Researchers Sasha Luccioni from Hugging Face and Boris Gamazaychikov of Salesforce demonstrated in a 2024 study that advanced AI models mimicking human reasoning can use up to 100 times more energy than simpler counterparts. The team tested 40 open-source models, including those from OpenAI, Google, and Microsoft. All models ran on identical hardware and faced the same tasks—from basic sports queries to complex math puzzles. The outcome? Models with active reasoning required, on average, a hundred times more electricity per 1,000 prompts than those without that feature. The reason: every reasoning step adds computational cycles, and those cycles demand energy. “This isn’t a marginal increase—it’s a chasm,” one of the report’s authors commented [1][2][3].
AI Models Reasoning Like Humans Consume Massive Power
The differences can be staggering. The stripped-down R1 model from China’s DeepSeek used just 50 watt-hours with reasoning turned off—about as much as a 50-watt bulb running for an hour. Once reasoning was enabled, the same model needed over 308,000 watt-hours, a more than 6,000-fold jump. Microsoft’s Phi-4 model consumed 9,462 watt-hours with reasoning, but only 18 without. Even OpenAI’s gpt-oss showed a significant rise: from 313 to 8,504 watt-hours, depending on settings [3][1].
As AI developers—essentially all major players—race to build models that reason step by step like humans, a key question emerges: can power grids keep up? In 2024, data centers worldwide consumed 415 terawatt-hours. The International Energy Agency projects that by 2030, this figure will exceed 945 TWh. AI alone is expected to drive more than 20 percent of the global increase in electricity demand, with fossil fuels still covering 40 percent of new consumption [1][2].
Neuromorphic Chips: Brain-Inspired Architecture, Real Savings
Neuromorphic processors may hold the answer. These chips mimic the architecture of brain neurons and can cut AI’s energy use by up to a thousandfold compared to conventional chips. Unlike traditional processors, where memory and computation are separate, neuromorphic chips combine both functions. The result? Data doesn’t shuttle between components, and it’s those transfers that eat up the most energy.
Innatera of the Netherlands and UK-based 42 Technology began collaborating in December 2024 to deploy the Pulsar chip—which processes sensor data at under one milliwatt—in commercial devices. Intel claims its Loihi 2 can perform AI tasks 100 times more efficiently and 50 times faster than classic CPUs or GPUs. IBM NorthPole, with 22 billion transistors and 256 cores, proved 25 times more energy-efficient than the NVIDIA V100 in image recognition tasks [3][4][5].
Breakthroughs in materials are emerging, too. In October 2025, a team at the University of Southern California showcased artificial neurons built from memristors operating on ions rather than electrons. The energy use: just 40 to 200 picojoules per impulse—a fraction of what conventional processors require [6][7][8].
The Race for Energy-Efficient Artificial Intelligence Accelerates
The neuromorphic chip market is set to grow from $4.89 billion in 2025 to over $76 billion by 2035, with annual growth topping 30 percent. BrainChip introduced the Akida processor in M.2 format, enabling rapid, low-cost AI deployment at the network edge. The company is already partnering with Raytheon on a contract for the U.S. Air Force. Sandia National Laboratories launched the SpiNNaker2 system, simulating 175 million neurons for nuclear deterrence research [4][5].
In November 2025, a team at the University of Texas at Dallas built a prototype neuromorphic computer that learns patterns with minimal computation—no costly cloud training required. Mercedes-Benz research suggests vision systems based on neuromorphic chips could cut energy use in autonomous vehicles by up to 90 percent. Analysts predict that by 2030, such chips will be found in 40 percent of IoT sensors [3][4].
Training large models like ChatGPT generates a huge carbon footprint—up to 552 tons of CO2, equivalent to a year’s energy use by 121 U.S. households. Yet it’s not training, but inference—everyday use of the model—that accounts for 90 percent of AI’s total energy consumption [2].
U.S. data centers consumed 183 terawatt-hours in 2024—over 4 percent of national usage, matching the entire country of Pakistan. By 2030, that figure is set to rise by 133 percent. If AI is to be widespread, not just for the wealthiest, brain-inspired chips are no longer a curiosity—they’re a necessity. The question is whether the industry can move fast enough before electricity bills overwhelm us.