If you’ve been keeping half an eye on AI data center power infrastructure lately, 800VDC has become pretty hard to ignore.

NVIDIA, Google and Microsoft are working through the Open Compute Project (OCP) to establish an 800VDC architecture for next-generation AI infrastructure. In July, OCP published its Solid-State Transformer Specification v0.3, laying out system requirements around medium-voltage AC to 800VDC conversion. NVIDIA later said that more than 80 equipment manufacturers and infrastructure companies were already developing products around the emerging architecture.

September brought two more SST headlines. On September 2, Infineon and Skeleton Technologies announced a partnership around solid-state transformers and high-power energy storage sidecars for AI data centers. Six days later, Enphase said its 4kW SST power modules were now being built in Texas for full-scale rack assembly and validation.

Solid-state transformers themselves aren’t exactly new. Researchers and power electronics companies have been working on the concept for years. The more interesting question in 2026 is why AI data centers are suddenly giving the technology so much attention.

AI Racks Are Getting Very Power Hungry

The appeal of 800VDC starts with a pretty ordinary electrical headache: current.

AI racks are becoming extremely power-dense, while all that electricity still has to move through limited space inside the data center. At the same power level, lower voltage means higher current. As current rises, conductors need to get larger, copper requirements increase, and busways, cabling and thermal design become harder to manage.

NVIDIA is already planning a row-level architecture supporting up to 2MW per row, using an overhead 800VDC busway. Its longer-term roadmap pushes the conversion further upstream, toward facility-scale DC power blocks capable of converting directly from medium voltage.

Raise the distribution voltage and the same amount of power can move at lower current. That is one reason OCP points to lower conductor and copper requirements when explaining the move toward higher-voltage DC distribution.

Then there is the obstacle course power has to run before reaching a GPU. Data center power architectures vary, but transformers, UPS systems, rectifiers and multiple DC/DC conversion stages are common along the way. NVIDIA has framed its 800VDC work around shortening this "grid-to-GPU" power path. At hyperscale, even relatively small conversion losses can add up quickly.

Which raises a pretty obvious question: if the racks ultimately want 800VDC, how much of the equipment between medium-voltage AC and that DC bus can be reduced or combined?

Enter the solid-state transformer.

What Is the SST Actually Trying to Remove?

Going 800VDC does not automatically mean using an SST.

Conventional transformers paired with rectifiers can feed a DC system, and Transformer Rectifier Units (TRUs) offer another route. OCP itself is working on a broader MV-to-DC architecture and has not crowned SST as the single answer. In fact, OCP explicitly refers to both TRUs and SSTs as facility-level building blocks for future DC-native AI data centers.

The attraction of an SST is its ability to combine medium-voltage conversion, electrical isolation and active power conversion within one system.

A stripped-down traditional path might look like this:

Medium-Voltage AC → Transformer → Low-Voltage AC → UPS / Rectifier → DC

An SST-based layout could look closer to:

Medium-Voltage AC → SST → 800VDC

How much equipment, floor space or energy this actually saves depends on the real system design. A clean block diagram does not guarantee a free lunch. But data center designers have good reason to investigate technologies that can reduce conversion stages or free up usable space, because compute density is becoming increasingly expensive to support.

This also explains why SSTs could remain technically interesting for years without displacing conventional transformers on a large scale.

Traditional transformers are extremely difficult products to dethrone. They have decades of field experience behind them, along with mature protection and maintenance practices. Operators know how they behave, how they fail and how to keep them in service for a very long time.

So SSTs have always had to answer a fairly brutal commercial question: why should a risk-sensitive operator pay more for sophisticated power electronics and accept a different set of failure modes?

AI data centers may finally have a convincing use case. Space is expensive, power density is exceptionally high, loads can change quickly, and the computing equipment at the end of the chain already runs on DC.

SST Is Starting to Move Beyond the Research Paper

What makes OCP’s work this year interesting is the companies sitting around the table. Google, Microsoft and NVIDIA are involved in the MV AC to 800VDC power conversion workstream, working on real system requirements including power quality, power smoothing, interfaces and interoperability. OCP is also engaging with UL Solutions, NFPA, IEEE and IEC on the safety certification and regulatory framework needed for deployment.

These are very practical engineering questions. Can equipment from different manufacturers work together? How should faults be handled? What will certification require?

September’s announcements pushed the conversation closer to physical hardware.

Infineon and Skeleton have signed an MoU to collaborate on SSTs and high-power sidecars for AI data centers. Infineon brings power semiconductor expertise, while Skeleton contributes high-power energy storage technology intended to help manage fast-changing AI loads. This remains a development program; neither company is claiming that a large commercial SST installation is already operating in a hyperscale facility.

Enphase, meanwhile, has actual hardware coming off a production line. On September 8, the company announced that the 4kW power modules for its IQ Solid-State Transformer were being built in Texas. Hundreds of these modules are intended to work together in full SST racks, with system capacities of up to 5MW. Enphase says the architecture is designed to convert medium-voltage AC directly to 800VDC and respond to dynamic AI loads on a sub-millisecond timescale.

To be clear, this does not mean a 5MW Enphase SST is already running commercially inside a hyperscale data center. The company says the modules are currently being used to assemble and validate full-scale IQ SST racks.

At this stage, real operating data will tell us far more than another polished SST concept rendering.

So What Happens to the Conventional Transformer?

In the short run? Probably very little.

NVIDIA’s own roadmap makes it clear that 800VDC is expected to enter the market gradually. Its near-term MGX-compatible 800VDC power rack is specifically designed to work inside existing AC data centers. Later stages move toward row-level conversion, followed by facility-scale DC power blocks capable of connecting much closer to the medium-voltage supply.

A facility operator is not going to rip out perfectly serviceable transformers, UPS systems and distribution equipment simply because a new architecture appears. Purpose-built greenfield AI campuses with extremely high power densities have much more freedom to experiment with MV-to-800VDC systems.

Even inside those new facilities, SSTs still have a difficult list of questions to answer. Conventional transformers can remain in service for decades. How will hundreds of semiconductor modules age under continuous megawatt-scale loading? What happens when a module fails? How quickly can it be replaced? How should DC faults and 800VDC arc-flash risks be managed? And will the gains in space and conversion efficiency justify the additional electronics, controls and thermal management?

For a data center operator, uptime still comes first.

Saving copper and freeing up floor space are attractive. But if maintenance teams are uncomfortable servicing the system, spare parts are hard to source, or the certification route remains unclear, adoption in mission-critical facilities will be much slower.

The SST story will become far more convincing when the headlines start moving from "another company is developing an SST" to real field experience: where the equipment is running, for how long, at what load, and with what reliability.

The effect of 800VDC on data center power architecture could eventually become a bigger story than the simple question of whether SSTs replace conventional transformers.

In today’s facilities, transformers, UPS systems, rectifiers and battery energy storage systems usually handle distinct parts of the power chain. A common DC backbone gives engineers more freedom to rearrange some of those functions. Energy storage can connect directly to the DC side, repeated AC/DC conversion stages can potentially be reduced, and some power-control functions can move further upstream. OCP is already discussing BESS integration, DC UPS functionality and future DC microgrids as part of its 800VDC architecture.

That leaves a genuinely interesting engineering question for future AI power systems: should transformation, rectification, energy storage and power-quality control continue to live in several separate pieces of equipment, or will some of those functions gradually be combined?

Nobody has a final answer yet.

What has changed in 2026 is that companies have started building the hardware needed to find out. OCP has published specifications. More than 80 companies are developing products around the 800VDC ecosystem. Semiconductor and energy-storage suppliers are entering SST development programs, and real SST power modules are now being assembled into full-scale validation racks.

The next milestones are fairly easy to define: the first large SST deployments inside operating AI data centers, real efficiency and reliability data, further revisions of the OCP SST specification, and clearer safety and certification requirements from organizations such as UL, NFPA and IEEE.

For Winley, this is the part of the 800VDC story worth watching most closely. As a transformer manufacturer, we are interested in how these new power architectures may change transformer specifications, system interfaces and the role of conventional magnetic equipment inside future data centers.

Conventional transformers are not going extinct tomorrow because AI data centers discovered 800VDC.

But the scale of AI infrastructure is making engineers revisit a question that used to have a much more predictable answer:

From the medium-voltage service entrance to the GPU, how many pieces of electrical equipment do we actually need?

Working on a data center power project?
Share your voltage, capacity and application requirements with Winley. Our team can help review the transformer options for your project.

[ Discuss Your Project ]

Related Products