Supermicro puts its H15 servers into production for AI and cloud
Supermicro says it has begun delivering its H15 range, built on sixth-generation AMD EPYC 9006 processors. The offering targets cloud, HPC, enterprise, and agentic AI deployments, with more than 20 systems.
Supermicro announced on November 11, 2026, that its H15 server range is now shipping, featuring more than 20 systems based on the sixth-generation AMD EPYC 9006 processors. The line targets cloud, high-performance computing, enterprise IT, and agentic AI deployments, a segment where compute density and bandwidth matter almost as much as the models themselves.
A power boost designed for inference nodes
The core of the matter is the capacity leap on the host nodes that feed GPU-accelerated systems. According to the technical white paper published in July 2026, some configurations reach up to 256 Zen 6 cores and 512 threads per processor, with 16 memory channels, DDR5-8000 or 12.8 GT/s MRDIMM, as well as CXL 3.1 extension. With MRDIMM, memory bandwidth reaches about 1.6 TB/s per socket.
Supermicro also claims to double PCIe bandwidth and increase memory bandwidth by 2.6 times compared to the previous generation. The company highlights a 1.8x gain in tokens per second for host nodes of leading-edge models, a metric that directly impacts inference throughput. In short, the server doesn’t flex for show: it’s there to feed the GPUs without making them wait.
A single generation, multiple deployment formats
The H15 range is not a homogeneous block. Hyper serves as the flagship dual-socket system for enterprise, AI inference, virtualization, and cloud. CloudDC targets cloud-scale environments with a single-socket architecture compliant with the Open Compute Project Data Center Modular Hardware System specification, an open standard for data center hardware.
For denser deployments, FlexTwin combines two dual-processor nodes with liquid cooling in a 1OU ORv3 chassis, designed for cloud-native and hyperscale infrastructures. GrandTwin adopts a 2U architecture with two nodes and single socket for object storage, cloud services, virtualization, and HPC. Multi-node designs share power, networking, and cooling to improve energy efficiency and fit more compute in the same rack.
Supermicro also adds Petascale Storage, a 2U flash platform for AI data lakes, large-scale analytics, and HPC, with up to 4.8 PB capacity per system. Finally, H15 SuperBlade uses a 6U chassis and supports one- or two-socket configurations, in air- or liquid-cooled versions.
And Helios, where does it stand?
The Helios and Helion platforms are not part of the H15 availability announcement. Supermicro and AMD only indicate they are working together to bring them to market. Helios targets training of leading-edge models and high-throughput inference with AMD Instinct MI455X GPUs, while Helion focuses on CPU performance and deployment flexibility with sixth-generation EPYC processors.
Product documents describe Helios as a 72-GPU rack system combining these accelerators with EPYC processors and AMD Pensando networking, for hyperscale AI clusters and large language model training. In this context, DCBBS, Supermicro’s Data Center Building Block Solutions architecture, acts as a broader integration layer: validated servers, storage, rack cooling, networking, power, site infrastructure, management software, validation, and on-site deployment services.
The bet is quite clear: to turn the refresh around EPYC 9006 into complete projects, from standalone servers to data centers. In a market where integrators seek to reduce back-and-forth between hardware and operations, this coherence often weighs as much as the technical specs.
Key points
- November 11, 2026: Supermicro says it is shipping the H15 range.
- More than 20 systems based on AMD EPYC 9006.
- Up to 256 Zen 6 cores per processor announced.
- Hyper targets enterprise, AI, and virtualization.
- Helios and Helion remain in preparation.
- DCBBS also covers cooling, networking, and deployment services.
By the numbers
- 20+ systems — H15 portfolio announced by Supermicro, November 2026.
- 256 cores — maximum per processor, according to the July 2026 technical white paper.
- 1.6 TB/s — memory bandwidth per socket with MRDIMM.
- 4.8 PB — maximum capacity per system for Petascale Storage.
- 72 GPUs — rack configuration described for Helios.