Every time electricity is converted from alternating current (AC) on the local grid to the direct current (DC) required by hardware, power is lost. In today's high-density AI data centers, that conversion happens multiple times between the facility walls and the GPU. Those losses add up quickly.
According to a report from the International Energy Agency (IEA), data center electricity demand surged 17% in 2025, with AI-focused facilities growing even faster. The report also estimates that power consumption from AI data centers will triple by 2030. Most data center power infrastructure was built for traditional workloads, not for the scale AI demands. The gap between what decades-old power systems were designed for and what they are now being asked to do is where efficiency, cost and sustainability challenges collide.
This article examines the challenges of AC to DC power conversion in AI data centers, especially for increasingly large workloads, and offers strategies to improve efficiency and sustainability.
Why the traditional power path strains under AI workloads
Data centers convert power multiple times before it reaches a single GPU. They take AC power from the grid and convert it to DC for battery backups. They then convert it back to AC for IT racks and, subsequently, to DC again to power individual servers and other hardware.
That process worked well when racks drew 10 kW each. However, it no longer works as those same racks now approach 1 MW, according to IEEE Spectrum. At AI scale, each conversion step is a significant source of energy loss, cost and heat that operators can no longer manage.
Sustainability is becoming a bigger issue
Reducing unnecessary conversion stages is one of the most direct levers operators have.
The sustainability stakes are just as significant. The IEA projects that global data center electricity consumption will reach 945 TWh by 2030, accounting for nearly 3% of global electricity use. More operators are signing corporate power purchase agreements (PPAs) for renewable energy, but those agreements don't fix conversion losses. PPAs make power consumption more palatable to investors and regulators without addressing the underlying inefficiency.
Reducing unnecessary conversion stages is one of the most direct levers operators have. Every watt recovered from conversion losses is a watt that doesn't need to be generated, transmitted or cooled.
The power infrastructure problem runs deeper
Conversion inefficiencies are only part of the challenge. The grid infrastructure feeding most data centers was built decades ago -- before the internet, let alone AI. Most of the U.S. power grid is roughly 30 years old and nearing the end of its life. Power generation has modernized with renewables and small modular reactors, but the transmission grid that connects generation to the facility is decades old.
New grid connections and additional generation capacity take years to come online. Meanwhile, rising AI workload demands are straining internal infrastructure and cooling systems, further pressuring already fragile grids.
Power architecture is gaining the most traction
Nvidia appears to be leading the charge in positioning 800 VDC power distribution as the optimal architecture for AI data centers. According to Nvidia, an 800 VDC facility would minimize energy losses by reducing the total number of conversion steps required across a network. It also enables more compact hardware placement and reduces copper and cabling use compared with current facility-level 480 VAC systems. The appeal of this approach is that it converts AC power to DC once and then distributes DC power directly to computing racks throughout the facility.
Schneider Electric offers another option for those not ready to commit to a full facility redesign. They suggest that rack-level 800 VDC systems provide an immediately viable path for any AI data center. This approach requires minimal architectural changes to existing AC-based infrastructure and lowers the cost of entry, making it a practical interim option.
New semiconductors offer another transition path
Silicon carbide (SiC)-based semiconductors are a key technology enabling the shift to higher-voltage DC distribution. SiC devices help address the safety, thermal management and standardization challenges that have historically slowed AC to DC transitions in data centers.
Compared with traditional silicon-based devices, SiC devices switch faster and operate at higher temperatures. This translates into smaller, more efficient power conversion stages, higher switching frequencies, and lower conduction and switching losses. The result is more compact power modules that support greater rack density without compromising thermal management.
Power standardization is still a challenge
DC power at the rack is a proven engineering approach, but standards supporting its use at industry scale are not in place. As a result, widespread DC deployment is not yet a routine model for most operators.
The Open Compute Project is working to close that gap. In its 2025 white paper, the organization called for broader industry participation in DC standardization efforts to ensure consistent adoption across operators of all sizes.
What operators can do now
The path forward doesn't require a complete facility overhaul. Data center operators can take a measured, phased approach:
Audit current power conversion losses. Map the full AC to DC conversion path in your facility and calculate efficiency losses at each stage. This will identify where the greatest gains are available.
Identify pilot candidates. Pinpoint the highest-density AI clusters in your facility as candidates for 800 VDC pilots before committing to a broader rollout.
Engage with standards bodies. Stay current on industry standardization efforts so your infrastructure plans align with where requirements are headed.
Evaluate SiC-based power hardware. Assess whether SiC-based conversion equipment meets your density and thermal requirements as you refresh or expand your power infrastructure.
Many operators are already shifting from passive energy consumers to active grid stakeholders. They're investing in infrastructure upgrades, exploring conversion-optimizing technologies, and integrating renewable energy and storage to support their facilities.
Power conversion is no longer a back-of-house concern. How operators deliver electricity from the grid to the GPU directly affects efficiency, cost, sustainability and competitive position. Operators who address it now will be better positioned as AI workloads and their power demands continue to scale.
Julia Borgini is a freelance technical copywriter, content marketer, content strategist and geek. She writes about B2B tech, SaaS, DevOps, the cloud and other tech topics.
