The AI data center power delivery architecture is undergoing its most fundamental transformation since the transition from lead-acid to lithium-ion UPS batteries a decade ago. On July 24, 2026, two parallel announcements — ON.Energy's strategic partnership with Crusoe Energy Systems to deploy 5 GW of UPS-grade battery storage for AI data centers, and Skeleton Technologies' collaboration with DG Matrix to develop 800VDC solid-state transformer (SST) power distribution — crystallized what industry insiders have been anticipating: the convergence of UPS, energy storage, and power conversion into a unified, medium-voltage DC power delivery architecture purpose-built for GPU cluster workloads. For engineers evaluating energy storage inverter compatibility, these developments define the technical specifications that next-generation power electronics must meet.
Overview of the Technology / News
ON.Energy, a US-based energy storage developer and EPC firm with over 400 MWh of deployed BESS projects, and Crusoe Energy Systems, the AI infrastructure company that has pioneered modular, low-carbon data center deployments, announced a partnership targeting 5 GW of UPS-integrated battery storage capacity for AI data centers across multiple US markets. The partnership's technical architecture is significant: rather than treating UPS and energy storage as separate subsystems (the conventional approach of deploying dedicated VRLA or Li-ion UPS batteries for sub-15-minute ride-through, separate from utility-scale BESS for energy shifting), ON.Energy and Crusoe are integrating these functions into a single battery energy storage platform capable of both millisecond-scale UPS protection and hour-scale energy shifting and grid services.
Simultaneously, Estonia-based Skeleton Technologies — known for its curved-graphene ultracapacitor technology that achieves power densities of 10-15 kW/kg — and DG Matrix, a Silicon Valley power electronics startup developing 800VDC solid-state transformer (SST) technology based on silicon carbide (SiC) semiconductors, announced a partnership to integrate Skeleton's ultracapacitor modules with DG Matrix's 800VDC SST platform. The combined architecture targets a medium-voltage DC (MVDC) power distribution topology operating at 800VDC nominal, eliminating the multiple AC-DC and DC-DC conversion stages that dominate current data center power architectures and collectively consume 6-10% of total data center energy as conversion losses.
These two announcements, while from different companies addressing different layers of the power delivery stack, converge on a single thesis: the AI data center of 2027 and beyond will be powered by medium-voltage DC distribution architectures that integrate UPS, energy storage, and power conversion into a unified platform — a departure from the AC-centric, functionally fragmented power delivery topology that has governed data center design for three decades.
Why This Development Matters
The transformation of AI data center power delivery matters because it addresses the single largest bottleneck to AI infrastructure scaling: power availability and power quality at the rack level.
GPU Load Fluctuation — The Hidden Power Quality Problem. Modern GPU clusters, particularly NVIDIA H100 and B200 systems organized in NVLink-connected pods of 8-32 GPUs, exhibit power draw fluctuations at two problematic timescales: (1) sub-millisecond current spikes during matrix multiply-accumulate (MMA) operations that can exceed the GPU's Thermal Design Power (TDP) by 30-50% for durations of 50-200 microseconds, creating voltage ripple on the rack-level DC bus that degrades GPU computational accuracy and accelerates silicon aging; and (2) second-to-minute-scale power ramps during AI training workload transitions — for example, when switching from forward propagation to backpropagation phases, cluster-level power draw can increase from 60% to 100% of rated capacity in under 2 seconds. Conventional UPS systems, designed for steady-state load protection with sub-second response times, cannot meaningfully address sub-millisecond GPU current spikes.
Energy Cost as AI Training's Dominant Variable Cost. A 100MW AI data center operating at 85% capacity factor consumes approximately 745 GWh annually. At an average industrial electricity rate of $0.07/kWh, annual energy costs exceed $52 million — often exceeding the amortized capital cost of the GPU hardware itself. The ability to time-shift energy consumption from peak to off-peak periods, using an integrated storage platform that serves double duty as UPS and energy arbitrage asset, can reduce effective energy costs by 15-25% — representing $8-13 million in annual savings for a 100MW facility.
Technical Deep Dive: 800VDC vs 380VDC vs 48VDC Architecture Comparison
Understanding the rationale behind the 800VDC architecture requires examining the voltage-level tradeoffs that have shaped data center power distribution design:
| Parameter | 48VDC (Legacy) | 380VDC (Current) | 800VDC (Next-Gen) |
| Typical Application | Telecom / Edge | Hyperscale DC | AI GPU Clusters |
| Distribution Losses | 8-12% | 4-6% | 1.5-3% |
| Cable Cross-Section | 120-185mm² | 50-70mm² | 16-25mm² |
| Max Rack Power | 15-20kW | 50-80kW | 150-200kW |
| Conversion Stages | 4-5 | 2-3 | 1-2 |
| Semiconductor Tech | Si MOSFET | Si IGBT/SiC | SiC MOSFET/GaN HEMT |
| Rack-Level DC-DC | Required | Optional | Eliminated |
The progression from 48VDC to 380VDC to 800VDC is not merely a voltage scaling exercise — it reflects a fundamental shift in the power conversion architecture enabled by advances in wide-bandgap semiconductors.
48VDC Architecture (Telecom Legacy). Originating from the telecommunications industry's -48VDC standard (established in the 1950s for lead-acid battery compatibility and personnel safety), 48VDC distribution suffers from I²R losses that scale inversely with voltage squared. A 20kW rack at 48VDC draws approximately 417A, requiring copper busbars with cross-sections exceeding 120mm² and producing distribution losses of 8-12% over typical 15-25 meter cable runs from the power distribution unit (PDU) to the rack. For AI GPU clusters drawing 80-200kW per rack, 48VDC is simply non-viable — the required conductor cross-sections would consume physical volume that exceeds the rack's available cable pathway space.
380VDC Architecture (Current Hyperscale Standard). Adopted by the Open Compute Project (OCP) and operational at scale in Facebook, Google, and Microsoft data centers, 380VDC (derived from the peak voltage of a 277V single-phase AC supply: 277V × √2 ≈ 392V, rounded to 380V nominal) reduces distribution losses to 4-6% at typical rack currents of 130-210A for 50-80kW racks. However, 380VDC still requires an intermediate rack-level DC-DC converter (typically 380V-to-48V) to power GPU voltage regulator modules (VRMs) that operate at 0.8-1.2V core voltages. This intermediate conversion stage adds 2-3% efficiency loss and approximately $8-12/kW of additional capital cost for the rack-level converters.
800VDC Architecture (Next-Generation AI Standard). DG Matrix's 800VDC SST platform eliminates the intermediate rack-level DC-DC conversion stage entirely. By distributing power at 800VDC — compatible with the DC-link voltage of a three-phase 480VAC rectifier (480V × √2 × √3 ≈ 1,176VDC, with 800VDC as a practical operating point for SiC devices rated at 1,200V) — the architecture enables direct connection to GPU VRMs through a single conversion stage. This topology is feasible only because of wide-bandgap semiconductors: SiC MOSFETs with 1,200V blocking voltage and sub-50mΩ on-resistance can switch at 50-100 kHz with switching losses 70-80% lower than equivalent silicon IGBTs, achieving converter efficiencies of 98.5-99.2% at the 800V-to-48V conversion stage.
Real-world Applications
The ON.Energy-Crusoe and Skeleton-DG Matrix architectures have immediate and transformative applications:
- Ultracapacitor-Battery Hybrid UPS Hierarchy. Skeleton's curved-graphene ultracapacitors, with power densities of 10-15 kW/kg and cycle life exceeding 1,000,000 cycles, handle sub-millisecond GPU current spikes (the "power quality" function), while the LiFePO4 battery array handles second-to-minute-scale power ramps during workload transitions and hour-scale energy shifting for peak shaving and grid services. This hierarchical architecture mirrors the analog and digital decoupling capacitors on a PCB, applied at the data center scale: the ultracapacitors are the "bypass capacitor" for GPU current transients, and the battery array is the "bulk storage capacitor" for sustained energy delivery.
- GPU Cluster-Aware Power Management. The integrated UPS-BESS platform, combined with Crusoe's AI infrastructure orchestration software, can implement GPU workload-aware power management: when training workloads enter their most power-intensive phases, the BESS pre-charges to full capacity; during lower-power inference or idle phases, the BESS dispatches stored energy to the grid for frequency regulation or capacity services, generating revenue. This "workload-aware dispatch" optimization can improve the net present value (NPV) of the integrated storage asset by 30-50% compared to standalone BESS operating on a fixed charge/discharge schedule.
- Grid Services Revenue Stacking. Because the integrated UPS-BESS platform maintains 85-95% state of charge during normal operation (reserving 5-15% for UPS ride-through), the platform can participate in multiple wholesale electricity market services simultaneously: frequency regulation (earning $15-25/MW-hour in PJM and ERCOT markets), spinning reserves ($5-10/MW-hour), and energy arbitrage (charging during low-price overnight periods, discharging during high-price afternoon peaks for $20-40/MWh spreads). For a 100MW / 400MWh platform, the combined revenue stack can exceed $8-12 million annually — sufficient to recover the incremental BESS capital cost within 4-6 years.
Industry Impact / Market Implications
Cannibalization of Standalone UPS Market. The ON.Energy-Crusoe integrated UPS-BESS platform directly challenges the $12 billion global data center UPS market, currently dominated by Vertiv, Eaton, and Schneider Electric. By combining UPS functionality with energy storage in a single, standardized battery platform, the integrated architecture reduces total cost of ownership (TCO) by an estimated 25-35% compared to deploying separate UPS and BESS systems. Established UPS manufacturers face a strategic dilemma: their revenue models depend on selling proprietary battery cabinets, power modules, and service contracts — a model that is fundamentally incompatible with the standardized, open-architecture battery platform approach enabled by LiFePO4 cell commoditization. For manufacturers of advanced battery management system BMS explained technology, the integrated data center power platform represents a new market segment with stringent requirements for multi-level hierarchical control spanning ultracapacitors, batteries, and grid services.
SiC vs GaN Semiconductor Battle in Power Conversion. The 800VDC architecture crystallizes the competitive dynamics of the wide-bandgap semiconductor market. SiC MOSFETs, with 1,200V and 1,700V voltage ratings and the ability to handle 100-200A continuous current in TO-247 and module packages, are the incumbent technology for 800VDC-to-48VDC conversion at the 10-50kW module level. GaN HEMTs, with superior switching speeds (dv/dt > 100 V/ns vs ~50 V/ns for SiC) and zero reverse recovery charge, offer higher efficiency at frequencies above 100 kHz but are currently limited to 650V rated devices — requiring a two-stage 800V-to-400V-to-48V topology. The GaN advantage in switching frequency translates to smaller magnetics and capacitors, reducing converter volume by 30-40% — a critical advantage in space-constrained rack-level installations. The competition between SiC's single-stage simplicity and GaN's higher-frequency density advantage will define the 800VDC converter market through 2030.
Future Outlook
The convergence of AI data center power delivery toward integrated, medium-voltage DC architectures will accelerate through 2027-2030, driven by three structural forces. First, the power density requirements of next-generation GPUs (NVIDIA Rubin platform, projected at 2.5-3.0 kW per GPU) will render the existing 380VDC infrastructure inadequate, forcing a migration to 800VDC or higher-voltage topologies. Second, the "power-as-a-service" business model — where data center operators contract for delivered power (kW and kWh) rather than owning power infrastructure — will favor integrated UPS-BESS platforms that maximize the revenue-generating potential of the storage asset through grid services participation. Third, utility grid interconnection queues — currently averaging 3-5 years for large-load connections in PJM, ERCOT, and CAISO territories — will make on-site energy storage essential for load management during the interconnection approval period, accelerating adoption of the integrated platform approach.
For the broader energy storage industry, the AI data center power delivery transformation represents both an opportunity and a competitive threat. The opportunity: AI data centers will become one of the largest demand drivers for utility-scale battery storage, with an estimated 50-80 GWh of integrated UPS-BESS capacity deployed globally by 2030. The threat: the consolidation of UPS and BESS functions into a single product category will blur the boundaries between historically distinct markets, enabling new entrants (including AI infrastructure companies themselves) to compete with established BESS integrators. For businesses evaluating home battery peak shaving savings strategies, the lessons from data center power architecture evolution — the move toward higher DC voltages, the integration of fast and slow energy storage, and the bundling of power quality and energy management into a single platform — will inevitably influence residential and C&I storage system design over the coming decade.
