# What machine learning engineers actually make, and how to make more

Source: https://www.techinterview.org/post/3233476387/what-ml-engineers-actually-make/
Updated: 2026-07-15 · techinterview.org

A machine learning engineer at Amazon and a machine learning engineer at OpenAI can carry the same title and take home compensation that differs by a factor of three. Levels.fyi put Amazon's MLE median near $265K in mid-2026 and OpenAI's near $800K. Same two words on the offer letter, very different deposit. That gap is the thing to understand first, because it points at where the pay actually comes from, and it isn't the base salary.

## Where the money is, and isn't

An MLE offer has four moving parts, and the recruiter usually leads with the lowest-variance one. Base salary for a working US MLE tends to land somewhere between $130K and $210K depending on level and metro, and it stops moving much once you clear the junior rungs. The annual bonus is typically 10 to 20 percent of base and often rides on company performance you have no say in. Equity is where that three-times spread actually lives. The sign-on bonus is a one-time cash patch, usually there to make a thin first year look competitive.

Equity is also where people misread offers. A four-year grant that vests evenly hands you a quarter of it per year, so a $400K grant is worth $100K a year, not $400K. Read the vesting schedule, because some companies back-load it: 5 or 15 percent in year one, the bulk in years three and four, with the sign-on covering the early gap. Pull the sign-on out and the steady-state number is lower than the four-year average the offer implies. The grant also expires around year four, and whether your comp holds after that depends on refresher grants that were never promised in writing.

At the frontier labs the equity isn't public stock, which changes the math again. OpenAI grants profit participation units; Anthropic and the other private labs grant equity with no open market to sell it in. The paper value can be large, and the quoted number assumes a valuation that may not hold, secondary sales are gated, and you can't liquidate on a random Tuesday. A Meta or Google grant in publicly traded stock you can sell the morning it vests is a different kind of asset than a private-lab unit worth a headline million.

## What each company pays for the same title

These are mid-2026 figures from Levels.fyi for people carrying the machine learning engineer title, base plus bonus plus equity. They move month to month, and the high ends are the richest companies at senior levels, not what a typical MLE sees.

| Company and MLE level band (Levels.fyi, mid-2026) | Entry-level total comp (USD/yr) | Senior/staff total comp (USD/yr) | Reported median (base + bonus + equity) |
| --- | --- | --- | --- |
| Amazon (L4 to L6) | ~$176K at L4 | ~$409K at L6 | ~$265K |
| Nvidia (IC1 to IC4) | ~$205K at IC1 | ~$331K at IC4 | ~$261K |
| Google (L3 to L7) | ~$199K at L3 | ~$743K at L7 | ~$290K |
| Apple (ICT2 to ICT6) | ~$191K at ICT2 | ~$528K at ICT6 | ~$401K |
| Meta (E3 to E6) | ~$187K at E3 | ~$786K at E6 | ~$430K |
| OpenAI (SWE L2 to L6) | ~$254K at L2 | ~$1.23M at L6 | ~$800K |

Two things are worth pulling out of that table. Meta's median reads higher than Google's at comparable seniority because Meta weights equity harder and refreshes aggressively. And the OpenAI row sits in its own tier, which is the AI hiring surge showing up in the data rather than a stable long-run rate. Treat that bottom row as the current top of the market, not the baseline you should expect.

## Level beats title, every time

The biggest single lever on your number is the level you're hired at, and levels don't line up across companies. Google L5, Meta E5, and Amazon L6 sit at roughly the same seniority even though the numbers differ. An L4 to L5 step at a big-tech company is often a jump of $100K or more in total comp, larger than any raise you'll negotiate inside a level. That's why the level line in an offer is worth pushing on harder than the base. Accepting a down-level to get in the door can cost you six figures a year until you climb back to where you already were.

## The title covers four different jobs

Machine learning engineer is a loose label, and which version you actually do sets a ceiling on the pay. The work splits roughly into ML infrastructure (serving, pipelines, feature stores), applied ML (training and shipping models against a product metric), ML platform, and research engineering, meaning the people who build and scale the training runs behind frontier models. Research-adjacent roles at the labs pay the most right now because very few people can debug a distributed training job at scale and everyone wants the ones who can. A data scientist who relabeled themselves MLE for the resume tends to get paid on the lower data-science curve. If you're aiming at the top of the range, the training-and-inference-at-scale end is where it sits in 2026.

## How people actually get here

The path into the role rarely starts at the role. Most working MLEs arrive from one of two directions: a software engineer who moved toward models, or a data scientist or researcher who moved toward production. The SWE-to-ML route tends to pay better because the systems skills (writing real services and handling data at scale) are exactly what applied-ML and infrastructure teams are short on. Modeling knowledge without the engineering to deploy it lands you closer to the data-science band. New-grad MLE roles exist but are thin; more often you get hired as a general SWE or data scientist and specialize once you're inside, then re-level or switch teams. That first internal move is usually where the comp curve steepens.

## What actually moves your number up

Changing companies is still the fastest raise. Internal bumps fight comp bands and budget cycles; an outside offer resets against the current market, which the AI surge has inflated. The skills carrying a real premium are specific: production LLM fine-tuning, RAG systems that survive contact with real traffic, distributed training, inference optimization. Recruiter benchmarks put those roughly $20K to $50K over a generalist MLE rate. Having watched a few lectures on transformers does not clear that bar, and classical ML on tabular data pays fine but isn't where the bidding war is.

Location still counts even in a remote market. Bay Area and New York bands run 15 to 30 percent over the same level in Austin or Atlanta, and most companies apply a geo cut when you relocate. Fully remote roles at large companies usually pay a second-tier metro band rather than the San Francisco one, so the dream of Bay Area pay from a cheap city is less common than the listings suggest.

Every figure here shifts, and in 2026 it shifts quickly. The AI comp bubble has pulled frontier offers up and dragged the broader market along, and a pullback is plausible. Don't treat any single number, including these, as your target. Pull the live distribution for your exact company, title, and level from Levels.fyi, sanity-check the outliers on Blind, and keep in mind that the flashy figures are the top of the range at the highest payers. The median working MLE outside the labs is nearer $250K to $300K total than a million.

Before you sign, ask the recruiter for the vesting schedule and the refresh policy in writing, rather than only the four-year headline. The companies that front-load equity and refresh well will hand it over without flinching. The ones that quote a large four-year total and go vague about years two through four are telling you something worth hearing.