Skip to the stories
Electronics & Semiconductors

800 VDC Could Recover Megawatts Lost to Conversion in AI Data Centers

  • Shaun Walsh
  • 847 reads

Cutting AC-to-DC conversion steps with 800 VDC could recover tens of megawatts per gigawatt, easing the power bottleneck that's slowing AI data center growth.

800 VDC Could Recover Megawatts Lost to Conversion in AI Data Centers

Power, not compute, is becoming the constraint on AI buildout. The power industry has three levers, according to the source: optimizing existing resources, moving data centers to DC distribution, and building new generation and delivery infrastructure designed for digital loads rather than analog ones. The argument is that all three have to work together, and the second lever is getting more attention than it used to.

The demand curve is steep

Data centers used an estimated 448 TWh of electricity globally in 2025, according to Gartner, and that figure is projected to more than double to 980 TWh by 2030. Goldman Sachs Research expects global data center demand to grow about 50% to 92 GW by 2027. U.S. demand alone is forecast to more than double, from 31 GW in 2025 to 66 GW in 2027. Deloitte estimates that power demand from U.S. AI data centers will grow more than thirtyfold by 2035, reaching 123 GW, up from 4 GW in 2024. McKinsey's central scenario calls for $5.2 trillion in global capital expenditure through 2030 to meet AI demand.

For the power sector, that's the largest demand-side growth event in a generation. Generation has taken most of the attention, but the source argues the faster route is converting to DC data centers and optimizing existing power delivery, which the Energy Efficiency Movement says could yield up to 20% more usable power.

Generation and interconnection are hitting physical limits

GE Vernova CEO Scott Strazik said on the company's July earnings call that it's now taking reservations for gas turbines to be delivered in 2031. Late last year, he expected to sell out reservations through 2030 by the end of 2026. That's a lead time measured in years, not quarters.

Interconnection is no better. Lawrence Berkeley National Laboratory's Queued Up report found more than 2,000 GW of generation and storage waiting to connect at the end of 2025, and time spent in queues is still rising.

PJM Interconnection runs the grid serving Northern Virginia, the world's largest data center market, and Virginia holds roughly 28% of U.S. operating data center capacity. PJM missed its reserve margin targets in its last two capacity auctions and planned a 6.8 GW backstop procurement for the 2028/29 delivery year, driven largely by data center demand. FERC approved only part of that plan in late September and sent PJM back to rework cost allocation and collateral rules. FERC Chairman Laura Swett called the last-minute filing a "mess," and PJM hasn't set a new timeline.

Pushback is growing too. A proposed bill, the Artificial Intelligence Data Center Moratorium Act, would pause new AI data center construction until federal AI safeguards are in place. Data Center Watch counts at least 75 projects worth about $130 billion blocked or delayed in Q1 2026, roughly matching all of 2025.

Every AC conversion step costs energy

Nearly every data center operating today is built around AC distribution. Utility power arrives as medium-voltage AC, transformers step it down, UPS systems provide battery-backed ride-through, power is distributed to each rack, and then each server converts it from AC to DC. Each of those conversions produces heat, adds a failure point, and wastes energy. Infineon estimates about 17% of an average data center's energy is lost across those steps, and new generation built on the same AC architecture would lose about the same share.

Nvidia's emerging 800 VDC platform attacks that chain directly. The company announced the architecture in May 2025 for 1 MW racks beginning in 2027, and its Vera Rubin DSX AI Factory reference design, released in March 2026, guides the transition in three stages:

  • The facility converts 13.8 kV medium-voltage AC directly to 800 VDC at the perimeter.
  • That power runs on 800 V busways to the compute rows, and racks accept two-conductor 800 V feeds.
  • At each rack, DC-DC conversion steps the voltage down to 54 V or 12 V, with a final step to the GPU's core voltage.

The efficiency gain comes from removing conversion stages. Rack-level AC-to-DC supplies disappear, along with the heat and failure points they add, and energy storage is built into the architecture.

What the recovered megawatts look like

In a 1-GW facility, losing 17% to conversion leaves about 830 MW for the IT load. Infineon estimates that specialized power semiconductors, which are central to 800 V designs, could cut those losses to about 9%, leaving 910 MW from the same utility feed. That's 80 MW recovered without adding generation. Nvidia's own estimate is more conservative: an end-to-end efficiency improvement of up to 5%, or roughly 50 MW in the same facility.

Software is a separate lever. Nvidia's DSX Max-Q, part of the Vera Rubin DSX platform, provisions power dynamically across the AI factory, and Nvidia says it lets operators deploy 30% more AI infrastructure within a fixed-power data center. Those are vendor figures, and they're worth testing against real deployments before planning around them.

Cost, copper and rack space

According to Nvidia's figures, as reported by IEEE Spectrum, 800 VDC distribution enables 85% more power to be transmitted through the same conductor compared with 415 VAC, reduces copper requirements by 45%, and lowers total cost of ownership by up to 30% at gigawatt-scale facilities. Nvidia also projects maintenance costs could drop by up to 70%, citing fewer power supply failures and lower upkeep labor. Smaller power equipment means more rack space returned to compute.

The timing is staged. Nvidia expects full-scale 800 VDC data center production in 2027, alongside its Kyber rack systems, and its MGX-compatible 800 VDC power rack for existing AC facilities arrives in the second half of 2026. Delta, Eaton, ABB, Schneider Electric, and Vertiv have all announced 800 VDC products or architectures. An Omdia report cited by IEEE Spectrum says higher-voltage DC data centers have already appeared in China, and the Mt. Diablo Initiative, a collaboration among Meta, Microsoft, and the Open Compute Project, is testing 400 VDC rack distribution.

Grid stability and the flexibility question

The grid side matters as much as the building side. Wood Mackenzie's 2025 survey put average delivery times at 128 weeks for large power transformers and 144 weeks for generator step-up units, with some orders stretching to four years. Against that backdrop, every percentage point recovered on a 1-GW campus is roughly 10 MW of generation that doesn't need to be built, permitted, interconnected, or energized.

Volatility is the other issue utilities are already watching. According to Nvidia, synchronized AI training can swing a rack's draw from roughly 30% to 100% in milliseconds, and across a data hall those swings can reach hundreds of megawatts within seconds. A DC architecture makes it easier to put energy storage where it does the most good, buffering swings before they reach the grid, and it integrates more naturally with batteries and solar arrays, which operate natively on DC.

Nvidia's DSX Flex software connects AI factories to grid services so they can adjust load dynamically and coordinate with onsite generation. Nvidia says it could unlock 100 GW of stranded grid power. Whether that holds up is an open question, but the direction matters: a facility that can flex its own demand is a different kind of grid asset than one drawing constant baseload.

What to watch

New gas-fired capacity, nuclear restarts, SMRs, behind-the-meter generation, and long-duration storage are all being pursued, and they'll take years. Without an architectural change, a typical data center keeps losing roughly 17% of its power to conversion before it reaches the GPU. My read is that the 800 VDC transition won't be universal by 2027, but the timeline is short enough that power engineers should be specifying DC-compatible busways, protection, and converters now. Utilities that learn to work with storage-buffered DC loads will have an edge in hyperscaler negotiations. And the teams that build DC-native infrastructure first will have the most room to scale racks from 100 kW to over 1 MW on the same footprint.

Source: Shaun Walsh · www.edn.com

From the same set