AI is getting cheaper at a pace no transformative technology has matched. That is the central finding of a Deutsche Bank report published on October 14, 2026.
According to the bank, the cost of reaching a specific level of AI performance has fallen by nearly 47% every quarter since 2023. Compounded over a year, that works out to a 13-fold reduction.
The numbers behind the deflation
Deutsche Bank benchmarked AI’s cost curve against other technologies that reshaped the economy. By the bank’s math, AI’s deflation is running about four times faster than the decline in DNA sequencing costs. It is roughly six times faster than gains in general compute. Against lithium-ion batteries, the gap widens to 18 times.
Separate analysis from Epoch AI, cited alongside the bank’s work, makes the point even more vividly. Epoch found a 725-fold drop in the cost of matching a given score on the GPQA Diamond benchmark. The comparison runs from early 2025 OpenAI models to later iterations.
GPQA Diamond is a set of difficult, graduate-level science questions used to test AI reasoning.
The trillion-dollar paradox
Falling unit costs have not translated into falling total spend. Deutsche Bank projects that capital expenditure on AI infrastructure by US and Chinese hyperscalers could approach $1 trillion in 2026. The bank frames that spending as essential to keep progress moving toward artificial general intelligence, or AGI.
The bank estimates AGI-level systems may need roughly 1,000 times the compute available today. That figure applies even in optimistic scenarios.
The costs that are not shrinking
Deutsche Bank also flags pressure points that cheaper models do not automatically solve. Two stand out: verification and energy.
Verification is the work of checking whether AI output is actually correct and reliable. The bank notes concerns that these verification and reliability expenses are rising.
Energy is the other constraint. According to the report, power limitations could hold back productivity gains before AI reaches superintelligence.
Background: why cost curves matter
The 2023 starting point anchors the analysis in the period when generative AI went mainstream and competition among model developers intensified.
What this means for the AI industry and investors
For model developers, if the cost of matching a given performance level falls 13x a year, any lead built on raw capability erodes quickly. Pricing power may increasingly depend on staying at the cutting edge, which brings the conversation back to that enormous capex bill.
The 1,000x compute estimate for AGI is perhaps the most sobering data point. If energy supply becomes the binding constraint, falling model costs may matter less than who can secure electricity and data center capacity.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.