Sean Lim, Manager at Auras Technology, delivered this session at OCP Global Summit 2025 in San Jose, California, arguing that cooling has become a foundational design constraint for AI infrastructure — not merely an operational consideration — and presenting Auras's complete liquid cooling product ecosystem as the answer. The talk framed the thermal challenge with market data (IDC projecting AI infrastructure spending growing from ~$42B in 2023 to ~$90B by 2028; McKinsey forecasting US data centers exceeding 12% of national electricity by 2030), charted the GPU TDP trajectory from NVIDIA H100 (~700 W) through projected Feynman (~2500+ W) and AMD MI-100 through MI-450 (~2000 W), and demonstrated that direct liquid cooling achieves PUE below 1.1 versus 1.6–1.7 for air cooling — saving $125,266 per rack over two years at 64-GPU density.

Key Takeaways

• IDC projects global AI infrastructure spending will nearly double from ~$42B (2023) to ~$90B by 2028, driven by GPU-centric compute models (source: IDC 2024 Worldwide AI Infrastructure Forecast).
• McKinsey forecasts US data center electricity consumption will exceed 12% of total national demand by 2030, up from less than 4% at the time of the presentation.
• Current AI accelerators (NVIDIA H100 ~700 W, B200 ~1000 W; AMD MI-450 ~2000 W) already exceed conventional air cooling limits; next-gen chips (NVIDIA Feynman) are projected to exceed 2500 W by 2027.
• Auras Technology's direct liquid cooling achieves PUE below 1.1 versus 1.6–1.7 for air cooling, saving 481,800 kWh and $125,266 per rack over two years at 64-GPU density.
• Auras's cold plate roadmap covers five thermal tiers from 700 W (copper, 0.2 mm fin) to over 5000 W (silicon wafer micro-channel, two-phase), with single-phase supporting up to 4500 W.
• Auras's 300 kW MGX liquid-cooled PSU uses PG25 coolant at 12 LPM, maintains ΔT of 13.56°C at 45°C ambient, supports hot-swap 150 kW modules, and integrates a liquid-cooled bus bar.
• Auras's in-row CDU supports 1.6 MW cooling capacity with 1+1 redundant pumps and real-time remote monitoring, completing a full chip-to-facility liquid cooling stack.
• The AI manifold features automatic leak-detection valves that isolate affected server trays without manual intervention, improving data center reliability.