Host Fabian Alefeld interviews Professor PS Lee, Head of National University of Singapore Mechanical Engineering & Program Director of STDCT, and founder of CoolestDC, about heat as a key constraint in AI infrastructure as data centers shift from CPU-heavy to GPU-dense systems with higher power density, hotspots, and bursty workloads. Lee explains why cooling must be treated as an integrated “chip to grid” system spanning cold plates, racks, CDUs, piping, controls, and operations, and discusses single-phase versus two-phase direct-to-chip liquid cooling and the added need for condensation in closed loops.
Professor Lee describes how additive manufacturing (AM) enables complex fin structures and unibody cold plates that reduce leakage risk, lower junction temperatures (~10°C), cut pumping power (60–70%+), and reduce material use, while requiring TCO evaluation. At NUS’s live testbed (22 liquid-cooled racks, ~500 kW), AM cold plates are being deployed and show promising thermal, power, and compute-performance improvements; future work includes two-phase cooling and broader system-level heat rejection, warm-water operation, and waste-heat reuse, plus a discussion of challenges and potential AM roles in space-based data centers.
02:13 Meet Professor Lee
05:26 Chip To Grid Cooling
10:22 Liquid Cooling Maturity
14:15 Two Phase Explained
18:12 Cold Plate Tradeoffs
20:40 AM Unibody Cold Plates
24:41 Benchmarking AM Gains
30:18 AM For CDUs
34:13 Inside STDC Testbed
38:09 Reliability And Leakage
39:00 Next Phase Roadmap
42:31 Orbital Data Centers
