AI chip performance per dollar grows 49% per year since 2023, Epoch finds
A new Epoch AI analysis of quarterly chip sales data finds that the performance purchasers get per dollar for AI accelerators has risen by about 49% annually since 2023, doubling every 1.7 years. The metric, tracking bit-operations per second per constant 2025 dollar, climbed from roughly 5.6×10^11 in Q1 2023 to about 1.4×10^12 by the end of 2025. Growth came in spurts: flat through mid-2024, then roughly doubling as Blackwell-generation chips became a majority of purchases. The trend matters because it makes compute cheaper over time, but the average is dragged down by Nvidia's high-margin GPUs, and the analysis ends in Q4 2025 with price estimates rather than definitive transactions.
AI chip performance per dollar grows 49% per year since 2023, Epoch finds
Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year. Performance per dollar rose from about 5.6 × 10 11 bit-operations per second per dollar in the first quarter of 2023 to about 1.4 × 10 12 by the end of 2025, in constant 2025 dollars.
Key takeaway
AI chip price-performance improves 49% yearly, doubling every 1.7 years, driven by new generations like Blackwell, reshaping compute cost assumptions for frontier AI development.
What happened
Epoch AI's analysis of quarterly chip sales data, drawing on its AI Chip Sales Hub, finds that the aggregate performance per dollar of purchased AI accelerators has grown by an average of 49% per year since 2023. From Q1 2023 to Q4 2025, the metric in constant 2025 dollars rose from about 5.6×10^11 to 1.4×10^12 bit-operations per second per dollar.
Growth occurred in spurts: performance per dollar stayed nearly flat until mid-2024, then roughly doubled as Blackwell-generation chips became a majority of new spending. While NVIDIA's GB300 is over 37 times more cost-effective than its 2016 P100, the overall average is lower because a large share of spending goes to high-margin Nvidia GPUs, and custom ASICs like Google's TPU v6e offer higher per-dollar performance.
Evidence
Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year.
Epoch · attributed
Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year.
Performance per dollar rose from about 5.6 × 10^11 bit-operations per second per dollar in Q1 2023 to about 1.4 × 10^12 by the end of 2025, in constant 2025 dollars.
Epoch · attributed
Performance per dollar rose from about 5.6 × 10^11 bit-operations per second per dollar in the first quarter of 2023 to about 1.4 × 10^12 by the end of 2025, in constant 2025 dollars.
Why it matters
Sustained 49% gains in compute-per-dollar mean AI labs can train larger models or reduce hardware budgets faster than expected, directly influencing scaling strategies and capital allocation.
Limits and uncertainties
The data series ends in Q4 2025 because sales estimates for later quarters are not yet complete.
The analysis models units sold rather than deployed, and uses representative purchase prices rather than rental or spot prices, which may not capture later price cuts.
Coverage before 2023 is sparse, limiting the historical context.
Practical implications
Hardware planners should assume roughly 49% annual improvements in price-performance, but note that observed gains may be lower for Nvidia GPUs due to high margins; consider custom ASICs for better per-dollar compute.
What to watch
Monitor Epoch's next quarterly update to see if the 49% growth rate continues, and watch for adoption of new chip generations like Blackwell and beyond, which could accelerate or decelerate the trend.
