The Surging U.S. Demand for Compute and Power — and How Galaxy Plans to Deliver

Introduction

The exponential growth in AI usage, alongside rising industrial energy needs, is fueling an unprecedented demand for power. However, the aging U.S. grid is struggling to deliver the high-density load required by modern industries. Most power transformers, circuit breakers, and transmission lines are over 25 years old, with many exceeding 30 years and wearing down fast, according to the U.S. Department of Energy’s infrastructure review. Significant transmission upgrades are needed to support the additional load coming online. The Electric Reliability Council of Texas (ERCOT), which manages the state’s power grid, estimates that meeting ever-growing power demand there will require a roughly $30 billion investment to lay thousands of miles of long-distance transmission lines.  

The U.S. relies on multiple independent electrical grids, including PJM, MISO, and ERCOT. Each operates under a different regulatory framework, making the development and interconnection of large-load energy campuses complex. As grid operators assess future supply, demand, and transmission constraints, interconnection queues continue to grow, making sites with approved access to large amounts of power rare and highly coveted. In most markets, the wait time to plug into the grid is now four years, according to real estate brokerage JLL.  

This report examines in depth the demand projections for data centers, the AI trends driving them, and the strong position of Galaxy’s flagship West Texas data center campus, Helios, to power the next generation of AI technology. 

Data Center Demand Projections and the State of the Market 

Several experts have projected a massive scale-up in demand for power from data centers. First, the U.S. Department of Energy (DOE) projects that data center power consumption will double or triple by 2028, rising from 176 terawatt-hours (TWh) in 2023 to between 325 TWh and 580 TWh in 2028.  Accordingly, McKinsey projects the capacity demanded from U.S. data centers to rise from 25 gigawatts (GW) in 2024 to more than 80 GW by 2030. For Texas alone, ERCOT projects data center energy demand to increase by 24 GW by 2031, with total load growing to 145 GW. This is a significant increase from ERCOT’s 2025 peak power projection of 87 GW. 

AI companies — such as hyperscalers, neoclouds, and AI labs — are rapidly leasing and building data center capacity to meet escalating computational needs. Despite skepticism among analysts that data center capex projections would remain high following the DeepSeek AI model’s apparent efficiency gains, hyperscalers including Amazon, Oracle, Google, Microsoft, and Meta have reiterated or increased their 2025 capex commitments. Collectively, these firms are expected to spend more than $360 billion, signaling sustained high demand for high-density data center space.  

We believe that the monetization potential of AI companies is fundamentally linked to the scale of data center infrastructure supporting their workloads. Greater infrastructure capacity – measured in critical IT load – enables greater computing power, which in turn allows for the development of more advanced AI models and supports more inference workloads, driving substantial revenue growth.  

For instance, training state-of-the-art models requires thousands of graphics processing units (GPUs), while large-scale inference allows companies to serve more users, enhancing monetization. This point is further reinforced by NVIDIA CEO Jensen Huang, who holds that the revenue-generating potential of a GPU-hosting “AI factory” is proportional to the number of tokens (units of data) it can generate and the speed at which they are produced (tokens per second). Consequently, access to large-scale power and suitable land for AI and high-performance computing (HPC) data centers will be extremely valuable as AI adoption continues to grow. As McKinsey notes, “If you’re looking to add 100, 200, 500 megawatts—or even a gigawatt—of power at a single campus, there aren’t that many places where you can just attach that to the grid.” The consulting firm’s research further emphasizes that “substantial new grid infrastructure will be needed to support the rising AI economy.”  

On the supply side, data center availability in primary markets has become severely constrained. According to real estate brokerage CBRE, “demand is outpacing supply in nearly every major market, driving elevated lease rates and prompting rapid market expansion,” with North America’s data center vacancy rate hitting an all-time low of 1.6% as hyperscale and AI occupiers race to secure power and capacity years in advance. As a result, over the past several months, new data center construction has started expanding outside traditional hubs such as Northern Virginia, to more rural areas including West Texas, where there is more power available. We believe ERCOT’s grid expansion initiatives and Texas’s business-friendly environment will differentiate the state as one of the fastest-growing regions for large-load consumers, positioning Texas to become a leading hub for data centers in the U.S. On April 24, the Public Utility Commission of Texas approved plans to construct the first 765 kilovolt (kV) extra high voltage transmission lines in the ERCOT region to support “the growing electricity needs of West Texas communities.” 

The Helios campus has an approved capacity of 800MW and an additional 2.7 GW of expansion capacity under study. At a total of 3.5 GW of potential gross capacity, Helios will be one of the largest single-site data center campuses in the world. This substantial capacity presents significant opportunities for AI companies to advance the development of smarter and more capable models. The chart below shows that the total potential critical IT capacity at Helios is larger than those of several primary and secondary data center markets combined.   

Key AI Trends Are Accelerating Power Demand  

This section explores how AI scaling laws are driving the demand for power and compute resources across the training and inference stack. Three distinct scaling laws govern AI performance: pre-training scaling, post-training scaling, and test-time scaling (also known as inference-time scaling). These scaling laws, widely validated and agreed upon across the industry, can be individually optimized to boost model performance, unlocking higher levels of intelligence—from perception-based AI and generative models to advanced agentic capabilities. 

Trends in AI Training 

One of major trends we’ve been observing in AI is that frontier models are becoming bigger, more computationally intensive, and more demanding of energy. According to the Stanford AI Index Report, the training compute required to train frontier models has doubled every five months, dataset sizes have doubled every eight months, and the power required has doubled annually. These findings are supported by data from Epoch AI, which indicates that the training compute required for frontier models grows at a rate of 4.2x per year, available training data expands at 3.5x per year, and the power needed for training grows at 2.1x per year.  

Furthermore, Epoch AI’s analysis attributes the 4.2x annual training compute growth in frontier models to three key factors: a 1.7x increase in training hardware quantity (more GPUs), a 1.5x increase in training duration, and a 1.4x improvement in hardware performance. The study concludes that training compute growth is driven by larger clusters, longer training duration, and better hardware performance.

Furthermore, Epoch AI notes that once models are trained beyond 10^24 total floating-point operations (FLOP), performance continues to improve with additional training compute. Across most benchmarks, accuracy rises by roughly 12 percentage points for every 10x increase in compute. On the GPQA Diamond benchmark—which evaluates advanced language models on their ability to answer challenging PhD-level scientific questions in chemistry, physics, and biology—each 10× increase in compute yields about a 14-percentage-point gain in accuracy, further underscoring the importance of scaling pre-training compute. 

Meanwhile, the power draw required to train frontier models has been rapidly increasing, with some recent models such as Grok 3 estimated to have required over 100 MW of power capacity to train. This is still a relatively small amount of capacity compared to the size of certain mega-campuses like Helios, showing the enormous potential capabilities that could be unlocked by training multi-hundred MW models. 

At Helios, we are excited to play a major role in scaling AI initiatives. As clusters grow larger and more energy-intensive, the Helios data center campus, with reliable access to power at scale, a vast land footprint, ample water supply, and low latency, is ideally equipped to meet the power demands of training next-generation frontier models. Our infrastructure is purpose-built to support billion-dollar clusters, delivering scalability and reliability for cutting-edge AI workloads. 

Trends in AI Inference 

Alongside growth in AI training, several key innovations in the inference stack have emerged that have fueled an increase in demand for compute resources. 

Reasoning workloads, unlike traditional LLM prompting, involve complex, recursive inference processes that enable models to solve problems by iteratively analyzing and synthesizing information. These models, often designed for such tasks as logical reasoning, decision-making, or code generation, break down queries into multiple steps, evaluating intermediate results and refining outputs through repeated computations. This recursive nature demands significantly higher computational resources, as the model processes multiple iterations of data, context, and logic, leading to exponentially greater compute and power requirements compared to simpler, single-pass LLM queries. Reasoning models have significantly increased the demand for compute at the inference level, requiring significantly more tokens than traditional large language model (LLM) prompting due to their complex recursive inference processes.  

A prime example is the DeepSeek-R1 model, a 671-billion-parameter mixture-of-experts (MoE) reasoning model optimized for math, coding, and logical tasks. DeepSeek-R1 leverages inference-time scaling, also known as test-time scaling or AI reasoning, where additional computational resources are allocated during inference to evaluate multiple potential outcomes. This approach has been shown to produce highly accurate results, particularly in complex tasks including generating optimized GPU attention kernels — a benchmark that evaluates numerical correctness (accuracy) and performance, including speed and resource efficiency, of attention mechanisms. This test enables AI models to optimize for the most relevant parts of query entries when reasoning. Below is a snapshot of the reasoning framework used by DeepSeek-R1 provided by NVIDIA. 

The reasoning framework applied to DeepSeek-R1 significantly improved the model’s accuracy. As shown in the chart below, the percentage of problems solved follows a logarithmic distribution; performance improves as inference time extends.  

Notably, inference-time scaling allows models to allocate more compute during inference, enabling exploration of multiple solutions, which results in optimized kernels but significantly increases compute consumption. This approach has shown the potential to enhance model intelligence, significantly increasing computational demand during inference and amplifying the need for robust, scalable power sources. 

Other Trends Fueling Demand for Power  

DeepSeek's recent breakthroughs in software and cost efficiency improvements have sparked debate over their implications for compute demand and data center growth. When combined with ongoing improvements in hardware performance, these developments have led some in the market to speculate that these efficiencies will temper power and data center demand. In reality, these efficiencies are unlikely to offset the rising compute and energy requirements driven by model complexity, scale and increasing AI adoption.  

According to Stanford University’s 2025 AI Index Report, leading AI companies have recently shifted their focus from solely scaling models during pre-training to now prioritizing data efficiency during training. This enables smaller models to achieve high performance with less data and lower training costs by training them on smaller high-quality datasets. The chart below compares the smallest models, measured by the number of parameters, capable of scoring above 60% on the Massive Multitask Language Understanding (MMLU) benchmark. 

The trend toward compact, high-performing models offers significant benefits, including faster and cheaper inference. This trend reduces the computational and financial barriers for users and organizations to run leading models. Combined with growing datasets, these efficiency gains are expected to enable the development of even more powerful and smaller models in the future, further democratizing AI development and increasing deployment across diverse use cases.  

As AI models become more efficient and compute resources grow more accessible, inference costs will continue to decline. This declining cost for compute will accelerate AI adoption, fostering new innovative use cases — and increasing demand for compute. The surge in compute demand following efficiency breakthroughs like DeepSeek exemplifies this trend. According to McKinsey, preliminary analysis of the compute power demand curve agrees that “efficiency gains may not substantially impact overall compute power demand over the long term.”  

The chart below illustrates inference price trends of some AI models across select benchmarks, showcasing improvements in efficiency and cost reductions that we believe are healthy for the industry and will drive AI adoption at an accelerated pace.  

Efficiency improvements that reduce compute costs and enable novel applications underpin the Jevons paradox, where these gains drive increased AI adoption and resource consumption. Research from Artificial Analysis shows that, although efficiency gains have lowered compute requirements for specific tasks, compute demand has outpaced these savings. Enhanced algorithms and training data enable smaller models to achieve greater intelligence, reducing compute needs by approximately one-tenth. Additionally, software and hardware efficiency improvements have cut compute requirements and costs by one-third each. However, scaling laws have driven larger models to require up to five times more compute per query. Innovations, such as reasoning models, are estimated to consume on average 10 times more tokens than single queries (up to 20 times in some cases), and the AI agent revolution is projected to generate 20 times more requests per use. This trend mirrors the Jevons paradox, where decreasing costs lead to significant increases in compute demand and, consequently, power consumption.

The demand for compute in AI workloads continues to surge. Breakthroughs such as DeepSeek and inference-time scaling are accelerating advances in model efficiency, complexity, and usage, driving up compute needs at an exponential pace. This rising demand is outstripping efficiency gains, fueling greater requirements for larger and more power-dense data centers.  

Helios is Positioned to Play a Critical Role in Training and Servicing Next-Gen AI Models

Helios is uniquely equipped to address the escalating demand for AI compute infrastructure. The Helios campus has the rare characteristics – reliable power at scale, high bandwidth and low latency fiber access, vast land and water resources, a skilled labor force, and proximity to leading academic institutions – that enable it to support substantial data center capacity growth.  

• Massive Power Capacity: Helios boasts 800 MW of approved power and an incremental 2.7 GW under study, totaling 3.5 GW of capacity at a single campus that rivals entire major data center markets like Northern Virginia or Atlanta. This extensive energy supply powers large-scale AI data centers and ensures scalability for future AI demands. The ability to train models at a single site that offers this much power can lead to significant efficiency gains and reduce the cost to train, thereby enhancing the economics of model development and monetization for AI companies. Moreover, concentrating massive power capacity at scale allows tenants to optimize infrastructure utilization, streamline operations, and ultimately lower the total cost of ownership across their deployments. 

• High-Speed Connectivity:  With dedicated fiber delivering 10–15 milliseconds (ms) of latency between Helios and Dallas, supported by two diverse paths designed for N+1 redundancy, Helios provides reliable, high-bandwidth performance. This low-latency design enables AI training and inference workloads, ensuring scalability and long-term resilience. 

• Expansive Land: With control of over 1,500 acres of flat, contiguous land, Helios is poised to grow to be one of the world’s largest data center campuses. 

• Sustainable Water Supply: Located above one of the country’s largest aquifers, Helios secures reliable water access to meet the cooling demands of ancillary data center equipment. 

• Skilled Workforce and Proximity to Academic Institutions: Helios benefits from proximity to a robust labor market in Lubbock, Texas, ensuring access to a skilled workforce capable of supporting advanced data center operations. 

The relentless growth in demand for AI compute shows no signs of slowing down. This sustained demand underscores the need for robust, scalable infrastructure to support increasingly power-hungry AI workloads. Helios is uniquely positioned to capitalize on this trend, with cutting-edge capabilities designed to meet the rising requirements for both training and inference. In addition, its extensive access to energy enables Helios to scale out efficiently. Helios stands ready to lead the charge in powering the next generation of innovation.