Overview: 2026 and the 800V Inflection

2026 is emerging as the year when data-center power architectures definitively turn toward high-voltage distribution. According to a Chinese industry analysis from EET China, NVIDIA, Google, Microsoft, Meta and Chinese hyperscalers including Meituan, Alibaba, ByteDance and Tencent are all pushing 800V DC power into deployment. The shift is not merely a bus-voltage change; it inserts a new high-voltage intermediate bus converter stage between the 800V DC bus and the traditional board-level conversion chain, forcing a re-examination of device technology, topology and packaging.

In conventional 48V architecture data centers, board-level power delivery typically converts 50Vdc to 12Vdc and then 12Vdc to 1Vdc through the intermediate bus converter and point-of-load stages. The 800V architecture adds an HV IBC stage that steps 800V down to 50Vdc. That added stage must handle higher input voltage, higher power density and higher conversion efficiency while meeting stricter reliability requirements. The result is a growing, durable demand opportunity for silicon carbide and gallium nitride devices, but also a more demanding set of engineering constraints.

Why HV IBC Is a New Device Battlefield

The HV IBC is not simply a higher-voltage version of an existing converter. Its primary side directly withstands the 800V bus, and common HV IBC topologies, generally open-loop LLC converters, require devices with low output capacitance, low gate charge and reverse-recovery charge approaching zero. These parameters directly affect switching loss, dead time, resonant behavior and the ability to push switching frequency higher without collapsing efficiency.

At the same time, the converter must fit within tight board space and deliver high power in a full-brick or similarly compact form factor. Thermal density, magnetic-component loss and high-current interconnection all become first-order design problems. The device choice therefore cannot be separated from packaging, thermal resistance, circuit layout and the maturity of the surrounding ecosystem.

SiC vs GaN: Trade-offs in the 800V Front End

The debate over SiC versus GaN is often framed as a simple competition, but the original analysis shows a more nuanced picture. GaN offers low on-resistance, low output capacitance, low gate charge and near-zero reverse-recovery charge, making it attractive for high-frequency LLC topologies. Its smaller die and lower losses can reduce PCB area, improve conversion efficiency and lower heat generation, which in turn reduces cooling volume. Those advantages align with the direction of AI data-center power design: higher frequency, smaller magnetics and higher power density.

SiC brings different strengths. It has high voltage tolerance and more robust avalanche and short-circuit withstand capability, and its long-term operating reliability has been extensively validated. Because the HV IBC primary side sees the full 800V bus, SiC remains an important candidate, especially where ruggedness and proven reliability are weighted heavily. Industry attempts to use SiC MOSFETs in 800V or +/-400V HV IBC designs have run into efficiency degradation as frequency rises, with power density and temperature rise failing to meet requirements. A SiC JFET cascode approach can achieve high-frequency LLC operation with efficiency close to GaN, but its ecosystem is less mature than GaN’s.

GaN’s Commercial Momentum

Among the companies interviewed, GaN appears to have gained the early commercial lead in HV IBC. Innoscience estimates that a 12kW HV IBC module uses eight 650V GaN devices and 32 100V GaN devices, implying a significant demand increase for low-voltage 100V GaN. Multiple GaN device suppliers have formed partnerships with end customers to develop HV IBC products operating above 1MHz and above 10kW per module. Innoscience says mainstream customers have adopted GaN HV IBC solutions and entered small-batch production, with mass production expected in the coming months.

AOS has launched HV IBC reference designs using both GaN and SiC, reflecting a multi-technology product strategy spanning silicon, SiC and GaN. For GaN-based 800V-to-50V designs, AOS uses top-side cooling packaging to address heat flux under extreme miniaturization, and develops practical reference designs to mitigate magnetic loss, noise and driver ringing at high frequency. Hongwei Aisi has introduced a 650V, 30mohm GaN product for HV IBC requirements and is collaborating with domestic universities on GaN HEMT-based three-stage AIDC power technology, covering device selection, prototype trial production and performance verification.

The Voltage-Route Debate: 50V, 12V and 6V

The industry is also debating which intermediate bus voltage will dominate: 800V-to-50V, 800V-to-12V or 800V-to-6V/5V. The companies largely agree that these routes are not mutually exclusive but correspond to different stages of GPU power evolution. For current power levels, 800V-to-50V has strong reuse of existing 48V infrastructure and is the mainstream choice. Hongwei Aisi expects that when single-card power reaches 2000W to 3000W, 800V-to-12V will gradually gain broader use. The 800V-to-6V/6.75V route offers the best full-link transmission loss and matches next-generation ultra-high-power GPUs, making it a long-term target.

AOS describes 800V-to-50V as the core architecture being pushed by AIDC industry leaders because it inherits the mature 48V ecosystem and balances conversion efficiency with development cycle. It sees 800V-to-6V as a future vertical-power-delivery topology that could break the high-current design bottleneck, but notes that it is technically very difficult and requires broader industry-chain investment. Innoscience favors 800V-to-50V for its ecosystem compatibility and low deployment risk, while acknowledging that 800V-to-12V and 800V-to-5V could eliminate a 48V-to-12V or 48V-to-5V conversion stage. The challenge is that feeding 1000A to 2000A from an off-board power-distribution board to the motherboard creates loss and cost pressure, while placing the IBC on the motherboard introduces 800V safety and power-density problems. Gallium Future adds that a 48V bus is easier to route on the PCB, carries lower safety risk and supports high transmission power with low heat, making 50V the widest-coverage, lowest-risk baseline during 800V penetration.

Engineering Bottlenecks: Thermal and Magnetic Losses

With the route debate narrowing, the real test is engineering execution. Gallium Future says the main HV IBC difficulty is thermal management: AIDC designs must deliver 10kW or more within a full-brick volume while ensuring device heat dissipation, a significant challenge for GaN. AOS points to a system-level barrier: as AIDC power systems move to higher frequency, magnetic components suffer high-frequency hysteresis loss, eddy-current loss and AC resistance from skin and proximity effects. These can offset the low-loss advantages of wide-bandgap semiconductors and widen the gap between ideal and actual system efficiency.

GaN’s advantages therefore must clear two thresholds: device-level thermal and packaging limits, and system-level magnetic and layout limits. Gallium Future has introduced an AC-to-48V intermediate bus solution rated at 12kW per module, with 49V/245A full-load output, 97.5% AC-DC conversion efficiency, 180% peak-load capability and hold-up time compliance. At the device level, it is reducing on-resistance per unit area through epitaxy and structure iteration, lowering thermal resistance through packaging, and running voltage and temperature accelerated aging tests beyond JEDEC standards. Innoscience’s all-GaN 12kW 800V-to-50V HVDC reference design uses 650V and 100V GaN devices in an all-GaN LLC architecture, achieving 98.63% peak efficiency and 98.18% full-load efficiency.

Market Outlook and Industry Impact

AOS expects CSP-driven +/-400V HVDC demand to begin fermenting in 2027, with AI-chip-leader-driven 800V architecture reaching full volume in 2028. It estimates that by 2028, HVDC architecture upgrades could generate more than RMB 600 million in output value for wide-bandgap semiconductors. That figure reflects a market with meaningful growth elasticity, but also one still in early industrialization. SiC and GaN are both in early exploration for HV IBC, and short-term dominance is difficult to declare. The longer-term direction, however, favors GaN if high-frequency operation becomes the absolute design priority, as AOS argues.

The industry impact extends beyond device vendors. The HV IBC transition touches magnetic materials, packaging, thermal engineering, safety certification and board-level power architecture. The 800V-to-50V route is likely to anchor near-term deployments because it preserves the 48V ecosystem, while 800V-to-12V and 800V-to-6V/5V routes will mature as GPU power rises. The HV IBC market is thus moving from a route debate toward engineering fulfillment. GaN has gained the early lead, but thermal dissipation, magnetic loss and reliability remain collaborative industry problems. The SiC-versus-GaN landscape will become clearer as 800V-to-50V volume ramps and lower-voltage routes continue to evolve.