xFusion brings 800V HVDC and supernodes for "token factories" to OCP 2026

On October 12, 2026, at the Open Compute Project Global Summit in San Jose, Chinese server maker xFusion (超聚变) said it treats computing and power supply as a single engineering problem. The company brought servers for AI workloads, supernode components and an 800V HVDC power system. The slogan on its stand was "算电协同", meaning coordination of computing and electricity. xFusion describes itself as a horizontal full-stack provider for the age of AI and data.

What exactly was shown

The available excerpt of the announcement indicates that the exhibit was built around three blocks: AI servers, supernode components, and power infrastructure running on 800 V direct current. The company also presented a white paper on "智算数能" (intelligent computing plus energy). The document's title is cut off in the source text, so its contents cannot be summarized.

Here, a supernode means a tightly coupled cluster of accelerators that behaves like one large machine for training and inference. Power per rack in such systems is growing fast, and the constraint is increasingly not chip availability but how much energy can be delivered to the rack and how much heat can be removed.

Why 800V HVDC

The whole industry is now discussing a move to high-voltage direct current in AI data centers. In 2025, NVIDIA publicly described an 800 VDC architecture for "AI factories", together with power partners. The logic is simple: at the same power, a higher voltage means lower current, which means thinner busbars and fewer conversion losses. That matters most when a single rack draws hundreds of kilowatts.

OCP remains the venue where such designs turn into open specifications. Showing 800V HVDC at this summit suggests that xFusion wants to fit into a common industry standard rather than offer a closed solution.

"Token factory" is a convenient frame for this logic. If you judge a data center by tokens per second per watt and per dollar of capex, power and cooling stop being auxiliary systems and become part of the cost calculation.

What is known and what is not

• Date and place: October 12, 2026, San Jose, OCP Global Summit.
• Company: xFusion (超聚变), formerly Huawei's server division, spun off in 2021.
• Shown: AI servers, supernode components, an 800V HVDC power system.
• Concept: "算电协同", computing and energy as a single system.
• Not disclosed in the source: rack power, number of accelerators per supernode, conversion efficiency, delivery timelines, names of chip partners.

How to read this announcement

For now, this is a bid for position, not a set of verifiable specifications. Data center operators and integrators should ask xFusion for specs: permissible power per rack, the conversion scheme from 800 V down to board level, compatibility with OCP open racks, and loss data compared with conventional alternating current.

Anyone planning AI sites two to three years out should already build in the option of moving to high-voltage direct current: feeder networks, zoning and maintenance safety requirements. For the market as a whole, something else matters more: even server makers now sell not just compute but also a way to power it. They will be judged on cost per token, not on peak teraflops.