One of the most important numbers in AI barely makes headlines: the cost of a given level of capability has fallen roughly a thousand-fold in about three years. What once took a frontier model and a large bill now runs on cheap hardware for cents. That collapse quietly rewrites what's worth building.
Where the savings come from
The cost drop isn't one breakthrough — it's a stack of them compounding: more efficient architectures (sparsity, better training), smaller models that punch above their weight (via distillation), and a deep bag of inference tricks — quantization, speculative decoding, smarter caching — that cut the cost of serving each token. Each shaves a slice; together they collapse the total.
When the price of a capability falls 1,000x, things that were absurd to build become obvious to build.
Why it changes strategy, not just budgets
Cheap inference doesn't only save money on existing products — it unlocks new ones. Features that were uneconomical at old prices (run a model on every request, every document, every frame of video) become routine. The constraint shifts from "can we afford to call the model?" to "what should we build now that we can call it freely?"
The catch worth naming
Cheaper per token can mean more tokens used, so total bills don't always fall — they often shift to bigger, more ambitious workloads. The discipline that matters now is spending the abundance well: right-sizing models, caching aggressively, and not reaching for a frontier model when a cheap one clears the bar. Abundance rewards the teams that stay deliberate.
