AI Gets 10 Times Cheaper Every Year, and That Is Why It Makes More Money
Anthropic cuts prices about in half while its costs fall even faster, and customers respond by finding far hungrier jobs for AI.
All-In with Chamath, Jason, Sacks & Friedberg · Anthropic S-1 Shocker, Greatest Business Ever? Google's $500B Wipeout, Midterm PredictionsFor Anthropic, doing a given job with artificial intelligence now costs about one tenth as much as it did a year ago, a pace that has held so far and could happen again if the trend continues.
The steady drop helps the business rather than squeezing it. David Sacks, an investor, put the company's gross margin, the share of each sales dollar left after the direct cost of running the models, at something like 85 percent.
The gains come from two directions. One is raw compute, which is scaling rapidly through better chips and more chips. The other is what Anthropic calls compute multipliers, things like algorithmic improvements, better curation of data, and better model designs that make the models more efficient, so they need far less computing for equal work.
When a task gets 10 times cheaper to run, Anthropic can cut what it charges by roughly half and still keep the rest as profit. Customers feel they got a bargain, and margins go up at the same time. Token prices, the fees charged for each small chunk of work the model processes, have fallen sharply for that reason, not because the company is losing pricing power.
Then comes the second half of the loop, a 19th century idea called Jevons Paradox. When coal became more efficient to use, people used more coal, not less. When AI gets cheaper, companies find more jobs for it.
In particular, the cheaper second and third tier models have become good enough for many businesses, so enterprises use them to solve more problems without paying for the most advanced system. And the newest jobs are hungrier still. Software agents that carry out multistep tasks use many more tokens than simple prompting, and continuous cyber defense uses many more tokens than an occasional security test.
The figure cited was 50 to 100 times as many tokens for those new uses, and Sacks agreed that new applications consume far more tokens. So the price per unit falls, but total demand and market size keeps growing.
Questions people ask
What exactly is a token?
A token is a small piece of text, roughly part of a word, that an AI model reads or writes. AI companies add up tokens to measure work, and they charge customers by the million tokens processed. More complex work, like an agent doing research, simply uses many more tokens.
Can efficiency really improve tenfold every year forever?
Probably not forever, but it has come from several sources at once: faster chips, more computers working together, better training data, and more efficient designs. Historically, computing has improved rapidly for decades as these advances stacked. Physical limits on power, chips, and data centers may eventually slow the pace.
Does Jevons Paradox show up outside AI?
Yes, it is common. More fuel efficient cars encouraged people to drive farther and buy bigger cars, so total fuel use did not fall as much as expected. Faster internet did not reduce data use, it enabled video streaming and cloud software. When a useful resource gets cheaper, people often invent new ways to consume it.