Grid to chip power efficiency: 800 V DC gains about 2 points, still unmeasured

Key takeaways

• For every 100 kW drawn from the grid, a ScienceShot stage model puts about 70 kW at the GPU core rails in a legacy AC design, 81 kW in modern AC with 48 V racks and 84 kW on the proposed 800 V DC path.
• The 800 V DC lead over modern AC is 2.3 points in the model. Single-factor checks on stages the two designs do not share move it between 1.85 and 2.71 points.
• No measured end-to-end comparison of 800 V DC and AC has been published. At the OCP Global Summit (12 to 15 October 2026), NVIDIA still listed an integrated test of transformer, storage and load as a future milestone.
• The efficiency figures in the summit decks ScienceShot read were requirements, vendor figures for single components, NVIDIA’s distribution model, or Google’s internal estimate against an unstated AC baseline.
• Power capping, in one NVIDIA example, and cooling can move useful work per facility megawatt by more than the conversion gap, so compare designs on validated workload per facility megawatt.

In a ScienceShot stage model, 100 kW taken from the grid leaves about 81 kW at the GPU core rails in a modern AC design with 48 V racks. The proposed 800 V DC path leaves about 84 kW, and a legacy AC design about 70 kW. As of 15 October 2026, nobody has published a measured end-to-end comparison of 800 V DC and AC. For a decision taken now, count no efficiency gain for 800 V DC. Judge it on rack current and copper, which our guide to 800 V DC for 1 MW racks works through.

The core rail is the supply of about 1 V or less that feeds the GPU chip itself. The model counts up to nine stages between the utility connection and that rail, and each one turns some power into heat. The OCP Global Summit, which ended on 15 October, added requirements and component figures. It added no measured result for a complete chain.

Where 100 kW from the grid goes

The model multiplies one efficiency per stage, from the high-voltage transformer to the voltage regulator beside the GPU. A legacy AC design loses 29.7 kW of every 100 kW on the way. A modern AC design with 48 V rack power loses 18.6 kW, and the proposed 800 V DC path loses 16.3 kW. Most of the 11-point step from legacy to modern AC comes from the rack power supply, the UPS (uninterruptible power supply, the battery-backed stage that rides through grid faults) and the voltage regulator.

The step from modern AC to 800 V DC works differently. It removes four stages: the medium-to-low-voltage transformer, the UPS, the low-voltage cabling and the AC rack power supply. In their place come a solid-state transformer (SST, a power-electronics converter that takes medium-voltage AC straight to DC), 800 V distribution and a rack converter from 800 V down to 50 V. The four removed stages pass 92.7% of their input in the model. The three new ones pass 95.3%. That difference, multiplied through the stages both designs share, is the whole 2.3-point gap.

On a 10 MW grid feed, 2.3 points is about 228 kW at the rails. That is close to two racks at 120 kW, the rack power NVIDIA uses in one of its summit examples. The model gives the 800 V path no energy-storage stage, although NVIDIA pairs the SST with storage in its plans. Any storage losses are not counted.

The table stops at the GPU core rail. It leaves out cooling, lighting, CPUs, memory, network and storage, so it is not a facility power budget. Reading 84 kW as “84% of a facility megawatt becomes AI work” would be wrong. “Mid-load” describes the intended scenario; the components were never measured at one common load point.

What the OCP summit added, 12 to 15 October 2026

The Open Compute Project (OCP) is the industry body that publishes open hardware specifications for data centres. ScienceShot read the slide decks of the five summit sessions closest to this question, and several more on 800 V DC, from the OCP recorded-content page. The session videos were not reviewed. Every efficiency figure in the decks is a requirement, a vendor figure for one component, a model or an internal estimate.

The requirement is the OCP SST specification, still at Revision 0.3.0, effective 22 June 2026. It asks for at least 98% efficiency between 50% and 100% utilisation, counting “power train loss only”. It lists two auxiliary supplies, one for controls and one for semiconductor drive and cooling, with no figure for what they consume. No newer revision was published at the summit. A separate proposal for 800 V modules on the server board targets above 98% from 25% to 100% of full load. Its workgroup starts in the fourth quarter of 2026.

Converter makers showed 800 V to 50 V shelves and modules quoted at 97.7% to 98.7%. Each is a peak or a full-load point for one component, and none states the auxiliary load or the test method. The model uses 0.975 for this stage, just below the full-load figures. Raising it to 0.98 would add 0.43 kW per 100 kW.

The only figure at the chip-to-grid boundary came from Google. Its session credits native low-voltage DC with more than 1% chip-to-grid efficiency, and a footnote calls the figures internal estimates against the current AC baseline. The deck states neither the baseline nor the test method. NVIDIA’s 2025 description of its 800 V architecture claims up to 5% better end-to-end efficiency than 54 V systems, also without a stated test setup. Google’s floor of 1% sits below the model’s 2.3 points and NVIDIA’s ceiling of 5% above, on different baselines, so neither confirms nor refutes it.

NVIDIA’s session on 800 V DC system performance requirements, listed in its summit schedule, dealt with electrical behaviour. It sets how the power system must handle GPU load swings to 150% of thermal design power (the rated sustained power) for 50 ms. Our article on why AI training makes the grid wobble explains where those swings come from. NVIDIA’s deployment deck lists “Demonstrate integrated SST + ESS + load performance” among milestones still ahead (ESS is energy storage).

OCP’s LVDC power distribution white paper moved to version 1.1.0 on 2 September 2026. It sets no numeric efficiency target. It says realised gains “depend on site-specific design, load profiles, and implementation details”. In its list of SST drawbacks it writes “Efficiency at scale to be proven”.

Levers larger than the conversion path

A summit deck from NVIDIA on power provisioning gives a sense of scale. Capping each GPU at 800 W instead of 1,200 W cut the power provisioned per rack from 120 kW to 90 kW, with the same throughput on one inference workload. That fits 1.33 times as many racks in the same power envelope. The same deck shows a static 540 kW budget of which 330 kW was used and 210 kW sat stranded (reserved but never drawn).

This is a vendor result for one workload at rack level, and it has nothing to do with 800 V. In that example the racks per power envelope rise by 33%. The conversion path moves rail power by 2.3 points in the model.

Cooling is the other large lever. PUE (power usage effectiveness) is total facility energy divided by IT energy, and cooling sits in that difference with the power delivery losses. Our EU PUE benchmarks by size and country show how far it varies between sites. None of it appears in the factor table.

The meter position also changes the answer. The Green Grid measures IT energy after all facility conversion, most likely at the output of the room power distribution units. By that rule the server power supply, the 48 V bus, the bus converter and the voltage regulator are IT load at every metering level, and a facility-side SST is overhead. The 800 V to 50 V stage counts as IT when the meter sits on its 800 V side, and as overhead when it sits in a separate power rack metered on the 50 V side. Two sites with the same total loss can therefore report different PUE.

The factor table, stage by stage

Each factor is an efficiency for one hypothetical path dedicated to an accelerator rail. The table does not split server input between the GPU and the other parts of the server. Every value is a scenario assumption. The label says whether a published requirement, a regulatory limit or a vendor curve informed it, and none of these measures this chain at one operating condition.

Stage	Legacy AC	Modern AC / 48 V	Proposed 800 V DC	Label	Basis and limit
HV/MV transformer	0.995	0.996	0.996	Assumption	Near the EU minimum peak efficiency index for large liquid-filled units of 3.15 to 40 MVA (99.52% to 99.72%, the efficiency at the most favourable load). Real loading is not known.
MV switching and distribution	0.999	0.999	0.999	Assumption	Allowance for switchgear and cable loss.
MV/LV transformer	0.994	0.995	none	Assumption informed by EU loss limits	EU maximum losses for a 2,500 kVA unit give 99.4% to 99.6% at 30% to 50% load. Unity power factor, transformer cooling excluded.
UPS	0.930	0.965	none	Assumption informed by vendor load curves	Legacy assumed, near the EU code of conduct floor of 92% weighted efficiency for double-conversion units above 200 kW. Modern is the mean of two vendor curves at 50% to 75% load.
PDU transformer	0.980	none	none	Assumption	Legacy stage only. A low-voltage unit, for which the EU transformer rules set no efficiency limit.
LV cabling	0.995	0.995	none	Assumption	Distribution loss.
AC rack or server power supply	0.900	0.970	none	Assumption informed by a requirement	Modern sits inside the OCP Open Rack V3 window (above 96.5% guaranteed and above 97.5% peak, at 30% to 100% load and 230 V to 277 V). Not a fleet average.
Solid-state transformer (SST)	none	none	0.980	Requirement	OCP SST Rev 0.3.0 floor at 50% to 100% utilisation, power train only. Auxiliary power not counted. Not a test result.
800 V distribution	none	none	0.997	Assumption	Inside the 99.5% to 100% that NVIDIA’s distribution model gives for busway layouts, 98.7% for the longest cabled case (a model, not a test).
800 V to 50 V rack conversion	none	none	0.975	Assumption informed by vendor peak figures	A vendor’s “up to 98%” for a 90 kW shelf, less 0.5 point. Summit component figures were 97.7% to 98.7% at peak or full load, auxiliaries not stated.
Low-voltage bus (12 V or about 48 V)	0.990	0.995	0.995	Assumption	Bus and connector loss.
Intermediate bus converter	none	0.975	0.975	Assumption informed by a vendor curve	Vendor peak above 97.5% over 30% of rated current. Not a universal operating point.
Core voltage regulator	0.880	0.910	0.910	Assumption	No primary mid-load curve found. Converter plus regulator (0.887) sits in the 84% to 89% full-load range of most published 48 V to 1 V prototypes above 200 A.
Power at the GPU core rail, per 100 kW in	70.25 kW	81.40 kW	83.68 kW	Product	Two decimals show the arithmetic, not measurement accuracy.
Electrical loss, per 100 kW in	29.75 kW	18.60 kW	16.32 kW	100 minus product	Cooling and other facility loads are not included.

Changing the modern AC voltage regulator from 0.91 to 0.89 or 0.93 moves that design to 79.6 or 83.2 kW. That one swing is larger than the gap between the two modern designs. Both use the same regulator, though, so the same change moves the gap only to 2.23 or 2.33 points. Shared stages make the uncertainty correlated.

The gap is sensitive to the stages the designs do not share. An SST at 0.975 instead of 0.98 cuts it to 1.85 points. A rack converter at 0.98 raises it to 2.71. A rack power supply at 0.965, the minimum the Open Rack V3 specification sets at 230 V to 277 V, raises it to 2.70, while a UPS at 0.97 cuts it to 1.86. Each is a single-factor check. No validated uncertainty model exists for this chain, so the checks give no confidence interval.

The transformer factors come from the EU ecodesign limits in Regulation 548/2014 and its amendment 2019/1783, evaluated at maximum permitted losses. The UPS floor is from the EU code of conduct for AC UPS. The rack power supply window is from the OCP Open Rack V3 48 V power supply specification, revision 1.0 (10 November 2022). The converter figures come from a converter maker’s white paper and a power vendor’s announcement. The regulator check uses a 2025 study of 48 V to 1 V converters from UC Berkeley and UC Santa Cruz.

How to compare two designs on your own site

For a site with a fixed power allowance, the buying metric is validated workload delivered per facility megawatt. Attach the benchmark, the quality target, the hardware and software, the meter point and the averaging period. A peak accelerator rating is a poor stand-in. Peak figures change with the number format, the sparsity assumption (skipping zeros in the arithmetic) and the product version. NVIDIA’s H100 table shows all three, and labels its figures preliminary.

System benchmarks are closer to the question. MLPerf Training measures how fast a system trains a model to a target quality. One result still covers one model at one scale, so check its system setup and power data before applying it to a site.

The facility step is one division: useful workload per facility MW equals system throughput per IT MW divided by PUE. Measure the PUE over the same run and at a stated IT meter boundary. For an energy PUE, compare total work with IT and facility energy over that interval. A single instantaneous power reading is not a run-average PUE, and an annual site PUE is not a conversion factor for a short benchmark run.

Our data-centre master plan estimates an annual PUE from a city’s hourly climate. That annual figure is the right input for a planning budget. The same boundary problem explains why one AI prompt can carry two energy figures.

A matched test of 800 V DC against AC would run the same load profile on both designs. It would meter power at the grid input, the IT input and the facility meter, include the auxiliaries and record workload throughput. No summit session reported one. The SST specification that such a test would rely on says of its verification load: “Detailed profiles will be provided in future versions.”

Sources

• OCP, Solid-State Transformer design specification, Medium Voltage to 800 VDC, Revision 0.3.0 (effective 22 June 2026), section 4.3 and clause 5.2.2
• OCP Data Center Facility, LVDC power distribution white paper, version 1.1.0 (2 September 2026)
• OCP Global Summit 2026, recorded content and slide decks (12 to 15 October 2026)
• NVIDIA, OCP Global Summit 2026 session schedule
• NVIDIA Technical Blog, 800 V HVDC architecture (May 2025, updated July 2025)
• NVIDIA Technical Blog, Hopper architecture in depth (H100 specifications)
• MLCommons, MLPerf Training benchmark
• The Green Grid, PUE: A Comprehensive Examination of the Metric (White Paper 49, via LBNL)
• Commission Regulation (EU) No 548/2014, ecodesign requirements for power transformers
• Commission Regulation (EU) 2019/1783, amending Regulation 548/2014
• European Commission JRC, Code of Conduct on Energy Efficiency and Quality of AC UPS, version 2.0 (2021)
• OCP, Open Rack V3 48 V PSU specification, revision 1.0 (10 November 2022), clause 3.8
• Converter maker white paper on bus converter modules (rev 1.1, November 2020)
• Power vendor announcement on 800 V DC products for the OCP Global Summit 2025
• Zou et al., 48 V to 1 V GaN switching bus converter, IEEE APEC 2025

@online{scienceshot-grid-to-chip-power-2026,
  author       = {Turner, Alex},
  title        = {{Grid to chip power efficiency: 800 V DC gains about 2 points, still unmeasured}},
  organization = {ScienceShot},
  date         = {2026-10-29},
  year         = {2026},
  url          = {https://scienceshot.com/post/grid-to-chip-power-efficiency-800v-dc}
}