Dr. Imran Latif
From Chip to Chiller: LLM Agents as the Next Efficiency Lever in AI Data Centers
Global Vice President, Technology and Innovation for Data Centers, Johnson Controls Intl., USA
Speaker Bio
Based in New York City, Imran Latif is the Global Vice President of Technology and Innovation for Data Centers at Johnson Controls International, where he leads multidisciplinary global teams driving cutting-edge technology integration and shapes strategic portfolios through M&A for JCI’s global data center business. He has structured and executed strategic acquisitions exceeding multiple billions in value, bringing deep expertise in technical due diligence, infrastructure integration, and post-merger value creation.
With over two decades of experience spanning the complete lifecycle of AI/HPC hyperscale data centers, Imran has led the design, build, and operational management of gigawatt-scale facilities for major hyperscalers, colocation providers, research institutions, and Fortune 100 enterprises worldwide. His technical leadership encompasses governance frameworks for AI deployments at unprecedented scale, extreme rack density optimization, comprehensive thermal management from direct-to-chip liquid cooling to full immersion systems, and GPU cluster architecture with high-bandwidth interconnects enabling seamless scaling across thousands of accelerators. He champions circular economy principles in data center design, bridging the critical gap between physical infrastructure and computational performance.
His research is featured in Lawrence Berkeley National Laboratory's 2024 US Data Center Energy Usage Report, a Congressionally mandated study shaping national energy policy. He optimizes critical metrics including tokens-per-watt and Power Usage Effectiveness for large-scale LLM training and high-performance computing workloads.
Previously, as Chief Operations Officer for Data Center Infrastructure at the U.S. Department of Energy's Brookhaven National Laboratory, Imran's leadership proved instrumental in supporting computational infrastructure for CERN's Large Hadron Collider experiments contributing to the Nobel Prize-winning Higgs boson discovery. He directed deployment of IBM's Blue Gene/Q supercomputer, ranked Top 5 globally on TOP500 and first on both the Graph500 and Green500 benchmarks. He also led infrastructure implementation for a 50-qubit superconducting quantum computer, managing extreme requirements including sub-millikelvin cooling and electromagnetic isolation for quantum coherence. During his tenure, he mentored graduate and post-graduate students on DOE-funded projects spanning AI, energy systems, and engineering disciplines.
Imran holds a Master of Science in Mechanical Engineering from the City University of New York and PhD in Artificial Intelligence from Florida Atlantic University, with research focused on optimizing infrastructure for AI workloads. He is a licensed Professional Engineer in New York and Texas and a frequent keynote speaker at Supercomputing (SC), Open Compute Project (OCP), and Data Centre Dynamics (DCD).
Presentation time
December 8, 2026
8:10 am - 9:20am EST
Abstract
AI racks now draw far more power than conventional air cooling was ever designed to handle, while grid constraints increasingly cap the total power a data center site can draw. This leaves one critical lever: optimizing the balance between compute and cooling. This talk presents a physics-grounded, chip-to-chiller simulation of a large GPU cluster used as a closed-loop testbed for LLM-based control agents.
A set of zero-shot language models, with no training or fine-tuning, were tasked with maximizing compute output per unit of energy consumed while operating under a hard thermal safety limit. Every model tested outperformed a static baseline, reducing cooling overhead while improving overall efficiency, with the strongest performers also demonstrating the safest operation.
The presentation will cover the simulator architecture, control strategy, and comparative results, along with the practical challenges of applying LLM agents to real-world thermal systems. It will also discuss the pathway toward deployment at much larger scale, where AI-driven clean cooling optimization could become an important tool for improving energy efficiency, thermal reliability, and compute capacity in next-generation data centers.