Don't worry. Be happy. Server Farms Will Learn for Themselves How to Be All Renewabley and "Green."

[NNadir]

Here's a fun paper everyone might enjoy, although, regrettably one needs access to read it:

Leveraging Deep Reinforcement Learning within Optimal Renewable Energy Strategies for Sustainable AI Data Centers Tianqi Xiao and Fengqi You Environmental Science & Technology 2026 60 (1), 522-536

From the introduction to the paper:

Generative artificial intelligence (AI) has seen unprecedented growth, driven by advancements such as ChatGPT (1) and DeepSeek. (2) With major investments such as the Stargate project, AI is poised to become a transformative force in society. However, unlike previous computing advancements, AI relies on energy-intensive servers that utilize graphics processing units (GPUs) and high-memory devices, resulting in energy demands up to ten times higher than traditional high-performance computing. (2−5) Given the large-scale deployment of AI servers, this shift is projected to reshape the data center industry. (6,7) By 2028, AI servers in the U.S. could consume over 300 TWh of electricity, three times the total power usage of all U.S. servers in 2023. (8) This demand is comparable to the annual electricity consumption of entire countries such as Spain or the combined usage of major U.S. industrial sectors such as paper manufacturing and food processing. While uncertainties remain, AI computing is set to make data centers one of the largest electricity consumers, disrupting the industry’s previously stable energy footprint. (8,9) In response, this study provides a pioneering assessment of AI computing’s operational efficiency and proposes solutions to address its decarbonization challenges.

The rapid expansion of AI presents severe sustainability challenges. (10,11) Efficiency improvements alone will no longer be sufficient. Power usage effectiveness (PUE), a key energy efficiency metric comparing total facility energy to server energy, has improved from 2.50 to 1.58 over the past two decades due to operational optimizations. However, the remaining efficiency gains cannot offset the rapid energy footprint growth of AI computing, even with state-of-the-art designs. (12−14) Given current limitations in renewable energy capacity, uncontrolled AI-related electricity growth may increase fossil-fuel reliance, leading to significant Scope 2 carbon emissions and water consumption. (15,16) Additionally, inference-based AI workloads could introduce peak loads and fluctuations (17,18) in response to varying demand, often requiring fossil-fuel-based grid stabilization and further increasing carbon emissions. Meanwhile, training-based AI workloads sustain high server utilization, resulting in extreme energy demands that challenge the feasibility of compensating for emissions through on-site renewable energy. Large-scale AI data centers also pose additional water sustainability risks, as cooling demands intensify in high-temperature, water-scarce regions with higher water usage effectiveness (WUE). (19−22) Moreover, according to the Science Based Targets Initiative (SBTi), (23,24) the data center sector must reduce carbon emissions by over 50% from 2020 levels to meet the 1.5 °C climate target. However, the challenges outlined above present significant barriers to achieving this sustainability goal.

Addressing the sustainability challenges of AI requires closing critical knowledge gaps as given below.

(1) Facility limits under AI-specific conditions: While traditional data centers have well-established efficiency benchmarks, (25−28) AI data centers operate at significantly higher utilization rates and demand far greater cooling capacity, approaching the limits of traditional air-cooled systems. (14,29) As climate change intensifies, ensuring energy and water efficiency under the system constraints becomes even more critical, especially in hot climates. (30,31) This necessitates a reassessment of existing efficiency strategies to support sustainable AI operations.

(2) Integration of AI data centers with the power grid and renewable energy: Strengthening grid interaction is essential for improving sustainability. (32) Cost- (33,34) and carbon-aware (35,36) operational strategies of data centers can reduce expenses and emissions while enhancing grid stability. Utilizing on-site wind and solar for data center power supply not only lowers carbon emissions but also mitigates Scope 2 water consumption risks. (19,20) Given the complexity of grid systems and renewable variability, it is crucial to evaluate how AI-focused data centers, on-site renewables, and grid infrastructure can be effectively integrated to ensure long-term sustainability...

Well, and if the wind doesn't blow and the data center is needed or night or covered with snow, there's, um, as always, batteries and "percent talk."

By employing an optimal cost-efficient on-site renewable energy adoption strategy to achieve a 50% reduction in carbon emissions, the estimated cost per unit of AI computing emission reduction ranges from $107 to $500 per ton. These optimal cost values are significantly influenced by the local availability of renewable energy resources. The findings offer valuable cost benchmarks for meeting industry decarbonization targets, such as those set by the SBTi, and underscore the need for strategic investment and siting decisions to capitalize on low-cost opportunities.

Data centers are designed to run 24/7 365 days a year. Who's fooling whom?

I am, I note, a supporter of computational power, and I believe that the use of data centers is inspiring a good strong look at sustainable energy of which there is one, and only one, type, nuclear energy.

Have a nice day tomorrow.