Compensation data in AI ages badly and varies wildly by market, funding stage and how badly the hiring team needs the person this quarter. What follows is the shape of the market rather than a salary table you should quote in an offer: the bands, the multipliers that actually move them, and the structures that get a good hire for less than the headline number.
The five bands, roughly
Base salary, US market, mid-2026, for a company that is neither a frontier lab nor a struggling seed startup. Total compensation runs higher wherever equity is meaningful.
- AI / LLM engineer, mid-level — the most liquid part of the market. Comparable to strong senior backend engineering, with a modest premium.
- AI / LLM engineer, senior with shipped production systems — a clear premium over general senior engineering, driven by scarcity of people who have operated these systems rather than built demos.
- Machine learning engineer with production ownership — similar to the above, higher where the role includes training and serving your own models.
- AI product manager — tracks senior product management, with a premium where the person can genuinely own evaluation and quality rather than only roadmap.
- AI researcher — the widest and highest band by a distance. Frontier-lab experience can multiply the rest of the market, and at that end compensation is mostly equity.
The distribution matters more than the midpoint: the gap between the median and the top decile is larger in AI than in any adjacent engineering discipline, because the top decile is defined by having done something very few people have done.
What actually moves the number
- Production evidence. Having operated an AI system with real users is the single largest legitimate multiplier. It is also the thing most easily verified — and most often assumed.
- Frontier-lab or AI-native employer history. Real signal, and frequently overpriced. You are buying exposure to good practice, not a guarantee of it.
- Scarcity of the specific skill. Evaluation and voice people are scarcer than general LLM engineers right now; RAG is closer to commodity than it was a year ago.
- Location. San Francisco sits above every other market, with New York behind it. The remote market has narrowed the gap without closing it.
- Your own urgency. The most expensive hires are the ones made in the last two weeks of a quarter.
Where teams overpay
Three patterns account for most of the waste we see:
- Hiring a researcher for an engineering job. The most expensive category in the market, doing work an AI engineer would do faster and enjoy more. It usually ends with the researcher leaving inside a year.
- Paying the pedigree premium for a job that does not need it. Ex-lab experience is worth a lot for frontier work and considerably less for shipping a retrieval feature over your own documentation.
- Buying seniority instead of evidence. A senior title from a company that never shipped AI is worth less than a mid-level engineer who has run one of these systems in production for a year.
The cheapest strong AI hire on the market is a very good engineer who has shipped one real LLM system and does not yet have the title to prove it.
Cheaper structures that work
- Hire for the skill, not the seniority band. Search by capability and read the evidence — plenty of people doing excellent AI work are priced at their old title.
- Contract-to-hire for the first system. A four-to-six week paid engagement to ship one thing tells you more than any interview loop, and both sides can walk away.
- Grow an internal person. A strong engineer already inside your codebase with three months of focused AI work often beats an external senior hire, and costs a fraction.
- Split the role. One evaluation-minded engineer plus a domain expert frequently outperforms one expensive generalist.
Budgeting the whole cost, not the salary
Salary is roughly two thirds of what an AI hire costs in year one. The rest: inference and infrastructure spend on whatever they build, which for a successful feature grows faster than anyone forecasts; evaluation and annotation work, which is real headcount even when it is contracted; and the hiring process itself, where the most expensive line is usually the eight weeks a role sat open.
The practical move is to shorten the open period. Search a pre-scored pool and shortlist from evidence instead of running a funnel from scratch.