AI salaries by country, 2026: the pay gap reshaping where models get built

The number people quote in Slack is almost never the number that hits a bank account. In the United States, Levels.fyi currently prints a median total compensation of $278,800 for the title “Machine Learning Engineer,” with a 25th percentile at $196,000, a 75th at $378,000, and a 90th at $495,000 (page updated 26 August 2026). That is self-reported offers, heavy on large tech and late-stage startups. It is not the Bureau of Labor Statistics, and it is not Bangalore.

The BLS median for computer and information research scientists — a related but broader occupation — is about $145,080. Glassdoor’s “AI engineer” base sits in a lower band around the mid-$130ks to low-$140ks. Those three facts can all be true at once. They measure different people. This article keeps the labels on.

What follows is a 2026 map of AI pay by country: what is measured, what is missing, why the United States still looks like a different labor market, and how to read an offer without lying to yourself. Every table names a source. Where sources disagree, we show the range instead of picking a flattering midpoint and calling it truth.

How a number is built (read this before the map)

Compensation is a stack. Mixing the layers is how Twitter threads go viral and how candidates get hurt.

1. Base salary — cash, taxed now, comparable across countries if you convert currency honestly.
2. Bonus — target percent of base; often 10–20% in big tech, variable at startups, sometimes zero.
3. Equity — RSUs at public companies (the US superpower); options at startups (a lottery ticket with a strike price). Four-year vest is still the default.
4. Location differential — the same level in New York, Austin, London, or Bengaluru is not the same cash.
5. Tax, visa, and cost of living — net, not gross. A $280k package in California is not a $280k life.

Levels.fyi’s US medians are total comp. Most UK, German, and Indian scoreboards that candidates actually see are cash, sometimes with a small bonus, rarely with liquid equity that looks like a Bay Area RSU grant. If you put them on one bar chart without saying that, you have done marketing, not reporting.

The United States, in numbers that have a URL

Source (2025–26)	What it measures	Printed figure	How to use it
Levels.fyi, MLE title, US	Median total comp (base+stock+bonus)	$278,800 median; $196k / $378k / $495k at 25/75/90	Best public window into large-tech ML offers. Selection bias toward people who file offers.
Levels.fyi, ML/AI SWE focus	Median total comp	about $245,000	Broader than the pure MLE title; still US-heavy.
Levels.fyi, Google MLE, US	Company-level TC	median about $288–290k; L3–L7 span ~$199k–$743k	One employer, not the market.
US BLS, computer & information research scientists	Occupational median wage	$145,080	Includes government, labs, and employers that never appear on Levels. Use it as a floor on “the whole country,” not as FAANG pay.
Glassdoor / similar, “AI engineer”	Mostly base, mixed titles	roughly $134k–$141k base in 2026 snapshots	Noisy titles. Do not average this with Levels TC.
Optiveum 2026 guide	Senior / staff bands, US	Senior often $190–260k; staff/principal TC $280–450k+; Bay senior base frequently from $225k with RSUs pushing TC over $400k	A recruiting firm’s bands. Directionally consistent with Levels at the top; not a census.
Placement blogs (staff/principal, 2026)	US total-comp ranges	Staff often quoted $500–800k TC; principal / L7+ $700k–$1.2M+ at the extreme	Treat as the right tail, not a median. We do not use these as country averages.

Two US markets live inside one passport. A senior ML engineer in the Bay Area or New York at a public tech company is playing a total-comp game. A senior ML engineer at a Midwestern insurer is playing a cash game that looks more like Germany than like Google L5. Country tables that dump “United States = $280k” are describing the first person and pretending they described the second.

Role ladder, not a single “AI salary”

Title inflation is how a $145k BLS median and a $279k Levels median coexist. Rough US 2026 shape, with the caveat that every row is a band, not a promise:

Rung	Years (typical)	What the work usually is	US 2026 ballpark	Where the number comes from
Junior / L3-ish	0–2	Fine-tunes, evals, glue	Often $120–170k base; TC $150–220k at tech firms	2026 recruiter guides; Levels L3 at big tech is the top of this
Mid / L4-ish	3–5	Own a model slice or a serving path	Levels MLE median lives here; mid TC often $230–380k in tech	Levels median $279k is the public anchor
Senior / L5-ish	5–8	Set the approach, on-call for the expensive failures	Cash $190–310k; TC $340–550k in the US tech distribution Optiveum/recruiters quote	Optiveum senior; recruiter TC bands — right-biased
Staff / L6	8+	Cross-team technical bets	TC commonly quoted $280–450k+ and, at public tech, $500k+; not a median of all staff jobs in America	Optiveum staff/principal cash-ish floor; Levels right tail
Principal / L7+	rare	Lab-defining bets	Public outliers $700k–$1.2M+ TC when equity cooperates	Right tail only. Do not average this into “the US.”

Outside the United States the ladder is shorter in money and longer in titles. A “staff” title in London or Bengaluru can be senior money in US units. Always compare scope (what breaks if you disappear) before comparing words on a business card.

United States is not one number

Levels’ national MLE median ($278.8k TC) is pulled upward by California and New York. Optiveum notes Bay Area senior base frequently starting around $225k, with RSUs pushing TC over $400k. That is a coastal public-tech story. The same title at a bank in Charlotte, a manufacturer in Stuttgart’s US plant, or a hospital system in the Midwest will look more like the BLS print than like a Google L5. When we say “the US gap,” we mean: the US tech-offer distribution vs other countries’ cash distributions. We do not mean every American with “machine learning” in a LinkedIn headline.

A PPP sketch — useful, not a leveller

Take a senior cash offer of $45,000 in Bengaluru (inside the $30–60k compilation band) and a US Levels median TC of $278,800. Nominal multiple: about 6.2×. Now cut the US number to cash-only for a fairer apples-to-apples — say a $180k base, which is plausible inside that distribution even if we do not pretend it is everyone’s base. Nominal multiple on cash: about 4×. PPP for India vs the United States often lands in a 3-ish range depending on the index and year. Even after a crude PPP haircut you are usually still looking at a multiple, not a rounding error. The remaining gap is equity + dollar savings + the option to move. That remaining gap is why work moves and why people still try to.

We will not publish a single “PPP-adjusted world ranking.” The indices disagree, rent is a city story, and imported laptops do not care about PPP. The sketch exists so nobody confuses a 9× tweet with a 9× life.

Tax, visa, and the parts of an offer that never trend

California and New York take a bite that a tweet never shows. A $280k TC in San Francisco is not $280k of choices; it is base taxed now, RSUs taxed at vest, plus rent that can erase a German senior salary by itself. Germany’s cash looks smaller and is often stabler, with health insurance that does not depend on staying at the company. The UK sits in between, with London rent doing California’s job at a smaller scale. India’s cash looks smallest and, at US captives, can still buy a local upper-middle life that a naive FX conversion hides — while not buying the same global mobility as a US W-2 plus equity.

Visa status is compensation. An H-1B holder comparing offers is not in the same market as a citizen who can walk across the street. Levels is full of people who can walk. Job boards in India are full of people who cannot. If you average them without saying so, you will invent a morality play.

Country map, 2026 — cash vs total comp

Currency conversions move. The table uses the bands as published, with a rough USD column so you can see order of magnitude. It is not payroll.

Market	Senior AI/ML band (published 2026)	~USD (order of magnitude)	Source	Mostly cash or TC?
United States	MLE median TC $278.8k; senior often $200–312k TC; Bay senior TC commonly $400k+ when RSUs hit	$200k–$400k+ at senior, median ~$279k on Levels	Levels.fyi; Uvik/Robert Half compilation; Optiveum	Total comp
United Kingdom	Senior £95k–£130k; lead/principal £140k–£180k+; some London AI bands £90–150k base / £110–200k total	roughly $120k–$165k senior; $177k–$228k+ lead	Optiveum; recruiter 2026 guides	Mostly cash; London equity exists but thinner than US public RSUs
Germany	Senior / MLOps €100k–€130k+; recruiter senior €85–140k base / €100–170k total	roughly $108k–$140k senior cash	Optiveum; 2026 EU guides	Cash; equity less liquid
India	Senior AI/ML often $30k–$60k in USD compilations; domestic prints vary widely by city and employer (AmbitionBox / Scaler-type boards)	$30k–$60k senior is the compilation band; Bengaluru FAANG/captive centers sit above it	Uvik 2026 compilation citing Scaler/AmbitionBox	Mostly cash; RSUs at US employers’ India offices exist but grants are usually smaller than US peers at the same level
Eastern Europe (compilation)	Senior AI/ML roughly $55k–$92k	$55k–$92k	Uvik 2026	Cash
Latin America (compilation)	Senior AI/ML roughly $58k–$85k; Colombia mid/senior prints lower	$58k–$85k senior band in that compilation	Uvik 2026	Cash

A 2026 comparison that took Levels’ US mid-level median (~$261k in one write-up) against an Indian mid-level print around ₹24.6 lakh (~$29k) called that a 9× gap. That arithmetic is real if and only if you accept both datasets and the same seniority. It is not a law of nature. Bengaluru offers at US tech captives can land well above ₹25 LPA. US non-tech ML jobs can land well below $200k TC. The gap is still large. It is not a single multiple.

Annual cost of an AI Techie

US Engineer: $280K/year
India Engineer + AI tools: $52K/year

Up to 5X cost advantage

— Rishi Bagree (@rishibagree) March 28, 2026

That post is the meme version of a real spread. $280k sits on top of the Levels MLE median. $52k sits inside the upper half of published India senior cash bands. “5× cost advantage” is a buyer’s sentence, not a worker’s. It also ignores: employer-side benefits in India, US benefits and equity risk, hours, on-call, and the fact that “AI tools” do not delete staff-level research. Use it as a temperature check. Do not paste it into a board deck as diligence.

Why the gap exists (the unromantic list)

Equity markets. Public US tech can grant RSUs that vest into a liquid stock. That is a compensation technology Europe and India have not fully copied. A $180k base in California plus $120k/year in vesting stock is a different product from €120k in Berlin with a tiny option pool.

Where revenue sits. Frontier labs and hyperscalers still book a disproportionate share of AI product revenue in the United States. They pay up for people who can move a training run or a serving stack this quarter. That demand is geographically concentrated even when the users are not.

Immigration and switching costs. H-1B, green cards, and spousal work authorization are part of US total comp in all but name. A worker who cannot switch employers freely will take a different deal than one who can. Levels data is full of people who can switch.

Purchasing power. $45k in Bengaluru is not $45k in San Jose. PPP closes part of the gap for local-currency life (rent, school, services) and almost none of it for imported goods, global school fees, or the option to move. We do not pretend PPP makes the offers equal. We also do not pretend a nominal 6–9× is the lived 6–9×.

Title inflation. “AI engineer” in 2026 can mean: fine-tune a 7B model; glue an API to a helpdesk; or own a training cluster. US Levels filers skew toward the last two rungs. National job boards mix all three. That alone can manufacture a fake gap.

Remote-from-anywhere, and the snapback

2021–23 produced a short fashion: SF-rate pay, any timezone. By 2025–26 the large employers had mostly returned to location bands. The labs that still pay a single global rate use it as recruiting theater and as a way to hire a few staff-plus researchers they cannot otherwise get. It is not the median policy. If an offer says “remote” and the band is San Francisco, get the location clause in writing. If it says “remote” and the band is “market,” ask which market. The difference is a house.

What actually moved in 2025–26

Inference and evaluation roles — people who keep models honest in production — saw faster posting growth than generic software engineering in several job-board cuts (see our fastest-growing AI jobs note). That is demand for a skill, not a promise that every prompt engineer clears $400k. Staff-level research compensation in the Bay Area can still clear seven figures when equity cooperates; that remains the right tail, and we will not write it as if it were the median.

Implementation work — RAG wiring, eval harnesses, support agents — did globalize. Bangalore, São Paulo, Warsaw, and a string of Eastern European cities show up in that layer because the work is spec-able and the US hiring loop is slow, not because the talent is a bargain in some moral sense. Captive centers of US firms in India pay above local product-startup cash and below US same-level TC. That sandwich is the actual market.

How to read your own offer

1. Split the PDF into base, bonus target, and equity. If equity is missing, you are looking at a European-shaped offer even if the letterhead is American.
2. Annualize RSUs on the vest schedule the company actually uses, not the four-year average they put in the pitch.
3. Name the city the band is built for. “Remote” is not a city.
4. Compare yourself to the same seniority in the same dataset. A BLS median vs a Levels 90th is a category error.
5. If someone quotes a 5× or 9× India–US gap, ask: same title, same year, cash or TC, which city, which percentile.

What we still cannot print as fact

There is no official global AI-salary census. Levels is opt-in. Glassdoor is opt-in. AmbitionBox is opt-in. BLS is a survey of occupations that do not say “LLM.” Recruiter blogs average the deals they see. We will not invent a single “world average AI salary.” We will not treat a seven-figure staff package as typical. We will not convert every rupee offer at this morning’s FX and call it a 2026 standard.

The pay gap reshaping where models get built is still this: training and high-end serving talent is paid in US total-comp units; a large share of implementation is paid in local cash units. That is enough to move work. It is not enough to write fairy tales about replacing a research org with a cheaper timezone.

Where the work actually sits

Follow the expensive jobs, not the press releases. Pre-training, post-training, and high-QPS serving still cluster where the people who have done it before already live — Bay Area, Seattle, New York, a London/Paris/Zurich research fringe, Toronto, a few Seoul and Beijing labs. Those seats are paid in the first table in this article. RAG systems, eval pipelines, data labeling ops, on-call for a chatbot, and “we wrapped a model in our CRUD app” sit wherever English-speaking engineers can be hired in weeks. That is the second table. A company that says it “moved AI to India” usually moved the second table and kept the first on a California offer letter. A company that says it “hires anywhere” usually still has a location band in the PDF.

None of this requires a morality lecture. It does require not mixing the tables. If your board believes a $52k fully-loaded India hire replaces a $280k US MLE because a tweet said 5×, you will ship a wrapper and call it a model. If your hiring manager believes every Bengaluru engineer is a $30k intern, you will lose the ones who already have US-captive offers. The data in this piece is enough to stop both mistakes. It is not enough to price a specific human. For that you still need a competing offer, a scope, and a city.

Sources

1. Levels.fyi — Machine Learning Engineer, United States (median TC $278,800; 25/75/90 as of 26 Aug 2026).
2. Levels.fyi — ML / AI software engineer focus (~$245k median TC).
3. Levels.fyi — Google MLE, United States (median ~$288–290k; L3–L7 span).
4. US BLS Occupational Outlook — computer and information research scientists, median ~$145,080 (via 2026 explainers citing BLS).
5. Optiveum, Machine Learning Engineer Salaries by Country 2025–2026 (US / UK / Germany senior bands).
6. Uvik, Software Developer Salaries & Rates by Country 2026 (US, UK, Germany, India, Eastern Europe, LatAm compilation).
7. Glassdoor / Built In 2026 snapshots for “AI engineer” base (lower than Levels TC; mixed titles).
8. Public X thread from @rishibagree, 28 Mar 2026, used as an example of how the gap is narrated — not as a dataset.

Updated August 2026. Figures are named-source ranges. If a number here does not have a source in the tables, it does not belong in a pitch deck.