AI Data Centers Are Secretly Exploding Your Electricity Bills

Maryland residents recently received a notice from their utility provider. Their monthly electric bills will increase by $14. Commercial customers in the Baltimore Gas and Electric zone face a 16% hike, adding roughly $170 to their monthly overhead. Further south, Dominion Energy projects residential bill increases in its service area could hit $255 per month by 2035. This is not standard inflation. This is a direct physical tax levied by artificial intelligence.

We track these developments to see the potential world in the next decade. The internet is flooded with shallow influencer takes regarding the cognitive capabilities of artificial intelligence. We ask a different question. When hyperscale technology companies possess near-infinite capital to build these systems, who actually pays the physical cost of their electrical grid expansion?

The PJM Interconnection operates the largest wholesale electricity market in the United States. It covers 13 states and the District of Columbia. PJM runs a forward capacity market to ensure enough power plants are available to meet peak demand three years in the future. In its 2026/2027 Base Residual Auction, PJM secured 134,311 megawatts of generation. The price cleared at the approved regulatory cap of $329.17 per megawatt-day for the entire footprint. This is a 22% increase from the previous year. In the prior auction, the Dominion zone, which services the incredibly dense data center hub of Northern Virginia, cleared at $444.26 per megawatt-day.

The independent market monitor for PJM explicitly stated that data center load growth is the primary reason for these market conditions. The monitor attributed a combined increase of $16.6 billion in capacity market revenues across two recent auctions directly to existing and forecast data center load. Data centers consume capacity faster than utilities can physically build it. The resulting scarcity drives up the marginal cost of every electron for every other participant in the economy.

A standard Google search uses approximately 0.03 watt-hours of electricity. An independent assessment by Epoch AI confirms a standard ChatGPT text query consumes roughly 0.34 watt-hours. That is a factor of ten increase. Advanced reasoning models consume drastically more power. Models like OpenAI's o3 require between 7 and 40 watt-hours per query. An academic study estimates that the GPT-4.5 model uses approximately 30 watt-hours for long prompts containing 7,000 words of input and 1,000 words of output. Image generation demands 20 to 40 times more energy than text. Video generation requires thousands of times more energy.

The math is terrifying at scale. ChatGPT possesses over 900 million weekly active users worldwide and processes roughly 2.5 billion requests daily. If global users execute one trillion standard queries a year at 0.34 watt-hours, the baseline inference energy cost is 340 gigawatt-hours. If just a fraction of those queries shift to advanced reasoning models drawing 30 watt-hours, the power demand explodes into the terawatt-hour range simply for basic text generation. Data centers consumed approximately 4.4% of total United States electricity in 2023. The Lawrence Berkeley National Laboratory projects this figure will reach between 6.7% and 12.0% by 2028. This represents a jump to 580 terawatt-hours. This is a vertical supply shock.

The underlying constraint is physics. For decades, the semiconductor industry relied on Dennard scaling. Formulated by Robert Dennard in 1974, the principle dictated that as transistors shrank, their power density remained constant. Engineers lowered the voltage while packing more transistors onto a chip. Performance increased while power draw remained flat. Dennard scaling ended roughly twenty years ago. Transistors became so small that current leakage caused unacceptable heat and reliability issues. Engineers can no longer reduce voltage proportionally.

We hit the wall of heat. Every new generation of artificial intelligence chips demands more absolute power. NVIDIA's H100 chip represented a massive leap for its Hopper architecture. The subsequent Blackwell architecture pushes the absolute limits of thermodynamics. The DGX B200 system doubles GPU-to-GPU bandwidth via NVLink 5.0, but the system draws approximately 14.3 kilowatts of power under full load. The raw speed generates intense thermal output.

Data centers must dissipate this heat. The industry measures this using Power Usage Effectiveness. This metric divides the total energy used by a facility by the energy delivered directly to the computing equipment. A perfect ratio is 1.0. Most artificial intelligence data centers operate with a ratio between 1.1 and 1.5. A ratio of 1.4 means that for every 100 watts used by the compute hardware, another 40 watts burn strictly on cooling pumps and facility overhead. The software cannot optimize its way out of this physical trap. Classical scaling laws dictate that training loss predictably decreases as you increase compute. Artificial intelligence has no natural efficiency counterbalance. Without Dennard scaling, more computation means a linearly higher power bill.

Grid operators cannot meet this vertical demand shock with intermittent wind and solar. Data centers require round-the-clock reliable power. They cannot operate on intermittent generation. If a grid operator attempts to meet a projected demand through renewable energy alone, equating 100 megawatts of electricity load with 100 megawatts of renewable capacity is mathematically false. For a data center contracting 100 megawatts of renewable power, 100 megawatts must be available every hour across all seasons. Because of the intermittency of renewable energy sources and seasonal variability, an installed renewable capacity equivalent to the contracted load fails immediately. A hybrid wind-solar power generation unit operates with an average capacity utilization factor of roughly 40%. A grid operator must install a nameplate renewable capacity of at least 222 megawatts to guarantee 100 megawatts of continuous data center output.

Hyperscalers recognize this physical bottleneck. Microsoft, Amazon, Google, and Meta accounted for nearly half of all clean power purchase agreement volumes globally in 2024. They execute a ruthless preemptive strike on available baseload generation. They use pristine balance sheets to lock in decades of guaranteed power. Global corporate clean power purchase agreement volumes actually fell by 10% in 2024. Smaller corporations pulled back due to tariff uncertainty and high prices. This leaves the technology giants to dominate the remaining supply.

Microsoft altered the energy market entirely in September 2024. The company signed a 20-year power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1. Microsoft will purchase the entire 835-megawatt electric generating capacity of the plant when it reopens in 2028. Financial analysts estimate Microsoft will pay between $110 and $115 per megawatt-hour. The average market power price sits around $50 per megawatt-hour. Microsoft is paying a 130% premium simply to guarantee reliable, uninterrupted supply. They value physical availability far above raw unit cost. Amazon executed a similar maneuver earlier in the year. Amazon Web Services purchased a 960-megawatt hyperscale data center campus adjacent to Talen Energy's Susquehanna nuclear power plant for $650 million. Talen supplies direct nuclear energy and expects to earn approximately $18 billion in revenue over the life of the contract.

Hyperscalers operate as synthetic utilities. Goldman Sachs estimates the required industry outlay to cover the United States green reliability premium for data centers will hit $34 billion by 2030. This premium is the cost difference between baseline combined cycle natural gas and green reliable sources like nuclear or solar paired with storage. For hyperscalers projected to generate a combined EBITDA of $1.29 trillion by 2030, this power premium represents just 2.7% of their earnings. They absorb the cost easily. Their adjusted cash return on capital invested drops by less than one percent. They successfully wall off the most reliable, cleanest baseload generation on the continent. We must ask what happens to the rest of the economy. What happens to the businesses that operate on single-digit margins and cannot sign a 20-year agreement at $115 per megawatt-hour?

Heavy industrials operate on massive operating leverage. Electricity constitutes a massive percentage of their cost of goods sold. When hyperscalers drive up wholesale power rates and absorb all cheap nuclear capacity, industrial balance sheets bleed cash. The unit economics are terrible for legacy operators.

We analyze the exposure of global chemical and steel manufacturers to understand the real-world erosion of unit economics. Dow Inc. operates massive chemical facilities. The company notes in its SEC filings that purchased feedstock and energy costs account for a substantial portion of total production costs. Dow purchases natural gas to generate electricity and buys electric power directly to supplement internal generation. The financial burden is already heavy. Dow recorded $823 million in pretax charges for operating pollution abatement facilities in 2024. High energy costs cripple regional competitiveness. Dow points out that Europe is currently the highest cost region in the world for their industry. This depresses their product demand entirely.

Steel manufacturing is notoriously energy intensive. Nucor Corporation relies heavily on electric arc furnaces. Electricity is the primary energy source for their scrap melting process. Nucor's total cost of products sold reached $28.6 billion in 2024. Any fluctuation in electricity pricing immediately impacts this massive line item. Nucor explicitly acknowledges that increases in energy costs that are not applicable to foreign competitors materially affect their business. To mitigate this, Nucor is investing in next-generation nuclear technology companies. They are attempting to secure the same clean baseload power that the hyperscalers are currently monopolizing.

The impact of power costs is highly visible in the European and Indian operations of Tata Steel. The company reported a consolidated turnover of INR 2,29,171 crore for the 2023-2024 financial year, but recorded a reported net loss of INR 4,910 crore. Tata Steel has a standalone energy intensity of 24.55 gigajoules per tonne of crude steel. Their total power and fuel expenses for the standalone entity reached INR 8,010 crore. In the United Kingdom, Tata Steel struggled with operational issues at the Port Talbot site. The deterioration of aging assets led to higher costs through increased energy usage. The company reported a restructuring and impairment disposal cost of £619 million. This included a £260 million full impairment of heavy end assets at Port Talbot. High power costs force heavy industrials to write off billions in legacy infrastructure.

If hyperscaler demand pushes baseload pricing up by 40%, the resulting cost of goods sold expansion destroys basis points of EBITDA. A steel manufacturer like Nucor, with electricity comprising 8% to 12% of its cost of goods sold, would face an estimated 230 basis point loss in EBITDA under a 40% price shock. A chemical producer like Dow faces a 290 basis point erosion. Heavy industrials cannot simply pass a 40% power price increase down to consumers without destroying demand. The steel and chemical markets are globally competitive. If United States or European power prices spike due to local data center density, foreign competitors utilizing cheaper state-subsidized coal power will capture market share immediately.

The physical grid is breaking under the weight of this transition. We track the Federal Energy Regulatory Commission interconnection queue to measure systemic friction. The numbers indicate a complete regulatory failure. Between 2021 and 2024, the United States interconnection queue grew from 1,400 gigawatts to over 2,000 gigawatts. As of early 2026, the backlog sits at 2,600 gigawatts. The median time to achieve commercial operation approaches five years. Project viability is collapsing. The withdrawal rate for projects entering the queue is nearly 80%. Developers face unpredictable delays and prohibitively high grid upgrade costs. PJM alone processed more than 170,000 megawatts of new generation requests since 2023, yet roughly 30,000 megawatts remain trapped in a transition queue.

This infrastructure deficit spans the globe. Europe has an estimated 1,700 gigawatts of renewable projects delayed in similar queues. Globally, over 3,000 gigawatts of projects are stuck. Meeting the projected data center demand requires an unprecedented expansion of national transmission networks. The grid cannot absorb the required capital fast enough.

A comprehensive study by the Union of Concerned Scientists modeled the electricity system with and without the projected data center growth. The counterfactual isolates the exact financial impact of the artificial intelligence boom. The results are severe. From 2026 to 2050, the cumulative wholesale electricity costs directly attributable to data centers range from $886 billion to $978 billion. This represents 18% of total United States wholesale electricity costs over that period. Without strict policy interventions, utilities pass this $978 billion cost onto residential and industrial ratepayers.

We are seeing the political blowback. Elections in Virginia and New Jersey swung on the issue of data center electricity costs. Citizens protest the heavy water demand, increased noise levels, and massive transmission lines cutting through residential zones. The Virginia State Corporation Commission was forced to create a new large-load class to protect everyday consumers from absorbing data center base-rate hikes. In India, the Central Electricity Regulatory Commission actively debates how to distribute the costs of grid integration as data center power demand jumps from 1 gigawatt to 13 gigawatts by 2032.

Data center expansion effectively acts as a regressive tax on domestic manufacturing. It erodes the fundamental discounted cash flow valuations of the S&P 500's industrial base. The repricing of the grid is permanent. An artificial intelligence query is not a digital abstraction. It is a physical event. It burns coal in Pennsylvania, boils water in a reactor in Illinois, and spins a turbine in Gujarat. The limit to artificial intelligence is not algorithmic data exhaustion. The limit is the cost of capital required to upgrade the global electrical grid, and the political willingness of the working class to subsidize it.

Investors must discount the future cash flows of any manufacturer lacking a firm, low-cost power purchase agreement. Heavy industrials must aggressively hedge their power exposure immediately. The power grid has become the ultimate arbiter of technological progress.