Next-Generation Data Center Power Architectures
Arch3 Sidecar, Arch4 HVDC, BESS, SST, and the Future of BBU Strategy
Abstract
Data center electrical architecture is undergoing a structural transformation. The combination of artificial intelligence training workloads, hyperscale rack densities approaching one megawatt, persistent transmission constraints, and the inadequacy of traditional alternating current distribution at scale has driven the industry toward externalized direct current architectures. This white paper analyzes that transition through the lens of two emerging configurations referred to industrially as Arch3 and Arch4. Arch3 relocates the power supply unit and battery backup unit functions from the server chassis to a sidecar adjacent to the rack, performing alternating-to-direct conversion immediately upstream of the IT load. Arch4 performs bulk alternating-to-direct conversion at the facility level and distributes high-voltage direct current at 800 volts to the rack. The work assesses primary market drivers, expected adoption timelines, the evolving role of battery energy storage systems and solid-state transformers, and the future of battery backup units within these architectures. The analysis is framed for The First Call Group’s advisory engagements with hyperscale operators, infrastructure investors, suppliers, and regulatory authorities. Findings indicate that Arch3 will dominate retrofit activity through 2028 while Arch4 emerges as the prevailing greenfield pattern between 2027 and 2030. Battery energy storage will transition from peripheral demand-management asset to integrated facility platform; solid-state transformers will displace conventional iron-core units in the highest-density segment; and battery backup units will persist in modified form but will lose their per-rack ubiquity. The paper provides reference architectures, capital allocation models, supplier landscape analysis, governance frameworks, and forty-six executive-grade figures.
Keywords: data center power architecture, Arch3, Arch4, sidecar, 800 V HVDC, solid-state transformer, BESS, BBU, AI infrastructure, hyperscale, capital deployment, governance.
Executive Summary
The data center industry stands at the most consequential power architecture inflection in three decades. Per-rack thermal design power has compressed two orders of magnitude into a single hardware generation, advancing from a typical 27 kilowatts in 2024 to disclosed roadmap targets of 600 kilowatts in the NVIDIA Rubin generation slated for the second half of 2026, 1,000 kilowatts in Rubin Ultra in 2027, and 2,000 kilowatts in the Feynman generation thereafter. Distributed alternating current power distribution, which has served data centers since the commercial mainframe era, cannot scale to those densities efficiently, safely, or economically. The industry is responding with two architectural patterns that this paper terms Arch3 and Arch4.
Arch3 is a sidecar architecture. The power supply unit and battery backup unit, historically housed inside each server chassis, are externalized to a power module mounted immediately adjacent to the rack. The sidecar performs alternating-to-direct conversion at the row level and presents a regulated direct current bus to the rack. Existing alternating current distribution upstream of the sidecar can remain unchanged, which makes Arch3 the natural retrofit pattern for the installed base of two-megawatt to fifty-megawatt halls that operators have already capitalized. The architecture supports rack densities in the range of eighty to one hundred fifty kilowatts and pairs naturally with single-phase direct-to-chip liquid cooling.
Arch4 is a facility-level architecture. Bulk alternating-to-direct conversion is performed at or near the substation, and high-voltage direct current at 800 volts (or bipolar plus and minus 400 volts where touch-safe constraints govern) is distributed to the data hall. The rack receives a direct current feed and converts to silicon core voltage through a single point-of-load stage. Arch4 is engineered for densities at and above 250 kilowatts per rack and is the architecture that the announced NVIDIA, Open Compute, and hyperscaler roadmaps converge on. Arch4 will not appear in meaningful brownfield volume because it requires changes upstream of the data hall that are difficult to retrofit. It will dominate the greenfield AI campus segment.
These two patterns are not mutually exclusive. The most plausible 2026 through 2030 estate at any large operator will run a mixture of Arch1, Arch2, Arch3, and Arch4 simultaneously. The Arch3 sidecar is well suited to phased deployments inside existing halls and will accumulate share quickly through 2028, peaking near 45 percent of new builds before being progressively displaced by Arch4 in greenfield. Arch4 will cross over Arch3 in new-build share around 2029 and will be the majority pattern for greenfield deployments by 2030. Legacy Arch1 and Arch2 share will continue to decline gradually as operators retire facilities or refresh halls but will persist in colocation, enterprise, and regulated workload segments through the early 2030s.
Battery energy storage systems are the connective tissue of both architectures. The same physical asset can serve at least eight overlapping functions: peak shaving, ride-through, frequency response on islanded generation, load smoothing for synchronized GPU collectives, black start, renewables firming, voltage and reactive support, and runtime offset for diesel and gas generators. Operators who size battery energy storage for one of these functions reliably under-build for the others. The single most important governance decision a hyperscaler can make in front-end engineering design is to define the dispatch hierarchy for battery energy storage explicitly, codify it in the energy management system, and treat the asset as a multipurpose energy platform rather than a backup battery.
Solid-state transformers are the second-most consequential equipment shift. A solid-state transformer collapses the medium-voltage to low-voltage step-down, the unit power supply rectifier, and the bus regulator into a single power-electronic stage with active fault current control, multi-port output, and direct integration with battery storage. Solid-state transformers are not yet at the production maturity of conventional iron-core transformers; lead times in early 2026 cluster between 26 and 52 weeks, capital costs run between 1.5 and 3 times the equivalent iron-core unit, and design lives are nearer 12 to 20 years than the 30 to 40 years of legacy gear. Despite those constraints, solid-state transformers are the only equipment class that simultaneously eliminates the conversion stage count problem, controls fault energy in 800-volt direct current bus systems, and provides a clean integration point for battery storage. Hyperscale operators are placing pre-production orders today.
Battery backup units do not vanish under Arch3 and Arch4 but their role and placement change profoundly. In Arch3 the per-chassis battery backup unit is externalized to the sidecar; one battery bank serves a rack rather than dozens of distinct one-rack-unit batteries. In Arch4 the rack-level battery backup unit becomes optional, retained where direct chip-level ride-through is required by the silicon vendor or to maintain millisecond stability against direct current bus excursions. The center of gravity for stored energy in next-generation architectures is the multi-megawatt-hour facility battery energy storage system, not the per-rack battery. The industry will continue to ship battery backup units through 2030, but unit volume per IT megawatt deployed will decline by an estimated 40 to 60 percent.
These transitions intersect three external constraints that any operator and any investor must internalize. The first is the United States transmission system. New high-voltage transmission additions have collapsed from approximately 1,700 miles per year in the early 2010s to fewer than 350 miles per year in 2020 through 2023. Generator interconnection queues exceed five years across most independent system operators and rate-regulated territories. Greenfield AI campuses cannot wait for the grid to catch up. They will be built behind on-site generation, behind battery storage, or in regions where flexibility-based interconnection programs such as the Federal Energy Regulatory Commission’s RM26-4 proposal and the Electric Reliability Council of Texas’s NPRR1325 and PGRR145 grant priority access to controllable loads. The second constraint is the equipment supply chain. Lead times for the largest medium-voltage transformers exceed 110 weeks; gas-insulated switchgear runs 85 weeks; even modular uninterruptible power supply units run 45 weeks. The third constraint is the workforce. Skilled labor shortages cluster precisely in the disciplines that Arch3 and Arch4 require most: solid-state power electronics, direct current protection coordination, battery energy storage commissioning, and operational technology cybersecurity.
For The First Call Group’s clients, this paper offers four sets of recommendations. Operators and hyperscalers should lock the architecture decision in front-end engineering design rather than deferring it to detailed design or construction; treat battery energy storage as a multi-purpose platform; pre-qualify at least two solid-state transformer suppliers per region; adopt skidded medium-voltage and low-voltage power blocks for modularity; and standardize commissioning at level four (integrated systems) as a minimum. Investors and owners should reward Arch4 readiness in capital allocation gates; treat the interconnection queue as an internal-rate-of-return variable rather than a project assumption; require a battery energy storage dispatch strategy in due diligence; insist on a living standards register before close; and price decommissioning into the base case capital model. Suppliers and original equipment manufacturers should invest in 800-volt direct current product lines now; publish solid-state transformer topology and Underwriters Laboratories 9540 evidence; codify field-service playbooks; develop firmware update governance; and partner on open standards, particularly through the Open Compute Project and International Electrotechnical Commission technical committees. Regulators and authorities having jurisdiction should codify 800-volt direct current National Electrical Code permitting paths; streamline UL 9540 acceptance procedures; recognize flexible-load priority through Federal Energy Regulatory Commission rulemaking; update arc-flash labeling for direct current systems; and publish liquid-coolant safety guidance reflecting two-phase and immersion methods.
The conclusion is that the next decade of data center power infrastructure will be defined less by incremental efficiency improvement than by structural reconfiguration. The operators, investors, suppliers, and regulators who treat that reconfiguration as an architecture decision rather than an equipment purchase will compound advantage; those who treat it as a procurement exercise will absorb the cost of rework. The First Call Group is publishing this analysis to equip our clients with the architectural, governance, and capital frameworks required to navigate that transition with discipline.
Full White Paper Below
