djbinder: AI Hardware Cost for Equivalent Tasks Far Below Human Wages
LessWrong author djbinder built an AI compute-cost dataset covering 484 tasks and 1,684 data points, finding that when AI reaches human-level performance, the median task's GPU cost is about 4 cents per hour of replicated human work—far below the US median wage of $25 per hour and the fully loaded $100–$300 per hour cost of skilled white-collar workers. Current API pricing is more than ten times the hardware cost, but the author expects long-run marginal costs to converge toward the hardware floor.
Doing a task is not the same as doing it cheaply. Frontier AI systems have already solved the Navier–Stokes Millennium Prize problem and completed tasks that would take months or years of human effort, but readers usually cannot answer a more basic question: how much does it actually cost to run one of these tasks once? If everything AI can do is expensive, then the scale at which it replaces human workers is limited by compute bottlenecks; if it is already cheap, the question becomes one of deployment speed rather than cost.
djbinder published this analysis on September 27, 2026, attempting to answer the question with data. He used GPT-6 Astra and Fable 5.1 to search and compile a database covering 484 unique tasks and 1,684 data points, comparing the compute an AI system needs to perform a task (measured in FLOPs) with the active time a human would take to perform the same task. The dataset draws on sources including Terminal-Bench, METR's time-horizon suite, SWE-Marathon, Epoch AI's benchmark runs, and GDPval. For each task, he takes the cheapest run that reached or exceeded human-level performance.
The core finding is that the compute required to perform a task is roughly linear in the time a human would take, with the median at about 10^11 FLOP per second of human time and 90% of tasks falling between 10^9 and 10^13 FLOP per second of human time. Using the H100's 8-bit floating-point peak rate, a rental price of about $3 per GPU-hour, and typical utilization of about 10%, the hardware cost per delivered FLOP is extremely low. The GPU cost of replicating an hour of human work for the median task is about 4 cents; even the most expensive tasks—the Navier–Stokes proof, a Pokémon Crystal play-through, and an ARC-AGI-3 environment—cost about $15 per hour, still below the US median wage of $25 per hour and far below the fully loaded $100 to $300 per hour cost of skilled white-collar workers such as research scientists and software engineers.