At GTC 2026, NVIDIA’s CEO offered a useful way to frame modern AI infrastructure: a “five-layer cake” that starts with energy, then moves through chips, platforms, models, and applications.
The important signal wasn’t the stack itself; it was what sat at the bottom. As AI scales, energy and power are primary design constraints. The data center is evolving from a general purpose cloud platform into something more specific: an AI factory, optimized to convert energy into tokens as efficiently as possible.
That shift has real implications for Electrical Original Equipment Manufacturers (EOEMs). But not the ones most headlines suggest.
AI infrastructure is scaling, but architecture still matters
NVIDIA now projects global AI infrastructure investment approaching $1 trillion by 2027, driven by hyperscalers, governments, and large enterprises racing to deploy next-generation capacity.
For EOEMs, the headline implication is straightforward: more AI means more capacity, more speed to deployment, and more modular electrical infrastructure.
But the architectural nuance matters. Not every AI deployment looks the same, and not all require radical electrical redesign.
AI factories change how performance is measured
In an AI factory, performance is increasingly evaluated by tokens produced per unit of energy, not just uptime or raw compute speed.
New AI platforms are optimized around this reality, driving:
- Higher rack level power densities
- Greater sensitivity to electrical efficiency
- Tighter coupling between power, thermal, and compute design
However, this does not immediately translate into wholesale changes across the electrical infrastructure stack.
The real question is where AC converts to DC
The architectural shift underway is often mischaracterized as “AC vs DC.” That framing is misleading.
The real question is where the conversion from AC to DC happens, and how that boundary moves over time.
For an EOEM, this matters because it defines what you build, what you buy, and what the facility delivers to your rack.
Today (Vera Rubin NVL72, per Schneider RD113): The boundary sits inside the rack. The facility delivers 480 VAC; rectification happens at the rack via 110 kW power shelves. Racks reach 227 kW, with liquid cooling doing the heavy lifting.
Next (Rubin Ultra / Kyber): The boundary moves out of the rack. Rectification shifts to a dedicated sidecar power rack next to each compute rack, which then delivers 800 VDC to the GPUs. Racks scale to ~1 MW. The compute rack becomes a DC consumer.
Why this matters for EOEMs: The boundary’s location determines where the AC-to-DC bill of materials lives, who owns thermal and safety engineering for it, and whether the rack design assumes AC input or DC input. Planning for both generations means planning for that
Why upstream DC is harder
As DC conversion moves farther upstream from next‑to‑rack to POD‑level, to facility‑level complexity increases sharply:
- Safety and arc‑fault considerations
- Grounding and fault isolation challenges
- Standards (UL / NEMA) lag
- Trades training becomes a gating factor
Because of this, rack‑adjacent rectification is the practical near‑term solution.
What 800 VDC means for EOEMs today
This is where the narrative often goes wrong.
For EOEMs, the near‑term implication of 800 VDC is not a forced pivot in product strategy.
In today’s architectures:
- Everything upstream of the power rack remains AC
- Traditional UPS, switchboards, switchgear, transformers, and protection systems still apply
- Existing EOEM portfolios remain fully relevant and in growing demand
This is precisely why you don’t see EOEMs rushing to redesign their entire portfolios around 800 VDC. They don’t need to. EOEMs can continue to:
- Deliver large amounts of AC capacity
- Support prefabrication and modular deployment
- Improve speed, density, and repeatability
- Scale proven architectures with confidence
All of that equipment feeds high‑power rectifiers located next to AI racks. This aligns with how customers are deploying AI infrastructure today.
In a follow-up, we’ll outline when 800 VDC actually becomes necessary, and what triggers that shift.
Go deeper
- Download: Our latest Reference Design Guide
- Explore: New AI Data Center Electrical Solutions
- Assess: Your 2026–2028 roadmap
Add a comment