What is the difference between 800V and 48V data center power?
800V DC carries a given amount of power at much lower current than 48V DC. That can reduce the conductor burden of dense AI racks and move power conversion out of valuable compute space. It does not mean processors run at 800V, or that every existing 48V installation needs replacement. Our judgment is that 800V deserves consideration when rack density creates a physical distribution constraint. An energy-saving retrofit needs a separate financial case.
The voltage labels refer to parts of the power path. A facility can distribute AC, convert it to DC at a rack, and use a lower-voltage busbar inside that rack. In an 800V design, higher-voltage DC travels farther toward the equipment before converters reduce it for the electronics. Describing the choice as two mutually exclusive kinds of data center hides those intermediate stages.
Nor is every system called 48V operated at exactly 48V. Murata's Open Rack V3 power shelf listing specifies 50V output and associates the product with the ORv3 48V PSU specification. Other architectures use 54V. These are related low-voltage distribution approaches, not interchangeable connector or operating specifications.
The useful comparison follows power between the same two electrical boundaries, with the same delivered load and redundancy. Comparing an 800V facility feeder with a 48V rack busbar without accounting for their different lengths and functions produces misleading savings estimates.
Why higher voltage changes the copper calculation
For a DC load, power equals voltage times current: P = V × I. Ignoring conversion and wiring losses, the current needed at the stated bus voltage is therefore P ÷ V.
| Delivered power | At 48V DC | At 54V DC | At 800V DC |
|---|---|---|---|
| 100 kW | 2,083 A | 1,852 A | 125 A |
| 250 kW | 5,208 A | 4,630 A | 313 A |
| 1 MW | 20,833 A | 18,519 A | 1,250 A |
These are aggregate idealized currents, rounded to the nearest ampere. Actual equipment divides power among feeds and converters. The table does not specify a single cable, connector, or rack design.
At equal power, the current at 800V is 6% of the current at 48V. Resistive conductor loss is I²R. If the total loop resistance stayed identical, the loss ratio would be (48 ÷ 800)², or 0.36%. Against 54V, the corresponding ratio is about 0.46%.
That is a comparison of one conductor path under fixed assumptions. It is not a forecast that an entire data center will consume 99% less electricity. GPUs still need power, converters still dissipate heat, and the 800V installation may use smaller conductors with higher resistance. A designer can spend the current reduction on lower losses, less copper, or some combination.
Copper mass also depends on route length, thermal limits, voltage drop, redundancy, and mechanical requirements. Each conductor estimate must include the outgoing and return paths. Applying the current ratio directly to the site's total copper bill ignores its AC wiring and downstream low-voltage conductors.
A practical design therefore needs a conductor schedule and installed-cost estimate for the proposed routes.
Why an 800V rack may still contain 54V power shelves
NVIDIA's second 800 VDC white paper describes an initial power-rack arrangement that converts AC to 800V DC beside the compute rack, followed by an 800V-to-54V shelf inside it. The design accommodates existing low-voltage compute hardware while changing how power reaches the rack.
The cost and conversion losses change again if 800V reaches closer to the processor and an intermediate rail is eliminated. Texas Instruments' March 2026 reference-design announcement describes an isolated 800V-to-6V converter followed by conversion from 6V to below 1V. TI reports 97.6% peak efficiency for the first converter. That figure covers one stage at its peak operating point, not the complete facility or every workload.
Two illustrative paths make the distinction clearer:
- A compatibility path: facility AC, side power rack, 800V DC delivery, 54V conversion, then downstream conversion for the processors.
- A lower-rail reference design: 800V DC input, 6V conversion, then processor voltage. Conversion upstream of the 800V input still has to be counted.
The measured efficiency depends on how the remaining converters perform. An added converter may enable cheaper installation or compatible hardware; eliminating one may require a new compute platform. A purchase comparison needs the efficiency curve of the complete proposed path, including light-load operation and redundant units.
How much can improved efficiency save?
Consider a modeled installation delivering a constant 10 MW at a common low-voltage output boundary after all conversion stages being compared. Assume the existing path is 94% efficient and the candidate is 96% efficient, including conductor losses and converter auxiliary power within that boundary. Both figures are scenario inputs, not measured specifications for 48V or 800V equipment.
Input power is output power divided by efficiency. The existing path draws 10 ÷ 0.94 = 10.638 MW. The candidate draws 10 ÷ 0.96 = 10.417 MW. The reduction is about 222 kW, or 2.08% of the original input.
At continuous operation and $0.10 per kilowatt-hour, the annual saving is:
(10 ÷ 0.94 − 10 ÷ 0.96) × 1,000 × 8,760 × $0.10 = about $194,000.
The table holds the 94% baseline, delivered load, tariff, and operating hours constant. It discounts five annual savings payments at 8%, with payments at year-end.
| Candidate path efficiency | Input reduction | Annual electricity saving | Five-year present value |
|---|---|---|---|
| 95% | 112 kW | $98,000 | $392,000 |
| 96% | 222 kW | $194,000 | $775,000 |
| 98% | 434 kW | $380,000 | $1,519,000 |
These values exclude cooling outside the measured path, demand charges, taxes, maintenance differences, and residual equipment value. Five years is the chosen comparison window, not a claim about electrical equipment life. Real loads fluctuate, so a project should apply efficiency curves to its load profile rather than assume this constant-load case.
If a retrofit carried a hypothetical $2 million incremental installed cost, the 96% case would have an electricity-only simple payback of about 10.3 years. Its five-year discounted electricity benefit would cover about $775,000 of that premium. Under these assumptions, a five-year investment case needs additional benefits, a lower installed premium, or a larger verified efficiency gain.
A capacity-constrained facility asks a different question. Holding the original 10.638 MW input constant, a 96%-efficient path could deliver about 10.213 MW at the output boundary, an increase of 213 kW. Using that capacity profitably requires compatible equipment, cooling, and paying demand. The electricity bill does not simultaneously fall by the full fixed-output saving. A financial model must specify whether it buys less electricity or uses the released capacity.
The distinction also applies to CPO versus LPO networking economics: a reduction in watts can have value as lower operating expense or as capacity for more computing, but counting both in full exaggerates the return.
What an existing facility has to pay to change
A side power rack can preserve the upstream AC installation, but the project still needs a place to put it. In its August 2026 architecture update, OCP makes the condition explicit: retaining upstream infrastructure depends on sufficient AC capacity and row space. It describes both local side-rack conversion and a longer-term route from medium-voltage AC to 800V DC.
The retrofit estimate should include the rectifiers, downstream conversion, distribution hardware, and protection, plus installation and commissioning. Battery arrangements may change. Work on an occupied row can also carry a scheduling cost: a cheaper design delivered late can postpone productive use of expensive compute equipment.
An AC system with years of useful life and spare capacity has a different replacement case from equipment due for renewal. Avoided future capital expenditure should enter the model only if it is actually avoided, at the date it would otherwise have been paid. The purchase price of equipment already owned is a sunk cost, although resale proceeds and future maintenance remain relevant.
NVIDIA's white paper addresses sustained DC arcing, grounding, fault isolation, and interlocked connectors. DC lacks AC's periodic current zero-crossings; a connector that can physically be removed is not permission to disconnect it under load. Required protection and certification must be included in the offered system and its cost.
TI's power-supply architecture paper explains how synchronized GPU activity requires fast converter response, coordinated current sharing, and energy storage that smooths the upstream load. Moving to 800V does not eliminate those requirements.
For an operator, the relevant reliability comparison includes the amount of compute affected by a fault and the time needed to restore it. Centralizing conversion changes that exposure. Fewer individual power supplies alone cannot establish a lower expected outage cost.
Which parts of the 800V rollout are ready?
NVIDIA's August 11 roadmap places the MGX-compatible power rack in the second half of 2026 and expects the row power center in 2027. Its facility-scale DC power block is a further deployment path. These are different products and schedules. An announced reference design does not establish fleet-wide availability or operating results.
OCP reports that Google, Microsoft, and NVIDIA are coordinating common requirements and have published Solid-State Transformer Specification v0.3. The specification gives suppliers common requirements; each offered assembly still needs appropriate certification.
Voltage terminology needs care as well. OCP distinguishes a side rack producing plus/minus 400V from a native 0-to-800V arrangement. A bipolar pair can have 800V between its rails, but its grounding, return arrangement, protection, and load interfaces require their own specifications. Equal headline voltage does not establish interchangeability.
For procurement, a usable delivery commitment identifies the complete system, supported compute hardware, approvals for the installation, factory acceptance criteria, and commissioning date. The public partner list provides much less purchasing certainty than those commitments.
Where adoption makes financial sense
We see the strongest early case in dense compute installations where the power equipment and conductors cannot fit economically within the available rack space. There, an 800V option may enable a planned deployment that an existing arrangement cannot accommodate. The relevant comparison includes alternatives such as distributing the compute across more racks, with their space and networking consequences.
For an existing installation that already meets its density target, measured operating savings have to carry more of the case. Our modeled two-percentage-point improvement produced about $194,000 of annual electrical savings for 10 MW delivered. That is material, but it leaves a $2 million retrofit dependent on benefits beyond a five-year electricity return.
Suppliers should be evaluated against the particular migration path their products serve. Side-rack rectifiers, rack DC/DC converters, protection devices, and facility power blocks enter projects at different stages. A sale can also displace another component the same supplier already sells. Higher 800V revenue is not automatically an equal increase in total revenue or profit.
Qualification work and repeatable installation can give a supplier an advantage. Open interfaces may also make bids easier to compare and increase competition. Investors need evidence of shipped systems, margins after support costs, and repeat customer orders before assigning a profit stream to a standards announcement.
An operator should request a complete installed quote and an efficiency measurement across agreed electrical boundaries, then price any density or availability benefit separately. Choose 800V when that combined result beats the best compatible alternative. A functioning lower-voltage installation with no density constraint needs a demonstrated return before it merits replacement.
Sources
- OCP: Murata ORv3 power shelf, 50V output and the 48V specification family
- NVIDIA: 800 VDC Architecture, Industry Alignment and Execution, white paper v2
- Texas Instruments: 800V-to-6V and processor-power reference designs, March 2026
- OCP: Joint 800V architecture and deployment update, August 2026
- Texas Instruments: Power-supply evolution and 800V architecture, March 2026
- NVIDIA: Power-rack and row-power-center roadmap, August 11, 2026
Current and financial tables are original InvisibleHill Research calculations. The financial case assumes a constant 10 MW delivered load, hypothetical path efficiencies, continuous operation, a $0.10/kWh tariff, and an 8% discount rate. Product announcements and reference designs are distinguished from independently measured deployment results.