MiTAC unveils full-stack infrastructure for agentic AI at SC26

MiTAC Computing is showcasing a full-stack infrastructure portfolio for agentic AI at SC26, combining liquid-cooled servers, disaggregated inference and software orchestration. The company says the approach addresses compute, thermal and data bottlenecks in enterprise AI deployments.
- MiTAC is showcasing its AI infrastructure ecosystem at SC26 in Chicago.
- The proposed Skymizer architecture separates GPU prefill from accelerator-based decode.
- The decode configuration provides 2 TB of aggregate accelerator memory.
- MiTAC is demonstrating liquid-cooled racks, storage systems and software orchestration.
- The Helion Agentic AI Rack is designed to scale to 32,768 AMD Zen 6 CPU cores.
Disaggregated architecture separates inference workloads
At Booth 4615 in Chicago, MiTAC is presenting an inference architecture developed with Skymizer that assigns compute-intensive prefill tasks to GPUs and memory-intensive decode tasks to Skymizer HTX301 accelerator cards. The design is intended to let enterprises provision each stage according to its workload while keeping model inference on premises.
The proposed configuration uses the 8U MiTAC G8825Z5 for prefill processing. The platform features AMD EPYC™ 9575F or 9755 processors and AMD Instinct™ MI350X GPUs, then transfers the resulting Key-Value cache to a decode server over high-speed networking.
For decoding, the air-cooled 4U MiTAC G4520G6 is proposed with Intel® Xeon® 6 processors and eight Skymizer HTX301-4S accelerator cards. Each card has 256 GB of memory, giving the system 2 TB of aggregate accelerator memory across eight PCIe slots. MiTAC says the collaborative configuration is designed to run large models, including the 2.8-trillion-parameter Kimi K3 large language model, for complex agentic AI workloads.
Software and storage complete the AI factory stack
MiTAC is also positioning its hardware alongside an enterprise software layer. Canonical Ubuntu provides the certified operating system, while the DDN Infinia solution is intended to process large data streams with low I/O latency and help prevent accelerator data starvation.
Through its collaboration with Rafay, MiTAC is offering GPU-as-a-Service orchestration to allocate GPU resources across distributed workloads. The company says this is designed to reduce idle capacity and improve resource use. Its MiCoreView™ software provides POD management, telemetry and unified visibility across the data center.
“Scaling Agentic AI requires cohesive infrastructure, not fragmented systems,” Rick Hwang, President of MiTAC Computing Technology, said in the announcement. “MiTAC provides the foundational full-stack solution to make sustainable AI factories a reality.”
Rack-scale systems target density and cooling
MiTAC is collaborating with AMD on the Helion Agentic AI Rack, an upcoming ORv3 44OU solution designed for rack-scale agentic architectures. The company says it will scale to 32,768 AMD Zen 6 CPU cores within an open-source stack, supporting hyperscale and enterprise environments.
At SC26, MiTAC is demonstrating a 52U liquid-cooled rack containing 12 G4826Z5 servers and 96 AMD Instinct™ MI355X GPUs. According to the company, the system provides 50% higher GPU density than conventional deployments.
A separate proof of concept with ZutaCore combines the G4826Z5 with Waterless Two-Phase Direct Liquid Cooling. MiTAC says the design supports thermal design powers exceeding 4,000W, heat flux over 250W/cm² and an annualized pPUE of 1.03. The company is also showing a DDN Infinia storage rack with seven GC68C-B8056 1U storage servers.
MiTAC says attendees can view the systems at Booth 4615 at McCormick Place from Nov. 17-19. The announcement reflects the company’s focus on giving enterprise and hyperscale customers integrated options across servers, racks, cooling, storage and workload management.
PR Newswire: MiTAC Computing Scales Agentic AI with Full-Stack Infrastructure at SC26

