AI chip performance per dollar grows 49% per year, Epoch AI finds
Epoch AI finds that the performance per dollar of AI chips purchased each quarter grew by an average of 49% per year since 2023, meaning each dollar buys roughly double the compute every 1.7 years. This improvement has been uneven: price-performance stayed nearly flat through mid-2024, then approximately doubled as Blackwell-generation chips became the majority of new spending. The calculation uses constant 2025 dollars and sales data from Epoch's AI Chip Sales Hub. While specialized chips like Google's TPU v6e offer higher per-dollar performance, most hardware spending currently targets Nvidia GPUs, which have lower ratios. The analysis models compute sold, not deployed, and uses representative purchase prices, so real-world efficiency may vary.
AI chip performance per dollar grows 49% per year, Epoch AI finds
Epoch AI: "Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year. At that rate, the performance a dollar buys doubles every 1.7 years."
Key takeaway
AI chip performance per dollar is doubling every 1.7 years, but the improvement comes in bursts tied to new generations, not steady gains.
What happened
Epoch AI released a data insight showing that the combined performance per dollar of AI chips purchased each quarter has grown by an average of 49% per year since 2023, in constant 2025 dollars. At this rate, the performance a dollar buys doubles every 1.7 years, according to the analysis.
The growth has been uneven: price-performance was nearly flat through mid-2024, then roughly doubled as Blackwell-generation chips became a majority of new spending. The analysis draws on Epoch's AI Chip Sales Hub and covers 24 chips with sales and price data from Q1 2023 to Q4 2025. For example, NVIDIA's GB300 costs about 5.5 times the P100's 2016 launch price but delivers roughly 200 times the performance, making it about 37 times more cost effective.
Evidence
Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year.
Epoch · attributed
Epoch AI: "Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year. At that rate, the performance a dollar buys doubles every 1.7 years."
Price-performance was nearly flat through mid-2024, then roughly doubled as Blackwell-generation chips grew to a majority of new spending.
Epoch · attributed
Price-performance was nearly flat through mid-2024, then roughly doubled as Blackwell-generation chips grew to a majority of new spending.
NVIDIA's GB300 costs about 5.5 times the P100's 2016 launch price, yet it delivers roughly 200 times the performance, making it about 37 times more cost effective.
Epoch · attributed
In 2025 dollars, NVIDIA’s GB300 costs about 5.5 times the P100’s 2016 launch price, yet it delivers roughly 200 times the performance, making it about 37 times more cost effective.
Why it matters
For operators, this means buying the latest generation can nearly slash compute costs in half within a couple of years, but timing purchases to generation shifts becomes crucial.
Limits and uncertainties
The analysis models the amount of units and compute sold rather than deployed each quarter.
It uses one representative purchase price per chip, which are purchase prices rather than rental prices.
Coverage starts in 2023 and ends in Q4 2025 because sales estimates for later quarters are not yet complete.
For Google and Amazon, chip prices are based on BOM modeling plus the partner's margin, not direct sales prices.
Practical implications
Operators should consider aligning hardware purchases with new chip generations to maximize performance per dollar, given the observed ~49% annual improvement.
Data center and AI budget planning should account for the rapid decline in cost per unit of compute, potentially allowing more compute for the same budget over time.
What to watch
Monitor Epoch's AI Chip Sales Hub for updated quarterly trends to see if the 49% growth rate persists or accelerates with future chip generations.
Watch adoption rates of Blackwell and subsequent chip generations to gauge if the uneven pattern of improvement continues.
