Application Use Feb 17, 2026
How to Test String Inverter with High-Power SAS Power Supply and Grid Simulator?
The string inverter is the core conversion unit that bridges PV strings and the utility grid. It converts...
How to Build MW Testing Power for AI Server?
The shift from LVDC to HVDC in AI servers brings complex new power testing challenges.
ActionPower's megawatt-class testing solutions are engineered to meet them。
Power Rating Scales Up as When Server Architecture Upgrades from LVDC to HVDC
HVDC AI Server Reaches 1MW and Here are the Whys and Challenges
The conventional data center relying on a 48V architecture - where power shelves and racks operate at 54 VDC to compensate for voltage drops - is not yet a legacy, but remains a mature option for standard server generations. Within this low-voltage architecture, GPU clusters are supported by power shelves equipped with hot-swappable power modules (handling either AC-DC or DC-DC conversion), with total shelf capacities typically ranging from 15kW to 33kW.
However, as LLM training and real-world workloads scale exponentially, this low-voltage architecture struggles to keep pace. The surging power demand results in an unrealistically high current draw. Delivering such immense power at 54V can generate thousands of amps per rack, thus the required size and weight of copper busbars and cabling are physically and economically impractical.
The advent of HVDC power architecture has thus emerged to resolve the contradiction between the surging server workloads that reach beyond 1MW per rack and the physical limitations of copper busbars and cabling. By converting AC to 800 VDC and routing it directly to the servers, this architecture reduces the number of power conversion stages.
Consequently, data centers avoid the unrealistic hardware costs associated with handling massive 1MW workloads at 48V, while significantly improving overall energy efficiency by minimizing power conversion losses.
ActionPower's MW-Class High Dynamic Regenerative Test Platform
Challenges for Power Testing Confronting the MW AI Server Workload
Testing the power delivery network for AI servers is not simply about power ratings or dynamics, but rather an interwoven, complex challenge
Programmable power testing for AI server has its focuses around validating whether the power distribution and subsystem can function stable, reliable and resilient against the AI server workload featuring shifting, volatile, unpredictable, and ultra-heavy current draw profile.
Programmable power, or to be more specifically the DC load simulator for such testing applications to replicate the server workload, sees inevitably the below requirements and functional features to achieve and comply with:
High-Dynamics and Fast Response
Replicating volatile power profiles and validating PDN resilience
Real-world AI server workloads exhibit unpredictably volatile power profiles with dramatic shifts in current draw. Consequently, for the electronic DC loads drawing current from the PSU, a high and fast slew rate, defined by di/dt, is a functional imperative. This is crucial for accurately replicating server workloads and validating the resilience of the Power Delivery Network (PDN).
High-Power Replication Capacity
Scaling programmable DC loads via master-slave architecture for MW testing
Even if building a physical 1MW AI rack solely for power testing is theoretically possible, replicating precise GPU workload functions remains a software challenge, and the hardware cost is prohibitive. Programmable DC loads, therefore, must be capable of being parallelled in a master-slave architecture to reach megawatt-level power, all while maintaining minimal accumulated latency and tight synchronization.
Thermal Management & Energy Efficiency
Utilizing regenerative loads to recycle energy and reduce thermal stress
Dissipating megawatt-level power as pure heat overwhelms laboratory cooling capacities and drives up electricity costs exponentially. Therefore, regenerative electronic loads - which recycle over 90% of the absorbed energy back to the grid - are now a strict necessity. To further manage extreme power densities, the testing instruments themselves may even increasingly migrate to liquid-cooled architectures, technically challenging for now though, and most probably in the near future.
Regenerative High-Dynamics DC Electronic Load for AI Server Testing
Introduction to ActionPower's PRL DC Load Simulator
Engineered for modern AI data centers, the PRL Series combines ultra-fast linear load dynamics (≤6μs response, up to 60 A/μs slew rate) with regenerative efficiency (recycling up to 87.1% of power to the grid). It is the ultimate testing ecosystem for both 50V/54V compute nodes and 800V/1000V HVDC architectures.
+
Cabinet Scalability
+
Inter-Cabinet Parallelation
Scale up from 5U 25kW Modular Chassis
A single 25kW, 5U PRL chassis serves as the foundational building block that delivers the core operating capabilities and technical features required to construct a megawatt-level AI workload simulator.
High-Power Cabinet Build-Up with 7 Units in Parallel
The PRL series features a highly scalable architecture. Up to seven 5U units can be paralleled within a modular cabinet system to deliver 175kW of power from a single platform.
8 Cabinets Master-Slaved for MW Power Scalability
Taking advantage of ActionPower's patented master-slave parallel technology, up to 8 fully loaded cabinets can be connected to build a massive 1.4MW AI server workload simulation platform.
Contact Us for MW AI Server Testing
Product Inquiry
Specification-matching testing and simulation solutions need discussion of both parties. Actionpower welcomes you inquiry to study together with us about how we can automate your testing processes.
Send InquiryTrade Shows & Events
These are the upcoming events and places you can find ACTIONPOWER and talk with us face-to-face!
Check Events