Will AI Run Out of Power? Projected Power Limits on Training Frontier AI Models, 2022-2033

By Robert Lara III

September 29, 2026

Abstract

In short: each year, the newest top AI model has needed about 3.7 times as much electricity to train as the one before it. The electricity available to AI in the US grows only about 1.2 times as much each year, so the pace of training growth will slow sharply. In 2026 the biggest training run used about $80 million of electricity. At the current pace it would need about $14.5 billion in 2030, more than the roughly $10.4 billion available to all AI in the US that year. That limit is reached between 2029 and 2032 for one company training one model, and between 2028 and 2030 if five companies each train one, if the growth stays between 2.5 and 4.5 times a year. If it is 2 times a year, the limits arrive in 2034 and 2031. Between 2032 and 2035, one run would cost more in electricity than the entire United States generates today.

This report follows four things from 2022 to 2033: the electricity bill for training, the power Texas and the US can supply to data centers, the chips and buildings that hold the GPUs, and the money in the business. GPUs are the specialized chips that train AI models.

The chips slow down first. Around 2031 a single chip gains only about 10% in transistor density a year, so most further growth has to come from putting more chips into each package and more packages into each data center.^[6] Data centers are limited by power, not floor space. The largest US campus could hold about 17 million top GPUs by floor area, but its power supply covers only about 213,000.^[11][30]

The money side moves on the same schedule. Token prices for flagship models fell from $37.50 per million in March 2023 to $7.00 in September 2026,^[12] while industry-wide token volume rose from 1.7 to 5.6 quadrillion a month in under a year.^[13] OpenAI lost about $20.9 billion on operations and $38.5 billion after all items in 2025, and targets breakeven around 2029 to 2030.^[15][16] Anthropic reported its first adjusted operating profit in Q2 2026.^[17]

These are projections that extend recent trends, not predictions of what will happen. The data shows that training runs cannot keep growing about 3.7 times a year past about 2030. It does not show that models stop improving. For example, models could get cheaper to run, agents could finish longer tasks, or smaller models could match today's largest, and none of those needs a bigger training run. More efficient chips do not change the picture much. Efficiency gains are already in the trend, and staying under the power limit would take about 3 times a year of extra efficiency, while chips are improving 1.1 to 1.8 times a year.^[18][19] The projection already includes the expected data center power build-out, about 52 GW more average load by 2030. Power plants beyond that would push the dates later, but a 10x boost, about 148 GW, adds only about 2 years.

When each limit binds

The first limits arrive between 2028 and 2032. The last, the whole US grid, arrives between 2032 and 2035. Every date assumes the training electricity cost keeps growing 2.5 to 4.5 times a year.

On my trend, the largest run in 2030 uses about 220 TWh of electricity, which is about 25 GW if spread over a full year, more than the roughly 15 GW available to all AI. That comes from the fitted trend used in the tables, which starts from about $97 million in 2026. Compounding the $80 million estimate directly gives about $14.5 billion, or about 21 GW. This report assumes a run lasts a year. A run that lasts six months needs twice that power, and one company's limit would arrive in 2029 instead of 2030. This is my own trend fit to estimates, not a published forecast. Other groups publish their own estimates, which I did not use, and they may differ.

These dates depend on two estimates I made: how much electricity the largest run used in 2026 ($80 million) and how fast that grows each year (3.7 times). If the 2026 figure were half as large, one company's limit would arrive in 2031 instead of 2030. If it were twice as large, it would arrive in 2029. If the growth were 2.5 times a year, it would arrive in 2032, at 4.5 times in 2029, at 2 times in 2034, and at 1.5 times in 2046.

This report does not model training data. Epoch AI estimates the usable stock of public human text at about 300 trillion tokens, and expects it to be fully used between 2026 and 2032, with a middle estimate near 2028.^[26] Ilya Sutskever, OpenAI's co-founder and former chief scientist, said in December 2024 that compute keeps growing but data does not, because there is only one internet, and that pre-training as we know it will end.^[28] Epoch's later analysis counts multimodal data and repeated use, and estimates the equivalent of 400 trillion to 20 quadrillion tokens available for training by 2030.^[27] Data is a separate limit from power, and its window overlaps the power limits in this report.

Why more efficient chips do not change this

More efficient chips do not change the dates by much. Efficiency is already in the trend: the 3.7 times a year growth in training electricity is measured after the efficiency gains chips delivered from 2022 to 2026. The dates only move if efficiency improves faster than it has.

One-time gains buy little time, because growth of 3.7 times a year uses them up quickly:

The power supplied to AI grows about 22.5% a year.^[1] To stay under it, run electricity would have to grow 1.2 times a year instead of 3.7, which takes about 3 times a year of extra efficiency on top of the historical pace.

Current chips are far below 3 times a year. The GB200 NVL72 delivers 3.4 times more FP8 compute per kilowatt than H100-based systems, about 1.8 times a year over roughly two years.^[19] Rubin is reported at 5 times the inference throughput of an H100 at 3.3 times the power, about 1.5 times per watt over four years, or 1.1 times a year.^[18]

Current chips are improving 1.1 to 1.8 times a year. Software and training methods also improve efficiency, including lower-precision math, models that use only part of themselves, and better use of chips. Those gains are already in the historical trend. Reaching the extra 3 times a year would need them to speed up as well, and I have no measured rate for them.

NVIDIA claims Vera Rubin needs one-fourth as many GPUs as Blackwell to train a large mixture-of-experts model, and delivers up to 10 times more agentic throughput per unit of energy.^[33][34] These are vendor claims on specific workloads, and one analysis puts the gain for dense models at 2 to 3 times Blackwell.^[35] If the vendor claims hold, chip efficiency is improving faster than the 1.1 to 1.8 times a year above.

How power cost and supply compare

The training electricity line grows from $0.3M in 2022 to a projected $870B in 2033, so the US power lines sit near the bottom of this scale. Hover over any point to see its exact value.

When training outgrows the power set aside for AI

One training run first exceeds the power available to AI in Texas in 2028 or 2029 and in the US in 2030. Five companies each training one run exceed it a year or two sooner.

The closest calls are US five runs at 20% in 2028 ($6.5B needed against $6.9B available) and Texas one run at 15% in 2028 ($1.31B against $1.27B).

The table compares one training run to all the power set aside for AI. But AI power also serves everyday use of finished models. Google and Meta reported that about 60% to 65% of their machine learning energy went to inference, which leaves at most about 35% to 40% for training.^[23] On that share, one company's run passes the limit in 2029 instead of 2030, and five companies' runs pass it in 2028 instead of 2029.

The US data center power in this report comes from Lawrence Berkeley National Laboratory's forecast, and data centers are running ahead of it.^[1] S&P Global puts 2025 at 313 TWh against the report's 235.^[22] Goldman Sachs, working from construction schedules, expects 66 GW in 2027 against the report's 40 GW, though it expects only about half of scheduled capacity to arrive on time.^[21] Suppose data centers stay 25% to 64% ahead of the forecast. On the training share above, one company's run then passes the limit in 2029 at the low end and 2030 at the high end, and five companies' runs still pass it in 2028. The 2033 limit set by the whole US electricity supply does not change.

A training run does not have to happen in one place. GPT-4.5 and Gemini 1.5 were each trained across several data centers.^[25] So the Texas and single-campus figures show local limits, not limits on a run. The limit that matters is the power available across the US.

This report first assumed AI stays at 15% to 25% of data center electricity.^[5] The International Energy Agency expects electricity for AI-driven servers to grow about 30% a year, against about 9% for other servers.^[29] On that, AI's share of data center power rises from 20% in 2026 to about 34% in 2030. With the rising share, one company's limit moves from 2029 to 2030 if training gets 40% of AI's power, and the other dates do not change.

What limits a GPU

A single chip's growth slows around 2031. After that, most gains have to come from adding chips to each package and packages to each data center.

Over 2022 to 2026, the transistors per package grew about 1.47 times a year and power per card about 1.36 times a year.^[7][8] Projected to 2036 on the same trend, that is 15.4 trillion transistors and 43.7 kW per card.

The reason for the slowdown is on the chipmakers' roadmap. Imec expects the A14 process in 2028, and after the A10 process around 2030 to 2031 the spacing between transistors stops shrinking. From then, density gains come only from smaller cell height, about 10% a year through the last roadmap node, A3, in 2038.^[6] At that rate a single die reaches roughly 320 billion transistors by 2036. Reaching 15.4 trillion would take about 48 dies in one package.

The shaded area is after 2031. Density per chip keeps improving at about 10% a year until the roadmap ends in 2038, but most of the growth now has to come from more dies and GPUs.

Data centers run out of power before space

A current top rack, the GB200 NVL72, holds 72 GPUs on about 7 sq ft of floor and draws about 120 kW.^[10] A Rubin rack is also 72 GPUs, at about 220 kW.^[9][30] The GPU count per rack is fixed, so power per square foot is what climbs.

Counting aisles and service space at an assumed 30 sq ft per rack, a hall holds about 2.4 GPUs per sq ft, so 10,000 sq ft holds about 24,000 GPUs and draws about 73 MW.

For Citadel, the 650 MW supply fills only about 1.2% of the floor space. The 5 GW at Hyperion is a build-out target, not power it draws today.

Sources put a Rubin rack of 72 GPUs at 190 to 230 kW, and this report uses 220 kW.^[30][31] NVIDIA has not published Rubin's power, so the rack figure comes from analysts.^[32] On that, the power at Tahoe Reno 1, Citadel and Hyperion supports about 42,500, 213,000 and 1.64 million GPUs.

Prices, demand and profit

Demand is rising fast while token prices have fallen, and the companies are still losing money. Flagship token prices fell from a median of $37.50 per million tokens (blended 3 input to 1 output) at GPT-4's launch in March 2023 to $7.00 in September 2026. Over the last 12 months flagship prices rose 101%, and the mean across today's 22 flagship models is $13.83.^[12]

Industry-wide, Goldman Sachs estimates token processing at 1.7 quadrillion a month in mid-2025 and 5.6 quadrillion in May 2026, and forecasts 47 quadrillion in 2028 and 120 quadrillion by mid-2030.^[13] Google alone rose from 9.7 trillion tokens a month in May 2024 to over 3.2 quadrillion in May 2026.^[14]

Projected to mid-2031, the flagship price falls to about $0.19 and industry volume reaches about 192 quadrillion tokens a month. The price line applies a 53% yearly decline, anchored to the latest actual value, and the volume line follows Goldman Sachs's forecast.

OpenAI's breakeven target falls in the same window as the first limits. Breakeven needs training spending to stop outgrowing revenue, so a cap on training growth would lower costs. It could also slow the capability gains that revenue depends on, and this report does not show which effect is larger.

Assumptions and limits

Every figure is an estimate or a trend extended forward. The main inputs are:

• Training electricity cost. The largest run's electricity cost is estimated at $0.3M (2022), $4M (2023), $8M (2024), $25M (2025) and $80M (2026), then extended at the fitted growth of about 3.7 times a year. Electricity is priced at $0.08 per kWh. Labs do not publish these figures, so the early values are estimates.
• US data center power. It grows from 192 TWh in 2024 to 649 TWh in 2030 (Lawrence Berkeley National Laboratory), then continues at about 22.5% a year.^[1]
• Texas power. Texas output is held at about $42.4B a year today and grows with ERCOT's adjusted forecast of about 50 GW of peak demand growth over five years, from an 85.5 GW record. Crypto mining (about 5 GW) is counted as non-data-center use.^[2][3]
• AI's share. AI is assumed to use 15% to 25% of data center electricity (an EPRI estimate for 2026), rising over time, as described below. The rest is cloud services, storage, streaming and similar work.^[5]
• Companies. Each company trains one run of the largest size per year. “Five runs” means five companies doing so.
• Agent capability. This is the task length AI agents complete with 50% reliability, measured in hours a human expert would take (METR benchmark). The 2026 value of 16 hours is the best reported result. Earlier years are approximate, and this report does not project it.^[4]
• Chips. Transistor counts are NVIDIA's flagship specifications, including Rubin's 336 billion.^[7][33] Rubin's 2,300 W is an analyst figure that NVIDIA has not published.^[30][32] The projections are exponential fits of about 1.47 times a year (transistors) and 1.36 times a year (power). The density outlook is imec's 2026 roadmap, and the 320 billion transistors per die applies its A14 to A5 shrink to Rubin's 168 billion per die.^[6]
• Data centers. The 30 sq ft per rack is my assumption. I applied it to the whole 7.2 million sq ft, which includes space that is not data hall, so 17.3 million GPUs is an upper bound. Campus sizes and power are as publicly listed, and Rubin GPU counts use 220 kW per 72-GPU rack, from analyst estimates.^[11][30][31]
• Prices and demand. Prices are the BenchLM median blended price of flagship models (3 input tokens to 1 output).^[12] The December 2022 GPT-3 price is from memory. Industry volume is Goldman Sachs's estimate. Other estimates are higher, including 11 quadrillion tokens a month from the I/O Fund.^[13]
• Profit. Figures come from reported and leaked financials, and sources use different measures. For OpenAI in 2025, the operating loss was reported at $20.9 billion and cash burn at about $9 billion, against the $38.5 billion GAAP loss shown above.^[15][16]
• Efficiency scenarios. The available power is 20% of US data center power growing 22.5% a year, and the extra efficiency is applied from 2027 to the fitted electricity cost path. The chip efficiency figures are as reported, not measured, and the two-year and four-year rates are my conversions.^[18][19]
• Model size. The size of a model, counted in parameters, is not the same as the electricity it needs. Some models use only a small part of themselves for each answer. DeepSeek V4 Pro, for example, has 1.6 trillion parameters but uses 49 billion for each token.^[24] The electricity figures in this report do not depend on parameter counts.

After 2031, each chip still improves about 10% a year, down from roughly 14% to 27% a year. Most of the growth in transistors per package comes from adding dies.^[6]

Agent task length rose from under a minute in 2022 to about 16 hours in 2026, and this report does not project it.^[4] It measures how long a benchmark task is, meaning a self-contained research or coding problem with a clear pass or fail. It does not mean AI could do that much human work. Building a complete system such as E*Trade or Adobe Creative Suite involves many linked products and constraints that no benchmark measures.