This OCP Global Summit 2025 executive session, co-sponsored by Lenovo and NVIDIA, features a panel of four experts exploring how Lenovo Neptune liquid cooling enables AI infrastructure at scale. Connor Malone (VP & GM CSP, Lenovo) moderates. Panelists: Chris Marriott (VP/GM Enterprise Platforms, NVIDIA), Pravin Patel (Distinguished Engineer & Chief System Architect, Lenovo), and Chris Coates (Principal Solutions Architect, Nscale). The session traces Neptune's IBM mainframe heritage, details its 6th-generation technical specifications, provides an operator perspective from Nscale, and rounds on the future of AI data center design — digital twins, pod/SU reference architectures, on-site power, heat reclamation, and 18-month refresh cycles.

Key Takeaways

• Liquid cooling has crossed from HPC niche to mainstream AI data center necessity as Grace Blackwell NVL72 racks reach 130 kW today and are expected to approach 1 MW within a few years.
• Lenovo Neptune's 6th-generation direct water cooling uses 45°C distilled water, copper piping, and aluminum cold plates to achieve 80% cooling efficiency on H100/H200/B200 GPU systems with no chemical handling risk.
• Neptune delivers 40% less energy, 4:1 server consolidation, and 6.1x compute throughput versus prior-generation air-cooled platforms; GPU junction temperature variation is kept below 1% with proper water flow.
• Lenovo's liquid cooling heritage traces to IBM mainframe practices of the 2000s and large-scale deployments at LRZ Barcelona in 2012, reaching 100% water-cooled solutions by 2020.
• Nscale reports water cooling via Lenovo Neptune has become more predictable than air, shifting the air-to-water split to approximately 90/10 and reducing deployment risk through blind-mate connections and end-of-row CDUs.
• NVIDIA's disaggregated inference architecture (Rubin CPX + Dynamo with separate prefill/decode racks) significantly increases inter-rack KV cache bandwidth demands and further drives the need for high-density liquid cooling.
• The future of AI data center design includes digital twins (NVIDIA Omniverse), pod/SU-level reference architectures, on-site power generation for gigawatt-scale facilities, and commercially viable heat reclamation tied to always-on industrial processes.
• Assuming an 18-month GPU refresh cycle (per US DoE guidance) argues for lightweight, disposable data center materials like fiber-reinforced polymer and minimum viable site builds over permanent full-featured structures.