xFusion scales enterprise AI from edge to liquid-cooled racks

At ISC 2026 in Hamburg, xFusion presented a four-tier hardware portfolio designed to move AI workloads from individual workstations to data-centre racks. The company showed systems for edge development, workgroup containment, corporate token processing and facility-level supernodes.

xFusion described the portfolio as a response to procurement issues it identified: hardware purchases that ignore physical operating limits and reliance on public APIs that can expose proprietary data. The firm outlined a sequential deployment path that scales processing capacity as workloads shift from local development to central inference.

The entry tier is the FusionXtation X3 8000 Gen2 edge workstation, aimed at engineers and specialised users running large local datasets. Configurations pair Intel Core Ultra processors with two professional GPUs, error-correcting DDR5 memory up to 256 GB and up to 8 TB of internal storage. Units include baseboard management controllers for remote IT access and four 40 Gbps Thunderbolt ports. xFusion presented production figures claiming up to 70 percent faster 8K rendering and up to 50 percent higher AI processing versus earlier hardware.

The second tier, the FusionXpark appliance, is intended for workgroups that must keep sensitive data isolated. Two FusionXpark nodes combined can run models of about 405 billion parameters on native CUDA environments. The systems ship with NVIDIA DGX OS and preinstalled toolchains. Administrators can route overflow workloads into DGX Cloud through native integrations.

For central corporate processing, xFusion displayed the TokenBox, an on-premises appliance for high-volume token generation and routine automation. A single TokenBox node is rated to run models up to 1.6 trillion parameters. The unit uses internal liquid cooling and operates at about 35 decibels during active computation. It ships with preinstalled software to reduce setup time compared with building custom server rooms.

The top tier covers data-centre infrastructure with liquid-cooled racks and supernodes. The FusionServer G6550 V8 inference server supports up to ten dual-width GPUs, while the FusionPoD platform manages cabinets rated for roughly 240 kilowatts each. Cooling elements include graphene pads and diamond cold plates with reported thermal conductivity near 1,200 watts per metre-kelvin. xFusion reported a partial power usage effectiveness of 1.06 and stated its low-loss core components reduce single-module operating expenses by about 15 percent.

xFusion also showed the FusionOne DFS storage solution for the compute clusters. In a demonstration a three-node array with 72 NVMe drives achieved sequential read throughput of 200 gigabytes per second. The arrays scale toward exabyte capacity and use erasure coding to reach reported storage utilisation of 94.1 percent.

Buyers at the exhibition raised practical concerns including noise, cooling, remote management and regulatory isolation. Technical specifications and product details are available from xFusion. The company did not present independent benchmarks at the show beyond the performance figures and efficiency metrics displayed during demonstrations.