Are DC data centers the solution to AI scale?

The power industry has three opportunities to solve AI power demand. The first is optimizing existing resources, as Crusoe and other companies are doing. The second is migration to DC data centers, which yield higher performance per megawatt. The third is new power generation and delivery infrastructure optimized for digital rather than analog loads. We need all three working in conjunction.
Data centers consumed 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. AI’s power needs will eclipse anything the industry has seen before, and every major forecast says so.
Goldman Sachs Research expects global data center demand to grow by about 50% to 92 GW by 2027, and U.S. demand 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 alone will grow more than thirtyfold by 2035, reaching 123 GW, up from 4 GW in 2024. McKinsey projects that, in its central scenario, $5.2 trillion in global capital expenditures will be required through 2030 to meet AI demand.
For the power industry, that’s the largest demand-side growth event in a generation. While power generation has received the lion’s share of attention, the fastest path to enabling AI infrastructure is converting to DC data centers and optimizing existing power delivery, which could yield up to 20% more usable power, according to the Energy Efficiency Movement.
New generation built on today’s AC architecture, meanwhile, would still lose about 17% of its energy to power conversion, according to power semiconductors supplier Infineon Technologies.
That focus on generation is already running into 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 said he expected to sell out reservations through 2030 by the end of 2026.
Interconnection waits that were years long before the AI boom have only gotten longer. 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 the time projects spend in queues is still increasing.
The grid also can’t buy capacity fast enough. PJM Interconnection runs the grid serving Northern Virginia, the world’s largest data center market. Virginia alone holds roughly 28% of U.S. operating data center capacity, according to an Axios article. 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 the plan in late September, sending PJM back to rework cost allocation and collateral rules. FERC Chairman Laura Swett called the last-minute filing a “mess,” and PJM has not set a new timeline.
The political limits are growing too. Members of Congress have introduced the Artificial Intelligence Data Center Moratorium Act, which would pause new AI data center construction until Congress enacts federal AI safeguards. At least 75 projects worth about $130 billion were blocked or delayed in Q1 2026, roughly matching all of 2025, according to Data Center Watch.
DC power unlocks more compute per gigawatt
Nearly every data center operating today is designed around AC distribution. Utility power enters as medium-voltage AC. Transformers step it down, feed it into UPS systems for battery-backed ride-through, distribute it to each rack, and finally convert it from AC to DC inside each server.
Each conversion generates heat, adds failure points, and wastes energy. Infineon estimates that about 17% of an average data center’s energy is lost across those steps.
The emerging 800 VDC platform Nvidia introduced changes the power architecture, enabling greater AI productivity with the same power generation. Nvidia 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 data centers through the transition. It works in three stages.
- The facility converts medium-voltage AC (13.8 kV) 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 primary advantage of 800 VDC is eliminating multiple power conversion steps, which is where the efficiency gains come from. Rack-level AC-to-DC power supplies disappear, along with the heat and failure points they add, and energy storage is built into the architecture.
At scale, that difference is enormous. In a 1-GW facility, losing 17% to power conversion leaves only about 830 MW for the IT load. Infineon estimates that specialized power semiconductors, central to 800 V designs, could nearly halve those losses to about 9%, putting 910 MW to work from the same utility feed. That’s 80 MW recovered without adding generation. Nvidia’s own estimate is more conservative, forecasting 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 its Vera Rubin DSX platform, provisions power dynamically across the entire AI factory. Nvidia says it lets you deploy 30% more AI infrastructure within a fixed-power data center.
Rewriting data center economics
Recovered megawatts are only part of the case for 800 VDC. The architecture also changes what it costs to build and run a data center. According to Nvidia’s figures, as IEEE Spectrum reports, 800 VDC distribution:
- Enables 85% more power to be transmitted through the same conductor, compared with 415 VAC.
- Reduces copper requirements by 45%.
- 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 labor costs for component upkeep. Power equipment shrinks, and rack space is returned to compute.
This isn’t theoretical, but it isn’t finished either. Hardware is arriving now, and Nvidia expects full-scale production of 800 VDC data centers in 2027, alongside its Kyber rack systems. For facilities already built around AC, Nvidia’s MGX-compatible 800 VDC power rack arrives in the second half of 2026.
Major power vendors, including Delta, Eaton, ABB, Schneider Electric, and Vertiv, have announced 800 VDC power 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.
Decoupling AI growth from grid constraints
This transition is urgent because power and grid bottlenecks are delaying AI data center development. In Wood Mackenzie’s 2025 survey, large power transformers averaged 128 weeks for delivery and generator step-up units 144 weeks, and some orders now stretch as long as four years.
That makes data center efficiency a grid strategy. On a 1-GW campus, every percentage point recovered is roughly 10 MW, generation that doesn’t need to be built, permitted, interconnected, or brought online.
The 800 VDC architecture also addresses volatility, a problem utilities are already worried about. According to Nvidia, synchronized AI training can swing a rack’s power draw from roughly 30% to 100% in milliseconds, and across a data hall those swings can reach hundreds of megawatts in seconds. Going to 800 VDC makes it easier to place energy storage where it’s most useful, buffering those swings before they reach the grid. It also simplifies integration with batteries and solar arrays, which operate natively on DC.
Nvidia’s DSX Flex software is another lever. It connects AI factories to grid services so they can adjust power use dynamically and orchestrate demand with hybrid onsite generation, and Nvidia says it could unlock 100 GW of stranded grid power. For utilities managing volatile load profiles from AI campuses, a facility that can flex its own demand is a fundamentally different grid asset than one that draws constant baseload.
Generation is important, but efficient consumption is better
Efficient consumption is the fastest way to add capacity. New gas-fired capacity, nuclear restarts, SMRs, behind-the-meter generation, and long-duration storage are all necessary and are being pursued aggressively, but they’ll take years to build. Without an architectural change, we’ll keep losing roughly 17% of a typical data center’s power to conversion before it reaches the GPU.
The shift to 800 VDC won’t arrive overnight, but Nvidia expects full-scale production in 2027, four years before a gas turbine reserved from GE Vernova today would arrive.
Three groups stand to gain the most:
- Utilities that learn to work with DC-native, storage-buffered loads will be better positioned to win hyperscaler business.
- Power equipment manufacturers building DC-native busways, protection, converters, and advanced film capacitors are entering a data center physical infrastructure market that Dell’Oro now projects at $120 billion by 2030.
- Operators deploying 800 VDC can scale racks from 100 kW to over 1 MW on the same infrastructure and return rack space to compute.
Physics favors DC, as AI’s synchronized loads and megawatt racks strain legacy architectures. Economics rewards it with less energy lost and more compute per gigawatt. And the leaders are already moving, building DC-native infrastructure to support the next generation of AI.
Shaun Walsh is chief marketing officer (CMO) at Peak Nano.
Related Content
- Data Center Power a Prominent Highlight of OCP 2026
- The transition from 54-V to 800-V power in AI data centers
- Data center power meets rising energy demands amid AI boom
- Infineon’s OktoberTech: Edge AI, SDVs, and the Road to 800 VDC
- Data center next generation power supply solutions for improved efficiency
The post Are DC data centers the solution to AI scale? appeared first on EDN.


