Executive summary
Research roles pay a $36k premium over generative-AI roles ($275,961 vs $239,453 avg), even though generative-AI has roughly 2.4x more openings (1,815 vs 760). Here's what the 3,556-posting compensation dataset tells us about where AI engineering compensation is actually flowing — and where agents and LLMs sit in the middle.
research — the top-paying skill tag with 100+ rolesllm — the most in-demand skill in the indexTop-paying skill tags
The top of the salary table is dominated by tags that take years to develop and don't show up in junior resumes. The 100-role minimum filters out small-sample noise so what's left is real, repeatable compensation signal.
| Tag | Role count | Avg salary |
|---|---|---|
research | 760 | $275,961 |
search | 717 | $271,750 |
reinforcement-learning | 591 | $270,450 |
distributed-systems | 1,498 | $257,128 |
infrastructure | 605 | $251,852 |
fine-tuning | 827 | $250,610 |
pytorch | 1,030 | $249,372 |
machine-learning | 641 | $246,384 |
llm | 2,636 | $245,362 |
deep-learning | 796 | $244,433 |
Most in-demand skill tags
Volume tells a different story than pay. The biggest tags are the ones every AI/ML team needs at least one of, not the ones the market pays the most for.
| Tag | Role count | Avg salary |
|---|---|---|
llm | 2,636 | $245,362 |
agents | 2,589 | $230,812 |
generative-ai | 1,815 | $239,453 |
cloud | 1,815 | $221,258 |
distributed-systems | 1,498 | $257,128 |
data-pipeline | 1,201 | $234,640 |
payments | 1,152 | $205,230 |
robotics | 1,126 | $208,925 |
healthcare | 1,073 | $211,220 |
pytorch | 1,030 | $249,372 |
The gap: where demand meets pay
Sweet-spot tags — high demand and high pay. llm (2,636 roles, $245,362 avg), generative-ai (1,815 roles, $239,453 avg), distributed-systems (1,498 roles, $257,128 avg), pytorch (1,030 roles, $249,372 avg), fine-tuning (827 roles, $250,610 avg). These are the skills the market is willing to pay above-median for and hire above-median volumes of. If you're choosing what to learn next, this is where leverage compounds.
High demand, lower pay. agents (2,589 roles, $230,812 avg), cloud (1,815 roles, $221,258 avg), data-pipeline (1,201 roles, $234,640 avg), payments (1,152 roles, $205,230 avg), robotics (1,126 roles, $208,925 avg). These tags are popular in job postings but compensation has not kept pace — usually because the labor pool is large or the work is more easily commoditized. Volume here doesn't translate to negotiating leverage.
The most useful frame for a job seeker isn't "highest paying" or "most posted." It's the intersection — tags where both axes are above the median. That's where you have negotiating leverage and a healthy supply of openings to negotiate against.
By experience level
The ladder skews heavily senior. Senior, lead, and principal roles account for 5,939 of 8,869 classified roles — 67.0% of the index. Junior makes up just 586 (6.6%). The salary endpoint reports a single weighted median of $215,000 (p25 $180,000, p75 $265,000, average $228,016); the heavy senior weighting is the main reason that median sits well above $200k. Note: the public /api/v1/stats endpoint exposes counts by level but not salary by level — to derive average salary per band, query the /api/v1/jobs endpoint with level= filters and aggregate locally.
| Level | Roles | Share |
|---|---|---|
| Junior | 586 | 6.6% |
| Mid | 2,344 | 26.4% |
| Senior | 3,016 | 34.0% |
| Lead | 2,363 | 26.6% |
| Principal | 560 | 6.3% |
Salary distribution
Of the 3,556 roles publishing salary, 2,642 (74.3%) fall inside the $150k-$300k core band — that's where the bulk of AI engineering compensation lives. 541 roles (15.2%) clear $300k, and 373 (10.5%) sit below $150k. The distribution is right-skewed: a meaningful tail above $400k (135 roles) pulls the average $228,016 above the median $215,000.
| Range | Roles | Share |
|---|---|---|
| Under $100k | 72 | 2.0% |
| $100k-$150k | 301 | 8.5% |
| $150k-$200k | 997 | 28.0% |
| $200k-$250k | 1,065 | 29.9% |
| $250k-$300k | 580 | 16.3% |
| $300k-$400k | 406 | 11.4% |
| $400k+ | 135 | 3.8% |
Practical takeaway
If you are choosing what to learn next. The pure-pay leaderboard (research, search, reinforcement-learning) clusters around research and infrastructure work that takes years to develop. The high-leverage practical picks are the sweet-spot tags — llm, generative-ai, distributed-systems — where demand and pay both sit above the index median. Generic machine-learning and data-science are not on that list, even though they remain common job titles. The market is rewarding specificity.
Methodology
This note is generated from the live aidevboard.com/api/v1/stats endpoint — a public, unauthenticated JSON API. The underlying index scrapes 560+ ATS sources (Ashby, Greenhouse, Lever, Workday, custom careers pages) on a daily cron, deduplicates by (company, title, location), classifies each role into a normalized skill-tag set with a rules-based parser, and stores employer-advertised salary ranges where present. Of 8,869 roles in the index, 3,556 disclose a salary range — the rest are silently excluded from the salary tables. Caveats: (1) salaries are advertised midpoints, not actual offers; (2) US-disclosure laws drive the highest disclosure rates so the dataset over-represents California, New York, Washington, and Colorado; (3) tag averages assume a role with multiple tags contributes its full salary to each tag's average; (4) the 100-role minimum on the top-paying table filters out small-n noise. This page auto-regenerates weekly.
Download raw data: The compensation-by-skill dataset is mirrored as a public gist — CSV · Markdown · view on GitHub. Auto-updated every weekly regeneration; canonical raw URLs are stable across revisions.
What's next
For the company-side view of this same dataset — who is doing the hiring, where the roles are concentrated, and how the workplace mix shakes out — see Q2 2026 AI Engineering Hiring Snapshot. For the infrastructure side — how many of the agents these engineers are building actually have a working MCP endpoint to talk to — see Q2 2026 MCP Ecosystem Health. For the engineering maturity ladder that separates the teams paying these premiums from the teams stuck in pilot, see Beyond the Prompt. Full reading paths at the Research Atlas.