NVIDIA's reference architecture for next-generation AI infrastructure distributes power within the rack at 800 volts direct current (VDC). Other hardware and facility designers are following. The technical rationale is sound: higher voltage reduces current, shrinks conductor cross-sections and weight, and improves efficiency at rack densities that were impractical just three years ago.

That convergence is an important milestone, but not the finish line. It is the starting line. So much of the conversation today focuses on what happens inside the rack. And for good reason. As AI workloads become more demanding, the electrical infrastructure within data centers must evolve to deliver unprecedented levels of power efficiently and reliably.

Consider a 600-kilowatt (kW) rack. At 50VDC – the conventional low-voltage (LV) standard – delivering that level of power requires roughly 12,000 amperes (A) of current. The conventional copper conductor sizes, weight, and thermal management demands at that current level are physically unworkable at scale.

Higher-voltage architectures such as 800VDC directly address this by reducing current by more than an order of magnitude, cutting conductor mass, lowering resistive losses, and enabling power distribution that can actually keep pace with next-generation rack densities.

But 800VDC solving the in-rack problem is itself an acknowledgement of a larger reality: the electrical systems that powered the cloud era were not built for the AI era. The world has a power delivery problem, and AI is exposing it faster than anyone anticipated.

Extending beyond the rack

Power still has to travel through feeders, busways, and campus distribution systems before it reaches the rack. Conventional copper conductors have a practical current density ceiling. As total facility power grows from tens of megawatts toward hundreds, the required conductor cross-sections, structural loading, and space needed to support copper distribution systems become increasingly difficult to accommodate within existing facility envelopes.

While 800 VDC is an important advancement inside the data center, it does not solve the broader problem of moving significantly more power to and within facilities already constrained by space, permitting, aging electrical infrastructure, and the long lead times required to build out conventional distribution systems at this scale.

AI's real bottleneck isn't compute. It's power

The grid constraints facing AI infrastructure are well documented – interconnection queues, transmission congestion, and permitting timelines that weren't designed for facilities drawing hundreds of megawatts. But the infrastructure challenge doesn't end at the utility fence.

Even on-site, the physical reality of delivering power at AI factory scale through conventional copper infrastructure is becoming a hard constraint. Conductor cross-sections, structural loading, electrical room footprints, and the labor-intensive sequencing required to install conventional distribution systems are increasingly the critical path in a facility build.

Higher voltages and solid-state conversion help – but they don't change the fundamental physics of moving high-current power through copper at scale. It extends to every layer of the power chain between the grid and the rack.

Beyond 800VDC: The next layer of AI infrastructure

Superconducting power delivery is where the infrastructure argument goes next.

Importantly, this is not an either-or proposition. 800VDC and superconducting power delivery are complementary at multiple layers of the power chain. Higher-voltage DC distribution optimizes how power is managed within the rack.

Solid-state transformers (SSTs) are accelerating that transition – whether fed directly from medium voltage or downstream of a conventional transformer at low voltage, they replace the traditional conversion and distribution stack with a single 800VDC output stage.

Low-voltage superconducting distribution takes that further – moving high-current 800 VDC feeders through the building at a fraction of the physical footprint and weight of conventional copper busway or cable bus, without the parallelization and space penalties that conventional conductors impose at AI factory power levels. And at the campus and utility scale, medium-voltage superconducting delivery addresses the upstream capacity problem entirely – ensuring enough power reaches the facility in the first place.

Across all three layers, this approach also reduces the installation labor and construction sequencing complexity that make conventional power distribution a frequent critical-path constraint in data center builds.

The industry has answered the question of what voltage should power the AI rack. The harder question – whether we can build the infrastructure to deliver enough power to the rack at the scale now being planned – remains open. That is the problem the next generation of power infrastructure needs to solve.

Because ultimately, the companies that succeed in the AI era will not just be those that build the most powerful computing systems. They will be the ones that solve the infrastructure challenge at every layer between the grid and the rack.

To learn more about VEIR’s superconducting solutions visit veir.com.