AI Data Centers Drive US Power Use to Record Highs as Grid Struggles to Keep Pace

AI and Energy Demand Reach Critical Juncture

The relationship between AI and energy has reached a pivotal moment as artificial intelligence reshapes electricity systems faster than many grids can adapt. US power consumption is expected to break records again in 2026 and 2027, after hitting a second consecutive annual high in 2025, according to the US Energy Information Administration1

1

. Total electricity use will rise from 4,195 billion kilowatt-hours (kWh) in 2025 to 4,268 billion kWh in 2026 and 4,391 billion kWh in 2027, driven largely by AI data centers and broader electrification1

1

.

AI-driven data centers now consume 67.7GW of electricity worldwide, representing approximately 1.9% of global generation, marking a 17% increase over the previous year1

1

. AI-focused facilities specifically increased their electricity consumption by 50% in 20251

1

. The International Data Center Authority warns that current grid capacity constraints are only the beginning, as AI fundamentally alters the trajectory of digital infrastructure power consumption1

1

.

The AI-Energy Paradox Emerges

A complex dynamic is unfolding that researchers call the AI-energy paradox. While AI increases electricity demand, it simultaneously offers tools to make global energy systems more efficient and resilient1

1

. Google reports that a median Gemini text prompt now uses about 0.24 watt-hours, after a 33-fold reduction in one year2

2

. Power consumption per AI task is declining by at least an order of magnitude annually, an unprecedented rate in energy history2

2

.

Yet this extraordinary efficiency improvement is racing against rising total demand. The IEA's central projection sees global data center demand roughly doubling from 485 terawatt-hours (TWh) in 2025 to 950 TWh by 2030, representing about 3% of global electricity2

2

. Without new grid construction, AI data centers could jump from consuming 6% of US power in 2025 to 12% by 20301

1

. Bloomberg projects that data centers will consume up to one-fifth of all power in the US by 2035, reaching up to 200 gigawatts3

3

.

Capacity Mining Offers Immediate Solutions

Before building new power plants, experts suggest examining how much grid capacity already sits unused. It costs about $5 and takes roughly 15 minutes to train an AI model that can help determine where America's next data centers should be built3

3

. Power systems are sized for the handful of hours a year when demand peaks, and the rest of the time, that capacity sits idle. Across most hours, we use only about half of existing grid capacity3

3

.

This capacity mining approach applies AI models to match data center demand against existing grid headroom, citing locations and operating profiles where available power and operational flexibility already exist3

3

. At one large investor-owned utility, calculations show that large loads, including data centers, can generate roughly $1 million per megawatt per year in new revenue, about $1 billion per gigawatt3

3

. Structured correctly, that revenue offsets fixed grid costs every other customer currently shoulders alone.

Power Systems and AI Infrastructure Converge

AI and electricity infrastructure are no longer separate stories. AI is entering the operating core of power systems, while power systems are becoming a binding condition for AI deployment4

4

. Industrial facilities that were once passive users of electricity are becoming autonomous energy managers through companies like Envision, continuously balancing on-site solar, battery storage and grid interaction without compromising production4

4

.

Battery storage solutions from companies like Hithium are addressing the gap between AI's rapid scaling and infrastructure buildout timelines. Their approach pairs long-duration storage at the energy source with lithium-sodium systems at the load side, providing grid-scale reliability and millisecond response on deployment timelines of one to two years rather than five to ten4

4

. Dr. Nazar Yi, Hithium Board Member and Vice President, notes that fast-response storage combined with long-duration flexibility provides the reliability that AI data center operators require4

4

.

Net Zero Planning Requires AI Update at COP31

Artificial intelligence was not in the room when the world designed its net zero targets, but at COP31 in Antalya this November, that needs to change2

2

. The incoming COP31 Presidency made electrification the flagship of its Action Agenda, proposing a collective goal to raise electricity's share of final energy demand from just over 20% today to 35% by 20352

2

.

The 2015 Paris Agreement invited countries to prepare long-term low-emission development strategies, and eighty have been submitted. However, the framework largely predates generative AI, and data centers are not consistently represented as a distinct end-use sector2

2

. The first structured attempt to model data center demand to 2050 within established climate-scenario frameworks finds a range of 1,800 to 5,000 TWh, a nearly threefold spread2

2

.

Grid Resilience Through AI's Transformative Role in Energy

PJM Interconnection, the biggest US grid operator serving a fifth of the population, is preparing proposals to close a widening gap between available electricity supply and demand, led largely by AI data centers1

1

. Similar pressures are being felt worldwide as the data center industry expands faster than many electricity networks can support1

1

.

AI's transformative role in energy extends beyond demand to becoming the means by which power systems see trouble, absorb it and recover. As grids carry more variable renewables and face extreme weather and cyber risks, grid resilience is becoming less about reserve capacity alone and more about speed of response4

4

. Energy optimization through AI enables systems that can sense, respond and adapt in real-time, fundamentally changing how transmission networks operate.

The question facing policymakers and industry leaders is whether AI's efficiency gains will outpace adoption growth, or whether Jevons' paradox will prevail. Long reasoning queries already consume roughly 13 times more energy than a standard prompt2

2

. Every siting approval, grid connection and power-purchase agreement embeds a view of AI's long-term electricity appetite, shaping infrastructure decisions that will last decades. Clean energy growth continues despite headwinds, suggesting that demand from companies, states and consumers, alongside falling technology costs, is supporting investment in renewables and clean energy infrastructure1

1

.