibility are exactly where manufacturer selection matters most right now. The large legacy switchgear OEMs are quoting the longest queues in the industry, which has opened the door for agile, vertically integrated manufacturers who engineer-to-order and control their own production. Two providers we represent and recommend for MV solutions:

As the manufacturers rep for these lines, Arkhon Power Group provides application engineering, budgetary pricing, submittal coordination, and factory access for both — contact us to discuss which platform fits your project's specification, schedule, and budget.

The Bottom Line for Anyone Specifying a New Build Today

Medium voltage moving closer to the rack is not a speculative trend — it's a direct, physics-driven response to rack densities that broke the assumptions the industry has run on for two decades. But the timeline for solid-state transformer maturity means most projects breaking ground this year and next will still be built on conventional MV/LV architecture. The engineering job right now isn't choosing the future architecture — it's designing today's conventional system so it doesn't become tomorrow's stranded asset.

That's exactly the kind of decision that benefits from bringing in a power distribution specialist early in the design process, before layouts, transformer orders, and busway routing are locked in.