Energy
YouTube ... Quora ...Google search ...Google News ...Bing News
- Energy-based Model (EBN)
- Case Studies
- Other Challenges in Artificial Intelligence
- Energy and Policy Considerations for Deep Learning in NLP | E. Strubell, A. Ganesh, and A. McCallum - College of Information and Computer Sciences & University of Massachusetts Amherst
- Energy Efficient Machine Learning and Cognitive Computing
- AI Is Throwing Battery Development Into Overdrive | Daniel Oberhaus - Wired ... Improving batteries has always been hampered by slow experimentation and discovery processes. Machine learning is speeding it up by orders of magnitude.
- SPIRAL Center uses artificial intelligence to make solar energy cheaper | Zachary Bahar - Johns Hopkins News-Letter
- Nuclear fusion is one step closer with new AI breakthrough | Tom Metcalfe - Livescience
- Using deep learning to develop a forecasting model for efficiently managing electric grids | Chung-Ang University
Energy and artificial intelligence meet in two directions at once. AI uses energy: every model trained and every prompt answered draws electricity inside a data center and turns it into heat. AI is also increasingly used on energy: forecasting wind and solar output, balancing electric grids, screening battery materials, steering fusion plasmas, and trimming the cooling bills of the very data centers that run it. This page covers both directions, along with the physics underneath them, because thermodynamics and entropy set hard limits on how cheaply any machine can compute.
For most of the 2010s, computing's appetite for electricity looked like a solved problem. A widely cited 2020 study in Science led by Eric Masanet estimated that the number of computing instances in the world's data centers rose about 550 percent between 2010 and 2018, while their energy use rose only about 6 percent. Better servers, virtualization, and large, efficient hyperscale facilities absorbed the growth. AI broke that pattern. The International Energy Agency (IEA) estimates that data center electricity use has grown about 12 percent a year since 2017, and Lawrence Berkeley National Laboratory (LBNL) found that US data center demand more than doubled between 2017 and 2023, driven largely by AI servers. In 2025 alone, according to the IEA, electricity use at AI-focused data centers jumped about 50 percent.
The people and organizations shaping this field fall into a few groups. Energy analysts, most visibly the IEA's Energy and AI team led by Thomas Spencer and Siddharth Singh, and LBNL's Center of Expertise for Energy Efficiency in Data Centers (Arman Shehabi, Sarah Smith, and colleagues), build the demand estimates that governments and utilities plan around. Measurement researchers work out how much energy individual models actually use: Emma Strubell and colleagues at the University of Massachusetts Amherst, David Patterson and colleagues at Google and UC Berkeley, Sasha Luccioni at Hugging Face, and the ML.ENERGY group led by Jae-Won Chung and Mosharaf Chowdhury at the University of Michigan. Hardware researchers such as Vivienne Sze at MIT design chips and algorithms that do more work per joule. Physicists, from Rudolf Clausius and Ludwig Boltzmann through Claude Shannon and Rolf Landauer to today's thermodynamic computing groups, explain where the energy ultimately goes and where the floor is.
The central mechanism is simple arithmetic: total energy equals energy per useful result multiplied by the number of results. Energy per AI task is falling at a pace the IEA describes as unprecedented in energy history, by at least an order of magnitude per year in recent years by its estimate. The number of tasks, and the energy intensity of new kinds of tasks such as step-by-step reasoning, video generation, and autonomous agents, is rising even faster. Economists call this a rebound effect, or the Jevons paradox: making something cheaper per unit can increase total consumption.
That is the paradigm shift. Computing energy used to be treated as an IT line item. It is now a question of grid planning, siting, and energy security, and the availability of power, grid connections, transformers, and turbines increasingly paces how fast AI can be deployed. At the same time, measurement is moving from operations per second to joules per token, per prompt, and per task, counted at the level of the whole facility and the grid that feeds it. Data centers are also being reimagined from passive loads into potential grid assets that can flex their demand. For related discussion, see Power (Management), Case Studies, and Other Challenges in artificial intelligence.
Every answer an AI gives is paid for in electricity and ends as heat. The energy cost of each task is falling faster than almost any technology in history, yet total demand keeps climbing, so the real questions are how much useful work we get per joule, where those joules come from, and whether AI helps the rest of the energy system more than it burdens it.
The Mechanism
Think of an AI service as a chain that converts electricity into answers, with losses at every link.
- Training. A frontier model is trained by running many thousands of accelerators (GPUs or TPUs) in parallel for weeks or months. Current flagship accelerators draw roughly 700 to over 1,000 watts each. Training energy is approximately the number of accelerators, times their average power, times the length of the run, multiplied by facility overhead. It is a one-time cost per model, but failed runs, experiments, and hyperparameter searches add to it.
- Inference. Every prompt runs the trained model again. Inference has two phases. In prefill, the model reads the whole prompt in parallel, which keeps the arithmetic units busy. In decode, it generates output one token at a time, and each step requires streaming the model's active weights out of high-bandwidth memory, so decode is limited by memory bandwidth more than by arithmetic. Energy therefore scales with the number of tokens processed and generated, with the number of active parameters, and inversely with how many user requests the system batches together. This is why a long "reasoning" answer can cost tens or hundreds of times more than a short reply from the same model.
- Host and idle overhead. Accelerators need host CPUs, DRAM, networking, and storage, and operators keep spare machines running to absorb traffic spikes and failures. In Google's 2025 measurement of Gemini, active accelerators accounted for only 58 percent of serving energy; host CPUs and memory took 25 percent, idle reserve machines 10 percent, and data center overhead 8 percent.
- Facility. Power conversion and cooling add overhead, summarized by power usage effectiveness (PUE): total facility energy divided by IT equipment energy. A PUE of 1.0 would mean zero overhead. Leading hyperscale fleets now report values close to 1.1 (Google reported 1.09 for 2025), while many older and enterprise facilities run considerably higher.
- Grid. Where and when the electricity is generated sets the carbon and water footprint. The same kilowatt-hour can carry very different emissions in different regions or at different hours.
- Embodied impacts. Manufacturing chips, servers, and buildings carries its own energy, emissions, and water use. As data center construction accelerates, this "supply chain" share is growing; Google reported its supply chain emissions rose 25 percent in 2025.
A compact way to write the accounting:
- <math>E_{\mathrm{facility}} = \mathrm{PUE} \times \left( E_{\mathrm{accelerators}} + E_{\mathrm{host}} + E_{\mathrm{idle}} + E_{\mathrm{network,\,storage}} \right)</math>
Per-prompt figures are usually computed by dividing the energy of a serving fleet over a time window by the number of prompts served in that window:
- <math>E_{\mathrm{prompt}} \approx \frac{E_{\mathrm{facility}}(\Delta t)}{N_{\mathrm{prompts}}(\Delta t)}</math>
Dependencies and conditions. The result depends on hardware utilization and batch size, model size and architecture (dense models versus mixture-of-experts models that activate only part of their parameters per token), numeric precision, prompt and answer length, hardware generation, the local climate (which affects cooling), and the grid mix at that place and hour.
Where uncertainty remains. Most frontier AI developers publish no per-query energy data, so independent estimates rely on assumptions about hardware, batching, and utilization. Measurement boundaries differ: some figures include idle machines and host power, others count only the accelerator. And industry announcements are usually stated in megawatts of planned capacity, which LBNL researchers caution does not convert directly into annual energy use.
The Physics
Energy in, heat out. Conservation of energy means essentially all electricity delivered to chips ends up as heat, which is why cooling is unavoidable. It also helps to keep power and energy straight: power is a rate (watts, gigawatts), energy is an amount (watt-hours, terawatt-hours). A 1 GW campus running around the clock for a year uses about 8.76 TWh. Grid planners worry about both: annual energy determines fuel and emissions, while peak power determines how much generation and transmission must exist.
Switching energy and the end of free scaling. The dynamic power of a digital chip follows approximately
- <math>P_{\mathrm{dyn}} \approx \alpha \, C \, V^{2} f</math>
where α is the fraction of circuits switching, C is the switched capacitance, V is the supply voltage, and f is the clock frequency. Through the early 2000s, Dennard scaling let each chip generation lower its voltage as transistors shrank, so energy per operation fell almost automatically. That stopped in the mid-2000s, when voltage could no longer drop without leakage currents exploding. Since then, efficiency gains have come from specialization, parallelism, lower numeric precision, and smarter data movement rather than from shrinking alone.
Data movement dominates. The most important number in AI hardware is not the cost of arithmetic but the cost of moving data. Mark Horowitz's widely used 2014 estimates for a 45 nm process put a 32-bit floating-point add at about 0.9 picojoules and a 32-bit floating-point multiply at about 3.7 pJ, while reading 32 bits from off-chip DRAM cost about 640 pJ, roughly 170 times the multiply. Moving data between memory and arithmetic units therefore costs far more than the math itself. Vivienne Sze's group at MIT built its accelerator designs around "dataflows" that reuse each value as many times as possible in small, nearby memories, and the same logic drives today's high-bandwidth memory, large on-chip caches, and research on compute-in-memory.
Precision. In the same estimates, an 8-bit integer multiply costs about 0.2 pJ, nearly 20 times less than a 32-bit floating-point multiply, and narrower numbers also mean fewer bytes to move. That gap explains the industry's push toward 8-bit and 4-bit formats for both training and inference, as long as accuracy holds.
Scaling relationships. Two rules of thumb connect model size to compute. Training a dense model with N parameters on D tokens takes roughly
- <math>C_{\mathrm{train}} \approx 6 \, N \, D \ \text{FLOPs}</math>
and generating each token at inference takes roughly
- <math>C_{\mathrm{token}} \approx 2 \, N_{\mathrm{active}} \ \text{FLOPs}</math>
Energy is compute divided by delivered efficiency η (useful FLOPs per joule, after utilization losses and overhead): E = C / η. As an illustration only, a model with 100 billion active parameters generating 1,000 tokens needs about 2 × 1014 FLOPs. At an effective efficiency between 1011 and 5 × 1011 FLOPs per joule, which is far below a modern accelerator's peak because decode is memory-bound, that works out to roughly 0.1 to 0.6 Wh. The order of magnitude matches recent production measurements for a typical text prompt. It also shows why reasoning is expensive: a model that "thinks" for 10,000 tokens instead of 300 does about 30 times more work.
The thermodynamic floor. Rolf Landauer showed in 1961 that erasing one bit of information must release at least
- <math>Q_{\min} = k_{B} \, T \ln 2</math>
of heat, where kB is Boltzmann's constant and T is the temperature. At room temperature that is about 2.9 × 10-21 joules. A 32-bit floating-point multiply at 3.7 pJ costs roughly a billion times that floor once you count all the switching, wiring, and clocking involved. Antoine Bérut and colleagues confirmed the Landauer bound experimentally in 2012 using a single colloidal bead in a laser-made double-well "memory," and noted that individual logic gates of that era were within about a factor of a thousand of it. Two caveats matter. The bound applies to logically irreversible operations, so reversible computing (Charles Bennett's idea) can avoid it in principle. And it is only approached in the slow, quasi-static limit; Bérut's measured heat approached the bound only for long erasure cycles. Modern stochastic thermodynamics, developed by David Wolpert and others, quantifies the extra cost of computing quickly. The practical lesson is that physics leaves enormous headroom: today's limits are set by engineering, not by the laws of thermodynamics. The Entropy section below explains why.
Cooling and heat transfer. Racks of the newest AI systems draw well over 100 kilowatts each, far beyond what air can carry away efficiently, which is pushing the industry toward direct liquid cooling. Water holds roughly 3,500 times more heat per unit volume than air for the same temperature rise. Evaporative cooling saves electricity but consumes water, so operators trade one resource against the other depending on climate and local water stress.
Grid physics. An electric grid must match supply and demand every second, and its size is set by the handful of peak hours each year. That is why a load that can shed or shift demand during those few hours is far easier to connect than one that cannot, and why researchers now study data center flexibility as seriously as data center efficiency. Synchronized training jobs can also make a campus's demand swing sharply within seconds, a behavior grid operators and reliability organizations have begun to study.
Why This Matters
- Energy infrastructure. Data centers are among the fastest-growing electricity loads, and they cluster: nearly half of US data center capacity sits in five regional clusters. A single campus can request as much power as a mid-sized city, forcing utilities to plan generation and transmission on multi-year timelines while AI projects move in months.
- Emissions and resources. The IEA expects renewables to meet about half of data center demand growth to 2030, with natural gas filling much of the rest, especially in the United States. How data centers are powered determines whether AI growth slows decarbonization. Water for cooling and electricity generation, critical minerals for chips and equipment (whose supply the IEA notes is highly concentrated), and embodied emissions from manufacturing all add to the footprint.
- Costs and fairness. Who pays for new grid capacity is a live regulatory question, and the IEA's 2026 update examined whether data centers are raising electricity prices. The answer differs by market and by how costs are allocated.
- Engineering direction. Energy is now a first-class design metric. Low-precision arithmetic, sparsity, mixture-of-experts models, efficient serving software, specialized accelerators, and liquid cooling are all responses to it.
- AI as an energy tool. Better forecasts, grid operations, materials discovery, and fusion control could make the whole energy system cleaner and cheaper. The IEA's analysis suggests that emissions reductions from widespread AI adoption could offset the emissions data centers add, but that outcome depends on adoption that is far from guaranteed.
- Conceptual stakes. Physics sets a minimum cost per erased bit, and today's hardware runs around a billion times above it for a typical arithmetic operation. That gap is the long-run opportunity, and new approaches such as thermodynamic computing are trying to exploit it.
- What would change if efficiency wins. If per-task efficiency keeps improving roughly tenfold per year and demand eventually saturates, the prediction that AI's carbon footprint will "plateau, then shrink" could come true at system scale, not just per model. If rebound effects win, AI becomes one of the largest new electricity loads of the century.
- What remains uncertain. The IEA's 2035 scenarios span 700 to 1,700 TWh of global data center demand, and LBNL's 2030 range for the United States spans 9.5 to 15.3 percent of national electricity. That spread reflects genuine uncertainty about adoption, efficiency, and physical bottlenecks, not just modeling noise.
Recent Developments
The past two years turned AI energy from a field of rough guesses into one with real measurements, official forecasts, and active policy fights. Here is what changed, what it shows, and what it does not.
How much electricity data centers use, and how fast it is growing
In April 2025 the IEA published Energy and AI, its first comprehensive global analysis of the topic, and launched an Energy and AI Observatory to track data center capacity and AI applications in the energy sector. It estimated that data centers used about 415 TWh in 2024, roughly 1.5 percent of global electricity, with the United States accounting for about 45 percent, China about 25 percent, and Europe about 15 percent. A typical AI-focused data center uses as much electricity as about 100,000 households, and the largest ones under construction will use about twenty times that. The IEA's Base Case projected demand more than doubling to about 945 TWh by 2030, slightly more than Japan's entire current electricity use, and about 1,200 TWh by 2035.
In April 2026 the IEA followed up with Key Questions on Energy and AI. Data center electricity use rose about 17 percent in 2025 to roughly 485 TWh, against roughly 3 percent growth in global electricity demand, and AI-focused data centers grew about 50 percent. The 2030 projection barely moved (950 TWh versus 945 TWh), but its makeup shifted toward AI: consumption at AI-focused sites is projected to roughly triple to about 465 TWh, approaching that of conventional data centers. The report also identified physical bottlenecks, including grid connections, equipment supply chains, and manufacturing capacity, as factors that could limit how quickly new projects come online.
In the United States, LBNL's December 2024 report found data center electricity use climbed from 58 TWh in 2014 to 176 TWh in 2023, about 4.4 percent of national electricity, and projected 325 to 580 TWh by 2028. LBNL's June 2026 update extended the outlook: 649 TWh in 2030 in its Reference Case, about 11.8 percent of US electricity, with sensitivity scenarios spanning 9.5 to 15.3 percent. The sensitivity cases are revealing. Fewer equipment installations gave 578 TWh in 2030; more GPU shipments gave 664 TWh; shorter AI chip lifetimes gave 590 TWh; but higher idle power and utilization of AI servers gave 782 TWh. How hard installed AI servers actually run matters as much as how many get shipped.
| Source | Scope | Recent estimate | Projection |
|---|---|---|---|
| IEA, Energy and AI (April 2025) | All data centers, worldwide | About 415 TWh in 2024, roughly 1.5% of global electricity | About 945 TWh in 2030; about 1,200 TWh in 2035 (Base Case); scenarios span 700 to 1,700 TWh in 2035 |
| IEA, Key Questions on Energy and AI (April 2026) | All data centers, worldwide, with AI-focused sites reported separately | About 485 TWh in 2025, up about 17%; AI-focused sites up about 50% | About 950 TWh in 2030, near 3% of global demand; AI-focused sites roughly triple to about 465 TWh |
| LBNL, 2024 United States Data Center Energy Usage Report (December 2024) | United States | 176 TWh in 2023, about 4.4% of US electricity (58 TWh in 2014) | 325 to 580 TWh in 2028, or 6.7% to 12% of US electricity |
| LBNL, United States Data Center Energy Usage Report: 2025 Update (June 2026) | United States | Revised bottom-up model based on equipment shipments | 649 TWh in 2030 (Reference Case), about 11.8% of US electricity; range 9.5% to 15.3% |
What these numbers do and do not establish. They are modeled estimates built from equipment shipments, assumed utilization, and facility efficiency, not meter readings. "All data centers" includes conventional cloud and enterprise computing, so none of these totals is the electricity used by AI alone. Earlier figures are not directly comparable either: in January 2024 the IEA projected 620 to 1,050 TWh in 2026 for data centers, AI, and cryptocurrency mining combined, a broader boundary than its later data-center-only series. And global shares understate local stress, because demand concentrates in a few regions.
Measuring a single prompt
For years the most quoted figure for a chatbot query was about 2.9 Wh, roughly ten times a conventional web search, a number used in EPRI's 2024 Powering Intelligence report and repeated widely. Measurements since 2025 suggest that a typical text prompt to an efficient production system uses far less, although the spread across models and tasks is wide.
- Google (August 2025). Google published a methodology and production measurements for Gemini. The median Gemini Apps text prompt in May 2025 used about 0.24 Wh, emitted about 0.03 g of CO2-equivalent, and consumed about 0.26 mL of water. Counting only active accelerator power, as many earlier estimates did, gives 0.10 Wh, a 2.4-fold difference that shows how much the measurement boundary matters. Google reported a 33-fold drop in energy per median prompt and a 44-fold drop in emissions over the preceding twelve months, with most of the energy gain attributed to model and software improvements and a smaller share to better hardware utilization. The study excludes training, networking to the user, and end-user devices. It was released as a technical report rather than a peer-reviewed paper, and critics noted that its water figure counts on-site cooling water but not water used to generate the electricity.
- OpenAI (June 2025). Sam Altman wrote that an average ChatGPT query uses about 0.34 Wh, without publishing a methodology. Epoch AI had independently estimated roughly 0.3 Wh for a typical GPT-4o query earlier that year.
- Mistral (July 2025). Mistral published a full lifecycle analysis of Mistral Large 2 with the consultancy Carbone 4 and the French environment agency ADEME, following ISO 14040/44 and the AFNOR Frugal AI methodology, with external review. Training plus 18 months of use produced about 20.4 kilotonnes of CO2-equivalent and consumed about 281,000 cubic meters of water. A 400-token response from its Le Chat assistant had a marginal footprint of about 1.14 g CO2-equivalent and 45 mL of water. The far larger water figure than Google's mainly reflects a wider boundary that includes upstream impacts, not necessarily a less efficient system. Mistral also found impacts roughly proportional to model size and proposed three standard metrics: total training impact, impact per request, and the ratio of inference to total lifecycle impact.
- Independent benchmarks. How Hungry is AI? (Nidhal Jegham and colleagues, 2025) combined public API speed data with data-center efficiency multipliers to estimate the footprint of more than 30 models. A short GPT-4o query came out at about 0.42 Wh, while reasoning models such as o3 and DeepSeek-R1 exceeded 33 Wh for long prompts, more than 70 times the most efficient small model tested. These are inferences from public data, not direct measurements. The ML.ENERGY Benchmark from the University of Michigan (NeurIPS 2025) takes the opposite approach, measuring open models directly on real hardware across 40 architectures and six tasks. It found that automated choices about serving configuration can cut energy by more than 40 percent in some cases without changing what the model computes.
What this establishes: for short text prompts on efficient, well-utilized production systems, energy per query is on the order of a few tenths of a watt-hour, and the old 3 Wh figure is outdated for that case. What it does not establish: total consumption, which depends on query volume; the long tail of heavy prompts; image, video, and agent workloads; or the many providers that publish nothing. No independent party has audited the company-reported figures.
Reasoning, video, and agents push the other way
Energy scales with tokens, and newer workloads generate many more of them. In December 2025 the AI Energy Score project, led by Sasha Luccioni of Hugging Face and Boris Gamazaychikov of Salesforce, tested 40 open models on 1,000 text prompts with reasoning switched on and off. Reasoning modes raised average energy use by well over an order of magnitude, and in extreme cases far more: one Microsoft Phi-4 reasoning model used about 9,462 Wh across the prompt set with reasoning enabled versus about 18 Wh with it disabled. Most of the increase came from much longer outputs. The study covers open models on standard hardware and was not peer reviewed, but it matches the IEA's 2026 observation that while energy per task is falling quickly, energy-intensive uses such as video generation, reasoning, and agentic tasks are spreading even faster. The practical implication is that routing simple questions to lighter models, or disabling extended reasoning when it is not needed, is one of the largest efficiency levers available to users and developers.
Training's footprint: revisiting the 2019 debate
The Strubell, Ganesh, and McCallum paper linked at the top of this page (ACL 2019) estimated that training a large Transformer with neural architecture search could emit about 626,000 pounds of CO2-equivalent, roughly five times the lifetime emissions of an average American car including fuel. It became the most cited number in the field and helped launch the study of AI's carbon footprint.
In 2022, David Patterson and colleagues at Google and UC Berkeley responded in The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. They argued that four practices (efficient model architectures, ML-specialized processors, efficient cloud data centers, and siting work in low-carbon regions) can cut training energy by up to 100 times and emissions by up to 1,000 times; that estimates for papers that did not report their emissions had been off by factors of 100 to 100,000; and that machine learning had held steady at under 15 percent of Google's total energy use for three years.
Both sides turned out partly right. Per-model efficiency improved dramatically. DeepSeek-V3, released in December 2024, reported training a frontier-class model on about 2.8 million H800 GPU-hours, far less than many had assumed was required. But the total footprint did not shrink. A 2025 position paper argued that rebound effects have kept overall AI energy use and emissions climbing, and companies' own reports bear that out (see below). Inference has also overtaken training as the dominant lifecycle cost for widely used models; the ML.ENERGY authors cite estimates that inference accounts for 80 to 90 percent of production machine learning compute.
Corporate disclosures and the rebound effect
Google's 2025 environmental report showed electricity use up 27 percent in 2024. It cut data center energy emissions 12 percent through clean power purchases and efficiency, but its overall reported emissions still rose 11 percent to about 11.5 million tonnes, 51 percent above 2019. The 2026 report showed electricity use up 37 percent in 2025, the largest annual increase in the company's history, with data centers consuming more than 42 TWh compared with 30.6 TWh the year before. Google reported a fleet PUE of 1.09, a 2 percent drop in operational emissions, record clean energy contracts of more than 12 GW, a 34 percent rise in water consumption to 10.9 billion gallons, and a 25 percent jump in supply chain emissions, which it attributed largely to construction and to chips manufactured on fossil-heavy grids in Asia. The company acknowledged that its AI buildout is currently accelerating faster than the grid is decarbonizing. Microsoft has likewise reported emissions roughly 23 percent above its 2020 baseline, largely because of data center expansion.
Two accounting details matter when reading these reports. "Market-based" emissions credit companies for renewable energy they purchase, even if the physical electrons on the local grid come partly from fossil plants; "location-based" figures reflect the actual local grid and are usually higher. And embodied emissions from building facilities and fabricating chips grow with every expansion regardless of how clean the operating electricity is.
Grid flexibility, new supply, and nuclear restarts
In February 2025, researchers at Duke University's Nicholas Institute, led by Tyler Norris, published Rethinking Load Growth. Analyzing the largest US balancing authorities, they estimated that the existing grid, which is sized for rare extreme peaks, could absorb about 76 GW of new load if those loads accepted curtailment averaging 0.25 percent of their maximum annual uptime, about 98 GW at 0.5 percent, and about 126 GW at 1 percent. Curtailment events averaged about two hours, and for most of the curtailed time at least half of the new load could keep running. The Electric Power Research Institute launched its DCFlex initiative in October 2024 to demonstrate flexible data center operation; Texas passed Senate Bill 6 in 2025, requiring large new loads to provide backup power and participate in emergency load reduction; and Google announced agreements in August 2025 with Indiana Michigan Power and the Tennessee Valley Authority to reduce some machine learning workloads during grid peaks.
The Duke study has critics. It is a planning screen that estimates theoretical headroom, not a guarantee that any specific project can connect, and some data center developers argue that forced curtailment conflicts with the economics of expensive, always-on AI hardware. Many prefer on-site generation or batteries, and shifting work between regions, over actually switching off.
On the supply side, technology companies have turned to firm, low-carbon power. Constellation Energy plans to restart the 835 MW reactor at the former Three Mile Island Unit 1, renamed the Crane Clean Energy Center, under a 20-year power purchase agreement with Microsoft, with a 2027 target and support from the US Department of Energy's loan program. In October 2024, Google signed an agreement with Kairos Power for small modular reactors totaling 500 MW by 2035, and Amazon announced an investment in X-energy. Google has also agreed to buy 200 MW from Commonwealth Fusion Systems' planned ARC fusion plant in Virginia, targeted for the early 2030s. Most of these resources will not deliver power for years, which is why natural gas continues to fill much of the near-term gap.
AI applied to the energy system
- Weather and renewables. Google DeepMind's GenCast, published in Nature in December 2024, generates probabilistic 15-day global weather forecasts at 0.25 degree resolution using a diffusion model. It was more accurate than the European Centre for Medium-Range Weather Forecasts' ENS ensemble on 97 percent of the targets tested and better predicted the combined output of wind farms worldwide. Each 15-day ensemble member takes about eight minutes on a single TPU v5. The test was a proof of principle using 10-meter wind speeds and the Global Power Plant Database, not an operational grid deployment.
- Fusion. In October 2025, Google DeepMind and Commonwealth Fusion Systems formalized a research partnership around TORAX, DeepMind's open-source, differentiable plasma simulator written in JAX. CFS uses TORAX to run large numbers of virtual experiments on its SPARC tokamak before it operates, while DeepMind applies reinforcement learning and evolutionary search to find efficient operating scenarios and, eventually, real-time control strategies. SPARC's first plasma is now projected for 2027, with net fusion energy (Q greater than 1) targeted after that. AI's role in actual reactor operation remains exploratory, and commercial fusion power is still years away.
More detail on these and earlier results appears under AI for Energy Systems.
New physics of computation
- Landauer's principle in a quantum many-body system (2025). Stefan Aimet and colleagues in Berlin and Vienna, writing in Nature Physics, used ultracold Bose gases as a quantum field simulator to track how entropy, mutual information, and dissipated energy evolve after a sudden change to the system. Their measurements matched quantum field theory predictions, extending the experimental reach of Landauer's principle from single bits to complex quantum systems.
- Thermodynamic computing hardware. Normal Computing reported in Nature Communications (2025) a small "stochastic processing unit" built from eight coupled electrical resonators that used thermal noise to perform Gaussian sampling and matrix inversion, and later taped out its first integrated chip. Extropic unveiled its X0 probabilistic chip and XTR-0 development platform in October 2025, and in July 2026 published an architecture paper with MIT's Isaac Chuang in npj Unconventional Computing projecting that a future device could match GPU-based models on a simple binarized image benchmark while using roughly 10,000 times less energy per sample. That figure combines chip measurements with circuit models and simulation; it is a projection for a toy benchmark, not a measurement on production workloads.
- Generative thermodynamic computing (2026). Stephen Whitelam of LBNL showed in Physical Review Letters that a physical system governed by Langevin dynamics can, in principle, act as a diffusion-style generative model, turning noise into structured images through its own natural time evolution. Training it amounts to minimizing the heat the process gives off. The demonstration so far is a digital simulation of such hardware.
These results are scientifically important but far from deployment. What is demonstrated: the physics works on small devices and in simulation. What is not: that such hardware can scale, tolerate manufacturing variation, and run workloads comparable to today's large models.
Disclosure and reporting rules
Under the EU AI Act, providers of general-purpose AI models have had to keep technical documentation since 2 August 2025 that includes the known or estimated energy consumption of their models, estimated from computational resources if not directly measured (Article 53 and Annex XI). The information goes to regulators and the EU AI Office rather than the public, the Commission's enforcement powers over these obligations began on 2 August 2026, and models placed on the market before August 2025 have until August 2027 to comply. Alongside regulation, voluntary methods are converging: Google's per-prompt methodology, Mistral's lifecycle metrics, the AI Energy Score ratings, and the ML.ENERGY Leaderboard all aim to make figures comparable.
Open questions
- Will per-task efficiency gains keep pace with rising use, or will rebound effects dominate through 2030 and beyond?
- How much AI load can realistically be flexible, and will grid rules reward flexibility enough to change how data centers are built?
- How much new demand will be met by clean power versus natural gas, and how fast can firm clean resources such as nuclear actually arrive?
- Can water use and embodied emissions be measured on consistent boundaries across companies?
- Will AI-enabled emissions reductions in the wider energy system materialize at the scale the IEA's analysis suggests is possible?
- Can thermodynamic or other physics-based hardware scale beyond small benchmarks?
- Will independent auditing of company-reported energy figures become standard?
AI for Energy Systems
The other half of the story is what AI does for the energy system itself. Machine learning tends to help in four ways: it forecasts things that are hard to model from first principles (weather, demand, equipment failure); it controls complex systems faster than human operators can (cooling plants, grids, plasmas); it searches enormous design spaces (materials, molecules, reactor operating scenarios); and it squeezes waste out of existing assets. The IEA's 2025 analysis concluded that emissions reductions enabled by AI could more than offset the extra emissions from data centers, but only if AI is adopted widely across the energy sector, and its 2026 update devoted one of its key questions to the barriers slowing that adoption, such as data access, skills, and regulation.
Forecasting and grid operations
Wind and solar output depends on the weather, so better forecasts let grid operators commit less backup generation and throw away less renewable power. The IEA highlights forecasting for variable renewables as one of the most mature AI applications in the power sector. GenCast (see Recent Developments) is the current state of the art for medium-range ensemble forecasts. At the scale of a single utility, deep learning load forecasting, such as the Chung-Ang University model linked at the top of this page, helps operators schedule generation more precisely. The World Economic Forum video below describes how combining hardware with software is helping India integrate renewable power for a very large population. For grid-side management concepts, see Power (Management).
Data center cooling
In 2016, DeepMind and Google's data center team showed that machine learning recommendations could cut the energy used for cooling Google's data centers by up to 40 percent, and the system later moved to direct, supervised control of cooling plants. Jim Gao, who led that work, describes it in the video below. It is a reminder that AI's own facilities are among the best targets for AI optimization.
Storage and materials discovery
Battery improvement has always been slowed by the pace of laboratory experiments, and machine learning is speeding up the search (see the Wired article linked above). In January 2024, Microsoft and the Pacific Northwest National Laboratory reported using AI models and cloud computing to screen about 32.6 million candidate materials, narrowing them to about 500,000 predicted to be stable and then to 18 promising candidates in roughly 80 hours. PNNL scientists synthesized one, a sodium-lithium solid electrolyte called N2116, and built a working prototype battery in under nine months; the material could use up to about 70 percent less lithium than comparable designs. It remains a proof of concept: long-term stability, conductivity, and manufacturability still have to be validated at scale. On the operations side, companies such as Stem (video below) pair storage with software that decides when to charge and discharge.
Fusion
In 2022, DeepMind and the Swiss Plasma Center at EPFL used deep reinforcement learning to control the magnetic coils of the TCV tokamak, producing and holding a range of plasma shapes (see the Livescience article linked above). In 2024, a Princeton-led team reported in Nature that a reinforcement learning controller on the DIII-D tokamak could forecast dangerous tearing instabilities up to about 300 milliseconds in advance and steer the plasma to avoid them. The 2025 DeepMind and Commonwealth Fusion Systems partnership around TORAX extends this work to scenario design for a machine aiming at net energy gain (see Recent Developments). AI accelerates the research loop; it does not change the underlying plasma physics or the engineering challenges of building a power plant.
Buildings and demand
Buildings consume a large share of electricity, and much of it is wasted on equipment running when it does not need to. Startups such as Verdigris (video below) use sensors and deep learning to identify that waste. Smarter demand also helps grids: loads that respond to prices or grid conditions reduce the need for peak generation.
|
|
|
|
|
|
Solar Energy
Solar power is the fastest-growing source of new electricity in much of the world, and it matters directly to AI: the IEA expects renewables, led by solar, to meet about half of the growth in data center demand through 2030. AI helps solar at several points in its life cycle.
- Materials discovery. Machine learning guides the search for better and cheaper photovoltaic materials, including perovskites, by predicting which compositions and processing conditions are worth testing. The Johns Hopkins SPIRAL Center work linked at the top of this page is one example.
- Forecasting. Models trained on satellite imagery, sky cameras, and weather forecasts predict solar output minutes to days ahead, helping grid operators plan around clouds and sunsets.
- Inspection and maintenance. Computer vision on drone and thermal imagery spots cracked cells, hot spots, soiling, and failed inverters across large solar farms faster than manual inspection.
- Design and sales. Software tools automate roof modeling, panel layout, shading analysis, and quoting, as described by The Solar Labs in the video below.
- Powering AI directly. Researchers are now comparing solar plus storage against grid power and on-site natural gas for supplying AI data centers; a 2026 Stanford Energy Seminar by Carnegie Mellon's Costa Samaras presented an open-source techno-economic model of these options across the continental United States.
For the physics of the Sun itself, see Sun / Solar.
|
|
Reducing Energy Consumption of AI
Because energy scales with arithmetic, data movement, tokens, and overhead, there are levers at every layer of the stack. Google's 33-fold reduction in energy per median Gemini prompt over a single year shows how large the combined effect can be, and Patterson and colleagues found that model, hardware, facility, and siting choices together can cut training energy by up to 100 times and emissions by up to 1,000 times. The table summarizes the main levers.
| Lever | How it saves energy | What the evidence shows | Caveats |
|---|---|---|---|
| Choose the right model | Smaller or distilled models do less arithmetic per token | Independent benchmarks find per-query energy varies by well over an order of magnitude across models | Smaller models may fail harder tasks or need retries |
| Use reasoning only when needed | Fewer generated tokens | Reasoning modes raised energy by well over an order of magnitude on average in the AI Energy Score tests | Some tasks genuinely benefit from extended reasoning |
| Lower numeric precision (INT8, FP8, FP4) | Cheaper arithmetic and fewer bytes moved | An 8-bit integer multiply costs a small fraction of a 32-bit floating-point multiply | Accuracy must be validated; not every layer tolerates low precision |
| Mixture-of-experts architectures | Only part of the model is active for each token | Separates total model size from per-token compute | Larger memory footprint and routing overhead |
| Batching, caching, and speculative decoding | Spread each weight read across many requests and skip repeated work | Better hardware utilization contributed part of Google's per-prompt gains | Larger batches can increase latency |
| Energy-aware serving configuration and GPU tuning | Pick batch sizes, parallelism, and clock or power settings for energy as well as speed | ML.ENERGY reports savings above 40% in some cases without changing the computation | Gains depend on latency targets |
| Specialized hardware | Accelerators and dataflows built around tensor math and data reuse | Google reports its Ironwood TPU is about 30 times more energy efficient than its 2018 TPU | Frequent hardware refreshes add embodied emissions |
| Efficient facilities | Lower PUE through better cooling and power conversion | Leading fleets report PUE near 1.1 | Evaporative cooling trades electricity for water |
| Carbon-aware and grid-aware scheduling | Run flexible work when and where power is clean and plentiful | Location and timing choices can cut training emissions dramatically | Interactive inference is hard to shift in time |
Measuring before optimizing. You cannot reduce what you do not measure. Open tools now make this practical: CodeCarbon estimates emissions from code as it runs, the ML.ENERGY Leaderboard and its open-source benchmark report measured inference energy for open models, and Hugging Face's AI Energy Score rates models on a standardized set of tasks. When comparing published numbers, check the boundary (accelerator only, or full serving stack and facility), the workload (prompt length, reasoning on or off), and whether the figure is measured or estimated.
Edge versus cloud. Vivienne Sze's talks below make the case for processing on the device when privacy, latency, or bandwidth demand it. On-device AI avoids network and data center overhead, but it runs on smaller, less specialized chips and battery budgets, so the most efficient place to run a task depends on the model size and how often it runs.
|
|
|
|
|
|
Entropy
Entropy is a measure of how many hidden ways something could be arranged while looking exactly the same from the outside. Equivalently, it measures how much information is missing about a system's fine details. Entropy underlies the second law of thermodynamics and connects physics to chemistry, life, the direction of time, and artificial intelligence. It also explains why every computer, including every AI system, must turn electricity into heat.
Beyond the messy room
The usual story says entropy is the universe's tendency to make a mess, like a tidy bedroom that's chaos again by Friday. That story gets it nearly backwards. A shirt on the floor and a shirt on a hanger have essentially the same entropy, because the shirt's molecules don't care where you think it belongs. Entropy measures something stranger and more useful.
What entropy actually is
Rudolf Clausius coined the word in 1865 while studying steam engines. He noticed that heat flows on its own only from hot to cold, and that every real engine loses some heat in a way that can never be recovered. He named the quantity that tracks this one-way loss "entropy" and summed up his science in two sentences: the energy of the world is constant, and the entropy of the world tends toward a maximum. Clausius knew what entropy did, but not what it was.
Ludwig Boltzmann answered that in 1877 by counting. A party balloon holds more than 100,000 billion billion air molecules, each with its own position and speed. You can't see any of them. You see only the balloon's size, pressure, and temperature. A staggering number of exact molecular arrangements, called microstates, would produce exactly the same balloon you observe, called the macrostate. Entropy is the count of how many microstates fit one macrostate. Boltzmann's formula is carved on his gravestone in Vienna:
- <math>S = k \log W</math>
Here S is entropy, W is the number of possible arrangements, and k (Boltzmann's constant) converts that count into ordinary physical units. One historical footnote: the compact form on the stone was actually written down by Max Planck around 1900, who also introduced the constant k. Boltzmann's own papers expressed the same idea in a different notation, but the insight was his.
This makes the second law almost obvious. Things move toward higher entropy for the same reason a shuffled deck almost never comes out in order: there are vastly more ways to be in a common state than a rare one. Nothing pushes the molecules. They wander, and wandering almost always lands them in the biggest crowd of possibilities.
In 1948, Claude Shannon added a second way to see it. Working on telephone lines at Bell Labs, he found that the measure of a message's uncertainty has exactly the same mathematical form as Boltzmann's entropy:
- <math>H = -\sum_{i} p_{i} \log_{2} p_{i}</math>
where pi is the probability of each possible message, and H is measured in bits. So entropy can also be read as missing information: the number of yes-or-no questions you'd need to ask to pin down a system's exact hidden state.
The cleanest definition is this: entropy measures how many hidden ways something could be arranged while looking exactly the same from outside. Equivalently, it measures how much you don't know about its fine details.
Why "disorder" is the wrong picture
Sometimes the arrangement with the most possibilities looks the most orderly. In 1957, two teams simulated boxes of perfectly hard bouncing balls, with no attraction between them at all. Once the balls filled about half the available space, they spontaneously formed a neat crystal. The crystal won because a jammed, jumbled crowd traps each ball in a tiny cage. In a crystal, each ball gets a slightly roomier cell to jiggle in, so the crystal has more hidden arrangements. This is entropic crystallization, and Pusey and van Megen confirmed it in 1986 with real plastic spheres in liquid.
Gravity flips the rule even further. A smooth cloud of gas under its own gravity is a low-entropy state, and its entropy rises as it clumps, heats up, and radiates light. A star isn't entropy losing. It's what rising entropy looks like when gravity is in charge. Galaxies, planets, and the ground under your feet are all entropy at work, not entropy defeated.
A law of odds, not a wall
Because the second law rests on counting, it's statistical. James Clerk Maxwell compared it to throwing a tumbler of water into the sea: you can't get the same water back out. That isn't strictly impossible, just so unlikely no one will ever see it happen.
At tiny scales, the odds aren't lopsided at all. The fluctuation theorem, developed in the 1990s by Denis Evans, E. G. D. Cohen, and Gary Morriss (1993) and by Evans and Debra Searles (1994), predicts exactly how often entropy will run backward. In 2002, a team at the Australian National University held a microscopic bead with a laser beam and dragged it through water. For stretches of up to about two seconds, the bead drew energy from the random jostling of the water and entropy went down, at rates consistent with the theorem's predictions.
In 2019, researchers in Brazil went further. Using two atomic nuclei inside chloroform molecules, they showed that when the nuclei were linked by quantum correlations, heat flowed spontaneously from cold to hot. Nothing was broken: the correlation itself was a store of order, and it got used up. Counted together, the heat and the information still obeyed the law. The lesson running through all of this is that information belongs on the same ledger as entropy.
Entropy and chemical bonding
Chemical bonding is a tug-of-war between energy and entropy (see Chemistry). When two atoms bond, they lose freedom. Two independent atoms become one molecule with fewer possible arrangements, so the atoms' own entropy drops. That seems to defy the second law, but forming a bond releases energy as heat, and that heat spreads into the surroundings and raises their entropy by more than the atoms lost.
Chemists track this balance with Gibbs free energy:
- <math>\Delta G = \Delta H - T\Delta S</math>
Here ΔH is the energy released or absorbed, ΔS is the change in entropy, and T is temperature. A reaction happens on its own when ΔG is negative. Temperature acts as the referee. The hotter things get, the more the entropy term counts, which is why heating melts solids, boils liquids, and eventually tears molecules apart. At high enough temperature, freedom beats bonding.
Some processes run almost entirely on entropy.
The depletion force
Sho Asakura and Fumio Oosawa showed in 1954 that large particles in a crowd of small ones get pushed together. They aren't attracted to each other; pairing up simply frees more room for the small particles to move. The inside of every living cell is crowded this way, and the depletion force helps proteins find each other and fold.
Liquid crystals
Lars Onsager predicted in 1949 that crowded rod-shaped molecules line up side by side, because pointing the same way gives each rod more room to move. That's the principle behind liquid-crystal screens.
The hydrophobic effect
Oil and water separate mainly because of water's entropy, not because they strongly repel. Water molecules next to oil get forced into restricted, cage-like arrangements. When oil droplets clump together, those water molecules are freed. This entropic push is a major reason proteins fold with their oily parts tucked inside, and why cell membranes assemble themselves.
Instant cold packs
These dissolve a salt that absorbs heat, so the reaction gets colder as it goes. It still proceeds because the dissolved ions have vastly more possible arrangements than the solid crystal did.
Entropy, life, and meaning
Life
In his 1943 Dublin lectures, published in 1944 as What Is Life?, Erwin Schrödinger argued that living things survive by "feeding on negative entropy." They stay organized by drawing order in from their surroundings and dumping disorder back out. He later admitted "free energy" would have been the more exact term, which is the same quantity chemists use for bonding.
Earth itself works this way. For every photon of sunlight that arrives, Earth sends roughly 20 dim infrared photons back to space. The energy in and out is nearly equal, but 20 small packets can be arranged in far more ways than one large one. That outflow of entropy pays for weather, rivers, rainforests, and the chemistry in your cells. Earth isn't breaking the second law; it's a machine for speeding it up. Every heartbeat and every thought raises the universe's entropy a little. You aren't an exception to the law. You're one of the ways it gets carried out.
Memory and time
The basic laws of physics barely distinguish past from future, so why do you remember yesterday and not tomorrow? Making any record, whether a mark in the brain or a footprint in sand, raises total entropy. So every memory lines up with the direction entropy increases. In 1985, Stephen Hawking argued that if the universe began to collapse, entropy and memory would reverse together. Don Page and Raymond Laflamme showed entropy would keep rising anyway, and Hawking later called this his greatest scientific mistake. His reasoning about the mind held up, though: the direction you feel time flowing is a fact about entropy recorded in you.
The gift at the beginning
Charles Lineweaver and Chas Egan found in 2010 that nearly all the universe's entropy is locked inside supermassive black holes. Run the clock backward, and the early universe had almost none. It was a hot, nearly perfectly smooth gas, which with gravity in charge is the least likely state. Roger Penrose estimated the odds of such a special start arising by chance at 1 in 1010123. No known law explains it; physicists simply have to assume it. And it won't last. As Lord Kelvin foresaw, when entropy reaches its maximum there will be no hot or cold left, no differences of any kind, and with no differences nothing can happen at all.
What that might mean
Physics describes how things happen, not why they matter, so meaning remains a philosophical question. The entropy picture does offer some perspectives people find compelling. Everything we value, such as thought, love, memory, and creation, is a process, and processes run on gradients. Entropy isn't the enemy of those things; it's what makes them possible. We also live in a narrow window between a beginning nobody can explain and an end nobody can avoid. Some find that bleak. Others find it makes the present more precious, since a finite window gives each moment weight. One way to put it: the second law isn't the law that destroys things. It's the reason anything happens at all.
Entropy and artificial intelligence
The connections between entropy and AI run deep, mostly through Shannon's information side of entropy, and increasingly through the physical side as well.
AI learns by reducing entropy
Large language models are trained by minimizing "cross-entropy," a measure of how surprised the model is by the actual next word in real text:
- <math>H(p, q) = -\sum_{x} p(x) \log q(x)</math>
Here p is the true distribution of text and q is the model's prediction. Learning amounts to shrinking the model's missing information about the world's patterns. Shannon also showed that entropy sets the limit on how far data can be compressed, and predicting well and compressing well turn out to be the same skill.
The vocabulary comes straight from physics
The math that turns a model's raw scores into probabilities is Boltzmann's distribution:
- <math>p_{i} = \frac{e^{z_{i}/T}}{\sum_{j} e^{z_{j}/T}}</math>
where zi is the model's score for option i. The "temperature" setting T on AI systems plays exactly the mathematical role that temperature plays in Boltzmann's distribution, with the scores standing in for negative energies. High temperature spreads probability across more options, producing more varied, higher-entropy output. Low temperature concentrates it, producing more predictable output. Early neural networks called Boltzmann machines were named after him, and John Hopfield and Geoffrey Hinton shared the 2024 Nobel Prize in Physics partly for this physics-inspired work. Models that define probabilities directly through an energy function are covered in Energy-based Model (EBN).
Image generators run entropy in reverse
Diffusion models were explicitly inspired by nonequilibrium thermodynamics; the 2015 paper by Jascha Sohl-Dickstein and colleagues that introduced them borrowed its framework directly from that field. During training, noise is added to images step by step until only static remains, the maximum-entropy state. The network then learns to run that process backward, sculpting structure out of noise. It's like learning to scoop Maxwell's tumbler of water back out of the sea, paid for with information and computation.
Thinking has an entropy cost
Leo Szilard showed in 1929 that knowing has a price, which is why Maxwell's demon fails. Rolf Landauer later showed in 1961 that erasing even one bit of information must release a minimum amount of heat:
- <math>Q \ge k_{B} T \ln 2 \approx 2.9 \times 10^{-21}\ \mathrm{J} \quad (T \approx 300\ \mathrm{K})</math>
This is no longer just theory. In 2012, Antoine Bérut and colleagues in Lyon trapped a single glass bead in a laser-made double well, used its two positions as a one-bit memory, and measured the heat released when they erased it. In the limit of slow erasure, the heat settled right at Landauer's bound. In 2025, a team working with ultracold atoms in Vienna and Berlin extended the test to a quantum many-body system, tracking how entropy, shared information, and dissipated energy evolve together.
Every computation pays an entropy tax, and real hardware pays far more than that minimum: a single 32-bit multiplication in modern digital logic costs on the order of a billion times the Landauer limit once all the switching and data movement is counted (see The Physics above). In this sense AI resembles life: it builds local order, in the form of organized knowledge and useful outputs, by consuming low-entropy electricity and exporting high-entropy heat. The difference is how much room for improvement remains between what today's chips spend and what physics demands.
Entropy as an uncertainty gauge
Entropy also helps AI systems know when they don't know. In 2024, Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal at Oxford published a method in Nature called semantic entropy. The idea: ask a language model the same question several times, group the answers by meaning rather than exact wording, and compute the entropy over those meaning groups. If the model keeps giving answers that mean different things, its semantic entropy is high and it is probably confabulating, inventing an arbitrary answer. The method flagged likely false claims even in generated biographies. It has limits: it only catches inconsistent errors, not errors a model makes confidently every time, and it costs extra energy because each question has to be answered several times.
Entropy collapse in reasoning models
Reasoning models are often refined with reinforcement learning, and there entropy plays a different role: it measures how much the model still explores alternative answers. A 2025 study by Ganqu Cui and colleagues found that the entropy of the model's output distribution typically collapses early in training, and that performance then plateaus. They fit an empirical relationship in which performance rises as entropy is used up, with the form R = −a eH + b, meaning that once entropy is exhausted, so is easy improvement. They traced the collapse to a covariance between action probabilities and their advantages, and proposed limiting updates on the tokens driving it. The result is a preprint, but the methods were adopted into widely used open-source training libraries. Entropy, in this setting, is a budget for curiosity.
Thermodynamic computing: using noise instead of fighting it
Conventional chips spend much of their energy keeping signals far above the thermal noise of their own transistors, so that every bit is reliably a 0 or a 1. Thermodynamic computing turns that around. Many AI tasks, especially generative ones, are really about drawing samples from probability distributions, and physical systems sitting in a heat bath naturally sample from Boltzmann distributions. Let the physics do the sampling, and you may save most of the energy.
Several groups are testing this. Normal Computing built a small "stochastic processing unit" of coupled electrical resonators that uses noise to draw Gaussian samples and invert matrices. Extropic builds chips of "probabilistic bits" from ordinary transistors that sample from energy-based models, and has proposed "denoising thermodynamic models" that break generation into a series of simpler sampling steps, much like a diffusion model. Stephen Whitelam at Lawrence Berkeley National Laboratory showed in 2026 that a physical system obeying Langevin dynamics could itself act as a generative model, with training equivalent to minimizing the heat the process gives off, which ties the design of generative AI directly to the second law. These ideas are promising but early: the hardware is small, the largest efficiency claims are projections for simple benchmarks, and none has yet run workloads comparable to today's large models. Details and caveats are in Recent Developments.
Summary
Across all of these areas, the same idea applies: count the hidden possibilities and follow where the information and heat flow. Atoms bond when doing so increases the universe's total possibilities. Life persists by speeding up that increase. Memory and time point the direction entropy rises. AI learns by reducing its own uncertainty while raising entropy in the wider world, and the electricity it consumes is the bill for that exchange. Physics guarantees the bill can never reach zero, but it also shows that today's machines pay far more than they must.
Resources
Additional Viewing
Key Questions on Energy and AI (International Energy Agency webinar, June 2026). IEA Director of Sustainability, Technology and Outlooks Laura Cozzi and the lead authors of the IEA's 2026 report present its findings, followed by a panel with executives from government and the technology and energy sectors. The session covers the revised data center demand outlook, how much efficiency can moderate growth, whether data centers are raising electricity prices, and the barriers to using AI across the energy sector. It is the most direct way to hear the reasoning behind the demand figures in Recent Developments. Watch on YouTube
Global Conference on Energy and AI (International Energy Agency, December 2024). A recording of the IEA's Paris conference, which the agency describes as the largest gathering to date of the technology and energy sectors on this topic and which laid the groundwork for its 2025 Energy and AI report. Policymakers, utilities, and technology companies discuss data center electricity demand, grid and supply constraints, and AI applications in energy. It is a useful baseline for how the debate looked just before the 2025 measurements and forecasts arrived. Watch on YouTube
AI-Driven Data Centers and Demand Impact: A Day on the ERCOT Grid (seminar, February 2026). Jack Steinhauser, President of Hunter Stuart Energy Advisors, uses the Texas grid operated by ERCOT to examine how large AI data center loads affect daily grid operations and demand. Texas is one of the fastest-growing data center markets and the setting for the large-load rules described under grid flexibility above, so the talk grounds that discussion in a real power system. Watch on YouTube
AI Is About to Drain America's Power Grid? with Andrew Chien (University of Chicago, Big Brains). Andrew Chien, a University of Chicago computer scientist and senior scientist at Argonne National Laboratory, explains why AI's electricity demand is outrunning decades of computing efficiency gains and why letting data centers build their own generation outside the grid raises new problems. He argues for data centers that cooperate with the grid, flexing their load to match variable clean generation, which connects directly to the Duke flexibility study and the grid physics discussed on this page. Watch on YouTube
The Stochastic Thermodynamics of Computation, David Wolpert (Joint ICTP-SISSA Colloquium). David Wolpert of the Santa Fe Institute explains how modern nonequilibrium thermodynamics extends Landauer's work to real computers, which are small, fast, and far from equilibrium. This foundational lecture shows why the Landauer bound is a floor that fast computers cannot reach, and what determines the extra energy they must spend, which underpins the Physics and Entropy sections of this page. Watch on YouTube
Thermodynamic Computing: Better than Quantum? Guillaume Verdon and Trevor McCourt (Extropic). Extropic's co-founders explain the idea behind thermodynamic computing: hardware that harnesses thermal fluctuations to sample from probability distributions instead of suppressing noise, and how this differs from quantum computing's need for near-zero temperatures. Recorded in 2024, before Extropic's first hardware platform and its peer-reviewed architecture paper, it should be treated as the founders' case for the approach rather than independent evidence; compare it with the caveats in Recent Developments. Watch on YouTube
References
Demand estimates and outlooks
- Energy and AI: Executive Summary | International Energy Agency April 2025 special report with global estimates of data center electricity use (about 415 TWh in 2024) and projections to 2030 and 2035, plus analysis of AI applications in the energy sector.
- Key Questions on Energy and AI | International Energy Agency April 2026 update reporting 17% growth in data center electricity in 2025, about 50% growth at AI-focused sites, rapid efficiency gains per task, and physical bottlenecks to expansion.
- DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers | US Department of Energy Announcement of LBNL's 2024 report: 176 TWh in 2023 (4.4% of US electricity), with 325 to 580 TWh projected for 2028.
- United States Data Center Energy Usage Report: 2025 Update | Lawrence Berkeley National Laboratory June 2026 bottom-up update projecting 649 TWh in 2030 (about 11.8% of US electricity), with a 9.5% to 15.3% range across sensitivity scenarios.
- Recalibrating Global Data Center Energy-Use Estimates | Eric Masanet and colleagues, Science (2020) Found global data center energy use rose only about 6% from 2010 to 2018 while computing instances rose about 550%, the baseline that AI growth later broke.
Per-prompt and per-model measurement
- Measuring the Environmental Impact of Delivering AI at Google Scale | Google (2025) Technical paper reporting 0.24 Wh, 0.03 g CO2e, and 0.26 mL of water for the median Gemini Apps text prompt, with a full-stack measurement boundary.
- Measuring the Environmental Impact of AI Inference | Google Cloud Summary of the Gemini methodology and the reported 33-fold energy and 44-fold emissions reduction per prompt over twelve months.
- Our Contribution to a Global Environmental Standard for AI | Mistral AI (July 2025) Full lifecycle analysis of Mistral Large 2 with Carbone 4 and ADEME, including training totals and per-response emissions and water.
- How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | Nidhal Jegham and colleagues (2025) Infrastructure-aware estimates for more than 30 models, showing reasoning models can use more than 70 times the energy of small models per long prompt.
- The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization | Jae-Won Chung and colleagues, University of Michigan (NeurIPS 2025) Open benchmark measuring inference energy for 40 model architectures and showing automated configuration savings above 40% in some cases.
- AI Reasoning Models Use Far More Power | Bloomberg via Insurance Journal (December 2025) Coverage of the AI Energy Score study of 40 open models comparing energy with reasoning enabled and disabled.
Training footprint debate
- The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink | David Patterson and colleagues, IEEE Computer (2022) Argues best practices can cut training energy up to 100x and emissions up to 1,000x, and that many published estimates were badly off.
- Position: Neglecting the Sustainability of AI is Fuelling a Global AI Arms Race | arXiv (2025) Argues that rebound effects have kept AI's total energy use and emissions rising despite per-model efficiency gains.
Corporate disclosures, grid, and supply
- Google's Electricity Use Grew 37 Percent in 2025 | DatacenterDynamics Summary of Google's 2026 environmental report: 42 TWh for data centers, PUE 1.09, rising supply chain emissions and water use.
- US Grid Has Headroom for Flexible Load | Utility Dive Coverage of Duke University's Rethinking Load Growth study estimating 76 to 126 GW of new load could connect with modest curtailment.
- Constellation Seeks Regulator's Help for 2027 Plant Restart | World Nuclear News Status of the Crane Clean Energy Center restart under a 20-year agreement with Microsoft.
AI for energy systems
- GenCast Predicts Weather and the Risks of Extreme Conditions | Google DeepMind (December 2024) Diffusion-based ensemble weather model published in Nature, including a wind power forecasting test.
- Bringing AI to the Next Generation of Fusion Energy | Google DeepMind (October 2025) Research partnership with Commonwealth Fusion Systems built around the open-source TORAX plasma simulator and SPARC.
- Accelerating Fusion Science Through Learned Plasma Control | Google DeepMind The 2022 deep reinforcement learning control of the TCV tokamak with EPFL, published in Nature.
- Avoiding Fusion Plasma Tearing Instability with Deep Reinforcement Learning | Jaemin Seo and colleagues, Nature (2024) Princeton-led work on DIII-D forecasting and avoiding tearing instabilities in real time.
- How Microsoft's AI Screened Over 32 Million Candidates to Find a Better Battery | Microsoft and Pacific Northwest National Laboratory (2024) AI-accelerated discovery of the N2116 solid electrolyte, which could reduce lithium use.
Hardware efficiency
- Computing's Energy Problem (and What We Can Do About It) | Mark Horowitz, ISSCC (2014) Source of the widely used energy-per-operation and energy-per-memory-access estimates showing data movement dominates.
- Efficient Processing of Deep Neural Networks: A Tutorial and Survey | Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel Emer, Proceedings of the IEEE (2017) Foundational survey of dataflows, data reuse, and hardware design for energy-efficient deep learning.
Thermodynamics, entropy, and computation
- Experimental Verification of Landauer's Principle Linking Information and Thermodynamics | Antoine Bérut and colleagues, Nature (2012) First experimental confirmation of the minimum heat released when erasing a bit.
- Experimentally Probing Landauer's Principle in the Quantum Many-Body Regime | Stefan Aimet and colleagues, Nature Physics (2025) Extends tests of Landauer's principle to a quantum field simulator of ultracold atoms.
- Experimental Demonstration of Violations of the Second Law of Thermodynamics for Small Systems and Short Time Scales | G. M. Wang and colleagues, Physical Review Letters (2002) The Australian National University optical trap experiment confirming the fluctuation theorem.
- Reversing the Direction of Heat Flow Using Quantum Correlations | Kaonan Micadei and colleagues, Nature Communications (2019) Demonstration that quantum correlations can drive heat from cold to hot while respecting the second law.
- A Larger Estimate of the Entropy of the Universe | Chas Egan and Charles Lineweaver, Astrophysical Journal (2010) Shows supermassive black holes dominate the universe's entropy budget.
- Thermodynamic Computing System for AI Applications | Denis Melanson and colleagues, Normal Computing, Nature Communications (2025) An eight-cell stochastic processing unit performing Gaussian sampling and matrix inversion with noise.
- Generative Thermodynamic Computing | Stephen Whitelam, Lawrence Berkeley National Laboratory (Physical Review Letters, 2026) Framework in which a Langevin system's natural dynamics perform diffusion-style generation, trained by minimizing heat emission.
- Thermodynamic Computing: From Zero to One | Extropic (October 2025) Company announcement of the X0 chip, XTR-0 platform, and denoising thermodynamic models; claims are the company's own.
Entropy in machine learning
- Deep Unsupervised Learning using Nonequilibrium Thermodynamics | Jascha Sohl-Dickstein and colleagues (2015) The paper that introduced diffusion models, framed directly in terms of nonequilibrium thermodynamics.
- Detecting Hallucinations in Large Language Models Using Semantic Entropy | Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal, Nature (2024) Entropy over meanings, rather than words, flags likely confabulations.
- The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models | Ganqu Cui and colleagues (2025) Documents entropy collapse during reinforcement learning of reasoning models and proposes methods to control it.
- The Nobel Prize in Physics 2024 | Nobel Prize Outreach Awarded to John Hopfield and Geoffrey Hinton for foundational work enabling machine learning with artificial neural networks.
Policy and reporting
- Annex XI: Technical Documentation for Providers of General-Purpose AI Models | EU Artificial Intelligence Act Requires documentation of known or estimated model energy consumption, which may be estimated from computational resources.