Two of the Index's curves run in opposite directions, and the distance between them can be worked out. Training compute for notable models doubles roughly every five months [5], which is 2.4 doublings a year, or about 5.3 times more compute. Machine learning hardware costs fall about 30% a year [6], so each unit of that compute costs about 0.7 of last year's price. Multiply the two and a run that stays at the frontier costs about 3.7 times more each year [16]. Cheap tokens and expensive training runs are one market seen from opposite ends.
Which end you stand on depends on whether your quality bar moves. Both headline reductions are pinned to a fixed level: 60% on MMLU for the parameter count [1], and GPT-3.5-equivalent accuracy of 64.8% for the price [3]. What has become nearly free is a 2022 grade of capability. For work whose bar was set then and has not moved, the build case is better than it was; for work that needs whatever is best this quarter, the arithmetic above governs.
That price fell more than 280-fold in eighteen months [3], which could argue for waiting rather than building, since a further drop might land before any in-house project ships. But inference prices have fallen between 9 and 900 times a year depending on the task [4]. A rate uncertain across two orders of magnitude offers no planning input; the level observed at a point in time is what can actually be used. The most recent price point in this record is October 2024 [3], in a report published on 7 April 2025 [12].
The board-deck version says unit costs collapsed, organisational AI use went from 55% to 78% of survey respondents in a year [19], and the response is to insource. It is incomplete in two places. Two-thirds of the 149 foundation models released in 2023 carried open weights [13], which is the part of the build case that holds, but on RE-Bench top systems beat human experts fourfold on two-hour tasks and lose two to one at 32 hours [11]. What is cheap is short work.
The legal side moved without waiting for either decision. Stanford's tally reports how many state AI laws passed, not what they require, and federal bills rose while passage stayed low [9]. Reported incidents reached 233 in 2024 [10], up from about 149 the year before [17]. Those are counts, so anyone sizing a compliance line is estimating from outside the Index, and the obligation follows the deployment: training your own model does not change which state a user sits in.
The decision this quarter is therefore narrower than the price curve implies. The buy case holds for anything measured against the current frontier, and the build case holds where the bar is fixed and the tasks are short. The consequence to name for next quarter is that both routes answer to the same state-by-state surface [9], and this record gives no figure for what answering it costs.
