Is AI Growing Exponentially? The Math and 2026 Data

Quick answer: Several measurable AI trends have grown exponentially, but AI as a whole has no single growth rate. Epoch AI reports that training compute for frontier language models has grown about 5 times per year since 2020, a doubling every 5.2 months. METR’s November 20, 2026 update estimates that the length of tasks AI agents complete with 50% success has doubled about every 3.6 months on its combined test suites. These are historical trends, not guarantees: real-world growth usually slows as it hits limits.

Facts checked November 22, 2026.

“AI is improving exponentially” is one of the most repeated claims in tech. Sometimes it is accurate; sometimes it is a figure of speech for “fast.” Mathematics gives a precise test, and a few formulas let you check any doubling claim yourself. Here is what the data show as of late 2026, and where the exponential story breaks down.

What “Exponential” Actually Means

A quantity grows exponentially when it multiplies by the same factor in every equal time period:

\[ y(t) = y_0 \cdot b^{t}, \]

where \(y_0\) is the starting value and \(b\) the growth factor per period. Linear growth adds the same amount each period; exponential growth multiplies. The doubling time \(T_d\) is how long it takes to double:

\[ T_d = \frac{\ln 2}{\ln b}. \]

Example: growth of 5 times per year gives \(T_d = \frac{\ln 2}{\ln 5} \approx 0.43\) years, about 5.2 months. A quick shortcut for small percentage rates is the rule of 70: growth of \(r\)% per year doubles in about \(70 / r\) years, so 7% a year doubles in about 10 years (exactly \(\ln 2 / \ln 1.07 \approx 10.2\)). For big rates the shortcut fails: 41% a year gives \(70/41 \approx 1.7\), but the exact answer is \(\ln 2 / \ln 1.41 \approx 2.0\) years. Review the basics in exponential growth and decay.

AI Trends That Really Are Exponential

TrendGrowth rateDoubling timeSource
Training compute, frontier language models (since 2020)About 5× per yearAbout 5.2 monthsEpoch AI, updated Nov 5, 2026
Training cost of frontier modelsAbout 3.5× per yearAbout 6.6 monthsEpoch AI
Length of tasks AI agents complete at 50% success (2019–2025)About 3× per yearAbout 7 monthsMETR, March 2025
Task length on METR’s newer combined suitesAbout 10× per yearAbout 3.6 months (95% range about 2.1 to 4.9)METR, Nov 20, 2026
AI chip performance per dollar, newly purchased chips (since 2023)About +49% per yearAbout 1.7 yearsEpoch AI

Check the METR rows: doubling every 7 months is \(2^{12/7} \approx 3.3\) times per year, and doubling every 3.6 months is \(2^{12/3.6} \approx 10\) times per year. Note that these are different measures: compute is an input, while METR’s time horizon is a measure of capability.

The METR Time-Horizon Trend

METR, an AI evaluation nonprofit, measures the 50% time horizon: the length of task, measured by how long it takes skilled humans, that an AI agent completes with a 50% success rate. In March 2025, METR reported that this horizon had doubled roughly every 7 months since 2019, with Claude 3.7 Sonnet at about 50 minutes. In its November 20, 2026 update, METR estimated the doubling time on its combined task suites at roughly 3.6 months, with the strongest public models (OpenAI’s GPT-6 Astra and Anthropic’s Opus 5.5) at roughly 40 hours, though with very wide uncertainty ranges (roughly 13 hours to well over 1,000 hours for the top model) because few tasks are that long.

What the time horizon is not: it is not how long the AI runs, and a 40-hour horizon does not mean an AI can reliably do a human’s work week. It means that on tasks that take skilled people about 40 hours, the agent succeeds about half the time. Is growth accelerating? The 3.6-month figure comes from newer, harder task suites, so comparing it directly with the older 7-month trend does not by itself prove acceleration; METR’s recent trend on its older suite was similar, about 105 days.

What would a 3.6-month doubling imply if it simply continued? As a hypothetical, over two years:

\[ 2^{24/3.6} \approx 2^{6.7} \approx 100. \]

A hundredfold increase in two years. At the older 7-month pace, two years gives \(2^{24/7} \approx 11\). The difference between those two answers shows how sensitive exponential forecasts are to the exact rate. METR itself stresses that its measurements are getting harder as models approach the longest tasks in its test set.

Why Exponentials Usually Become S-Curves

No real quantity grows exponentially forever. Populations, new technologies and benchmark scores typically follow a logistic curve, or S-curve:

\[ f(t) = \frac{L}{1 + e^{-k(t – t_0)}}. \]

Early on, it looks exactly like exponential growth. At the midpoint \(t = t_0\) it reaches \(L/2\), and then growth slows as it approaches the ceiling \(L\). Benchmark scores are capped at 100%, so they must bend. Epoch AI’s FrontierMath leaderboard showed a top score of 95% on Tiers 1–3 (version 2) in November 2026, up from a few percent when the original benchmark launched in late 2024; the versions differ, but either way the score cannot keep rising exponentially (see FrontierMath explained). A flattening score does not mean math ability has stopped improving; it means the test is running out of room, which is why Epoch added unsolved problems. The hard question is never “is it exponential now?” but “where is the ceiling, and when do we reach the bend?”

For AI compute, the ceilings are physical and financial: electricity, chip factories, data and money. Epoch AI projected in 2024 that training runs of about \(2 \times 10^{29}\) operations would likely be feasible by 2030, with power the most likely bottleneck.

Reading Exponential Charts

  • Log scale. Exponential trends look like straight lines on a logarithmic y-axis, where each gridline is 10 times the one below. A straight line on a log chart means steady exponential growth; a curve bending down means slowing.
  • Exponential vs power law. AI scaling laws are power laws in model size, not exponentials in time (see AI scaling laws explained). Inputs grow exponentially; returns per unit of input diminish.
  • Cherry-picked windows. Starting a chart at a quiet period makes any recent jump look explosive. Check the full history.

More statistics checks are in how to read AI benchmarks.

So, Is AI Growing Exponentially?

ClaimVerdict
“Compute used to train AI is growing exponentially.”Yes, about 5× per year since 2020.
“AI agents can handle exponentially longer tasks.”Yes on METR’s measure so far, with wide uncertainty at the frontier.
“AI intelligence is exponential.”Not a measurable claim: there is no single scale for intelligence.
“It will continue forever.”No real-world exponential does; the question is when it bends.

Common Misconceptions

“Exponential means very fast.” It means multiplying at a steady rate. Growth of 1% a year is exponential too.

“A doubling trend lets you predict the future exactly.” Small changes in the doubling time give huge differences after a few years: compare 11× and 100× in two years above.

“Every trend must become an S-curve.” Logistic curves are a useful model of growth with limits, not a law of nature.

“If it is an S-curve, progress will stop soon.” S-curves can be stacked: when one approach saturates, a new one (for example, test-time reasoning after pretraining scale) can start a fresh curve.

Watch: Exponential Growth Explained

If you learn better by watching, these videos are worth your time.

Exponential growth and epidemics (3Blue1Brown). The clearest visual explanation of exponential versus logistic growth and log scales.
AI can’t cross this line and we don’t know why. (Welch Labs). How AI performance tracks compute along a straight line on log-log axes.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. Growth of 5 times per year has what doubling time?
Found a bug? Let us know!
2. Using the rule of 70, about how long does 7% annual growth take to double?
Found a bug? Let us know!
3. If a quantity doubles every 3.6 months, about how much does it grow in 2 years?
Found a bug? Let us know!
4. What does an exponential trend look like on a log-scale chart?
Found a bug? Let us know!
5. What does a METR time horizon of 40 hours mean?
Found a bug? Let us know!

Frequently Asked Questions

Is AI growing exponentially?

Some measures are. Training compute for frontier models has grown about 5 times per year since 2020, and METR’s AI task-length measure doubled about every 7 months from 2019 to 2025, and about every 3.6 months on METR’s newer test suites in 2026. But there is no single measure of intelligence, and real exponentials eventually slow.

What is the doubling time of AI?

It depends on the measure: about 5.2 months for frontier training compute (Epoch AI) and about 3.6 to 7 months for METR’s time horizon of tasks AI agents can complete.

What is the METR time horizon?

The length of tasks, measured by how long skilled humans take, that an AI agent can complete with a 50% success rate. In November 2026 METR estimated about 40 hours for the strongest public models, with wide uncertainty.

Will AI growth slow down?

Real exponential trends usually become S-curves as they hit limits. For AI, likely limits include electricity, chip supply, data and cost, though new methods can start fresh growth curves.

How do you calculate doubling time?

Divide ln 2 by the natural log of the growth factor per period. For 5× per year, ln 2 ÷ ln 5 ≈ 0.43 years. For small percentage growth rates, the rule of 70 gives a quick estimate: 70 divided by the percent rate.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Is AI Growing Exponentially? The Math and 2026 Data - Effortless Math"?

No one replied yet.

Leave a Reply