27 Training Costs Should Continue To Fall 75% Per Year According to Wright’s Law, improvements in accelerated compute hardware should reduce AI-relative compute unit (RCU) production costs by 53% per year, while algorithmic model enhancements could lower training costs further by 47% per year. In other words, the convergence of hardware and software could drive AI training costs down by 75% at an annual rate through 2030. AI Training Hardware Cost AI Software Training Cost Using Neural Networks E Actual $ / RCU Predicted $ / RCU Actual Compute Estimated Compute C $100,000.00 1.000 N E G I L $10,000.00 L 0.100 E T N $1,000.00 I ) * ) AL e e 0.010 I $100.00 C ays cal I cal D S F S - I g g T $ / RCUo $10.00 o AR (L TFS (L 0.001 $1.00 0.000 $0.10 $0.01 0.000 0 1 100 10,000 1,000,000 100,000,000 0 1 100 10,000 1,000,000 Cumulative RCUs Produced Cumulative RCUs Produced (Millions) (Log Scale) (Millions) (Log Scale) *TFS-Days is a measure of compute required to train a model. Wright’s Law states that for every cumulative doubling of units produced, cost will fall by a constant percentage. Sources: ARK Investment Management LLC, 2024. This ARK analysis is based on a range of data sources as of Jan. 9, 2024, which are available upon request. Forecasts are inherently limited and cannot be relied upon. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.