# How AI Can Solve Its Own Energy Crisis | Varun Sivaram | TED Source: https://www.youtube.com/watch?v=p8Ed8pDlAmM Recap page: https://rapidrecap.app/video/p8Ed8pDlAmM Generated: 2025-12-12T16:36:38.935+00:00 --- ## Quick Overview Varun Sivaram argues that Artificial Intelligence (AI) data centers can become the greatest ally to the electric grid by implementing spatiotemporal flexibility, allowing them to dynamically adjust power consumption to align with renewable energy availability and grid stress, thereby preventing blackouts and enabling cleaner energy integration without massive infrastructure buildouts. **Key Points:** - On a blistering hot day in Phoenix, Arizona (September 2025), AI data centers drove up demand on the power grid, but flexible AI actually helped buck the trend. - AI data centers can reduce power consumption by 25% during peak demand periods, providing relief to the grid, as demonstrated in a simulation. - The inherent energy demands of AI, especially Large Language Models and advanced chips, are massive, with US data center energy demand potentially soaring from 4% today to 12% by 2030, equivalent to adding Germany's entire power grid. - The solution lies in Spatiotemporal Flexibility: Temporal Flexibility allows AI workloads (like training or simulation) to pause/slow during peak stress and resume when power is abundant; Spatial Flexibility allows workloads to move across geographies (like the US or UK) to areas with cheaper, cleaner power. - The speaker's team at Emerald AI demonstrated that flexible AI workloads can maintain high job performance (above 90% for low flexibility, 50% for high flexibility) while absorbing peak grid demand. - This flexibility allows utilities to avoid building expensive new infrastructure, as exemplified by the fact that 100 GW of currently unused grid capacity is available across the country for flexible AI data centers. - By acting as 'giant shock absorbers' that can ramp consumption up and down quickly, AI can better integrate intermittent clean energy sources like solar and wind, driving down energy costs. ![Screenshot at 00:34: The presenter displays a graph illustrating how flexible AI power consumption \(green line\) can dip significantly below the total grid power demand \(red line\) during the peak demand period, demonstrating the relief AI can provide to the grid.](https://ss.rapidrecap.app/screens/p8Ed8pDlAmM/00-00-34.png) **Context:** Varun Sivaram, introduced as an energy executive and clean energy diplomat, presents at TED Countdown on the growing energy demands of Artificial Intelligence (AI) and how this growth threatens grid stability, particularly in hot areas like Phoenix, Arizona. He frames the conflict between increasing AI power consumption and the need for a reliable, clean power grid, proposing that AI itself holds the key to resolving this tension through flexible scheduling and geographic load shifting. ## Detailed Analysis Varun Sivaram argues that the massive energy demands of the burgeoning AI revolution, fueled by energy-hungry chips and large language models, pose a significant threat to the stability of the electric grid, citing reports that US data center demand could reach 12% of total power demand by 2030, equivalent to Germany’s entire grid. He highlights data showing that in places like Phoenix, Arizona, this demand peaks on the hottest days when air conditioning demand is already maxed out, risking grid failure and blackouts. However, Sivaram proposes that AI itself can be the solution through 'Spatiotemporal Flexibility.' Temporal flexibility involves intelligently pausing or slowing down non-urgent AI workloads (like training or simulation) when the grid is stressed and ramping them back up when clean, cheap power (like solar) is abundant, all while maintaining acceptable job performance thresholds (e.g., 50% performance for highly flexible workloads). Spatial flexibility involves moving workloads across geographies, utilizing existing transmission capacity across the US and internationally via high-speed fiber optic networks to areas where power is currently clean or cheap. He cites 100 GW of currently available, unused grid capacity across the US that could immediately support flexible AI data centers. By coordinating these flexible AI workloads, the overall energy footprint can be managed to smooth out peak demand, avoid costly infrastructure upgrades, and accelerate the integration of intermittent renewable energy sources like solar and wind. ### The AI Energy Problem - AI workloads (training, LLMs) demand massive energy, threatening grid stability during peak times like a hot Phoenix day in September 2025 - AI power consumption could hit 12% of US demand by 2030, equal to Germany's grid - Media reports highlight grid strain and rising energy bills due to AI. ### The Solution - Spatiotemporal Flexibility: Temporal flexibility means pausing/slowing AI workloads during grid stress (peak hours) and resuming when power is cheap/clean - Spatial flexibility means moving workloads across the country or globe via fiber optics to regions with lower demand or cleaner supply. ### Demonstrated Impact - Simulations show that flexible AI can reduce peak power load by 25% for the required hours while meeting job performance targets (e.g., low flexibility maintains >90% performance) - This allows AI to become the grid's 'greatest ally' and 'giant shock absorber,' integrating renewables like solar and wind. ### The Future Vision - Emerald AI is building software to orchestrate this flexibility, showcasing a reference design for a 'Power-Flexible AI Factory' that connects seamlessly with the grid, proving that this coordination is possible without waiting years for new infrastructure. ![Screenshot at 00:08: Varun Sivaram taking the stage at TED Countdown to discuss the energy demands of AI.](https://ss.rapidrecap.app/screens/p8Ed8pDlAmM/00-00-08.png) ![Screenshot at 00:17: A graphic showing the projected peak grid power demand \(red line\) versus AI power consumption \(green line\), illustrating the potential for AI to reduce the peak load.](https://ss.rapidrecap.app/screens/p8Ed8pDlAmM/00-00-17.png) ![Screenshot at 02:45: A slide displaying news headlines from Bloomberg, Newsweek, and MIT Technology Review highlighting the growing power demands and fossil fuel dependence associated with the AI revolution.](https://ss.rapidrecap.app/screens/p8Ed8pDlAmM/00-02-45.png) ![Screenshot at 04:41: A chart showing the hourly electricity demand on the Arizona Public Service \(APS\) system from July 2024 to July 2025, demonstrating seasonal peaks.](https://ss.rapidrecap.app/screens/p8Ed8pDlAmM/00-04-41.png) ![Screenshot at 10:29: A map illustrating potential geographic flexibility, highlighting Arizona, Illinois, Virginia \(US\), and the UK as locations for flexible AI data centers.](https://ss.rapidrecap.app/screens/p8Ed8pDlAmM/00-10-29.png)