Development of Arch3 and Arch4
Over the past decade, data‑center power architecture has undergone a quiet but profound transformation. As compute density accelerated, first through cloud virtualization and later through AI GPU clusters, the traditional model of distributing AC power into the data hall space and relying on thousands of small, server‑integrated power supplies began to show its age. Operators faced rising inefficiencies, thermal constraints, and operational complexity, all of which signaled that the legacy rack‑level power path was no longer aligned with the demands of modern workloads. This pressure set the stage for a new generation of power topologies designed to streamline conversion, reduce losses, and support far higher rack densities.
Arch3 emerged as the first major response to these challenges. Rather than attempting a full architectural overhaul, Arch3 rethought the placement and consolidation of power components at the rack level. By moving PSUs and BBUs out of individual servers and into a shared sidecar, Arch3 centralized AC‑to‑DC conversion and battery backup while preserving compatibility with existing AC distribution systems. It represented a pragmatic evolution: an architecture that improved efficiency and serviceability without requiring operators to rebuild their electrical rooms or abandon familiar operational models.
As AI workloads began to dominate data‑center design, however, even Arch3’s improvements proved insufficient. The rise of 80‑ to 200‑kilowatt racks and the expectation of even higher densities exposed the limits of low‑voltage DC distribution and the inefficiencies of maintaining AC throughout the white space. These pressures catalyzed the development of Arch4, a topology that reimagines the entire power path by performing bulk AC‑to‑DC conversion at the facility level and distributing high‑voltage DC directly to rows or racks. Arch4 is not an incremental improvement but a foundational shift, aligning the electrical architecture with the physics and scale of next‑generation compute.
Arch3
Arch3 represents a transitional stage in the evolution of data‑center power architecture—one that modernizes the rack‑level power path without requiring a full reengineering of the facility’s electrical backbone. In this model, components traditionally embedded inside each server, such as power supply units (PSUs) and battery backup units (BBUs), are relocated into a dedicated “sidecar” positioned adjacent to the rack. By centralizing these elements, Arch3 eliminates the inefficiencies inherent in thousands of small, independent PSUs and replaces them with a more efficient, shared conversion system. The result is a cleaner, more predictable power path that reduces conversion losses and frees up valuable space inside the rack for compute hardware.
This externalization of power components also brings significant operational advantages. Maintenance becomes far simpler when technicians can service batteries or power modules from the sidecar rather than opening individual servers. Thermal management improves as well, since heat‑generating power electronics are removed from the compute airflow path. These changes collectively support higher rack densities and more consistent thermal profiles, which are increasingly important as AI and GPU‑heavy workloads push traditional rack designs to their limits.
Despite these improvements, Arch3 remains fundamentally tied to legacy AC distribution within the white space. Power still enters the room as AC, and although it is converted to DC before entering the rack, the distribution voltage remains low—typically 48 to 54 volts. This low‑voltage DC architecture limits how far Arch3 can scale, because higher currents translate into larger conductors, greater copper losses, and more heat. For many operators, however, Arch3 offers an attractive balance: it delivers meaningful efficiency gains and operational improvements while remaining compatible with existing AC‑based facilities.
In practice, Arch3 is best understood as an evolutionary rather than revolutionary step. It modernizes the rack‑level power environment, supports denser compute, and reduces operational overhead, but it does not fundamentally change the electrical architecture of the data center. For operators seeking incremental improvements without the disruption of a full redesign, Arch3 provides a practical and achievable path forward.
Arch4
Arch4 marks a decisive shift toward a fully modernized, high‑efficiency power architecture built around facility‑level AC‑to‑DC conversion and high‑voltage DC (HVDC) distribution. Instead of delivering AC into the white space and relying on localized conversion near the rack, Arch4 performs bulk conversion at the facility’s front end and distributes power as HVDC—typically 380, 600, or even 800 volts—throughout the data hall. This approach dramatically reduces the number of conversion stages in the power path, improving overall electrical efficiency and reducing the thermal and mechanical burden on downstream equipment.
The use of HVDC fundamentally changes the economics and physics of power distribution. Higher voltage means lower current for the same power level, which in turn reduces conductor size, minimizes resistive losses, and lowers heat generation. These characteristics make Arch4 uniquely suited to the extreme rack‑level power densities emerging in AI and HPC environments, where individual racks may draw 100 to 300 kilowatts today and are projected to exceed 500 kilowatts within the decade. Traditional low‑voltage architectures simply cannot move this much current efficiently, whereas HVDC distribution handles it with far greater ease.
Another defining advantage of Arch4 is its natural compatibility with DC‑native power sources. Technologies such as solid‑oxide fuel cells (SOFCs), large‑scale battery systems, photovoltaic arrays, and solid‑state transformers all produce or operate on DC. In an Arch3 or legacy AC environment, these sources require multiple conversion stages—DC to AC, then AC back to DC—each introducing losses and complexity. Arch4 eliminates much of this overhead by allowing these sources to feed directly into the HVDC bus, creating a more integrated, efficient, and resilient power ecosystem.
Arch4 is not a retrofit solution; it requires a purpose‑built facility and new safety, protection, and operational practices. HVDC distribution introduces different arc‑flash characteristics, new grounding strategies, and specialized breakers and connectors. For operators willing to embrace this architectural shift, the payoff is substantial: Arch4 delivers the highest efficiency, the cleanest power path, and the most scalable foundation for next‑generation compute. As AI workloads continue to reshape data‑center design, Arch4 is rapidly emerging as the architecture most closely aligned with the power and density demands for that need.