AI’s Energy Problem is an AI Problem: Resolving the Grid Crisis Through Constrained Optimization
June 29, 2026
The rapid proliferation of artificial intelligence has initiated an unprecedented technological transition, yet it has brought to light a critical physical limitation related to energy. As data centers expand globally, the computational power required to train and run frontier models is placing an immense strain on electrical grids.
Recently, at the GStar AI Summit 2026 in Vietnam, Professor Laurent El Ghaoui, Vice-Provost of Research and Innovation at VinUniversity, addressed this challenge. His core thesis presents a compelling paradox: while AI is a primary driver of the current energy crisis, advanced AI is also a critical mechanism for resolving it.
Beyond Hyperscaling: The Real Optimization Challenge
The conventional industry response to AI’s energy appetite has focused heavily on physical expansion – building larger hyperscale facilities, improving cooling systems, and engineering faster silicon. However, this hardware-centric scaling strategy is rapidly encountering physical limitations. The consequences are already visible globally: strained power grids, water-resource conflicts, displaced electrical loads, and localized community pushback. As Prof. El Ghaoui noted, “We’re borrowing against a budget we haven’t fully audited.”
The core challenge is not simply generating more energy but intelligently managing the resources we already possess. We are shifting toward renewable energy sources like wind and solar, which are inherently intermittent and weather dependent. At the same time, power generation is decentralizing across millions of localized nodes, such as household solar panels and battery storage units, while consumer demand remains highly volatile.
Managing a dynamic system of this complexity cannot rely on intuition or traditional spreadsheets. Matching supply, demand, storage, and pricing across seconds and seasons represents a massive, multi-dimensional optimization problem.
The Role of Advanced Computing in Grid Management
To address these challenges, the scientific community must apply the same computational rigor used to build frontier AI models to the physics of the energy grid. We need advanced algorithms designed to operate under strict, real-world physical boundaries.
Prof. El Ghaoui highlighted several critical areas where AI must evolve to manage resource constraints:
- Foundation Models for Time Series: Just as Large Language Models (LLMs) predict the next word in a sentence, foundation models designed for time-series data can learn to predict complex weather patterns, grid loads, and price fluctuations across highly distributed networks.
- Learned Solvers Over Cold Solves: When grid optimization problems repeat daily with identical structures but different input data, machine learning algorithms can “learn” from previous solutions. This dramatically speeds up computation, enabling the real-time adjustments necessary to prevent grid failures.
- Coordinated Systems: Physical grid management requires high reliability. Research shows that centralized computational coordination can achieve massive efficiency gains over uncoordinated, localized systems.
“AI created this problem, so AI should solve it. It’s the most interesting optimization problem on the table right now” Prof. El Ghaoui emphasized.
From Theory to Action: VinUniversity’s Research Frontier
At VinUniversity, the intersection of advanced mathematics, artificial intelligence, and environmental sustainability is a core institutional priority. Through the Center for Environmental Intelligence (CEI), VinUni researchers are actively developing the methodologies needed to transition these mathematical concepts into practical, real-world tools.
CEI’s research agenda is structured around three key pillars:
- Smart Cities and Energy Efficiency: Developing data-driven frameworks to balance energy loads and optimize local smart grids.
- AI-Powered Materials Discovery: Utilizing machine learning to accelerate the design of advanced materials, such as high-energy sodium-ion batteries, to support sustainable energy storage.
- Sustainable AI Development: Addressing the carbon footprint of computer science itself by researching resource-efficient serverless computing architectures for model training.
By focusing on these areas, VinUni aims to create widely applicable computational tools that address global climate and energy challenges.
HORIZONS 2026: Optimization for Sustainability
Reflecting this commitment, VinUniversity will host HORIZONS 2026: Optimization for Sustainability from July 1-3, 2026. This prestigious international conference will bring together global researchers, industry leaders, and innovators to discuss state-of-the-art optimization and machine learning theories and how to translate them into practical tools for sustainable systems.
Key session topics include:
- Large-scale learning and optimization for resource efficiency.
- Machine learning architectures designed for resource-constrained environments.
- Reinforcement learning and advanced control for intelligent energy systems.
The challenges of the AI energy transition require deep, interdisciplinary collaboration. Through forums like HORIZONS 2026, VinUniversity continues to support the development of computational tools that promote sustainable, long-term technological progress.
More on HORIZONS 2026: https://horizons2026.vinuni.edu.vn/
