Updated Feb. 5, 2026

Trends in Artificial Intelligence

Frontier AI systems are advancing rapidly from increases in compute, hardware performance, algorithmic efficiency, and investment. This dashboard explores those dynamics.

What drives AI progress? The story is dominated by scale. Training AI systems with more compute, power and data has consistently led to better performance. Since 2010, the compute used to train notable AI models has increased 4.5× per year. Meanwhile, researchers have made the underlying algorithms far more efficient — each year, the same performance can be achieved with 3× less compute.

This massive scale-up in training compute comes from three sources: deploying more chips in parallel, running training for longer, and leveraging increasingly powerful AI processors. The consequences are striking. Training costs are climbing by 2.5× annually, while power requirements double each year. Today’s cutting-edge AI training runs consume tens to hundreds of megawatts — comparable to a medium-sized power plant. These trends appear set to continue through 2030.

Show growth values in

LLM inference prices

40 ×/year
2 months
2 OOM/year

The cost to inference an LLM at a fixed level of performance has fallen rapidly, but unevenly across tasks.

The cost to inference an LLM at a fixed level of performance has been halving every 2 months.

The cost to inference an LLM at a fixed level of performance has fallen by 2 OOMs per year.

Model Performance
90% CI 10× to 900× 1 to 4 months 1 to 3 OOM

Compute stock growth

2.3 ×/year
10 months
0.36 OOM/year

The total computing power of the stock of NVIDIA chips is growing at a rate of 2.3×/year.

The total computing power of the stock of NVIDIA chips is doubling every 10 months.

The total computing power of the stock of NVIDIA chips is growing by 0.36 OOMs per year.

AI Companies
90% CI 2.2× to 2.5× 9 to 11 months 0.34 to 0.40 OOM

Training compute

5 ×/year
5.2 months
0.7 OOM/year

Training compute for frontier language models has been growing at 5× per year since 2020.

Training compute for frontier language models has been doubling every 5.2 months since 2020.

Training compute for frontier language models has been growing at 0.7 OOMs per year since 2020.

Training Runs
90% CI 4× to 6× 4.6 to 6.0 months 0.6 to 0.8 OOM

Algorithmic progress

÷ 3.0 ×/year
7.6 months
0.5 OOM/year

Pre-training compute efficiency is improving at roughly 3.0× per year.

Pre-training compute efficiency is doubling roughly every 7.6 months.

Pre-training compute efficiency is improving by roughly 0.5 OOMs per year.

Training Runs
90% CI 2.8× to 4.4× 5.6 to 8.1 months 0.4 to 0.6 OOM

Largest AI data center

600,000 H100e

The largest known AI data center has computing power equivalent to 600,000 NVIDIA H100 chips.

Data Centers
90% CI 400k to 900k H100e

FLOP/s per dollar

1.37 ×/year
2.2 years
0.14 OOM/year

AI chip performance per dollar has improved by 37% per year.

AI chip performance per dollar has doubled every 2.2 years.

AI chip performance per dollar has improved by 0.14 OOMs per year.

Hardware
90% CI 1.30× to 1.45× 1.9 to 2.6 years 0.11 to 0.16 OOM
Model Performance

The cost to inference an LLM at a fixed level of performance has fallen rapidly, but unevenly across tasks.

The cost to inference an LLM at a fixed level of performance has fallen rapidly, but unevenly across tasks.

The inference price of LLMs has fallen dramatically in recent years, while performance has rapidly improved. The rate of decline varies dramatically, with costs falling by between 9× and 900× per year depending on the performance milestone.
Learn more

Model capabilities

15.5 ECI/year

Model capabilities have grown faster since early 2024.

Model Performance
90% CI +13 to +18 ECI

SOTA vs. consumer hardware models

8 months

Frontier AI performance becomes accessible on consumer hardware within a year.

Model Performance

LLM context windows

30 ×/year
2.4 months
1.5 OOM/year

LLM context windows have grown by 30× per year since 2023.

LLM context windows have been doubling every 2.4 months since 2023.

LLM context windows have grown by 1.5 OOMs per year since 2023.

Model Performance
90% CI 10× to 50× 2.1 to 3.6 months 1.0 to 1.7 OOM

Data on AI

Benchmarking

Track and compare frontier model performance across 40+ evaluations
AI Companies

The total computing power of the stock of NVIDIA chips is growing at a rate of 2.3×/year.

The total computing power of the stock of NVIDIA chips is growing at a rate of 2.3×/year.

The computing power of the total stock of NVIDIA chips has grown at 2.3× per year, doubling every 10 months, since 2019, based on estimates of NVIDIA sales using its data center revenue, and estimates of the longevity of AI chips.
Learn more

OpenAI revenue growth

3.2 ×/year
7.2 months
0.51 OOM/year

OpenAI's annualized revenue has grown by 3.2×/year since 2024.

OpenAI's annualized revenue has been doubling every 7.2 months since 2024.

OpenAI's annualized revenue has grown by 0.51 OOMs per year since 2024.

AI Companies
90% CI 3.0× to 3.4× 6.8 to 7.6 months 0.48 to 0.53 OOM

Frontier inference compute

30 %

Only 30% of OpenAI's compute spending in 2024 was used on inference.

AI Companies

Frontier AI investment

$170 billion

Frontier AI labs have raised more than $170 billion.

AI Companies

Data on AI

Companies

Data on key organizations and labs driving AI development worldwide.
Training Runs

Training compute for frontier language models has been growing at 5× per year since 2020.

Training compute for frontier language models has been growing at 5× per year since 2020.

The amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models has grown by a factor of ~10,000.
Learn more

Algorithmic progress

÷ 3.0 ×/year
7.6 months
0.5 OOM/year

Pre-training compute efficiency is improving at roughly 3.0× per year.

Pre-training compute efficiency is doubling roughly every 7.6 months.

Pre-training compute efficiency is improving by roughly 0.5 OOMs per year.

Training Runs
90% CI 2.8× to 4.4× 5.6 to 8.1 months 0.4 to 0.6 OOM

Most compute used in a training run

5e26 FLOP

The largest known training run was Grok 4, at around 5e26 FLOP.

Training Runs

Training cost

3.5 ×/year
7 months
0.5 OOM/year

The cost to train frontier language models has increased by 3.5×/year since 2020.

The cost to train frontier language models has been doubling every 7 months since 2020.

The cost to train frontier language models has increased by 0.5 OOMs per year since 2020.

Training Runs
90% CI 2.8× to 4.4× 6 to 8 months 0.4 to 0.6 OOM

Report

Training Compute

Training compute of frontier AI models grows by 4-5× per year.

Data on AI

Models

Database of AI systems tracked by training compute, org, and release date.
Data Centers

The largest known AI data center has computing power equivalent to 600,000 NVIDIA H100 chips.

The largest known AI data center has computing power equivalent to 600,000 NVIDIA H100 chips.

Meta’s Prometheus cluster has an estimated 700 MW of power capacity and $20 billion in capital costs, making it the largest known AI data center. Microsoft’s Fairwater Wisconsin data center is expected to be nearly 10x more powerful, at 5.2M H100-equivalents by September 2027.
Learn more

Build time

2.1 years

Gigawatt-scale data centers can be built in about 2 years.

Data Centers
90% CI 1 to 3.6 years

Cost per gigawatt

$30 billion

An AI data center with 1 gigawatt of total facility power costs roughly $30 billion to build.

Data Centers
90% CI $20B to $40B

Country with the most computing power

75 %

The US holds three-quarters of global GPU cluster performance.

Data Centers

Data on AI

Frontier Data Centers

AI data centers tracked via satellite imagery and permit data.
Hardware

AI chip performance per dollar has improved by 37% per year.

AI chip performance per dollar has improved by 37% per year.

The compute performance you can buy for a dollar has improved by roughly 40% per year across over 20 AI accelerators released between 2012 and 2025. Much of this is driven by manufacturers introducing more powerful and more expensive chips — the GB300 costs nearly 9× the P100's release price, but delivers about 24× the performance per dollar.
Learn more

Memory bandwidth

1.28 ×/year
2.8 years
0.11 OOM/year

Memory bandwidth of GPUs has been growing by 28% per year since 2008.

Memory bandwidth of GPUs has been doubling every 2.8 years since 2008.

Memory bandwidth of GPUs has been growing by 0.11 OOMs per year since 2008.

Hardware
90% CI 1.24× to 1.33× 2.4 to 3.2 years 0.09 to 0.12 OOM

Adoption of numerical formats

3-4 years

Widespread adoption of new numeric formats took 3-4 years in past cycles.

Hardware

FLOP/s per watt

1.34 ×/year
2.4 years
0.13 OOM/year

GPUs have become 34% more energy efficient per year since 2008.

GPUs energy efficiency has doubled every 2.4 years on average since 2008.

GPUs have become 0.13 OOMs more energy efficient per year since 2008.

Hardware
90% CI 1.29× to 1.39× 2.1 to 2.7 years 0.11 to 0.14 OOM

Report

Trends in hardware

FLOP/s performance in 47 ML hardware accelerators doubled every 2.3 years.

Data on AI

ML Hardware

Key data on AI accelerators, such as GPUs and TPUs.

Trusted by leaders at OpenAI, DeepMind,
and governments worldwide

Need deeper insights? Our team offers custom research and advisory services.

Book a consultation