The stronger the AI computing power, the harder the power supply: From 48 V to 800 V, how can SiC support power delivery for data centers and communications?

5. The dynamic demands of AI and communication power supplies must be validated in layers. Fast transients at the chip level are primarily handled by board-level regulation and decoupling; upstream power supplies experience port-level dynamics shaped by the load circuit and energy-storage components. Power supply dynamic testing should employ customer-defined port load waveforms; for communication projects, battery connection, parallel operation, and hot-swap scenarios must also be covered. High-voltage DC branches further require attention to pre-charging, discharging, circuit interruption, protection selectivity, and fault isolation. Long-duration backup power supplied by batteries and short-duration hold-up provided by DC-link capacitors must be separately defined and verified. IX. What R&D truly needs is a set of verifiable, deliverable results. Verification Item Key Observables Acceptance Criteria Double-pulse VDS/VGS peak values, Eon/Eoff, ringing, and false turn-on Specific device limits, design margins, and actual commutation conditions Full-range efficiency Losses under light, medium, full, and varying input loads Customer efficiency curve and test conditions Soft-switching range Whether commutation completes, residual VDS, reverse conduction time Input, output, load, and temperature ranges Continuous thermal operation Case temperature, driver board & magnetic component temperature rise, junction temperature estimation Thermal model, rated values, and project lifetime targets Dynamic load Voltage droop, overshoot, recovery time, and bus voltage variation Customer-defined port dynamic specifications Abnormal conditions & protection Startup, brownout recovery, short-circuit, interruption, and fault feedback Protection coordination and system-level requirements Parallel/Redundant operation Current sharing, single-module removal, and standby module engagement System continuous power supply requirements EMI & insulation Common-mode/differential-mode noise, insulation, leakage-related metrics Applicable product standards and customer certification requirements A beautifully smooth full-load efficiency curve cannot substitute for comprehensive operational boundary validation. Final designs must consistently yield acceptable results across all customer-specified operating conditions—input, load, temperature, and abnormal states. X. Hengde New Energy’s core value as a Tier-1 authorized distributor: Integrating component selection into the project workflow. AI data center and communication power supply customers face concurrent pressures: R&D schedule, product reliability, and procurement planning. BASiC Semiconductor’s SiC MOSFETs and SiC modules, together with Bronze Sword Technology’s domestically produced SiC module driver boards, only become evaluable solutions when mapped precisely onto specific power units and verification requirements. As a core Tier-1 authorized distributor for both manufacturers, Hengde New Energy Co., Ltd. serves as a centralized interface for customers to align component and driver requirements. During project execution, we recommend jointly specifying the following four deliverables—the scope of support to be confirmed during project discussions: • A component selection table: Clearly specify bus voltage range, branch current, switching frequency, operating temperature, package type, and candidate part numbers—ensuring BASiC Semiconductor SiC MOSFET comparisons are conducted under identical conditions. • A driver compatibility table: Specify candidate Bronze Sword driver boards, output gate voltage, source/sink capability, isolated power supply specs, interface definitions, and protection timing—avoiding mismatches between modules and drivers selected independently. • A verification checklist: Detail test conditions, observables, and acceptance criteria—from double-pulse tests through full-system dynamic validation—explicitly indicating which items have already been measured and which remain pending. • A supply confirmation table: Cross-check original manufacturer part numbers, product revisions, traceability documentation, sample & volume production requirements—and confirm actual inventory levels, quotations, and lead times.

As AI data centers shift toward 800 V DC power distribution, what improvements and challenges arise for energy efficiency and deployment?

When reading this table, note two points: First, it indicates ‘deployment readiness’, not ‘mass delivery’—a gap remains involving customer validation, production ramp-up, and on-site commissioning. Second, the three phases are not mutually exclusive; 48 V will persist in legacy applications, and the next three to five years will see coexistence of all three architectures. Engineering efforts are not starting from zero. In a deployment case made public in October 2026, an intelligent computing center operated by a provincial telecom carrier adopted 800 V DC power supply across its entire building. This raised power supply efficiency from 92%–94% to 98.5%, reduced copper usage by 45%–60%, halved footprint, and cut construction costs by ~30%. Such projects typically occur in newly built campuses—where DC bus architecture, protection configurations, and rack-level power supplies are implemented holistically from the outset. Retrofitting existing data halls often proves more costly than new construction, explaining why architectural transitions appear as a smooth line on paper but unfold building-by-building on site. VII. Six Requirements for Testing and Verification Power architecture changes demand corresponding updates to validation methodology. The following six items—ranked by importance—are directly tied to this transition: (1) Plan measurement ranges for 1500 V—not just today’s 800 V. Equipment lifespans typically span five to eight years, while this architectural iteration cycle is clearly faster than the last. Purchasing instruments rated only up to 800 V may render them obsolete for new projects within three years. (2) For DC measurements, check the first word in the bandwidth specification. Instruments labeled ‘DC, 0.5 Hz–’ treat DC as a valid operating point within their range—their DC accuracy, zero-point drift, and thermal drift are all specified in datasheets. Those labeled only ‘0.5 Hz–’ derive DC values via calculation; DC accuracy and long-term stability fall outside formal specifications. When measuring 800 V bus DC components or ripple RMS values, these two instrument types may yield results differing by orders of magnitude. (3) Common-mode voltage is a hidden bottleneck. At 800 V bus voltage, measurement channels endure high ground-referenced common-mode voltages. A differential probe’s common-mode rejection ratio (CMRR) and common-mode voltage rating determine whether accurate measurement is possible—not merely whether physical connection is feasible. Wiring methods, shielding, and grounding must be redesigned specifically for DC systems; applying AC-based empirical rules risks significant errors. (4) Dynamic response must capture the complete process. Load steps occur on millisecond timescales; power supplies must respond within milliseconds. Sampling rate determines whether you observe envelope trends or individual switching cycles. A 200 kS/s sampling rate suffices only for millisecond-scale trend envelopes; to resolve overshoot and ringing within individual switching cycles, sampling rates must increase by over one order of magnitude—and sufficient memory depth is required to store pre-step, step, and post-recovery segments simultaneously. (5) Efficiency measurements must be synchronized and time-aligned. Input and output signals must be sampled simultaneously; otherwise, calculated efficiency errors stem from timing misalignment—not instrument precision. Inter-channel time alignment error matters more than absolute accuracy in this context. (6) Ripple and harmonics must be analyzed separately. On the rectifier side: low-frequency harmonics dominate, with characteristic orders determined by rectifier pulse count. On the DC side: switching-frequency ripple prevails, spanning tens of kHz to MHz. These two frequency bands differ by three orders of magnitude. Few single instruments cover both high-order harmonics and high-frequency ripple adequately. Test plans must explicitly assign instruments per side—and specify the highest harmonic or ripple order each instrument covers.

Envision Deploys First 800V System to Deliver Green Power for AI

14 hours ago

Envision Group deployed the world's first commercial 800V DC system to power AI data centers using renewables, reducing infrastructure costs and copper usage while integrating energy storage and grid coordination.

bloomberg

2026-11-01

Envision Deploys First 800V System to Deliver Green Power for AI

Envision Deploys First 800V System to Deliver Green Power for AI

FAQ