Overview

Sanan Optoelectronics has disclosed a combined materials-and-topology effort aimed at the efficiency ceiling of power conversion in AI data centers. According to a report published on Xueqiu on November 2, 2026, the company's subsidiary Hunan Sanan spent three years developing a low-resistance silicon carbide (SiC) substrate that holds average resistivity at 11 mΩ·cm — roughly half the 20 mΩ·cm figure typical of conventional volume production — while simultaneously meeting production-grade benchmarks for stacking faults, dislocations, micropipes and surface flatness. The company pairs that substrate work with a soft-switching converter topology it calls S²I-ARCP, positioning the two together as a route to higher efficiency in 800V-class AI data center power architectures.

The framing matters because server power supplies have become a first-order determinant of data center power usage effectiveness (PUE) and operating cost as AI compute infrastructure scales. Mainstream supplies are now expected to meet the 80 PLUS Titanium threshold of at least 96% efficiency, a target that conventional silicon devices are approaching only with difficulty.

The Substrate Problem: Low Resistivity Versus Crystal Quality

The industry difficulty Sanan describes is a trade-off rather than a single metric. Lowering substrate resistivity reduces the series contribution to a device's on-resistance, but aggressive doping in SiC has historically come at the cost of crystalline quality — stacking faults, dislocations, micropipes and surface flatness all tend to degrade. Sanan claims to have broken that coupling: 11 mΩ·cm average resistivity alongside metrics it says are benchmarked against volume production standards. The company characterizes the result as resolving the "low resistance and high quality are mutually exclusive" problem that has constrained the category.

The sensitivity is meaningful at the system level. According to the company, every 1 mΩ·cm reduction in substrate resistivity translates into a 0.2%–0.6% reduction in device on-resistance. That is a modest-sounding lever in isolation, but in a server power supply it propagates into lower heat generation, smaller heatsinks and the ability to run at higher switching frequencies — which in turn shrinks the transformers and capacitors that occupy much of a PSU's volume. The cumulative effect pushes power modules toward both higher efficiency and smaller form factors.

S²I-ARCP: Making Soft Switching Practical

SiC's fast switching is also its liability. High dv/dt amplifies EMI and insulation constraints, so capturing the material's benefit in AI data center applications requires co-design across device, topology and gate drive rather than a drop-in replacement. Sanan's answer is S²I-ARCP — a single shared inductor auxiliary resonant commutated pole — in which one shared auxiliary inductor serves all three phases, reducing the magnetic component count relative to per-phase implementations.

The company reports that an 800V/15kW prototype built on its SiC MOSFETs achieved wide-load efficiency above 99.6% and can be cooled by natural convection. Under matched conditions at 30 kHz and 10 kW, device temperature was reduced by roughly 50°C, with dv/dt controlled and set at 6.7 kV/µs. Sanan also says the work covers soft-switching optimization of planar SiC MOSFET parameters including Rg,int, Ciss and Coss, and clarifies how the Coss×Rdson figure of merit should be traded off when selecting devices for high-frequency LLC/CLLC resonant DC/DC stages in server power, automotive on-board chargers and telecom power.

Device Portfolio and Packaging

Sanan has assembled a SiC MOSFET matrix aligned to the soft-switching topologies dominant in server supplies. For AI data center solutions the company lists coverage spanning 2300V at 35 mΩ; 1200V at 14, 20, 26, 32 and 40 mΩ; and 750V at 9, 14, 18, 27, 35 and 50 mΩ. Available packages include TOLL, TOLT, X2PAK and QDPAK. The portfolio is positioned at PSU LLC/CLLC resonant DC/DC stages as well as solid-state transformer and HVDC soft-switching scenarios, with the company emphasizing a combined advantage in low on-resistance and the Coss×Rdson figure of merit.

Market Context and Customer Traction

Sanan states that its AI data center products are already shipping in volume to leading server power and data center infrastructure companies in China and abroad, and that it has signed long-term agreements with multiple customers. Through the supply chains of those server power and infrastructure partners, the company says its devices have reached AI data center deployments at end customers including NVIDIA, Google, Microsoft and Amazon — a claim that, if sustained, would place Sanan across most major global AI compute and cloud infrastructure platforms.

Implications for AI Data Centers

The significance lies less in any single number than in where the gains compound. A 96%-plus Titanium requirement leaves little headroom for conventional silicon, and the residual losses in a PSU are distributed across conduction, switching and magnetics. Reducing substrate resistivity attacks the conduction term at its source; soft switching with controlled dv/dt attacks the switching term while relaxing EMI filtering; and higher permitted switching frequency shrinks the magnetic and capacitive content. Together these move efficiency, thermal density and volume in the same favorable direction — the combination that matters when power shelves are packed into racks alongside accelerators.

Future Outlook

Sanan frames the strategy as a shift "from faster SiC to SiC that understands the system," betting that wide-bandgap materials leadership will increasingly be judged on system-level outcomes rather than on raw device speed. Several questions remain open from the published account: the report does not disclose substrate volume capacity, yield, long-term reliability data, or the specific efficiency figures of shipping PSUs built on the platform. Nor does it quantify the cost premium of the 11 mΩ·cm substrate relative to the 20 mΩ·cm baseline. What the disclosure does establish is a coherent technical thesis — that substrate resistivity, device figure of merit and resonant topology must be optimized jointly — and a product matrix already aimed at the 800V power architectures that AI data centers are converging on.