800 VDC is a distribution voltage, not a chip voltage. Nothing on the GPU package runs at 800 volts — the accelerator still wants under a volt, and the board still wants 12 V or 48 V. What changes is the distance over which power travels at high current. In the architecture every GPU rack uses today, AC arrives at the row, an in-rack power shelf converts it to 54 VDC, and a copper busbar carries that 54 V to each compute tray across the width of the rack. At 800 VDC, the AC-to-DC conversion moves out — to a rectifier at the building perimeter, or to a power rack standing beside the compute racks — and the busbar carries 800 V instead, with a DC-to-DC converter inside each tray dropping it to what the silicon needs.

That is the entire idea, and it is worth saying plainly because most coverage dresses it up: the voltage went up so the current could come down, and the current had to come down because copper is heavy and I²R loss scales with the square of current. Everything else in this article — the topology war, the standards gap, the 3% versus 5% — follows from that one substitution and the decision about where to put the converter.

Why does 54 V stop working at a megawatt?

Two independent limits arrive at roughly the same place, and the vendors describe them with different numbers because they are describing different failures. NVIDIA's blog is specific about the engineering limit: "as racks exceed 200 kilowatts, this approach begins to hit physical limits," and it names three — space, copper and conversion count. Google's principal engineer Madhusudan Iyengar, speaking at the 2025 OCP EMEA Summit in Dublin, described the same busbar as having "served us very well between 10kW and 100kW a rack," while Google's own blog predicts that machine learning "will require more than 500 kW per IT rack before 2030." Both statements are on the record, and they are not in conflict: 200 kW is where 54 V becomes expensive to engineer, 500 kW is where it stops being buildable at any price.

The space limit is the one that reads as a punchline. A GB200 NVL72 or GB300 NVL72 rack uses "up to eight power shelves" to feed its compute and switch shelves. Scale that same 54 V approach to a megawatt and, in NVIDIA's words, "power shelves would consume up to 64 U of rack space for Kyber at MW scale, leaving no room for compute." Sixty-four rack units is most of a cabinet. The alternative NVIDIA names is equally unappealing: "a dedicated rack of power supplies for every compute rack" — which is a real architecture, and it is what several operators do today, at the cost of a whole cabinet of rectifiers per cabinet of GPUs.

So the honest summary of the 54 V era is that it did not fail on efficiency. It failed on floor space and on kilograms, which are the two things a data hall has a hard quantity of.

What is the copper actually worth?

Here the vendor numbers and first-principles arithmetic can be checked against each other, which is rare enough to be worth doing carefully. NVIDIA states that 54 VDC in a single 1 MW rack "requires up to 200 kg of copper busbar," and that "the rack busbars alone in a single 1 gigawatt data center could require up to 200,000 kg of copper." Those two figures are internally consistent at 1,000 racks per gigawatt. The question is whether 200 kg per rack is the right order of magnitude, and it is — with a caveat about design current density that changes the answer by a factor of two in either direction.

Take the conventional busbar design rule of roughly 1 A/mm² for naturally cooled copper in a rack (manufacturers publish ampacity tables within a factor of two of this; forced-air and liquid-cooled busbars run higher). One megawatt at 54 V is 18,519 A, so each conductor needs about 18,519 mm² of copper — 185 cm². A metre of that weighs 166 kg, and a two-conductor busbar (feed and return) weighs 332 kg. NVIDIA's 200 kg therefore corresponds to roughly 0.6 metres of two-conductor run, or a metre at a higher current density. Either way, the number is right and the implication is stark: the copper in a 54 V rack busbar is a structural element, not wiring.

Same power, two voltages120 kW rack (GB200 NVL72)1 MW rack
Current at 54 VDC2,222 A18,519 A
Current at 800 VDC150 A1,250 A
Copper per metre, 2 conductors @1 A/mm² — 54 V≈40 kg≈332 kg
Same at 800 V≈2.7 kg≈22 kg
I²R loss per metre — 54 V≈75 W≈622 W
I²R loss per metre — 800 V≈5 W≈42 W

Everything in that table is our arithmetic from P = V·I, ρ = 1.68×10⁻⁸ Ω·m for copper and 8.96 g/cm³, at a constant 1 A/mm² current density, and every row moves by the same 14.8× because holding current density constant makes cross-section, mass and loss all scale linearly with current. That is the useful property: the physics is one number, repeated. It also explains why the industry's language slides between 14×, 15× and 16× — the ratio is 800/54 = 14.8 and everyone rounds it differently.

Two vendor claims deserve separate treatment because they are frequently quoted as if they were the same claim. NVIDIA says 800 VDC busways transmit "85% more power … through the same conductor size" versus 415 VAC, and separately that copper requirements fall by 45%. The 45% we can reproduce: 1 MW at 415 VAC three-phase is about 1,385 A per phase, and an AC busway carries that across three phases plus a neutral, so the copper is roughly 4,155 amp-conductors against 2,500 for a two-wire 800 VDC feed — a 40% reduction, which brackets NVIDIA's 45% once you allow for how the neutral is sized. The 85% we could not reproduce from P = V·I at any conductor count; it is best read as a measurement of a specific busway product at its rated ampacity, where AC busways lose usable capacity to skin effect, proximity effect and reactive losses that DC does not have. Quote it as a product claim, not a law.

If the busbar burns 0.6 kW per metre, where does 5% come from?

This is the part of the story that gets skipped, and it changes how you read every efficiency number attached to 800 VDC. Take the 1 MW rack: 622 W per metre of busbar. Give each rack one metre and a gigawatt is 1,000 racks, so the entire busbar loss across a gigawatt of 54 V racks is about 0.6 MW — 0.06% of the load. Replace every busbar with 800 V and you recover 0.58 MW. You do not recover 50 MW.

So the headline benefit cannot be resistive loss in the busbar, and it is not. It is the number of times the power is converted. In a conventional chain the power goes through a medium-voltage transformer, a UPS that rectifies to DC and inverts back to AC, a row or rack PDU, an in-rack AC-to-54 VDC power shelf, and finally a board-level converter. Each stage is a loss, and the in-rack stage is the one with the worst conditions: an 80 PLUS Titanium server PSU is required to hit 96% efficiency at 50% load on 230 V, which means a megawatt rack's power shelves are dissipating on the order of 40 kW of heat inside the cabinet at best-case efficiency, before part-load derating. The 800 VDC chain deletes the UPS's double conversion and the in-rack AC-to-DC stage: one rectifier at the perimeter, one DC-to-DC in the tray, one converter on the board.

Put the two published numbers on the same footing and the difference resolves. Google's figure — "improves the end-to-end efficiency by ~3%" — is for the Mt Diablo sidecar, an AC-to-DC power rack bolted next to existing compute racks in a hall that still distributes 480 VAC to the row. It removes the in-rack AC-to-DC conversion and keeps everything upstream. NVIDIA's figure — "up to 5% improvement in end-to-end power efficiency" — is for converting 13.8 kV AC directly to 800 VDC at the data center perimeter, which also deletes the row-level AC distribution and the UPS's inverter. Different boundaries, different numbers, both defensible. If someone quotes you 5% for a sidecar retrofit, they are quoting a new-build figure against a brownfield project.

What is that worth in money? At 1 GW of IT load, 5% is 50 MW and 3% is 30 MW. At an assumed $70/MWh — roughly the US industrial average, and an assumption we are labelling as one — 50 MW continuous is about $30.7M a year and 30 MW is about $18.4M. That is a real number and it is the reason this transition is happening. It is also 3–5× smaller than the number implied by quoting "half the power is lost in conversion" in the same breath, which nobody credible does but plenty of slideware implies.

One more consequence worth flagging because it trips up reporting: PUE cannot see this improvement, and can move the wrong way. PUE is total facility power divided by IT power. If IT power falls 5% and the cooling and electrical overhead falls with it — which it roughly does, because there is 5% less heat to reject — the ratio is unchanged by construction. If the overhead does not fall, because pumps and fans are fixed, PUE gets slightly worse while the electricity bill gets better, since you shrank the numerator's biggest term. A site that swaps to 800 VDC and reports an unchanged PUE has not failed; it has used a metric that was never designed to measure a change in the efficiency of the IT load itself.

Which 800 volts? Two wires or three.

The industry has converged on "800 VDC" the way it converged on "USB-C": the voltage on the label matches and the implementations do not. There are two live topologies, and the difference is a safety decision before it is an engineering one.

NVIDIA: monopolar +800 VDCOCP Mt Diablo: bipolar ±400 VDC
Conductors2 — positive and return3 — +400 V, −400 V, midpoint
Highest voltage to ground800 V on every conductor400 V on every conductor
Where DC is made13.8 kVAC → 800 VDC rectifier at the building perimeter; DC-to-DC inside the rackAC-to-DC sidecar power rack beside the compute racks; Mt Diablo 2.0 for native 800 V
Claimed end-to-end gainup to 5%~3% (sidecar)
Stated advantageFewer conductors, simpler cabling and protectionInsulation and clearance rated at 400 V; components borrowed from the EV supply chain
Documented statusFull-scale production with Kyber racks in 20270.5 specification draft published for industry feedback in May 2025; targets 1 MW per rack

The reasoning behind ±400 V is the most instructive thing in the whole transition, because it is a supply-chain argument wearing a physics coat. Google's blog says the quiet part: choosing 400 VDC "allows us to leverage the supply chain established by electric vehicles (EVs), for greater economies of scale, more efficient manufacturing, and improved quality and scale." In a bipolar system no conductor is more than 400 V from ground, so clearances and insulation can be specified at 400 V instead of 800 V, using components that already ship by the million in EVs and industrial drives — at the price of a third conductor that carries no net power. NVIDIA's single 800 V rail makes the opposite trade: two wires instead of three, and every insulation system in the chain rated for the full 800 V.

Both are being built, and OCP's own position is that they will coexist: Mt Diablo's sidecar spec covers ±400 VDC today, with "a Mt Diablo 2.0 … coming later to support native 800V," and OCP describes two deployment paths — an LVDC side power rack next to the compute, or direct medium-voltage-to-800 VDC conversion feeding a whole data hall. For a buyer, this is a genuine standards risk and it belongs in the contract: a hall wired for one topology is not a hall wired for the other, and the transformer-rectifier or solid-state-transformer gear that feeds it is a long-lead item.

What breaks first: safety, spares, or the workforce?

Safety and standards, and the vendors say so. NVIDIA's blog is unusually candid for a launch document: "Deploying 800 VDC at the facility level introduces new challenges in safety, standards, and workforce training." That is three problems, and only the first is about physics. A DC arc does not self-extinguish at a current zero crossing the way an AC arc does, which is why DC protection gear is a specialised product rather than a catalogue item; OCP's answer is to borrow from electric vehicles and industrial power — connectors, protection devices, breakers "and solid-state breakers that clear faults in milliseconds" — and to adapt them for AI halls. Until that gear is listed and standardised, an 800 VDC hall is a bespoke electrical installation with a bespoke approval path, and the electricians who wire it have spent their careers on AC.

Redundancy granularity, which nobody markets. The old architecture had dozens of small PSUs per rack and N+1 or N+N sharing, so a failure cost you a module and a maintenance ticket. Centralising conversion produces a shared rectifier or a row power centre — the obvious efficiency win and the less obvious reliability trade, because the failure unit is now larger. NVIDIA acknowledges the direction of the work rather than claiming it is finished: "Innovations will still be required to fully deliver possible overcurrent protection reliability and maintenance gains," and "fault detection and serviceability in VDC systems is a key area for innovation." When a vendor writes that in the launch blog, believe it. Your specification should ask what the row rectifier's redundancy is, what a failure takes down, and whether the answer is one rack or thirty-two.

The rack still has power supplies. "No PSUs in the rack" is shorthand for no AC-to-DC PSUs. Every 800 VDC design still converts, twice — 800 V to a tray rail inside the rack, and the rail to the sub-1 V the silicon wants. What changes is that the converters are DC-to-DC, they are smaller, and they no longer need to be galvanically isolated from a mains supply in the same way. Teams that budget for "PSUs disappear" are budgeting for the wrong line item: the count falls, the per-unit value rises, and the spares you stock become expensive.

The thing that improves quietly: heat location. Moving 40 kW of conversion loss out of a 1 MW cabinet does not delete it, it relocates it — from a sealed, liquid-cooled box where every wasted watt competes with the GPUs for cooling capacity, to a power hall where it can be air-cooled or liquid-cooled on its own terms. Given that the constraint in 2026 is thermal density rather than floor area (our rack-density piece runs that arithmetic), relocating 4% of the load out of the rack is worth more than the 4% itself.

What can you actually buy in 2026, and what should you specify?

The production date is unambiguous and it is not now. NVIDIA: "Full-scale production of 800 VDC data centers will coincide with NVIDIA Kyber rack-scale systems in 2027." Everything shipping in 2026 is either 54 VDC inside the rack, AC to the row, or a sidecar that makes DC next to the rack. NVIDIA demonstrated the retrofit shape at GTC 2025 — "an 800 V sidecar to power 576 of the Rubin Ultra GPUs in a single Kyber rack" — and the OCP ecosystem is standardising the interfaces in public, with more than 80 partners named in OCP's own post. That is a lot of commitment and very little installed base.

Which means the decision in front of most buyers this year is not "800 V or 54 V." It is whether the shell you sign for in 2027 can host either. Six questions do most of the work:

1. What voltage does the busway carry, and who owns the rectifier? A hall built for 415 VAC distribution to the row cannot be converted to perimeter 800 VDC without new electrical infrastructure; the sidecar exists precisely to avoid that, at the cost of the 3% instead of the 5%. Ask which path the design assumes and price the difference rather than accepting "800 V ready."

2. How many watts per rack, at what rack unit count? A megawatt rack needs the 64U of power shelves to have gone somewhere. If the answer is "a sidecar," you have spent floor area; if it is "perimeter rectifiers," you have spent a long-lead electrical package.

3. What is the redundancy granularity of the DC plant? N+1 racks is not the same failure domain as N+1 modules. Get the answer in writing with the single-point-of-failure list.

4. Which topology, and is the second one supported? ±400 VDC and +800 VDC are not interchangeable, and the equipment that feeds them differs. If your accelerator supplier ships one and your facility team standardises on the other, you will be buying a converter you did not budget for.

5. What is the DC protection and listing story? Solid-state breakers that clear faults in milliseconds, DC-rated disconnects, arc-flash labelling at 800 V, and staff trained for DC. This is the item most likely to delay energisation, and it is not on the vendor's critical path — it is on yours.

6. Does the efficiency number you were quoted include the conversion you kept? 3% is a retrofit against existing AC distribution; 5% is a new build with a perimeter rectifier. Neither is "the saving from 800 volts." Both are "the saving from deleting specific stages," and the stages are named in the design documents.

The pattern here is the same one that runs through this series. Compute was the first bill, memory the second, the fabric the third, and optics turned out to be half of the third. Power was never a bill at all — it was a constraint that arrived as a queue position and a transformer lead time. What 800 VDC changes is that power distribution is now a design decision with a vendor ecosystem, which is progress of a sort: the constraint has become procurable. It has also become something you can get wrong for a decade, because the voltage you wire into a hall outlives the GPUs you install in it by at least two generations.