AI cost crash speeds up Moore’s Law
The economics of artificial intelligence are undergoing a transformation so rapid it defies easy comprehension. New research suggests that the price of AI capabilities has plummeted by thousands of times in recent years—a rate that outpaces any other major transformational technology seen in the last century.
This dramatic deflation isn’t a slow trend; it’s an exponential sprint. The findings indicate that the cost of artificial intelligence is dropping by just under 50% every quarter, translating to a reduction of 13 times the cost annually. This pace is faster than the evolution of lithium batteries, DNA sequencing, and even the computational advancements that benefited from Moore’s Law throughout the 20th century.
The implications of this cost collapse are profound. While falling prices undoubtedly boost adoption, they introduce a difficult philosophical question for the world’s leading AI labs: what is the incentive to constantly race toward frontier AI development if the cost of achieving superiority is constantly diminishing? If a competitor can offer a nearly identical capability for less money, the traditional model of building and maintaining customer loyalty becomes incredibly tenuous.
This creates a fascinating conundrum for innovators like OpenAI and Anthropic. It becomes challenging to imagine large profits when the advantage of being the most capable model is fleeting. The race is no longer just about raw intelligence; it’s about the ruthless efficiency of cost management.
The benchmarks reveal a nuanced picture of this accelerating cost reduction. When tracking the cost of achieving specific AI performance levels across various disciplines, the progression isn’t a smooth line. Some metrics, like the cost of executing complex math or chess puzzles, showed more regular improvements. Others, such as frontier mathematics, experienced sharp drops followed by noticeable plateaus.
Furthermore, the rate of cost reduction appears to be slowing. Initially, cost falls measured 66% per quarter across several benchmarks, but that rate has noticeably decelerated over time. This slowdown aligns with broader economic pressures: the cost of energy and compute hardware has skyrocketed, making it harder to offer cheaper services when the raw materials for AI intelligence are becoming vastly more expensive.
This brings a critical caveat into focus: the possibility of “Benchmaxxing.” AI developers could be incentivized to tune their models specifically to perform well on certain benchmarks, rather than achieving true generalized intelligence. While researchers attempted to mitigate this by using randomized elements, the data suggests that optimizing for specific tests—whether they are generalized or not—is occurring. This warrants a degree of skepticism when interpreting these phenomenal cost savings.
Ultimately, the shift in the AI landscape is not solely a matter of price tags. Consumers and businesses must weigh the promise of cheaper models against the need for trust. Switching providers requires faith that new services are reliable, secure, and won’t drastically change their pricing structure down the line. The future of AI adoption depends not just on how cheaply we can run a model, but on the trust we place in the system we choose.
Credit
