The compression at the median did not come from AI pay falling. Over the tracked period the L3 ML/AI median moved from $375K to $383K, about 2.1 percent, while the non-AI comparator went from $279K to $297K, about 6.5 percent [3][1]. Levels.fyi reads its own monthly series as "plateaued / mild compression. ML/AI flat, non-AI catching up" [4]. The premium is narrowing because the rest of the field is being repriced upward, which is what an arbitraged skill looks like on the way to becoming ordinary.
Inside large established employers that repricing has already happened. At Google, Meta and Nvidia the ML/AI tag adds only low to mid single digits on top of an already high baseline, according to the analysis, while the frontier units in those same companies, Google DeepMind and Meta's superintelligence group, pay in frontier-lab bands because they bid against the labs for the same people [11]. The tag buys almost nothing; the box on the org chart buys a great deal.
The clean test is a job with no AI content in it whatsoever. The median US tech recruiter sits at about $155K across 1,367 submissions, about $217K at Google, and about $580K at OpenAI [9]. Google's recruiter premium works out at 1.40x, larger than the 1.29x an L3 ML/AI engineer now commands over a non-AI peer [2]. At OpenAI the recruiter median is 91 percent of that company's own median engineer package [3]. No amount of sourcing skill explains that number. Employer capacity to pay does.
The dispersion figures say it more coldly. The gap between the ML/AI 90th-percentile-to-median ratio and the broad software equivalent went from 0.07 to 0.17, roughly two and a half times wider [6][4]. The top-decile stretch is not happening to software pay generally. It is happening in one specialization, and inside that specialization at a short list of employers whose engineer medians are around $635K and $665K and whose L4 to L5 packages reach $1.5M to $3M [8].
Two caveats are load-bearing, and the source states both. The OpenAI recruiter median rests on seven submissions and is flagged as directional [10]. A large share of frontier-lab compensation is equity marked at private valuations, which Levels.fyi calls "paper, not liquidated cash," drawn from voluntary submissions rather than any employer's payroll [12]. Levels.fyi also refused to print a figure for the "AI Engineer" title, citing fewer than about 30 US L3 data points per half-year skewed junior [7]. Worth noting, given how much hiring copy already treats that title as if it had a market price.
What survives is the unexciting finding. ML engineers have held roughly 40 percent over data scientists at the same level across four full half-year periods with no drift, which Hakeem Shibly of Levels.fyi calls a durable, skill-based gap rather than a labeling quirk [5]. On the latest trailing-year cut, L3 ML/AI stands at $367,398 against $243,768 for data-focused engineers [2]. That is the part a compensation plan can be built on. The seven-figure tail is a bidding contest among a few buyers, settled in a currency nobody has cashed yet.
