AI training clusters can swing GPU load from near zero to hundreds of megawatts and back again several times per second, and that swing pattern is what is pushing hyperscalers toward 800 VDC distribution and toward UPS or grid-isolation hardware that can absorb it before it reaches the interconnection point.
Quick take
800 VDC power architecture is the choice when a facility can absorb higher upfront switchgear complexity in exchange for lower conversion losses and smaller equipment footprint under NVIDIA Vera Rubin-class GPU density.
External UPS or grid-isolation hardware, not the DC bus voltage itself, is what protects the grid from GPU load swings and satisfies ERCOT-style ride-through rules.
Grounding and protection design still diverges between 800V DC and ±400V DC topologies, so the choice is not yet standardized across vendors.
800 VDC power architecture: the efficiency case
The clearest field evidence comes from a California Energy Commission-funded demonstration at the San Diego Supercomputer Center, where Alderbuck Energy’s Nexus Power Unit will convert 12 kV AC-to-800 VDC ahead of deployment in an AI data center.
Alderbuck’s stated targets, evaluated against a conventional 480 V AC data center power architecture, are a 4% energy-efficiency improvement, more than 50% reduction in power equipment footprint, and more than 50% faster power-system installation.
Those are vendor targets, not measured results. The SDSC unit still has to pass factory testing and hardware-in-the-loop validation at UC San Diego’s $42 million NSF-funded DERConnect facility before it reaches a live data center.
Separately, LG Uplus and LS Electric signed a memorandum of understanding on July 21, 2026, to jointly develop and standardize an 800-volt direct-current distribution system aimed at NVIDIA’s Vera Rubin platform, where a single NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs.
That partnership signals commercial intent from a telecom-turned-data-center operator and a power-equipment maker, but it is a testing and standardization agreement, not a shipped product architecture the market can benchmark yet.
UPS and grid isolation under AI load swings
The efficiency argument for 800 VDC is separate from the reliability argument. IEEE Spectrum’s reporting on grid instability treats them as distinct problems: data centers could account for 3 to 4 percent of total global consumption within this decade, but the harder issue is how synchronized GPU load patterns behave on the grid, not the total draw.
ON.Energy and Crusoe’s newly announced partnership addresses that second problem directly, with a 5GW deployment of ON.Energy’s AI UPS technology across US hyperscale campuses, and commissioning expected to commence this year and extend into 2027.
ON.Energy’s AI UPS is a medium-voltage unit installed outside the data center, positioned inline between the grid and the facility so GPU load variations are absorbed before they reach the interconnection point, rather than being smoothed out only after they strain grid equipment.
Each 3.5MW unit combines a bidirectional power conversion system, batteries, a second power conversion system, and a transformer, with batteries rated for up to eight hours of backup power, according to the reporting cited in the ON.Energy coverage.
LVRT compliance and the ERCOT baseline
ON.Energy’s AI UPS has been validated against ERCOT’s Large Load Interconnection requirements, including compliance with Nodal Operating Guide Revision Requests (NOGRR) 282 and ERCOT’s Low Voltage Ride-Through protocols, which matters because ERCOT’s Large Load Working Group is separately evaluating how to accommodate multiple 800 VDC facility designs.
LVRT compliance is a grid-code requirement, not an 800 VDC feature. A facility can run 800 VDC internally and still fail ride-through rules if the interconnection hardware in front of it cannot hold the connection during a voltage fault, which is the gap UPS and isolation hardware are built to close.
Grounding, topology choice, and what isn’t standardized yet
ABB’s whitepaper on the Open Compute Project’s disaggregated power rack architecture frames the open design question as 800V DC versus ±400V DC topologies, each carrying different grounding, protection, and UL/NEC compliance implications.
That split matters for procurement because the internal rack-level DC topology is not yet converged even as facility-level 800 VDC distribution moves toward pilots, so evaluating one vendor’s 800 VDC claim does not automatically validate compatibility with another vendor’s topology.
What to verify before committing
Choose an 800 VDC power architecture path now when the workload is a new-build, GPU-dense facility where footprint and installation-speed targets like Alderbuck’s are worth underwriting as unproven vendor claims pending SDSC field data.
Avoid committing to a single vendor’s 800 VDC topology when the facility also needs OCP Sidecar-style rack disaggregation, since grounding and protection choices between 800V DC and ±400V DC are still open per ABB’s own framing.
Grid-facing reliability is a separate purchase decision from bus voltage. ERCOT-style validation, as ON.Energy has pursued through NOGRR 282, is the credential to ask any UPS or isolation vendor for, regardless of whether the facility runs 800 VDC or a conventional architecture.
Before signing, ask for measured, not targeted, energy-efficiency and footprint numbers from the SDSC demonstration once Alderbuck’s factory testing and DERConnect validation are complete, since the 4% and more-than-50% figures are pre-deployment targets.
Confirm which grid code the UPS or isolation vendor has actually been validated against. ERCOT’s NOGRR 282 and LVRT protocols are one baseline, but other interconnection regions may set different ride-through terms not covered in this evidence.
Finally, verify which DC topology a rack-level vendor commits to, 800V DC or ±400V DC, before assuming compatibility with a facility-level 800 VDC distribution plan, since ABB’s own comparison treats the two as distinct engineering paths.