AI’s Compute Boom Is Testing the Power Grid in North America

The accelerating adoption of artificial intelligence is driving an unprecedented surge in electricity demand, as data centres running large language models and other AI workloads consume power at growing rates. Generative AI systems such as OpenAI’s ChatGPT and services like Perplexity are computationally intensive, with each query or task drawing significantly more energy than traditional online operations. A single prompt to ChatGPT has been estimated to consume about ten times more electricity than a standard Google search. While individual AI queries are minour, the aggregate impact is enormous: one analysis suggests that ChatGPT’s processing of 78 billion prompts in a year uses around 226.8 gigawatt-hours (GWh) of electricity, costing nearly $30 million in power and requiring substantial water for cooling. This insatiable computing appetite is forcing tech companies and power providers to confront a new reality - the growth of AI may soon be limited not by chips or data, but by the capacity of the electrical grid.

The Soaring Energy Needs of AI Compute

Across the globe, data centers already consume an estimated 415 to 460 terawatt-hours (TWh) of electricity per year, which is roughly 1.5–2% of global electricity consumption. AI is set to be the primary driver of future growth. The International Energy Agency projects that by 2030, data centre electricity use will more than double to approximately 945 TWh annually, roughly equivalent to the power used by Japan, with AI-related workloads accounting for the majority of that increase. In advanced economies, data centres are expected to contribute over 20% of all new power demand this decade. Nowhere is this trend more pronounced than in the United States, where running servers for AI is on track to consume more electricity in 2030 than all US heavy industries (steel, cement, chemicals, etc.) combined. Data centers are poised to account for nearly half of the US electricity demand growth through 2030, a significant shift that reflects the rise of AI from a niche technology to a ubiquitous utility.

This AI compute boom is evident in the soaring power usage of tech companies’ cloud infrastructure. The data centres of Google consumed over 30.8 TWh of electricity in 2024, up 113% from 2020. Such growth far outpaces overall power demand trends. Industry analysts at Goldman Sachs warn that US data centers could draw 8% of the nation’s electricity by 2030, nearly triple their share in 2022 – a rate of increase “not seen in a generation”. Even Canada, with its smaller tech sector, could see data centers account for roughly 14% of national power capacity by 2030 if all proposed projects proceed. Importantly, AI-specific servers (such as those used for model training and inference) are a growing subset of this footprint. Researchers estimate that AI workloads already account for 15-20% of data center energy use today, and their share is rising rapidly.

Grid Strain: When Compute Demand Meets Finite Power Supply

This explosive growth in energy-hungry computing is straining electrical grids and infrastructure planning. Utilities and grid operators are scrambling to keep up with the increasing demand. In the US, forecast models for electricity demand have been sharply revised upward to account for a surge in new data centers. Grid planners nearly doubled their five-year load growth outlook once surging data center demand (driven by generative AI) was factored in. Between 2024 and 2026 alone, data centers are expected to contribute approximately one-third of the increase in US power consumption. Every data center is essentially an industrial-scale electricity customer, typically with a capacity of 200-500 megawatts per large facility, enough to power a small city. The heat generated by thousands of servers also necessitates intensive cooling, adding to energy usage (and even significant water consumption for cooling towers). A recent study found that a single 20-50 question conversation with an AI like ChatGPT can indirectly consume around 500 millilitres of water for cooling, equivalent to a typical bottle of water, due to the heat generated by its computations. Multiply such interactions by millions of users, and the resource burdens mount quickly.

A central data center in Phoenix, Arizona (Iron Mountain). Such facilities require a substantial amount of electricity, and clusters of data centers can strain local power infrastructure. In fast-growing hubs like Phoenix, now one of the largest data center markets in the US, officials are balancing the economic benefits of new data centers against concerns over grid capacity and water use.

Around North America, pockets of intensive data centre growth are testing the limits of local grids. One prominent example is Northern Virginia, which has the world’s highest concentration of hyperscale data centers (housing about 35% of global cloud capacity). The regional grid operator, PJM, has warned that “unprecedented data center load growth” in northern Virginia could soon soak up all remaining transmission capacity. Planned facilities in that area would require the equivalent of several new large power plants to supply them. In other words, the cluster of server warehouses springing up in Virginia’s “Data Center Alley” is outpacing the ability to build new power generation and grid infrastructure. Similar stories are unfolding elsewhere: Phoenix, Arizona, now the second-largest data center hub in the US, risks overtaxing its electrical grid (and depleting limited water supplies) as more server farms come online. In Silicon Valley and Dallas, the high concentration of data centers has prompted utilities to upgrade substations and feeder lines. Even in Canada, Alberta, Quebec, and Ontario face difficult choices as proposals for new data centers compete with the electrification of transportation and heavy industry for grid capacity.

The fundamental challenge is that power infrastructure expands much more slowly than AI demand. Building a large data center typically takes 1-2 years, but upgrading grid capacity, including new transmission lines, transformers, or power plants, can take 5-10 years or more due to regulatory, physical, and financial hurdles. As Brian Venturo, co-founder of cloud provider CoreWeave, explained, a typical substation in an industrial area might supply a total of 30 MW, with perhaps 5 MW allocated for a data center and the rest serving factories and businesses. Now, AI data center operators are asking for 500 MW in the exact location. “I want 500,” Venturo recounts of current requests – a hundred-fold jump that necessitates entirely new substations and high-voltage lines. Critical equipment, such as grid transformers, often must be ordered years in advance, and skilled labour to construct high-voltage infrastructure is in short supply. In short, the surge in demand is outpacing the speed at which the grid can be expanded or upgraded, creating a potential bottleneck for the AI industry.

Industry Responses and the Future of Sustainable Compute

Facing these constraints, the AI and cloud computing industries are rapidly adjusting their course. Power has become a top strategic priority for companies like OpenAI, Google, Microsoft, Amazon, and challengers in the AI space. This year has seen a flurry of deals and investments aimed at securing long-term energy supply for AI workloads. In May, Microsoft signed one of the world’s largest-ever corporate clean energy agreements, partnering with Brookfield to guarantee renewable power for its data centers. Google is developing software to schedule AI tasks in regions at times when renewable energy is abundant (for instance, shifting some computations to locations with strong midday solar output or nighttime wind). Perhaps most eye-catching, cloud providers are turning to nuclear energy for a reliable, carbon-free baseload: Microsoft is helping restart a shuttered atomic reactor at Pennsylvania’s famous Three Mile Island plant specifically to feed its AI data center needs. Around the same time, Amazon struck agreements to develop small modular nuclear reactors (SMRs), and Google invested in a startup building advanced nuclear plants (Kairos Power) – both securing future dedicated power sources for their server farms. As Oracle’s chairman, Larry Ellison, remarked on this trend, “I mean, my God, nuclear reactors to run data centers – are you kidding me?” It may sound drastic, but it highlights the critical importance of reliable power to the AI ecosystem’s continued growth.

Major data center operators are also doubling down on efficiency and grid support measures. Many hyperscale facilities are pushing the limits of cooling technology and design to improve Power Usage Effectiveness (PUE), ensuring that as much electricity as possible is directed toward computing rather than overhead. Efforts are underway to use AI itself to optimize data center cooling and workload scheduling for energy savings. Some data centers are becoming more flexible grid citizens: enrolling in demand response programs, installing on-site batteries, and even leveraging their backup generators to support the public grid at peak times. These steps can help alleviate stress on the system. However, efficiency alone won’t offset the sheer scale of new demand. Even with aggressive optimizations, experts predict global data center energy use could reach ~1,000 to 1,300 TWh by 2030 under high-AI growth scenarios – more if unforeseen AI breakthroughs ignite further compute hunger. This means power sector investment must rise in parallel. Utilities in the US and Canada are now planning substantial new generation, from wind farms to gas peaker plants, explicitly citing data center and AI loads as justification. Policymakers are beginning to incorporate “AI load” into their long-term resource plans. In effect, the trajectory of AI development is now intertwined with energy infrastructure planning.

Looking ahead, scaling compute sustainably will require close collaboration between the tech industry, utilities, and governments. Investors in both sectors are taking note: venture capital and infrastructure funds are increasingly evaluating power availability as a key opportunity for AI startups and cloud expansions. For tech investors, supporting innovations in energy-efficient chips, cooling systems, and distributed computing could help mitigate the bottleneck. For energy investors and utilities, the AI boom presents an opportunity to secure new high-value customers – if they can build capacity in time. Notably, some analysts suggest that if managed wisely, AI’s growth could even spur beneficial grid upgrades and renewable projects that ultimately strengthen the overall system. For example, the steady 24/7 demand of data centers can complement intermittent solar and wind resources, potentially enabling more clean energy on the grid if properly orchestrated.

The coming years will test whether North America’s power systems can keep up with the AI revolution. Thus far, the trend is clear: compute and power demand are converging, and electricity has become the “fuel” for AI progress. Suppose the grid can keep up through new energy generation, smarter distribution, and perhaps game-changing technologies like small nuclear reactors (SMRs). In that case, AI’s proliferation can continue unhindered (and with a smaller carbon footprint). If not, the expansion of AI may slow or shift to regions with more available power. The limits of AI are no longer just algorithmic or hardware-based – they are increasingly tied to the capacity of our energy infrastructure. For tech and energy leaders alike, ensuring the lights stay on for the AI era is now a shared imperative.