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 (4.4% of US power) and 192 TWh in 2024. 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. The IEA says AI-focused data centers grew 50% in 2025 alone. 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, and Grok 4 an estimated 310 GWh, 240× more.

More numbers #

• 75 TWh is about 1.8% of the 4,195 TWh 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 , 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 for a median Gemini text prompt and0.34 Wh for an average ChatGPT query. But ChatGPT alone handled2.5 billion prompts a day in mid-2025, and reasoning and agent tasks can usehundreds to thousands of times more per query.

Yes, but GPUs are getting more efficient #

Nvidia pitched Blackwell as up to 25× less energy per inference than Hopper. Vera Rubin, in full production since CES 2026, claims up to 10× more tokens per megawatt than GB200, and CoreWeave measured 10× on DeepSeek-R1. 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, about 45,000× in eight years.

The overall hardware trend is 1.28–1.40× more efficient per year (Epoch, Epoch), and the IEA says it is gradually slowing.

More numbers #

• Rubin's racks are fully liquid-cooled, with a PUE of 1.1 vs ~1.25 .
• 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× .
• Every chip trend before it slowed too. Computations per kWh doubled every 1.6 years until 2000, then every 2.6 . Single-thread speed went from +52% a year toabout +3% .
• Physics also sets a ceiling: CMOS chips top out around 207× an H100's efficiency , 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 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 as the internet was in the '90s, and the internet went from 14% of US adults in 1995 to 68% in 2005. At twice the speed, AI's next ten years look like the internet's twenty, which reached 86% by 2015. 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 (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 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.

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, 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. llm.c reproduced the 1.5B model in 24 hours for $672 in 2024. By March 2026, nanochat beat it in 1.65 hours, for about $48: 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. 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, and by late September they were using ~7× the tokens humans did. OpenClaw wakes itself on a 30-minute heartbeat by default, and Jensen Huang predicts a future with 100 billion agents. Agents are going proactive and are growing in use massively. Some of that 7× is cheaper than it looks: ~90% of agent tokens 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 , a persistent agent that runs scheduled automations. OpenClaw has391k GitHub stars , and its default heartbeat means 48 unprompted sessions a day.
• ChatGPT Pulse researches overnight before you ask, Claude Code's scheduled tasks fire"even while the session is idle" , and one Cursor customer runs2,000+ agent runs a week with no manual prompting.
• Microsoft already has ~40 million agents registered.
• Each agent task is heavy. Agents use 4× the tokens of chat and multi-agent systems 15×, Claude Code agent teams use~7× , and reasoning output is growing5× a year .
• Meta employees burned 60 trillion tokens in 30 days , ~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 in Q4 2025. The top rung is Taalas's HC1, 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 than CPUs (Nvidia's figure).
• Google's first TPU managed 30–80× the TOPS/W of the CPUs and GPUs of its day, and Ironwood is~30× the first Cloud TPU .
• AWS rates Trainium3 at 4× the efficiency of Trainium2 , and transformer-only Etched just raised$700M at a $21B valuation . TrendForce expects ASIC unit shipments topass GPUs in 2026 .
• The HC1 is TSMC 6 nm, 815 mm², ~200–250 W per card . It streams ~17,000 tokens/s to a single user, where an H200 manages~230 .
• 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 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 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. 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 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 , while light used per person rose 6,500× and total use 25,000×. Spending on light has held near0.72% of GDP for 300 years , 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% . That ended once GPUs arrived.

• "Jevons paradox strikes again!" posted Satya Nadella the week DeepSeek R1 landed. Frontier models now cost3–18× more a year 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× , yet AI data-center load still rose50% .

• The evidence is thin, and one study found demand inelastic across models .

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". ERCOT alone has 474 GW of large loads waiting, yet the most new data-center load the US grid has connected in one year is ~10 GW. Every part of the supply chain is backed up: gas turbines take ~6 years, big transformers 2½–4.

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 , and Alphabet's CFO calls it"a supply-constrained environment" .
• Northern Virginia's data-center vacancy is 0.2% , 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 , 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 .
• GE Vernova has 116 GW 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 waitsover 5 years from grid request to operation.
• Builders are routing around the grid with on-site gas (124 GW announced , 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 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, and the IEA's estimate works out to ~230 TWh 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 : 192 TWh in 2024, +22%, +29%) × AI's 15–25% share (EPRI ), upper end because AI-focused load grew 50% in 2025 (IEA ). 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 FLOP/s per W trend;Epoch leading hardware +40%/yr;IEA 30–40%/yr, slowing. Ceiling:Ho et al. CMOS limit ≈ 207× H100 (±1 SD: 41–1,037×), divided by Blackwell's ~1.6× over H100. Historical decay:Koomey 1.6 → 2.6-yr doubling;Hennessy & Patterson .
Adoption	2% → 68–86% of households (34–43×)	a16z / PNC paid-AI households, April 2026. Bounds: US internet adoption in 2005 and 2015 (Pew ), since AI is spreading ~2× as fast (St. Louis Fed )
Model efficiency	2.6–10.8× / yr, halving every 3–10 yrs	GPT-2 cost curve via nanochat (~900× in 7 years);densing law (density doubles every 3.5 months). The half-life is our judgment, informed by price declines slowing as a capability ages (Epoch ), thedata wall , and gains tied to compute scale (Gundlach et al. ). 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 ). 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. ) 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 ,SynMax ); the grid's best single year connected ~10 GW (BNEF );EIA 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 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.

source & further reading

html.non.io — original article