AI's Growing Thirst for Power and Water

Global data centers consumed approximately 415 terawatt-hours (TWh) in 2024, representing about 1.5% of the world's total electricity usage. In the U.S. alone, data centers accounted for over 4% of national electricity consumption in 2024, a figure that could climb to as high as 12% by 2028. Projections indicate that by 2030, global data center electricity consumption could double to 945 TWh, an amount equivalent to the current energy demand of Japan. The water required for cooling these facilities is equally staggering. Training OpenAI's GPT-3 alone is estimated to have consumed 700,000 liters of fresh water. In 2022, Microsoft's water usage surged by 34% to nearly 1.7 billion gallons, while Google used 5.56 billion gallons, largely for cooling its data centers. By 2027, global AI-related water demand could reach between 4.2 and 6.6 billion cubic meters, which is more than half the annual water withdrawal of the United Kingdom. Many new data centers are being constructed in water-stressed regions, creating significant operational and reputational risks. Nearly one-third of data centers currently under construction are in areas projected to face greater water scarcity by 2050. This trend is creating a tension between corporate "water positive" pledges, like those from Meta and Amazon Web Services, and the escalating water demands of their expanding infrastructure. In response, a dual approach of efficiency and offsetting is emerging. AI-driven automation is now used to optimize data center cooling and energy consumption in real-time, with Google reporting a 40% reduction in cooling energy by using DeepMind's AI. For unavoidable emissions, companies are increasingly turning to carbon credits and direct air capture solutions. Google has contracted over $100 million in carbon removal credits, and SAP has secured 33,500 tonnes of removal credits through 2034. The resource intensity of AI is also shaping API design and developer practices. An "AI-first" API design philosophy is emerging, prioritizing unambiguous, machine-readable contracts that allow autonomous agents to operate efficiently without human interpretation. This shift requires creating parallel API surfaces built specifically for AI, moving beyond simply documenting existing systems to fundamentally redesigning the data access layer for machine consumption. As enterprises adopt AI, governance frameworks are being established to manage its environmental, social, and governance (ESG) impacts. Regulatory frameworks like the EU AI Act and voluntary standards from organizations like NIST are pushing for greater transparency, requiring companies to disclose the environmental footprint of their AI systems. This is leading to the integration of sustainability criteria directly into AI development and procurement processes. Large-scale AI adoption is forcing enterprises to re-evaluate their infrastructure strategies, as seen in Lumen's partnership with Anthropic to provide high-capacity networks for AI workloads. Case studies from companies like Walmart show that AI-driven optimizations in logistics can lead to significant cost savings and emission reductions, saving approximately $75 million and cutting nearly 72 million pounds of CO₂ in one fiscal year. This demonstrates a tangible link between operational efficiency, cost reduction, and sustainability goals in enterprise AI deployment. The development of agentic AI workflows, which can autonomously perform tasks, further intensifies the need for sustainable infrastructure and efficient API design. These systems require low-latency, high-bandwidth connectivity and rely on well-structured, predictable APIs to function effectively. As enterprises experiment with these autonomous patterns, the underlying resource consumption of each API call becomes a more critical factor in both cost and environmental impact.