AI compute costs are falling faster than almost any technology in history — so why hasn't that made mass automation happen yet?

This explores the gap between two clocks: how fast AI is getting cheaper, and how fast cheaper AI turns into automated work. The corpus has strong data on the first clock and an important correction on the second. The short version is that price is falling faster than for almost any technology on record, yet price alone does not decide when a job gets automated.

Start with the speed. Epoch AI tracked what it costs to hit a fixed score on five benchmarks and found that cost dropping about 47% per quarter since 2023, roughly 13× per year. That is about four times faster than DNA sequencing got cheaper and 54 times faster than electricity did historically. One example: a result that cost OpenAI's o3 $0.30 was later matched for $0.0004 How fast is AI capability becoming cheaper to produce?. The decline also isn't steady. Prices fall fastest, around 66% per quarter, right after a capability first becomes state of the art, then slow to about 32% per quarter within two years. Epoch reads this as a short-lived monopoly premium that competition wears away Do AI performance prices fall faster when capabilities debut?. So a newly possible task gets cheap fast at first, and the later savings come more slowly.

Here is what you might not expect: being technically able to do a task and being worth automating are different thresholds. Svanberg and colleagues built a full cost model for computer-vision tasks. They found only 23% of the wages paid for vision work were economically attractive to automate. Most tasks sat below the minimum scale needed to pay back the fixed costs of building and running a system, even when the AI could already do the job Which vision tasks actually make economic sense to automate?. Falling model prices should push that share up over time. But the bottleneck there was fixed cost per deployment, not the cost of each model call, so cheaper inference alone doesn't remove it. The bill below the model points the same way. In a one-gigawatt data center, servers make up about 60% of annual cost and energy only about 7%, so the price floor is tied to how fast hardware gets paid off and replaced What actually drives AI data center costs?. Smaller models trained on real multi-step work can reach the low-cost end of the cost-versus-performance curve Does model efficiency matter more than peak capability for real work?. This matters because in long tasks, cost and delay add up across every step, not just one answer.

Automation also won't arrive evenly. Firms more exposed to AI replace online freelance workers faster and more cheaply than other firms do Do firms substitute labor for AI at different rates?. That points to companies building internal know-how that compounds, rather than one cost threshold that flips a whole economy at once. And price-per-benchmark-point may itself mislead. Benchmarks favor neat, auto-graded tasks. Evaluations on messy, long-running real work, with costs reported, sometimes show more capability than benchmarks do and sometimes less Do automated benchmarks hide what frontier AI systems can really do?.

On timelines specifically, the corpus is thin. The only dated forecast covers one sector, AI research itself. Kwa's simple 8-parameter model predicts more than 99% automation of AI R&D by mid-2032 Can simpler models predict AI R&D automation timelines accurately?. A separate critique argues that claims of compressing years of progress into months rest on unproven assumptions, such as whether skill on small tasks carries over to research that really matters Could automated AI research compress years of progress into months?. No note here gives a date for economy-wide automation. What the corpus does suggest is a reason to care whatever the date turns out to be. If automation comes as gradual, firm-by-firm replacement rather than a sudden tipping point, it can quietly remove a safeguard: institutions stay aligned with human interests partly because they depend on human workers. That erosion may be hard to notice until it can't be undone Does incremental AI replacement erode human influence over society?.