Subscribe to Our Digest
The GB200 NVL72 rack draws approximately 120 kilowatts. Its successor, the Vera Rubin NVL144, targets around 200 kW. At those densities, the 48-volt power architecture that served the cloud era well has become the bottleneck, and the industry has effectively agreed that 800 VDC is the answer.
NVIDIA’s 800 VDC reference design, formally presented at GTC 2026 in March, collapses the multi-stage AC-to-DC conversion chain into a single facility-level conversion, then distributes 800 VDC directly to the rack. Texas Instruments demonstrated a two-stage 800 V to processor-level voltage architecture at the same event. Higher voltage reduces I²R losses and copper mass substantially, making 800 VDC the only credible path for rack densities above roughly 100 kW. OCP ORv3 HPR shelves, Delta’s 660 kW in-row racks, and Eaton’s Beam Rubin DSX platform are the immediate commercial embodiments of this transition.
The 48-volt distribution standard that Google popularized through Open Compute was a genuine engineering advance. At 20 to 30 kW per rack, it worked well.
Power loss in a conductor equals current squared multiplied by resistance. Doubling voltage at the same power level halves the current, cutting conductor losses by a factor of four. At rack loads approaching 100 kW, the current through a 48-volt bus becomes impractical. Connectors, bus bars, and cable cross-sections all scale with current, not power. Delivering 120 kW at 800 V requires approximately 150 amperes. At 48 V, the same load requires 2,500 amperes. The OCP ORv3 HPR specification already hit this limit at 48 V; connector standards had to be revised just to handle 1,000-ampere outputs at the shelf level, as documented in the TE Connectivity OCP ORv3 power solutions guide
NVIDIA published its 800 VDC architecture reference design in March 2026, formalizing an approach the company had been steering ecosystem partners toward for 18 months. The core principle: convert AC utility power to 800 VDC once at the facility’s power distribution unit, then distribute 800 VDC directly to racks. This eliminates the intermediate AC-to-48V conversion stage in conventional architectures.
Texas Instruments moved from reference design to silicon at GTC 2026, held 16-19 March. The TI and NVIDIA joint announcement described a complete 800 VDC architecture requiring only two conversion stages from 800 V to processor voltage, down from three or four in the previous generation.
Liteon announced an 800 VDC rack and 110 kW power shelf for the NVIDIA MGX platform; Delta Electronics unveiled a 660 kW in-row 800 VDC power rack co-developed with NVIDIA, paired with a 480 kW battery backup unit; Eaton’s Beam Rubin DSX was specifically engineered for the NVL72 architecture.
The GB200 NVL72 at approximately 120 kW per rack is the current design point. According to Entogo’s 2026 AI data centre power equipment analysis, the Vera Rubin NVL144 targets around 200 kW per rack, with OCP 2025 reference designs already contemplating 1 MW rack clusters.
A 200-rack AI cluster at NVL144 densities draws approximately 40 MW from a single hall. Mechanical and electrical infrastructure designed for the 10 to 15 kW per rack average of a 2019-era hyperscale facility requires complete reassessment: floor loading, busduct ratings, cooling water flow rates, and UPS sizing all change fundamentally.
A conventional data centre converts AC power through a UPS, distributes at 480 VAC, then converts again to 48 V inside the rack. Each conversion stage runs 95 to 98 per cent efficient. Four stages in series compound to roughly 81 to 92 per cent overall efficiency, meaning 8 to 19 per cent of input power becomes heat before reaching the processor.
The 800 VDC model compresses this to two stages. At 98 per cent efficiency per stage, end-to-end efficiency reaches approximately 96 per cent. At 40 MW of IT load, that saves roughly 1.6 MW in cooling load. The ORv3 HPR architecture achieves 97 to 98 per cent efficiency at the shelf level.
A self-build hyperscale operator can specify 800 VDC from the design phase. For colocation operators the challenge is more acute: existing facilities were built for 48 V or 480 VAC distribution, and tenants are now requesting high-density AI suites within buildings not designed for 800 VDC.
The upgrade path exists; 800 VDC can be introduced zone by zone in a live facility. But it requires switchgear replacement, significant busduct work, and phased tenant communications. Building discrete AI-density halls as additions to existing colocation campuses is increasingly common and sidesteps the retrofit problem entirely.
For colocation operators, assess which current or planned high-density suites will receive GPU cluster tenants expecting 800 VDC compatibility. If that transition is within 18 months, design work must start now. Switchgear and busduct with 800 VDC ratings carry lead times of 40 to 60 weeks for custom specifications.
For self-build operators evaluating a greenfield AI campus, the NVIDIA reference design is the starting point. Engage the power electronics ecosystem for the specific rack density you are designing around. The GB200 and Vera Rubin design points differ enough that specifying for one and deploying the other creates costly rework.
In both cases, OCP ORv3 HPR specifications are the baseline for procurement conversations. Request measured test data at your target rack density, not data sheets.
Can an existing data centre be retrofitted from 48 V to 800 VDC without a full rebuild?
A zone-by-zone retrofit is technically feasible, but involves replacing switchgear, transformers, and bus duct serving the new zone, adding rack power shelves and UPS equipment rated for 800 V, and ensuring adequate arc-flash protection. It is a significant capital project, not a firmware update.
What happens to existing GPU racks rated for 48 V when a facility introduces 800 VDC?
Those racks continue to operate on their existing feed. The 800 VDC infrastructure serves new high-density zones only. Both voltage classes coexist within a facility provided the electrical design maintains appropriate separation.
What does the two-stage TI conversion architecture mean for processor voltages?
Modern AI processors require 0.7 to 0.9 V at the die. TI’s architecture converts 800 V first to an intermediate bus voltage such as 48 V, then to core voltage in a second stage. Two stages at 98% efficiency each deliver approximately 96% end-to-end efficiency, significantly better than the 81 to 92% achievable with three or four stages.
Are 800 VDC battery backup units available at commercial scale?
Delta’s 480 kW BBU, co-announced with its 660 kW in-row rack at GTC 2026, is the most visible commercial offering at this density. The market is early-stage; only a handful of vendors have products validated for architectures above 200 kW.
How does 800 VDC affect electrical safety certification requirements?
Distribution above 600 V triggers different arc-flash and insulation requirements under NEC and IEC 62368. Facilities must use cable and connectors rated for 800 V and ensure maintenance procedures account for the higher potential. Most hyperscale operators working on 800 VDC designs are engaging their authorities having jurisdiction early to align on inspection protocols.