As artificial intelligence scales to unprecedented heights, the global power industry faces a monumental demand-side growth event that threatens to outstrip traditional energy infrastructure. According to projections from major financial institutions and research firms, data centers are consuming vastly more electricity than ever before, pushing power grids to their absolute limits and forcing a fundamental rethinking of how electricity is delivered to compute infrastructure.

To solve this surging power demand, industry experts argue that the power sector must rely on a three-pronged approach. The first strategy involves optimizing existing resources, an effort already underway by specialized companies. The second strategy centers on migrating to direct current (DC) data centers, which yield significantly higher performance per megawatt. The third strategy requires developing new power generation and delivery infrastructure explicitly optimized for digital rather than analog loads. Industry leaders emphasize that all three approaches must work in conjunction to prevent widespread grid instability.

The sheer scale of the coming demand is reshaping long-term energy forecasts. Data centers consumed an estimated 448 terawatt-hours (TWh) of electricity globally in 2025, according to data from Gartner, and that figure is projected to more than double to 980 TWh by 2030. This trajectory highlights that AI’s power needs will eclipse anything the industry has witnessed before, a conclusion shared by every major industry forecast.

Goldman Sachs Research anticipates that global data center demand will grow by approximately 50% to reach 92 gigawatts (GW) by 2027. Domestically, the trajectory is even steeper, with U.S. demand projected to more than double from 31 GW in 2025 to 66 GW in 2027. Looking further ahead, Deloitte estimates that power demand from U.S. AI data centers alone will multiply more than thirtyfold by 2035, climbing from 4 GW in 2024 to an astonishing 123 GW. To support this expansion, McKinsey projects that its central scenario will require $5.2 trillion in global capital expenditures through 2030.

For the power industry, this represents the largest demand-side growth event in a generation. While expanding power generation has captured the lion’s share of public and corporate attention, industry analysts note that the fastest path to enabling AI infrastructure lies in converting to DC data centers and optimizing existing power delivery. According to the Energy Efficiency Movement, these improvements alone could yield up to 20% more usable power from existing feeds.

Conversely, building new generation facilities reliant on traditional alternating current (AC) architecture continues to introduce systemic inefficiencies. Power semiconductor supplier Infineon Technologies notes that traditional generation and distribution setups still lose roughly 17% of their energy strictly to power conversion processes before the electricity ever reaches the processors.

Compounding these efficiency losses, the traditional focus on building new generation is running squarely into physical and logistical limits. During a corporate earnings call, GE Vernova CEO Scott Strazik revealed that the company is already taking reservations for gas turbines scheduled for delivery as late as 2031, following expectations that reservations through 2030 would sell out completely.

Interconnection wait times, which were already frustratingly long prior to the AI boom, have only continued to expand. A report titled Queued Up from the Lawrence Berkeley National Laboratory found that more than 2,000 GW of generation and storage capacity was waiting in queues to connect to the grid at the end of 2025, with the average duration projects spend waiting still on the rise.

Are DC data centers the solution to AI scale?

Grid operators are similarly struggling to procure capacity quickly enough to meet the demand. PJM Interconnection, the regional transmission organization managing the power grid for Northern Virginia—the world’s largest data center market—has faced acute pressures. Virginia alone hosts roughly 28% of all operating U.S. data center capacity. Reflecting the strain, PJM missed its reserve margin targets in its two most recent capacity auctions and was forced to plan a 6.8 GW backstop procurement for the 2028-2029 delivery year to account for surging data center loads.

Regulatory friction has further complicated these efforts. The Federal Energy Regulatory Commission (FERC) approved only a portion of PJM’s backstop plan, instructing the grid operator to rework its cost allocation and collateral rules. FERC Chairman Laura Swett characterized the last-minute filing as a regulatory mess, and PJM has yet to establish a new timeline for approval.

Political constraints are also gaining momentum. Members of Congress have introduced the Artificial Intelligence Data Center Moratorium Act, a legislative proposal designed to pause the construction of new AI data centers until federal safeguards are formally enacted. According to tracking from Data Center Watch, at least 75 projects representing approximately $130 billion in investments were blocked or delayed in the first quarter of 2026 alone, a figure that roughly equals the total delays recorded across the entire previous year.

DC Power Unlocks More Compute Per Gigawatt

Nearly every data center operating today is built around traditional AC distribution. Utility power typically enters facilities as medium-voltage AC, where heavy transformers step it down, feed it into uninterruptible power supply (UPS) systems to provide battery-backed ride-through capabilities, distribute it across server racks, and finally convert it from AC to DC inside each individual server.

Every single one of these conversion steps generates heat, introduces additional points of failure, and wastes valuable energy. Infineon estimates that roughly 17% of an average data center’s total energy consumption is lost across these transitional steps alone.

The emergence of the 800-volt direct current (800 VDC) platform introduced by Nvidia fundamentally alters this power architecture, enabling significantly greater AI productivity without requiring proportional increases in power generation. Nvidia formally announced the architecture for 1-megawatt racks beginning in 2027, supported by its Vera Rubin DSX AI Factory reference design released to guide data centers through the hardware transition.

The primary advantage of transitioning to an 800 VDC system is the elimination of multiple power conversion steps, which directly drives efficiency gains. Rack-level AC-to-DC power supplies disappear from the hardware stack, removing the heat and potential failure points they traditionally introduce, while energy storage capabilities are built natively into the architecture.

At scale, the impact of these efficiency gains is transformative. In a 1-GW facility, losing 17% of incoming power to traditional conversion leaves only about 830 MW available for actual IT workloads. Infineon estimates that specialized power semiconductors central to 800 V designs can nearly halve those losses to approximately 9%, successfully putting 910 MW to work from the exact same utility feed. This recovery yields an additional 80 MW of usable power without requiring the construction of a single new generation facility. Nvidia’s internal estimates are similarly impactful, forecasting an end-to-end efficiency improvement of up to 5%, or roughly 50 MW in an equivalent facility.

Are DC data centers the solution to AI scale?

Software optimization serves as an additional lever for efficiency. Nvidia’s DSX Max-Q software, integrated into the Vera Rubin DSX platform, provisions power dynamically across the entire AI factory. According to the company, this software allows operators to deploy up to 30% more AI infrastructure within a fixed-power data center footprint.

Rewriting Data Center Economics

Recovered megawatts represent only part of the business case for adopting 800 VDC infrastructure. The architecture also reshapes the overall capital expenditures and operational costs associated with building and running massive data facilities. According to figures compiled by Nvidia and reported by IEEE Spectrum, 800 VDC distribution streamlines physical infrastructure requirements significantly.

Industry projections indicate that maintenance costs could drop by as much as 70%, driven by a reduction in power supply failures and lower labor requirements for component upkeep. As bulky power distribution equipment shrinks in physical footprint, valuable rack space is successfully returned to core computing tasks.

This transition is no longer merely theoretical, though full industry-wide adoption remains ongoing. Advanced hardware is arriving on the market, with Nvidia anticipating full-scale production of 800 VDC data centers by 2027, alongside the rollout of its Kyber rack systems. For existing facilities already built around legacy AC infrastructure, Nvidia’s MGX-compatible 800 VDC power rack was slated to arrive in the latter half of 2026.

Major power equipment vendors, including Delta, Eaton, ABB, Schneider Electric, and Vertiv, have announced their own 800 VDC power products and supporting architectures. A report from Omdia notes that higher-voltage DC data centers have already begun appearing in China, while the Mt. Diablo Initiative—a collaborative effort among Meta, Microsoft, and the Open Compute Project—is actively testing 400 VDC rack distribution models.

Decoupling AI Growth from Grid Constraints

The urgency behind this architectural shift stems from the reality that power and grid bottlenecks are actively delaying AI data center development. In a survey conducted by Wood Mackenzie, delivery lead times for large power transformers averaged 128 weeks, while generator step-up units stretched to 144 weeks, with some custom orders extending out to four years.

These prolonged timelines make data center efficiency an effective grid strategy. On a 1-GW campus, every percentage point of energy recovered represents roughly 10 MW of generation that does not need to be built, permitted, interconnected, or brought online.

Furthermore, the 800 VDC architecture directly addresses electrical volatility, a growing concern for utility operators. Nvidia points out that synchronized AI training workloads can cause a rack’s power draw to swing dramatically from roughly 30% to 100% in milliseconds, creating aggregate fluctuations across an entire data hall that reach hundreds of megawatts within seconds. Transitioning to 800 VDC makes it substantially easier to place energy storage directly where it is needed most, buffering these severe load swings before they ever reach the broader electrical grid. It also simplifies the physical integration of on-site batteries and solar arrays, which operate natively on direct current.

Are DC data centers the solution to AI scale?

Nvidia’s DSX Flex software provides another critical operational lever by connecting AI factories directly to grid services. This connectivity allows facilities to adjust their power consumption dynamically and orchestrate demand in coordination with hybrid on-site generation. According to Nvidia, this capability could effectively unlock up to 100 GW of stranded grid power. For utility providers managing increasingly volatile load profiles from major AI campuses, a facility capable of flexing its own demand represents a fundamentally different and more manageable grid asset than one drawing a constant baseload.

Generation is Important, But Efficient Consumption is Better

While expanding efficient consumption is the fastest way to add operational capacity, industry leaders emphasize that a diverse energy strategy remains essential. New gas-fired capacity, nuclear reactor restarts, small modular reactors (SMRs), behind-the-meter generation, and long-duration storage are all necessary components that are being pursued aggressively. However, these traditional and alternative generation projects will take years to build and commission. Without an immediate architectural change in how data centers consume power, facilities will continue losing roughly 17% of their electricity to conversion before it ever reaches a GPU.

The broader shift toward 800 VDC will take time to saturate the market, but Nvidia expects full-scale production by 2027—four years ahead of the delivery timeline for a gas turbine reserved today from GE Vernova.

Ultimately, physics strongly favors direct current as AI’s synchronized loads and high-density megawatt racks continue to strain legacy architectures. Economics rewards this transition with reduced energy waste and significantly more compute output per gigawatt. As industry leaders accelerate the deployment of DC-native infrastructure to support the next generation of artificial intelligence, the power sector is finding that smarter consumption is just as vital as new generation.

Shaun Walsh is chief marketing officer (CMO) at Peak Nano.

By Nana Wu