Yes, But

US data centers consumed 192 TWh in 2024 and are projected to reach roughly 300 TWh by 2026, with AI accounting for about 75 TWh a year, according to Berkeley Lab, EPRI and IEA figures cited in an analysis of AI's electricity demand. The analysis notes that inference represents about 65% of AI-server energy in 2024 and is heading to 80%, while Nvidia claims its Blackwell platform delivers up to 25× less energy per inference than Hopper and Vera Rubin up to 10× more tokens per megawatt than GB200. Single sites are now nuclear-plant sized, with xAI's Colossus 2 drawing 946 MW and Microsoft's Fairwater Wisconsin planned at 2.3 GW.

Yes, but… Electrical use of AI I was on a fishing boat for the last 4 days, which brings in a mixed crowd. One of the discussions that came up was energy usage of AI, and I found myself saying "yes, but..." quite a bit. There's a lot of nuance and factors that go into the future energy usage of AI, and I want to talk through them. AI is already a grid-scale load US data centers used 176 TWh in 2023 https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers 4.4% of US power and 192 TWh in 2024 https://escholarship.org/uc/item/33m6w3x0 . Berkeley Lab's reference case then has them growing 22% in 2025 and 29% in 2026, to roughly 300 TWh. EPRI estimates that 15–25% of that is AI https://restservice.epri.com/publicattachment/96784 . The IEA says AI-focused data centers grew 50% in 2025 alone https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions . Put together, that is about 75 TWh a year for AI in 2026. Training dominates the headlines despite inference using more , since the single up-front cost raises eyebrows. GPT-3 took 1,287 MWh https://arxiv.org/abs/2104.10350 , and Grok 4 an estimated 310 GWh https://epoch.ai/data-insights/grok-4-training-resources , 240× more. More numbers - 75 TWh is about 1.8% of the 4,195 TWh https://www.eia.gov/outlooks/steo/report/elec coal renew.php the US consumed in 2025, or the electricity of roughly 7 million homes. - Single sites are now nuclear-plant sized: xAI's Colossus 2 draws 946 MW https://epoch.ai/graphs/largest-ai-data-centers-by-power-capacity , and Microsoft's Fairwater Wisconsin is planned at 2.3 GW. - The power drawn by frontier training runs grows 2.2× a year https://epoch.ai/blog/power-demands-of-frontier-ai-training , and the largest single runs could need 4–16 GW by 2030. - Inference is most of the bill: about 65% of AI-server energy in 2024, heading to 80% https://escholarship.org/uc/item/33m6w3x0 . - A single prompt is tiny: 0.24 Wh https://arxiv.org/abs/2508.15734 for a median Gemini text prompt and 0.34 Wh https://blog.samaltman.com/the-gentle-singularity for an average ChatGPT query. But ChatGPT alone handled 2.5 billion prompts a day https://techcrunch.com/2025/07/21/chatgpt-users-send-2-5-billion-prompts-a-day/ in mid-2025, and reasoning and agent tasks can use hundreds to thousands of times https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary more per query. Yes, but GPUs are getting more efficient Nvidia pitched Blackwell as up to 25× less energy https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing per inference than Hopper. Vera Rubin, in full production since CES 2026 https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer , claims up to 10× more tokens per megawatt https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/ than GB200, and CoreWeave measured 10× on DeepSeek-R1 https://blogs.nvidia.com/blog/vera-rubin/ . On Nvidia's own chart, energy per token on a 1.8T-parameter MoE fell from 17,000 J on Pascal to 0.4 J on Blackwell https://mcomputers.cz/wp-content/uploads/2025/11/08 Koen van den Berg NVIDIA-Blackwell -powering-the-next-generation-AI-Factories-EN-1.pdf , about 45,000× in eight years. The overall hardware trend is 1.28–1.40× more efficient per year Epoch https://epoch.ai/blog/can-ai-scaling-continue-through-2030 , Epoch https://epoch.ai/data-insights/ml-hardware-energy-efficiency , and the IEA says it is gradually slowing https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf . More numbers - Rubin's racks are fully liquid-cooled, with a PUE of 1.1 vs ~1.25 https://developer.nvidia.com/blog/scaling-token-factory-revenue-and-ai-efficiency-by-maximizing-performance-per-watt/ . - Much of that headline gain is lower precision FP16 → FP4 , bigger NVLink domains and software, not the silicon itself. At fixed precision, Rubin's FLOPs per watt are roughly level with Blackwell's, and dense-model gains are closer to 2–3× https://blog.barrack.ai/nvidia-rubin-vs-blackwell-rent-now-or-wait/ . - Every chip trend before it slowed too. Computations per kWh doubled every 1.6 years until 2000, then every 2.6 https://www.electronicdesign.com/technologies/microprocessors/article/21802037/energy-efficiency-of-computing-whats-next . Single-thread speed went from +52% a year to about +3% https://dl.acm.org/doi/10.1145/3282307 . - Physics also sets a ceiling: CMOS chips top out around 207× an H100's efficiency https://arxiv.org/abs/2312.08595 , and dropping from FP16 to FP4 has already used most of the number-format gains. - So we let the trend fade as it nears that wall, ~130× from Blackwell range 26–650× . Ten years then yields 6–14× instead of 12–29×. Hold usage at today's level, and AI's electricity still falls every year, just less steeply. Yes, but AI adoption is only at 2% a16z's State of Markets II https://a16z.com/state-of-markets-ii/ shows only ~2% of US households paying for an AI subscription as of April 2026 PNC transaction data . Adoption is going to skyrocket in the next decade. AI is being adopted about twice as fast https://www.stlouisfed.org/on-the-economy/2024/sep/rapid-adoption-generative-ai as the internet was in the '90s, and the internet went from 14% of US adults in 1995 https://www.pewresearch.org/internet/2014/02/27/part-1-how-the-internet-has-woven-itself-into-american-life/ to 68% in 2005 https://www.pewresearch.org/internet/fact-sheet/internet-broadband/ . At twice the speed, AI's next ten years look like the internet's twenty, which reached 86% by 2015 https://www.pewresearch.org/internet/fact-sheet/internet-broadband/ . If paid AI follows the same kind of S-curve, from ~2% of households to 68–86%, usage rises 34–43× , more than cancelling the hardware gains. Yes, but models are getting more efficient The densing law https://www.nature.com/articles/s42256-025-01137-0 Xiao et al., Nature Machine Intelligence , 2025 finds that the capability per parameter of open LLMs doubles about every 3.5 months: the same score from half the model, every season. The preprint https://arxiv.org/abs/2412.04315 clocks a 266.7× price drop for equal performance, from GPT-3.5 Dec 2022 to Gemini 1.5 Flash Aug 2024 . These improvements had a direct impact on energy usage. Google cut energy per Gemini prompt 33× in a single year https://arxiv.org/abs/2508.15734 . Obviously these densing gains won't last forever. A decade of it gives an obviously absurd result: at the low end, all of America's AI running on less power than one house. The reality is requirements for a given capability slow sharply as that capability ages https://epoch.ai/publications/the-plunging-price-of-thought , but there are still tremendous gains to be made. More numbers Andrej Karpathy's GPT-2 reruns are a great way to view the progression of this. OpenAI's 2019 training run cost roughly $43,000 https://github.com/karpathy/nanochat/discussions/481 . llm.c reproduced the 1.5B model in 24 hours for $672 https://github.com/karpathy/llm.c/discussions/677 in 2024. By March 2026, nanochat beat it in 1.65 hours, for about $48 https://github.com/karpathy/nanochat : roughly 900× cheaper in seven years, or 2.6× a year. On fixed hardware, the community speedrun to GPT-2-small loss fell from 45 minutes to 1.13 https://github.com/KellerJordan/modded-nanogpt . For the estimate here, we took the 2.6× a year Karpathy to 10.8× a year densing law and let it fade, with the yearly gain halving every 3–10 years. Yes, but TAM is larger than the population Agents, not people, are now the 1 user of tokens. On OpenRouter, agent API keys out-used humans for the first time on February 6, 2026 https://openrouter.ai/data/newsletter/cache-is-crucial , and by late September they were using ~7× the tokens humans did. OpenClaw wakes itself on a 30-minute heartbeat https://github.com/openclaw/openclaw/blob/main/docs/gateway/heartbeat.md by default, and Jensen Huang predicts a future with 100 billion agents https://thenewstack.io/huang-semiconductor-tenfold-ai-agents/ . Agents are going proactive and are growing in use massively. Some of that 7× is cheaper than it looks: ~90% of agent tokens https://openrouter.ai/data/newsletter/cache-is-crucial are cached re-reads of the same prompt. Weighted by compute, agents are about 1.5–2× human usage today. We assume agent compute per adopter grows 2–20× by 2036, or about 1.07–1.35× a year. More numbers - The top public app by tokens on OpenRouter is Hermes Agent https://openrouter.ai/apps , a persistent agent that runs scheduled automations. OpenClaw has 391k GitHub stars https://github.com/openclaw/openclaw , and its default heartbeat means 48 unprompted sessions a day. - ChatGPT Pulse https://mobilesyrup.com/2025/09/26/chatgpt-launches-new-briefing-feature-called-pulse/ researches overnight before you ask, Claude Code's scheduled tasks fire "even while the session is idle" https://code.claude.com/docs/en/costs , and one Cursor customer runs 2,000+ agent runs a week https://releasebot.io/updates/cursor with no manual prompting. - Microsoft already has ~40 million agents https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 registered. - Each agent task is heavy. Agents use 4× the tokens of chat https://www.anthropic.com/engineering/multi-agent-research-system and multi-agent systems 15×, Claude Code agent teams use ~7× https://code.claude.com/docs/en/costs , and reasoning output is growing 5× a year https://epoch.ai/data-insights/output-length . - Meta employees burned 60 trillion tokens in 30 days https://the-decoder.com/meta-employees-compete-for-token-consumption-on-an-internal-ai-leaderboard/ , ~700M each. - The 2–20× assumption is a small fraction of 2026's pace, and it stacks on top of human adoption. Yes, but hardware is getting more optimized There's a ladder of ways to serve a model, each more efficient and less flexible than the last: CPUs, then GPUs, then TPUs, then fixed-function ASICs. Most AI still runs on GPUs: Nvidia shipped 67% of new AI compute https://epoch.ai/graphs/ai-chip-market-share-by-designer in Q4 2025. The top rung is Taalas's HC1 https://taalas.com/?p=251 , which hard-wires Llama 3.1 8B into the transistors themselves and gets roughly 3× a fully loaded GPU's tokens per joule. We move ASICs from ~25% of AI power today to 40–70% by 2036, at 1.5–20× a GPU's tokens per watt: up to ~13× more AI from the same power . More numbers - GPUs are about 42× more efficient on inference https://blogs.nvidia.com/blog/2023/05/21/gpu-energy-efficiency-nersc/ than CPUs Nvidia's figure . - Google's first TPU managed 30–80× the TOPS/W https://research.google/pubs/pub46078/ of the CPUs and GPUs of its day, and Ironwood is ~30× the first Cloud TPU https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/ . - AWS rates Trainium3 at 4× the efficiency of Trainium2 https://press.aboutamazon.com/2025/12/trainium3-ultraservers-now-available-enabling-customers-to-train-and-deploy-ai-models-faster-at-lower-cost , and transformer-only Etched just raised $700M at a $21B valuation https://thenextweb.com/news/etched-700m-series-d-21-billion-jane-street . TrendForce expects ASIC unit shipments to pass GPUs in 2026 https://www.trendforce.com/presscenter/news/20251030-12762.html . - The HC1 is TSMC 6 nm, 815 mm², ~200–250 W per card https://www.nextplatform.com/2026/02/19/taalas-etches-ai-models-onto-transistors-to-rocket-boost-inference/ . It streams ~17,000 tokens/s to a single user, where an H200 manages ~230 https://heise.de/-11185112 . - Per joule the gap is smaller than that 74× suggests, because GPUs serve hundreds of users at once: a fully loaded H200 pushes ~17,700 tokens/s https://nvidia.github.io/TensorRT-LLM/1.2.0rc0/developer-guide/perf-overview.html on the same model. Taalas itself claims 10×. - The cost is flexibility: the model is frozen in metal with 3/6-bit quantization, and a new one takes ~2 months. AMD agreed to buy Taalas https://ir.amd.com/news-events/press-releases/detail/1296/amd-acquires-taalas-to-advance-compute-solutions-for-rapidly-growing-ai-inference-market in August. Yes, but Jevons paradox In 1865 William Jevons noticed that efficiency doesn't save fuel: when Scottish ironworks cut coal per ton of iron to under a third, Scotland's coal use rose tenfold https://www.econlib.org/library/YPDBooks/Jevons/jvnCQ7.html . AI is following the same pattern. Google cut energy per prompt 33×, yet its monthly tokens went from 9.7 trillion to 3.2 quadrillion https://blog.google/innovation-and-ai/sundar-pichai-io-2026/ in two years. Adoption and agents are already counted above, so the question here is how much more gets re-spent on top. AI's own record in 2025–26 implies 75–85% : most of every efficiency gain the earlier sections banked comes back as load. More numbers - Jevons' line: "it is wholly a confusion of ideas to suppose that the economical use of fuel…" means using less of it. - Between 1800 and 2000, the price of light in the UK fell to under 1/3,000th https://orca.cardiff.ac.uk/id/eprint/27209/ , while light used per person rose 6,500× and total use 25,000×. Spending on light has held near 0.72% of GDP for 300 years https://iopscience.iop.org/0022-3727/43/35/354001 , which amounts to roughly 100% rebound. - The one real counterexample: from 2010 to 2018, data-center compute grew 6.5× while energy rose just 6% https://eta.lbl.gov/publications/recalibrating-global-data-center . That ended once GPUs arrived. - "Jevons paradox strikes again " posted Satya Nadella https://fortune.com/2025/01/27/microsoft-ceo-satya-nadella-deepseek-optimism-jevons-paradox the week DeepSeek R1 landed. Frontier models now cost 3–18× more a year https://arxiv.org/abs/2511.23455 to run as savings get spent on bigger ones. - Where 75–85% comes from: in 2025–26 tokens grew 7× and energy per task fell ≥10× https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary , yet AI data-center load still rose 50% https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions . - The evidence is thin, and one study found demand inelastic across models https://arxiv.org/html/2601.10088 . Yes, but the grid can't be built that fast Right now AI is limited by supply, not demand. Satya Nadella says Microsoft has "a bunch of chips sitting in inventory that I can't plug in" https://www.datacenterdynamics.com/en/news/microsoft-has-ai-gpus-sitting-in-inventory-because-it-lacks-the-power-necessary-to-install-them . ERCOT alone has 474 GW of large loads https://www.ercot.com/files/docs/2026/07/29/ERCOT-Senate-July-29-Panel-1-Assessing-The-Grid.pdf waiting, yet the most new data-center load the US grid has connected in one year is ~10 GW https://about.bnef.com/insights/data-centers/six-things-to-know-about-bnefs-new-us-data-center-capacity-outlook/ . Every part of the supply chain is backed up: gas turbines take ~6 years https://www.woodmac.com/press-releases/gas-turbine-prices-soar-195-as-market-faces-supply-demand-crisis/ , big transformers 2½–4 https://pv-magazine-usa.com/2026/05/11/u-s-transformer-market-faces-severe-supply-constraints-as-lead-times-extend-to-four-years/ . We cap AI at 10–25 GW of new capacity a year , which works out to 650–1,700 TWh by 2036 . Past ~1,700 TWh, it stops being a question of demand. The turbines, transformers and wires don't exist. More numbers - Amazon's Andy Jassy expects not enough capacity to meet 2026 demand https://datacenterknowledge.com/infrastructure/amazon-lifts-ai-infrastructure-spending-to-220b-as-demand-outpaces-capacity , and Alphabet's CFO calls it "a supply-constrained environment" https://www.marketbeat.com/instant-alerts/alphabet-q2-earnings-call-highlights-2026-07-22/ . - Northern Virginia's data-center vacancy is 0.2% https://www.cbre.com/insights/books/north-america-data-center-trends-h1-2026 , and 80% of what's under construction is already leased. - About 90% of ERCOT's waiting loads are data centers. Dominion has 54 GW under contract https://www.pjm.com/-/media/DotCom/committees-groups/subcommittees/las/2026/20260929/20260929-item-5ai---dominion energy.pdf , yet its data-center load grows ~0.5 GW a year. - PJM's capacity auction hit its price cap for the third year running and still came up 6.8 GW short https://www.pjm.com/-/media/DotCom/about-pjm/newsroom/2026-releases/20260714-pjm-capacity-auction-procures-138318-mw-of-generation-resources.pdf . - GE Vernova has 116 GW https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000147/gevpressrelease2q26.htm of gas-turbine orders and reservations against ~20 GW a year of output, and turbine prices are up 195% since 2019. A new power plant now waits over 5 years https://www.publicpower.org/periodical/article/backlog-power-plants-seeking-transmission-grid-connection-eased-somewhat-2025-lbnl from grid request to operation. - Builders are routing around the grid with on-site gas 124 GW announced https://about.bnef.com/insights/data-centers/six-things-to-know-about-bnefs-new-us-data-center-capacity-outlook/ , half of it speculative , but that hits the same turbine queue. - 10–25 GW a year runs from today's pace up to roughly double the entire US generation build. Even the high case falls short of the 50 GW by 2028 https://anthropic.com/news/build-ai-in-america Anthropic says the industry needs. So, how much? With all seven factors taken into account, the range for 2036 is . The high end . Published forecasts sit inside this range Berkeley pegs it at ~290–460 TWh https://escholarship.org/uc/item/33m6w3x0 , and the IEA's estimate works out to ~230 TWh https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary for the US . Methodology & assumptions For each year t 0 = 2026 , E t = E₀ · A t · G t · η t ^ R−1 . A is the adoption multiplier, a logistic curve from 2% to the target share reached in 2036. G is agent compute per adopter, following an S-curve from 1× today to its 2036 value. η is the combined efficiency gain: GPU × model × the mix of ASIC and GPU compute. R is the rebound, meaning the share of efficiency gains spent on more usage 0 = none, 1 = all . The GPU and model trends fade instead of compounding forever: each one's yearly gain halves every h years, so its total gain levels off at rate^ h/ln 2 . For GPUs, h is chosen so that total gain levels off at the physical ceiling. Finally, load can never exceed a supply ceiling that rises roughly linearly from today's 75 TWh to C in 2036. The band is the minimum and maximum across every combination of the range endpoints below, and the midpoint uses the middle of each range. | Input | Range | Basis | |---|---|---| | E₀, US AI electricity 2026 | 75 TWh | ~300 TWh data-center load LBNL 2025 update https://escholarship.org/uc/item/33m6w3x0 : 192 TWh in 2024, +22%, +29% × AI's 15–25% share EPRI https://restservice.epri.com/publicattachment/96784 , upper end because AI-focused load grew 50% in 2025 IEA https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions . History 2018–2025 uses the same data-center totals × an AI share rising from ~4% to ~20%. These are estimates. | | GPU efficiency | 1.28–1.40× / yr, fading toward a 26–650× ceiling half-life ≈ 7–18 yrs | Epoch https://epoch.ai/blog/can-ai-scaling-continue-through-2030 FLOP/s per W trend; Epoch https://epoch.ai/data-insights/ml-hardware-energy-efficiency leading hardware +40%/yr; IEA https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf 30–40%/yr, slowing. Ceiling: Ho et al. https://arxiv.org/abs/2312.08595 CMOS limit ≈ 207× H100 ±1 SD: 41–1,037× , divided by Blackwell's ~1.6× over H100. Historical decay: Koomey https://www.electronicdesign.com/technologies/microprocessors/article/21802037/energy-efficiency-of-computing-whats-next 1.6 → 2.6-yr doubling; Hennessy & Patterson https://dl.acm.org/doi/10.1145/3282307 . | | Adoption | 2% → 68–86% of households 34–43× | a16z / PNC https://a16z.com/state-of-markets-ii/ paid-AI households, April 2026. Bounds: US internet adoption in 2005 and 2015 Pew https://www.pewresearch.org/internet/fact-sheet/internet-broadband/ , since AI is spreading ~2× as fast St. Louis Fed https://www.stlouisfed.org/on-the-economy/2024/sep/rapid-adoption-generative-ai | | Model efficiency | 2.6–10.8× / yr, halving every 3–10 yrs | GPT-2 cost curve via nanochat https://github.com/karpathy/nanochat ~900× in 7 years ; densing law https://www.nature.com/articles/s42256-025-01137-0 density doubles every 3.5 months . The half-life is our judgment, informed by price declines slowing as a capability ages Epoch https://epoch.ai/publications/the-plunging-price-of-thought , the data wall https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data , and gains tied to compute scale Gundlach et al. https://arxiv.org/abs/2511.21622 . This is the least certain input. | | Agents G | 2–20× by 2036 | Compute-weighted agent:human ratio ~0.6 → 1.8 on OpenRouter in 2026 raw tokens 1 → 7×, ~90% cached: OpenRouter https://openrouter.ai/data/newsletter/cache-is-crucial . Global ratio assumed ~0.3–1 today, growing 10–45%/yr to ~2–30 an overall multiplier of ~3–30×, or 2–20× relative to today . | | Rebound R | 0.75–0.85 | Rebound beyond the adoption and agent growth already modeled, calibrated to 2025–26: energy grew ~1.5× = users ~2× × agents ~1.3× × efficiency ~10–33×^ R−1 . Lighting's ~100% Tsao et al. https://iopscience.iop.org/0022-3727/43/35/354001 includes new uses, which are counted separately here. | | Supply ceiling C | 650–1,700 TWh in 2036 | 10 GW a year today's pace to 25 GW a year about double the whole US generation build of new AI capacity, at 70–80% utilization. Calibration: data-center capacity added was 6.4 GW 2024 , 8.5 GW 2025 and ~14–16 GW in 2026 Goldman https://www.publicpower.org/node/30450 , SynMax https://blog.synmax.com/vulcan/2026-wont-deliver-21-gw-of-data-centers ; the grid's best single year connected ~10 GW BNEF https://about.bnef.com/insights/data-centers/six-things-to-know-about-bnefs-new-us-data-center-capacity-outlook/ ; EIA https://www.eia.gov/todayinenergy/detail.php?id=67205 2026 planned additions 86 GW nameplate, ~18 GW average include only 6.3 GW of gas. | | ASIC share / advantage | 25% → 40–70% of AI power; 1.5× → 1.5–20× tokens/W | Today: Epoch https://epoch.ai/graphs/ai-chip-market-share-by-designer Q4 2025 share TPU 20%, AWS 5% , assumed ~1.5× GPU efficiency Trainium3's "4×" is vs Trainium2, not GPUs . 2036: the ASIC edge ramps linearly; the upper end matches Etched/Groq-class claims, above Taalas's own 10×. The mix is weighted by power: tokens per watt = 1−s + s·k. | What this leaves out: on-device inference, AI load outside the US that serves US users, grid and supply-chain limits on how fast data centers can actually be built, and any electricity AI might save elsewhere. Every input is an extrapolation. Treat the band as a map of the arguments, not a forecast.