Supermicro’s AI, HPC and Liquid Cooling Roadmap: Explained
Supermicro has used this year’s Supercomputing Conference (SC25) in St Louis to showcase a broad set of technologies aimed at meeting the rapidly evolving thermal and density requirements of data centres built for AI and high-performance computing.
The portfolio spans rack-scale platforms, liquid-cooled systems and updated Data Center Building Block Solutions (DCBBS) based on Nvidia’s latest processors, reflecting how operators are rethinking both cooling and architecture to manage increasingly power-hungry workloads.
Shifting requirements in AI and HPC design
Speaking at the event, Charles Liang, President and CEO of Supermicro, positions the new releases as part of a wider transition across compute-intensive infrastructure.
“Supermicro continues to lead the industry in delivering complete, next-generation infrastructure solutions in close collaboration with our technology partners,” he says.
“At SC25, we are demonstrating our high-performance DCBBS architecture, direct liquid cooling, and rack-scale innovations, which empower customers to deploy AI and HPC workloads faster, more efficiently, and more sustainably.”
The latest platforms form part of an industry-wide shift towards hardware that can support larger training clusters, higher thermal loads and greater GPU density.
Supermicro’s launch lineup includes systems built around Nvidia GB300 NVL72 and HGX B300 architectures, along with air and liquid-cooled variants designed for different rack configurations.
Rack-scale systems and emerging liquid-cooling standards
At rack scale, Supermicro highlighted a liquid-cooled Nvidia GB300 NVL72 solution that integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs, each GPU equipped with 279 GB of HBM3e memory.
These systems are designed to be deployed alongside liquid distribution units and rear door heat exchangers, which the company is presenting as integral rather than optional in future high-density builds.
Other launches included a 4U HGX B300 liquid-cooled rack, a 1U Nvidia GB200 NVL4 server and a compact developer workstation built on the same GB300 architecture. Multi-node designs were also on display, including liquid-cooled 8U 20-node and 6U 10-node SuperBlade systems, along with the 2U FlexTwin platform, which can capture up to 95% of generated heat.
These systems reflect a broader trend towards direct-to-chip cooling as operators navigate rising chip TDPs and tighter rack power envelopes. Rear door heat exchangers, liquid-to-air sidecar cooling distribution units and closed-loop dry towers featured prominently, with capacities ranging from 50 kW to 200 kW depending on the configuration.
Integrating compute, storage and cooling
The company’s DCBBS architecture continues to bring together compute, storage, networking and cooling into modular designs that can be adapted to different facility layouts.
This approach is intended to reduce deployment times and streamline infrastructure choices at a moment when operators are reassessing mechanical and electrical systems to accommodate AI workloads that exceed traditional limits.
For dense clusters, SuperBlade remains one of Supermicro’s longest-running product families. The newest X14 generation supports both air and direct-to-chip liquid cooling and integrates switches for InfiniBand and Ethernet, which are frequently required for tightly coupled HPC and AI training environments.
FlexTwin targets similar use cases with a multi-node design optimised for compute-heavy applications such as research modelling, financial simulations and engineering workloads. Each node supports liquid cooling, high-performance CPUs and flexible networking up to 400G per node.
Density, variation and workload-specific choices
Several other families were also updated, including BigTwin, MicroBlade and MicroCloud, which offer operators different density and power profiles depending on their needs. BigTwin reduces component duplication through shared power and fans, while MicroBlade continues to serve customers seeking dense single-socket systems.
Storage systems received attention too, with Supermicro demonstrating all-flash, high-density hardware for software-defined storage platforms, as well as workstation-class systems designed for rackmount use in secure or remote environments.
Supermicro’s announcements at SC25 show how vendors are preparing for an environment where liquid cooling, integrated thermal systems and modular rack-scale designs become standard rather than specialised options.
For data centre operators planning new builds or retrofits, these developments highlight the growing need to align facility design with next-generation compute requirements.
