800 VDC power architecture for AI data centers
Arete Intelligent Power builds a native 800 VDC power architecture for AI data centers. It carries power from the medium-voltage service to the compute rack with fewer conversion stages than a legacy AC chain, and holds the supply steady through the load swings AI workloads create — without a centralized UPS room. Three configurations cover the range: in-rack power shelves for facilities that keep AC to the rack, an HVDC power rack that retrofits an existing AC data center, and medium-voltage direct conversion through a solid-state transformer.
Where the legacy chain runs out
Racks now draw more power, more suddenly, and more continuously than the legacy grid-to-load chain was built for. A native 800 VDC block delivers more usable capacity per megawatt and per square foot, rides through volatile GPU loads without a separate UPS, and reaches energized capacity through fewer vendors and less to interconnect.
The constraint is rarely the chip. It is everything between the utility service and the chip — the transformers, the switchgear, the uninterruptible power supply room and the distribution that has to be sized, sited, procured and energized before a single rack draws load.
Three configurations, one architecture
The three differ in where conversion happens and what the facility already has, not in the architecture underneath them.
- AC LPS and Power Shelf Solutions — three-phase facility AC carried to the rack and converted once, at the compute. No DC bus to design.
- HVDC Power Rack Solutions — conversion beside the compute rack, facility AC in and a high-voltage DC bus out, so an existing AC data center can adopt 800 VDC without rebuilding the power room.
- MV Direct Conversion — medium voltage converted straight to the DC bus through a solid-state transformer, with no line-frequency transformer in the path.
What stays steady when the load does not
AI and HPC racks swing their draw faster than the upstream network can follow. Arete's load power smoother sits alongside the supply and instantly supplies or absorbs transient power, so the profile the utility sees stays flat while the compute does not.
Because it is coupled alongside the supply rather than carrying load power, nothing critical depends on it. The same function appears at the DC bus in the high-voltage configurations, and inside the shelf as built-in supercapacitor storage in the in-rack ones.
Common questions
What specifiers ask first
How does a native 800 VDC architecture increase usable capacity?
A native 800 VDC block delivers more usable capacity per megawatt and per square foot than the legacy AC chain, because fewer conversion stages sit between the utility service and the chip and each stage that is removed is both a loss and a footprint.
Counting conversions between the utility and the DC bus that feeds compute: a conventional AC path and an HVDC path both take two. Medium-voltage direct conversion takes one.
What handles GPU load transients without a centralized UPS?
A load power smoother with supercapacitor storage. It monitors the supply directly and instantly supplies or absorbs transient power, holding a stable current profile upstream while the compute load swings.
In the in-rack configurations the same capability is built into the power shelf itself, with 20 ms hold-up time at 100% load, so there is no separate capacitor bank to site.
Can an existing AC data center move to 800 VDC without rebuilding the power room?
Yes — that is what the HVDC Power Rack is for. It sits beside the compute rack, takes three-phase facility AC in and puts a high-voltage DC bus out, so the transition happens at the row rather than at the service entrance.
It supports either high-voltage DC architecture — 800 VDC or ±400 VDC — and can carry battery or supercapacitor storage in the same enclosure.
What is the conversion path from medium voltage to the chip?
Today, in both the AC and HVDC configurations, the utility's medium voltage is stepped down to low-voltage AC and then rectified to DC — two conversions before the bus that feeds compute.
With a solid-state transformer the medium voltage is converted to the DC bus in a single stage, with no line-frequency transformer in the path. The transformer is also bi-directional, so power flows both ways.
Which configuration fits your site?
A few questions about how power reaches your equipment. You will see the configuration we would propose, and you can request the documentation for it in the same step.