LLMs werden radikal billiger
Perhaps the single most striking fact about the business-to-business AI market is the speed at which the price of intelligence has fallen.
To be sure, plunging prices in some form have been seen with other technologies.
The deregulation of telecommunications in the 1990s cut the price of a long-distance phone call by about 90 percent over 15 years.
Gordon Moore, who would later be a cofounder of Intel, predicted back in 1965 when working for Fairchild Semiconductor that the number of transistors on a chip would double every two years.
Moore’s Law delivered roughly a hundredfold reduction in the cost of computing per decade.
The LLM market has compressed an even larger price decline, which might fairly be characterized as a price collapse, into just two years.
Warum sinken die Preise?
Three forces drive these declines.
First, advances in model architecture and training techniques allow creators to achieve the same level of intelligence with fewer neural network parameters.
This translates into less computation required per query.
Second, improvements in hardware efficiency and serving infrastructure enable inference providers to deliver a given model at lower marginal cost, reducing the price of inference even while holding model capability constant.
Finally, competition—both between closed-source providers and from new open-source models—puts downward pressure on prices.
Unternehmen nutzen die sinkenden Preise, um bessere, vorher zu teure, Modelle zu verwenden:
The implications of this price decline for firm adoption are profound.
Tasks that were prohibitively expensive to automate a year ago are now economically viable, and the set of applications for which AI is cost-effective is expanding rapidly.
At the same time, firms shift token consumption to more intelligent models, so that the average per-token price paid remains relatively stable.
Firms are not simply buying the same intelligence more cheaply—they are upgrading to more capable models as prices fall.
The intelligence of models in use has risen steadily, with the token-weighted median intelligence score increasing from 0.30 in January 2025 to 0.44 by year-end.
This pattern is consistent with firms using price declines to purchase better, not just cheaper, AI.
The dynamic echoes the history of personal computers and smartphones—modest price declines paired with order-of-magnitude capability gains—except that for LLMs, both prices and capabilities have moved sharply in the buyer’s favor.
Aus der neu publizierten, im Journal of Economic Perspectives veröffentlichten Studie »The Emerging Market for Intelligence: How Firms Buy and Sell AI«.
Morgen kommt Teil 2 dazu.