The growth of AI computing has significant implications on data center architecture. Not in the least on its power infrastructure. As it stands now, 800V DC will be the solution to power density challenges that current architectures simply cannot solve.
As artificial intelligence workloads push rack power densities from today’s 150 kilowatts toward 400kW and even 1 megawatt per rack, the traditional 48-volt power distribution model is reaching its physical limits. We recently had the chance to sit down with Rob Bunger, Global Director of Data Center Solution Architecture at Schneider Electric, to hear more about this. The current 48V architecture becomes impractical at these extreme densities.
The challenge is straightforward: a 150kW rack already requires eight power feeds. Scaling to 1MW using the same architecture would demand 32 large power feeds per rack, consuming so much space that little room remains for actual compute equipment. This reality is driving the industry toward 800-volt DC distribution as the path forward.
The limitations of current power architecture
Today’s AI racks, including Nvidia’s GB200 and GB300 designs, rely on AC power feeds entering the rack where power supplies convert 400V AC to 48V DC. This voltage has been a telecom standard for decades because it’s touch-safe and can deliver substantial power with the right number of supplies.
However, this approach breaks down mathematically at higher densities. Even exploring alternatives like 600V three-phase AC (used in Canada) or 480V (common in the US) provides only incremental improvements. The fundamental problem remains: delivering massive amounts of power through conventional AC distribution requires an untenable number of conductors and takes up too much physical space.
Bunger notes that while medium-voltage AC solutions exist, safety requirements prevent such high voltages from being located near IT equipment. This creates a gap between what facility power can deliver and what rack-level equipment can safely accept.
Why 800V DC specifically
The choice of 800V DC is not arbitrary. This voltage level already has established standards in the EV charging industry, providing a proven foundation for data center applications. The voltage is high enough to deliver substantial power through fewer, smaller conductors, yet there are existing power supply technologies that can handle the conversion.
Importantly, off-the-shelf power supplies already exist that can take 400V DC input (or plus/minus 400V, which equals 800V with a center point) and convert it to 48V. These converters are more compact than traditional AC units, offering better power density even before considering the reduced cabling requirements.
The industry is also developing 800V to 12V converters that can mount directly on servers, though these are currently quite large at about 1kW capacity. This progression suggests a future where 800V DC could potentially be distributed directly to server-level components, eliminating another conversion step.
Sidecar deployment model
The initial deployment strategy for 800V DC involves “sidecar” systems. That is the term used for narrower racks positioned next to IT equipment that house the AC-to-DC conversion and energy storage. These sidecars, capable of delivering 660kW to over 1MW, are designed to arrive with the IT equipment as a complete system.
This approach allows data center operators to adopt 800V technology without rewiring their entire facility for DC distribution. Operators simply land traditional 200-amp AC feeds into the sidecar, which handles all the conversion internally. From the data center’s perspective, 800V remains contained within the IT system rather than becoming a facility-wide electrical challenge.
A single sidecar can typically support one to three high-density racks, depending on the specific power requirements. For a 400kW IT rack, one sidecar provides sufficient capacity with room to spare. This modular approach also addresses practical deployment concerns like floor loading and door heights, as these systems can be quite substantial in size.
Centralized DC distribution evolution
While sidecars solve the immediate deployment challenge, they represent a transitional architecture. As data centers deploy larger numbers of 800V racks, for example 20 or 50 Vera Rubin or Blackwell systems at 800kW each, the economics shift toward centralized conversion.
In this more mature architecture, DC UPS systems located in the facility’s gray space or service hallways convert AC to 800V DC, then distribute power via busway or cabling to the racks. This approach is more economical at scale and simplifies the overall system, though it requires more substantial facility modifications.
The shift toward centralized DC also changes the ratio of data hall space to supporting infrastructure. As IT density increases, the supporting facility infrastructure, (things like transformers, switchgear and cooling systems), demands proportionally more space, even as the compute footprint shrinks.
Safety engineering and testing
Unlike 48V DC, which is touch-safe, 800V DC presents serious safety challenges that require extensive engineering. Schneider Electric operates test labs in both the US and France dedicated to fault testing, deliberately creating fault conditions to verify that breakers and other protective equipment perform correctly.
Proper breaker coordination becomes critical too. The system must ensure that the lowest-level breaker trips during a fault rather than taking down the main breaker to the entire facility. This “selectivity” is more complex with DC than AC, requiring careful analysis of braking curves and fault characteristics.
Schneider has also engineered live-swap capabilities for serviceable components in their DC UPS systems, using shutters and other safety mechanisms to allow operators to remove and replace power modules while the system remains energized. This maintains the hot-swappable paradigm that data center operators expect from other infrastructure like CDUs (coolant distribution units).
Software and modeling tools
Managing 800V DC systems requires sophisticated software tools since electrical phenomena remain invisible to operators. Schneider’s ETAP electrical network modeling software now includes DC modules that allow designers to model breaker behavior, test what-if scenarios, and verify that proposed changes maintain safety and reliability.
The company validates these models against real-world testing, ensuring that simulations accurately reflect actual component behavior under DC fault conditions. This digital twin approach allows operators to understand how their electrical system will respond to changes without testing in the live production environment.
Advanced metering and monitoring will also become more critical in DC systems. Schneider has upgraded firmware in metering systems to detect phenomena like subsynchronous oscillation that can occur in DC systems but remain completely invisible without proper instrumentation. This intelligence layer should provide the notifications and visibility operators need to manage what they cannot directly observe.
Deployment timeline and market adoption
Bunger expects pilot deployments of 800V DC systems to begin within the next year, initially in sidecar format. These pilots will likely involve bleeding-edge companies—neo-clouds or hyperscalers willing to test new technology before scaling broadly.
The technology is arriving faster than many expect, driven by Nvidia’s aggressive roadmap of releasing new chip architectures annually. The company has communicated clearly with the ecosystem about power requirements, allowing infrastructure providers like Schneider to prepare solutions in advance.
However, 800V DC will not become universal in the near term. Over the next three to five years, it will serve the highest-density AI workloads rather than replacing all data center power architecture. The technology addresses a specific segment of the market where current solutions simply cannot scale.
Open questions about the future
While the path to 800V DC appears clear for the next several years, longer-term uncertainty remains. The assumption that GPU-based AI will continue driving power density increases could shift if alternative compute architectures emerge. Technologies like photonics or specialized AI accelerators might offer similar performance at lower power levels.
There’s also discussion of more distributed computing architectures that could reduce the concentration of power at any single location, even if total compute capacity continues growing. Such shifts could change the economics of 800V DC deployment.
Bunger acknowledges these uncertainties while maintaining focus on the clear near-term need: “Right now the clarity that we have on the roadmap is there will be 800 volt loads coming out and it’s our responsibility that they can be powered safely and reliably and make it as easy as possible for operators to do that.”
Also read: Scaling at speed: How AI is rewriting the blueprint for the modern data center