Overview: SST Efficiency Crosses 98.7% at the OCP Global Summit

The OCP Global Summit in San Jose has become the clearest lens through which to watch how hyperscale operators are redistributing design authority across the data center supply chain. If Hot Chips is about chips and architectures, and NVIDIA GTC is about AI compute platforms, OCP's distinguishing mandate is integration — turning silicon, servers, racks, networking, power delivery, cooling and software into infrastructure that can be deployed at scale. Under the now-common refrain that AI ends in compute and compute ends in electricity, OCP's center of gravity has shifted over three years from open server hardware toward open AI infrastructure and system-level co-design for AI factories.

At this year's event, the NVIDIA MGX AI Factory Ecosystem zone carried the strongest signal: 800V DC power architecture has moved from proof-of-concept to engineering-ready. Among the seven companies selected for NVIDIA's SST ecosystem showcase, several Chinese firms appeared, and one — Zhizhan Technology, co-exhibiting with Luxshare Precision — pushed system efficiency to 98.7%.

Technical Details: Inside the NovaLink 35kV / 10MW Liquid-Cooled SST

Zhizhan Technology's NovaLink is a 35kV / 10MW solid-state transformer built from 24 standardized liquid-cooled power cells. At the cell level, each unit delivers 420 kW of output in a 216 L envelope, achieving 1.94 kW/L power density including 35 kV insulation, and 99.0% full-chain cell efficiency. The cells use Zhizhan's self-developed ZPAK SiC power modules, co-integrating high-frequency power conversion, medium-voltage insulation and liquid cooling. Built-in auxiliary supply, communications and control redundancy allow the system to keep running after a single power-cell fault.

At the system level, the numbers are equally aggressive: 98.7% overall efficiency, greater than 1 MW/㎡ power density, 10 MW total output in a 9.6 ㎡ footprint, native 800 V and 1500 V DC outputs, and compliance with the OCP SST V 0.3.0 specification. The modular liquid-cooled cell design is intended to simplify both maintenance and volume manufacturing.

From Concept to Engineering-Ready: NVIDIA's 800VDC V2.0 Whitepaper

Roughly two months before the summit, NVIDIA chief electrical engineer Mike Tu published the 800VDC Architecture for Next-Generation AI Infrastructure whitepaper — the V2.0 revision. Unlike V1.0, which argued the merits of high-voltage DC, V2.0 completes the leap from technical justification to standardization: it formally establishes 800 VDC as the unified power foundation for gigawatt-scale AI compute factories, lays out a phased build-out path, and classifies solid-state circuit breakers (SSCB) as a near-term requirement while locking solid-state transformers (SST) in as the long-term endgame. The OCP showcase made the practical consequence explicit.

Why Rack Power Density Is Breaking Legacy Topologies

Traditional data center power chains are long: medium-voltage AC steps down through line-frequency transformers to low-voltage AC, passes through UPS systems for repeated AC-DC-AC conversion, and finally reaches server power modules. Every conversion adds loss. Goldman Sachs research cited in the original report notes that as NVIDIA's DGX NVL8 draws 3.7 times the power of an A100, and as single-rack density climbs from 30 kW toward 1000 kW by 2030, the conventional grid-transformer-UPS-PDU-server chain becomes an energy liability — roughly 1 to 1.5 kWh wasted for every 10 kWh entering the facility. A conventional five-stage architecture delivers only 75%–80% full-chain efficiency, whereas an SST plus 800V HVDC direct-supply architecture can reach 85%–90%.

  • Efficiency: Zhizhan reports 98.7% system efficiency with native 800 V and 1500 V DC output.
  • Density: SiC-based production SSTs can shrink to one-quarter to one-sixth the volume and weight of conventional designs; a 10kV/1MW transformer needing 50 ㎡ can be replaced by 1–5 ㎡, a 90% reduction.
  • Materials: An 800 V DC architecture can cut copper busbar usage by 45%, distribution footprint by 80%, and free up 30% more usable cabinet space.
  • Speed: SST response moves from seconds to milliseconds, with bidirectional power flow, harmonic suppression and active voltage regulation.

Market Context: A Capital Wave and a Fast-Growing SST Race

China's CAICT projects global compute capacity exceeding 16 ZFlops by 2030, with a 2022–2030 CAGR above 40% and intelligent compute passing 30% of the total. JLL expects global data center capacity to roughly double from 103 GW to 200 GW by 2030. Dell'Oro Group forecasts global data center capex approaching the $1 trillion milestone in 2026 and reaching $1.7 trillion by 2030, while Goldman Sachs estimates cumulative capex around $7.6 trillion between 2026 and 2031. On the SST side specifically, Huatai Securities expects the global market to climb from RMB 2.82 billion in 2026 to RMB 101.04 billion in 2030; Industrial Securities is more optimistic, projecting a RMB 271 billion off-rack power market by 2030 with SST penetration around 40% of installed capacity.

Industry Impact: The Chinese Advantage in Medium-Voltage Engineering

The competitive field is already crowded. Delta partnered with Meituan and Qinhuai Data in November 2025 on an SST-based intelligent DC power system — described as the first publicly disclosed commercial deployment — with 1 MW per power cabinet, 1 ㎡ footprint and 98.5% efficiency, followed in March 2026 by a containerized SST DC mobile intelligent computing center. Sungrow's EnerNeo converts 10kV–13.8kV medium-voltage AC directly to 800 V DC at 98.5% efficiency with 50% land savings and 99.999% availability. Guosheng Securities argues that stable SST operation demands field experience, and that crossing from 10 kV to 35 kV is substantially harder. Chinese firms have the deepest track record in UHV DC transmission and flexible DC distribution grids, having already validated grid-forming control, medium-voltage rectification and DC protection on the utility side.

Implications for AI Data Centers: From Load to Grid Node

Because SSTs are inherently multi-port, photovoltaic and storage DC sources can attach directly to the DC bus without separate AC/DC stages. That reframes the AI data center from a passive electricity consumer into an integrated source-grid-load-storage node — precisely the interface capability the AI era needs most.

Future Outlook: The Decisive Battle Is Architectural, Not Hardware

TTIR analysis suggests the real determinant of market position is upstream: who enters the AI giants' architecture design systems first. Guojin Securities breaks the barriers into product specification participation, medium-voltage system engineering, and high-frequency power electronics productization. NVIDIA's roadmap points to 600 kW racks with Rubin Ultra in 2027 and beyond 1 MW with the Feynman platform in 2028, though reported delays to the Kyber rack introduce uncertainty. The pace of standardization will decide which SST vendors convert engineering readiness into installed base.