MiTAC Computing Technology Corp. and AMD jointly presented liquid-cooled, OCP-compliant AI cluster solutions at OCP Global Summit 2025 in San Jose. Charles Chen (MiTAC) outlined three core challenges — high power density, infrastructure reliability, and sustainability — then introduced MiTAC's 64-GPU in-rack DLC solution and 256-GPU multi-rack pod with leak detection management. AMD's Mark Orthodox (Sr. Director, Business Development) presented AMD's annual Instinct GPU roadmap through MI400/Helios (2026) and MI500 (2027), the ROCm open software stack, and AMD's co-founding of the Ultra Ethernet and Ultra Accelerator Link consortia. MiTAC closed by detailing its 3,000-rack/month global manufacturing capacity and end-to-end rack design, integration, and service capabilities.

Key Takeaways

• GPU power is projected to exceed 2,000W per GPU with full GPU servers over 16,000W, and data center energy consumption is forecast to reach 945 TWh by 2030 — doubling 2024 levels — making liquid cooling a necessity for AI clusters.
• The global data center GPU market is forecast at 14.2% CAGR from 2025 to 2034, reaching $183B by 2034 — an 8x increase within a decade.
• MiTAC's 64-GPU in-rack DLC solution uses eight 4U G4826Z5 AMD MI355X servers in a 48U EIA rack (H 2258mm x W 800mm x D 1400mm) with a 4RU in-rack CDU, backend 64x400G and frontend 32x400G leaf switches, and a zero-U rear PDU.
• MiTAC's 256-GPU pod scales to four 64-GPU racks with centralized management, storage, and switching, using an in-row CDU topology preferred for large-scale CSP deployments.
• MiTAC's RMC leak management uses water leak detection cables and IPMI commands to auto-shutdown servers within 5 or 30 minutes of threshold breach; Phase 2 will replace the current AMD CPU-based RMC with a lightweight dedicated controller.
• AMD's annual Instinct roadmap: MI300A/X (2023) → MI325X (2024) → MI350 series inc. MI355X (2025) → MI400 series / Helios (2026) → MI500 series (2027); MI355X delivers 1.5x HBM vs. Vera Rubin.
• AMD Helios platform (Venice CPU + MI400 GPU + Vulcano 800G Pensando NIC) debuted at OCP 2025 is OCP compliant and compliant with Meta's OW rack standard released the same week.
• AMD Pensando Pollara is the first UEC-ready NIC delivering up to 20% more scale-out performance; AMD co-founded the Ultra Ethernet Consortium and UAL consortium to open previously closed AI networking ecosystems.
• MiTAC manufactures 3,000 racks/month across 13+ MW of global facilities (North America, EMEA, SEA, China), supports up to 480 VAC and 800 Gb/s networking, and offers end-to-end design, integration, burn-in validation, and on-site service.