ARTICLE

Powering the Future of AI: The Data Center Energy Crunch

AI has an energy problem. Here’s what’s behind it.

by:
Energize Capital, Contributions from Alexandra Lum

AI is changing the technology sector as we know it. Within climate solutions, artificial intelligence has reached far beyond ChatGPT, improving wind farm maintenance through AI-powered visual recognition technology, streamlining residential solar permitting, identifying key battery materials through mapping technologies, and so much more. But despite all the benefits of these new technologies, AI also poses a key challenge: it requires an immense infrastructure investment in data centers, and these data centers require a massive amount of power.

Our latest Deep Dive Report explores the causes and implications of this unprecedented power demand, from site selection constraints to operational efficiency challenges. As investors in the energy transition, we see this evolving landscape as a catalyst for breakthrough innovations in renewable energy and efficiency technologies—creating new opportunities for climate tech to shape the future of computing infrastructure. Watch for upcoming Energize research examining the emerging solutions meeting these challenges head-on.

A New Era for the Power Grid

The scale of data center power demand is unprecedented. Third-party data centers contracted last year alone will require 5.8 GW of new power—equivalent to adding another New York City to the U.S. grid. Individual facilities have grown in size from 30 MW a decade ago to between 500 MW and 2 GW today, while total capacity under construction has surged 10-fold since 2019 to over 3,000 MW. This surge marks a new era: after years of flat electricity demand growth, AI's computing needs could drive U.S. total power demand up by 15-20% over the next decade.

This demand surge has revealed vulnerabilities across the entire energy ecosystem, which is already strained by aging transmission infrastructure, extreme weather conditions, and increasingly complex energy generation patterns with the expansion of renewables. For data center operators, energy is their largest operational expense, driving them to explore new solutions across development, operations, and compute management to access reliable energy sources and optimize their energy consumption. While AI has tremendous potential to transform countless industries, its future hinges on solving the very energy problem it creates. The data center energy challenge requires massive innovation, and we believe climate software is poised to be a key contributor.

The Evolution of the Modern Data Center

Data centers themselves are not new; they have served as the backbone to search engines, cloud storage, and machine learning tools across the computing industry for decades.  But what is new is the advent of generative AI and consumer-facing large language models (LLMs) like ChatGPT, Claude, Perplexity, or CoPilot, which have captured the world’s attention through their vast flexibility and wide applications, driving the AI arms race to gain market share and find business opportunities.  

As the structure and applications of AI have evolved, data centers have evolved as well:

1. Data centers are upscaling quickly to meet a new era of data.  As large-scale, pre-trained "foundation" models take over the AI landscape, software companies have increasingly relied on vast data sets to train their systems. This has unlocked the usage of a near-exponential explosion of data. The analytical structures built upon these models are becoming ever-more advanced and complex. To handle this new load, data centers have needed to scale computing capacity and size.

2. Today’s AI is significantly more energy-intensive. Due to the complexity of today’s models, the power needed to run LLM-based queries is significantly higher, with a ChatGPT query requiring six to 10 times the power of a Google search. Where data centers in 2022 typically required 200 MW, today, hyperscalers like Microsoft and Google are signing and searching for sites that can support 1 GW—akin to the energy use of San Francisco.

3. New AI use cases are changing data center placement. While existing data centers are being retrofitted to accommodate the AI boom, new construction is also at an all-time high. The locations of these construction projects are also starting to diverge based on use case. Where AI training—which is dependent on scale, specialized hardware, and network bandwidth—occurs on massive data center campuses in remote areas, real-time user interactions—reliant on low latency and memory bandwidth—are being powered by data centers near major metro areas to maintain a fast response rate.

Data center developers, operators, and software companies are now scrambling to adapt to these changes, searching for solutions to drive sustainable scale, meet power requirements, and improve user experience.

Identifying Opportunities: Hurdles & Solutions

Challenge 1: Data Center Development

As hyperscalers compete intensely to win a share of the AI market,  developers are racing to bring new projects online. But in order to do so, they must handle the complex process of site selection and the growing hurdle of securing reliable power access.

The first challenge begins with site selection. Developers face a range of obstacles including land ownership, zoning requirements, resource availability, environmental considerations, and—most critically—power availability. Until recently, they've relied on legacy solutions and disparate data sets to find available land near adequate power sources, working with traditionally outdated and fragmented information about substation capacity, available power, and land ownership. This manual process has created a compelling opportunity for software platforms to streamline site selection and transform the traditionally time-intensive and costly process.

Once a site is selected, data center developers face their next hurdle: obtaining sufficient power supply. Though some facilities incorporate onsite generation and storage systems, these typically can't meet the massive energy demands of large-scale data centers, causing developers to turn to the grid for secure, reliable baseload power. The issue, however, is that our aging energy infrastructure faces severe limitations on transformer and grid capacity, causing a major interconnection backlog. An influx of power generation sources, combined with these data center requests, have caused the interconnection queue to increase sharply—in 2023, requests increased by 30% to reach 2.6 TW. As a result, developers today can expect a wait of three to five years to connect to the grid. In order to keep up with the pace of AI growth, the sector will be looking to new tools to streamline infrastructure development, permitting, and interconnection processes.

Technologies We’re Watching: data aggregators for site selection, permit application and management platforms, grid optimization and interconnection software, power flow modeling and simulation software

Challenge 2: Data Center Operations

Once a data center has been brought online, the challenge for the rest of its useful life is operational efficiency. Effectively managing computing infrastructure and energy use is essential to recoup capital expenditure, maintain uptime, extend asset lifespans, and improve operational reliability. Power remains central to this challenge, as energy, IT load, and cooling represent an operator's largest operational costs.

The challenge today lies in finding new ways to drive efficiency. While hyperscalers like Google, Microsoft, and AWS have achieved industry-leading Power Usage Effectiveness (PUE) ratings of around 1.1, the broader industry has plateaued, with PUE hovering between 1.5 and 1.6 over the past six years. This gap indicates that while further efficiency gains are possible, the industry is struggling to achieve them. Beyond PUE, data centers also face challenges with Power Capacity Effectiveness (PCE), which measures how effectively a facility utilizes its total power capacity. PCE remains low across the industry, largely due to practices that prioritize resilience and uptime over efficiency: stranded capacity from unused installations, over-provisioning of resources, fragmented facility layouts, and outdated equipment.

As data centers convert to high-performance computing to reach AI goals, both PUE and PCE have come under increased scrutiny, challenging operators to find new sophisticated innovations to offset power-hungry software. In addition to updated legacy infrastructure and improving fragmented project layouts, digital solutions have emerged to optimize cooling systems, coordinate workloads, and improve overall data center design, focusing specifically on operational efficiency.

Technologies We’re Watching: AI-enabled industrial control systems, data center infrastructure management (DCIM), demand flexibility and response platforms, water and cooling optimization systems, autonomous control solutions

Challenge 3: Compute Management

In addition to managing the physical elements of a data center, software companies, including hyperscalers, are working to make the computational processes themselves as efficient as possible to save energy and costs. Foundational computational processes—such as hardware orchestration, storage, processing, and data transfers—consume substantial energy and require innovative breakthroughs to improve efficiency.

While machine learning was initially built for small, fine-tuned models tailored to specific use cases, the advent of LLMs has now caused a dramatic shift to structures that are more versatile, robust, and computationally intense—requiring more power and representing a larger carbon footprint. Programs like ChatGPT require an extremely broad data set and increasingly complex computational processes; training ChatGPT-3 alone consumed 1,387MWh and emitted 502 tons of CO2—equivalent to 130 U.S. homes' annual emissions. While individual AI inference (response generation) uses less energy per request, its high usage volume dominates energy consumption, accounting for 80-90% of AWS's machine learning computing demand and 60% of Google's machine learning energy use.

Already, software models have made large strides in improving computational capacity, while also optimizing compute-to-power per server system ratios. Even so, efficiency gains have lagged. Since the advent of the chip industry, computing power has followed Moore’s Law—the observation that computing power would double roughly every two years while costs remained constant (including energy and cooling costs). This progress has now slowed, partially due to physical limitations of chips, the evolving complexity of these systems, and inconsistent utilization rates of Graphics Processing Units (GPUs).  

To achieve efficient, profitable operations, hyperscalers must now find new ways to streamline their computational processes—through compression, data management, and processing—to not only create quicker models but to cut energy costs and reduce carbon emissions.

Technologies We’re Watching: distributed computing platforms, cloud-based data warehouses, distributed data storage, edge computing systems, vertical AI applications for energy and industrials

The Technology Opportunity

The AI revolution presents a massive global opportunity, ushering in a new era of computational ability that can prove to be life-changing to science, technology, the environment, and beyond. But to meet this new era, data centers face a difficult task: providing massive computational scale while maintaining profitability and reducing carbon footprint. Across the entire data center lifecycle, from site selection to design to operational management, there will emerge new opportunities for new tools, smarter analysis, and innovative solutions. Climate software is uniquely positioned to seize this opportunity, bringing deep expertise in grid-related development, efficiency optimization, and edge computing to solve data centers' most pressing challenges. The race to an AI-driven future is on—and it hinges on revolutionizing how we generate, move, and manage our energy resources.