800V DC Link: Powering the Next Generation of AI Data Centers
- 2 days ago
- 3 min read
As artificial intelligence continues to scale at an unprecedented pace, one challenge is becoming impossible to ignore: power.

With high-performance GPUs such as the NVIDIA B300 driving rack densities to 100kW–300kW and full deployments exceeding 10MW traditional data center power infrastructure is reaching its limits. To meet these demands, a new architecture is emerging as the industry standard: the 800V DC Link.
At Amaryllo, where we focus on high-performance infrastructure and modular data center innovation, this shift represents more than just an upgrade, it’s a fundamental transformation in how AI data centers are designed and powered.
The Problem with Traditional Power Architecture
Most legacy data centers rely on an AC-based power flow:
Grid (AC) → UPS → PDU → Server → GPU
While widely adopted, this architecture introduces several inefficiencies:
Multiple AC/DC conversions lead to energy loss
Increased heat generation requires more cooling
Complex systems reduce scalability and flexibility
As AI workloads continue to grow, these inefficiencies translate directly into higher operational costs and physical limitations.
Introducing the 800V DC Link
The 800V DC Link simplifies power delivery by converting electricity once and distributing it as high-voltage DC:
Grid (AC) → AC/DC → 800V DC Bus → DC/DC → GPU
This streamlined approach is purpose-built for high-density AI environments.
Why 800V DC Is a Game Changer
Higher Efficiency
By reducing the number of power conversions, overall system efficiency can reach 96%–98%, significantly lowering energy waste.
Reduced Heat and Power Loss
Higher voltage means lower current for the same power output. This results in:
Less heat generation
Reduced cable losses
Improved system reliability
Simplified Infrastructure
The DC-based architecture:
Eliminates the need for traditional UPS systems
Enables direct integration with battery storage
Supports cleaner, modular system design
Enabling the Architecture: Advanced Power Engineering
The transition to 800V DC is made possible by advancements in power electronics and energy conversion systems. These technologies enable:
Stable AC to DC conversion at scale
Efficient DC bus distribution
Seamless integration with energy storage
Precision control for high-load environments
This is not just a component upgrade, it’s a complete rethinking of power infrastructure for AI.
How It Works in a Modular Data Center (MDC)
In a modern AI modular data center, the architecture is typically divided into two key components:
Power Container – The Energy Core
This unit is responsible for:
Converting incoming grid power into 800V DC
Managing energy storage systems
Distributing power through a DC bus
Compute Container – The AI Engine
This unit contains:
High-density GPU servers
Rack-level DC/DC converters
Direct liquid cooling (DLC) systems
Together, these components create a highly efficient, scalable, and modular infrastructure designed for AI workloads.
A Practical Example: 10MW AI Deployment
In a 10MW modular data center, the power flow typically follows:
Utility power enters at medium voltage (e.g., 22.8kV)
Voltage is stepped down to 415V
Converted into 800V DC
Distributed via a centralized DC bus
Delivered to compute containers
Converted to lower voltages for GPU operation
Each rack can support up to 300–500kW, making this architecture ideal for next-generation AI applications.
Why This Matters Now
AI is fundamentally changing infrastructure requirements. Power density is increasing faster than traditional systems can handle.
Legacy AC-based designs struggle with:
UPS inefficiencies
Complex distribution layers
Excessive heat generation
In contrast, 800V DC architecture offers:
Greater energy efficiency
Simplified deployment
Seamless scalability
Lower total cost of ownership
At Amaryllo, we see the 800V DC Link not just as an innovation, but as a necessary evolution for AI infrastructure. As modular data centers become the preferred approach for rapid deployment and scalability, integrating DC-based power systems will be critical to unlocking their full potential.
The future of AI data centers will not be built on legacy power systems. For organizations planning 10MW+ AI deployments, the decision is clear: Design for efficiency from the start—build with DC, not legacy AC.


