Amazon produced somewhere north of 2,400 open machine learning and applied science requisitions in the United States as of Q2 2026 — a number that, when combined with its historically high new-grad conversion rate from internship to full-time, makes it the largest employer of entry-level ML engineers in the country outside of Google and Microsoft. The company almost never appears in the "best first AI job" conversation. That is an information failure, not a talent judgment. The gap closed further in 2025 when Amazon lifted its base salary cap and restructured equity grants at the lower levels, bringing new-grad ML offers into a range that now competes directly with Big Tech peers. For the Class of 2026 who screened Amazon out early, the org is worth a second read.
What Amazon AI Actually Is
The phrase "Amazon AI" covers at least four structurally distinct organizations, and conflating them is the first mistake most candidates make.
Amazon AGI Labs is the newest and most research-forward unit. The SF Lab, seeded in 2024 with a team from Adept and led by David Luan as VP of Autonomy, focuses on foundational agentic AI — systems that can take actions in digital and physical environments. Luan's team is small by design: the initial hiring brief described a target of "a few dozen" researchers, drawn not only from ML but from physics, mathematics, and quantitative finance. That framing is deliberate. The SF Lab is explicitly not another model training organization; it is building the reasoning and action layer above the model. For new PhDs whose research sits at the intersection of agent systems and reinforcement learning, it is the most credible frontier research entry point Amazon has ever operated.
Amazon Science is the broader research organization, publishing across NLP, computer vision, robotics, automated reasoning, and quantum technologies. It runs the applied science internship program that feeds full-time Applied Scientist roles. PhD students entering through Amazon Science typically land at L5 — one level above the standard new-grad SDE entry point — reflecting the research credentialing premium Amazon has formally embedded in its leveling rubric.
AWS AI services is where volume lives. The teams building Bedrock (Amazon's managed foundation model API layer), SageMaker AI (the ML platform that powers an estimated 100,000-plus enterprise customers), and the underlying AI infrastructure are collectively the largest technical workforce in the AI-adjacent space at Amazon. Bedrock, which reached general availability in September 2023 and has been the centerpiece of AWS's generative AI go-to-market since, is currently hiring across model evaluation, inference optimization, and developer tooling. These are not research roles — they are the engineering roles that keep frontier model APIs running at enterprise scale. For an ML engineer who wants production exposure in year one, the Bedrock team's scope is unmatched by any pure-play lab.
Alexa AI — now operating as Alexa+ under a generative AI rebuild — went through a painful restructuring in late 2023 that eliminated several hundred positions as the organization pivoted away from the original rule-based architecture. What emerged is a team now focused on large-scale conversational AI, personalization, and the integration of Amazon Nova models into the Alexa product surface. The team is hiring again. The 2024 restructuring cleared the technical debt; the 2025 engineering buildout is filling the roles that the new architecture requires.
Amazon ML University: The Internal Development Engine
Amazon ML University (MLU) was founded internally in 2016, predating the current wave of AI investment by several years. It was opened to the public in 2020 and was formalized as part of Amazon's Upskilling 2025 initiative — a $1.2 billion commitment to train 300,000 employees across nine programs through 2025. MLU is the most technically intensive of the nine.
The program delivers graduate-school-level ML training through ten advanced courses, each structured as a 36-hour curriculum delivered in three-hour blocks over two weeks. Topics include deep learning, reinforcement learning, probabilistic graphical models, and ML production systems. The instructors are Amazon ML scientists — more than 400 of them, per Amazon's public disclosures — which means the curriculum reflects the actual production problems Amazon's systems face, not a theoretical survey of the field.
For a new-grad ML engineer arriving at L4 with a computer science or statistics background but limited production ML experience, MLU is a structured acceleration path that most employer-side alternatives cannot replicate at scale. Google has internal training infrastructure. Microsoft has it. Few others do. The program is free for all Amazon employees and does not require manager approval to enroll — a structural detail that matters for a 22-year-old who does not yet have the organizational capital to negotiate training resources.
MLU also functions as an external recruiting signal. Amazon has made the course materials publicly available and launched an Educator Enablement Program in 2022 to support university faculty adoption of the curriculum. The network effect is intentional: CS professors who use MLU materials know what Amazon trains for, and students who complete MLU-adjacent coursework arrive at the hiring process better prepared for Amazon's technical screen than candidates who have not.
The Compensation Reset
Amazon's compensation structure was the primary reason skilled graduates historically chose Google, Meta, or a frontier lab over Amazon. The base salary cap sat at $160,000 for years — a ceiling that constrained offer competitiveness at every level below L7. In February 2022, Amazon raised that cap to $350,000 for corporate employees. The move more than doubled the base ceiling and restructured the proportion of compensation that could be delivered in cash rather than back-loaded RSUs.
The practical effect for new-grad ML engineers entering at L4 in 2025 and 2026: base salaries in the $145,000 to $175,000 range, depending on location and team, with total first-year compensation landing between $195,000 and $240,000 when signing bonus and initial RSU vest are included. Seattle and Bellevue roles track the lower half of that band; New York and San Francisco roles track the upper half. These figures are sourced from Levels.fyi public submission data for L4 ML engineer and SDE roles at Amazon through Q1 2026 and from ValidGrad's published Amazon compensation guide.
The RSU structure is the variable that candidates consistently undermodel. Amazon's equity vesting schedule is back-loaded: 5 percent vests at the end of year one, 15 percent at the end of year two, 40 percent in year three, and 40 percent in year four. The practical effect is that first-year total comp, inclusive of a signing bonus that is specifically structured to offset the low early equity vest, is meaningfully lower than total comp at years three and four. A candidate who accepts an Amazon L4 offer of $210,000 in year one and remains for four years will, at a flat stock price, realize materially more in years three and four than in years one and two. At a rising stock price — AMZN appreciated approximately 27 percent over the calendar year 2024 — the back-loaded structure becomes an accelerant rather than a penalty. The candidates who leave Amazon after two years for a competing offer are making a specific bet that the competing offer's equity will outperform Amazon's back-half vest. That bet has been wrong more often than not over the past five years.
PhD graduates entering as Applied Scientist I (typically L5) see a higher floor: base salaries in the $170,000 to $200,000 range and total year-one compensation between $250,000 and $320,000, per the same Levels.fyi dataset. MS graduates typically enter at L4; PhD graduates enter at L5. Amazon applies that distinction consistently across teams, which means the leveling premium for a PhD at Amazon is structurally enforced rather than negotiated case-by-case.
The Entry Points
There are three distinct routes into Amazon's ML engineering org for new graduates.
SDE / MLE new-grad track via Amazon University Talent Acquisition (AUTA). This is the primary volume channel. AUTA runs structured on-campus recruiting at universities with strong CS and engineering programs, with the primary requisition cycle opening in late summer for the following graduation class. The track is open to BS, MS, and PhD graduates across software development and ML engineering roles. Candidates who complete Amazon internships are prioritized — per AUTA public materials, return offers are the default outcome for interns who meet bar, which in practice means the intern class is pre-screened at significant scale. For non-intern candidates, the AUTA process runs through an online application followed by a four-to-six round technical screen, with the loop including a Leadership Principles assessment that Amazon weights materially in hiring decisions at all levels.
Applied Science internship and new-grad track. For PhD candidates, Amazon Science runs a separate Applied Science Intern program targeting PhD students with backgrounds in ML, NLP, computer vision, automated reasoning, robotics, and quantum technologies. The published base pay range for the 2025 Applied Science internship was $136,000 to $222,200 annualized, depending on geography — figures drawn directly from Amazon's job posting, which by law must disclose pay ranges in jurisdictions including California, New York, and Washington. Full-time Applied Scientist I (L5) offers to PhD graduates who perform in the internship typically follow within two to four weeks of the final intern review.
AGI SF Lab direct applications. The SF Lab operates a non-standard recruiting process. Given its small target headcount, it is not running AUTA volume recruitment. The team solicits direct applications at agi-sflab-careers@amazon.com for candidates whose profile does not map neatly onto standard job postings — explicitly including people from outside ML who bring relevant quantitative or systems thinking backgrounds. For a new PhD whose research sits at the intersection of agent reasoning and a non-ML discipline, this is the highest-leverage application channel Amazon currently operates.
The interview process across all three tracks includes Amazon's standard technical assessment (data structures, algorithms, ML system design) plus the Leadership Principles behavioral screen. The LP component disproportionately filters candidates who prepare for it as an afterthought. The candidates who clear the Amazon interview bar at L4 and L5 in a single loop have typically done the same preparation work for the LP component as they did for the coding screen. That is not a cultural observation. It is a pass-rate observation.
Why Graduates Overlook Amazon — and Why the Calculus Has Shifted
Three reasons dominate the candidate-side data for why Amazon gets screened out early in the job search:
First, Amazon is not a frontier lab. It does not publish headline model benchmarks under its own brand the way OpenAI, Anthropic, and Google DeepMind do. Amazon Nova, released in late 2024 as a suite of six foundation models, is a credible frontier offering — but its launch was quieter than a comparable OpenAI or Anthropic announcement, and the media surface that candidates track did not amplify it proportionally. Candidates building their mental model of "where important AI work happens" from NeurIPS paper authorship lists and Twitter engagement are systematically undercounting Amazon's output.
Second, the Alexa restructuring created reputational drag. The late 2023 layoffs, widely reported as several hundred Alexa engineers, left a visible negative signal in the candidate research process. The follow-on story — that those layoffs cleared the path for a generative AI rebuild that is now hiring — has not fully penetrated the campus recruiting conversation. Recruiter sourcing from Amazon's Alexa AI team in 2025 is describing the org as net-additive in headcount relative to the post-restructuring floor, but that signal has not reached the level of the 2023 reduction news in candidate awareness.
Third, Amazon's compensation structure requires a four-year model to evaluate accurately, and most new graduates are modeling year-one numbers. The company that appears to pay $210,000 loses the comparison to the company that appears to pay $240,000 — even if the four-year cumulative value at Amazon is higher for a candidate who holds and the stock performs. The comp reset in 2022 narrowed the year-one gap, but it did not eliminate the structural modeling problem. Candidates who build a four-year model rather than a year-one comparison are more likely to accept Amazon offers when they have them.
The countervailing data is straightforward. Amazon employs more ML engineers than any US company outside Google and Microsoft. The infrastructure they run — Bedrock, SageMaker, the recommendation systems underlying Amazon's retail operations, the Alexa+ conversational layer — operates at a scale and reliability threshold that no frontier lab has yet had to meet. The production ML credential that an L4 engineer builds in three years on the Bedrock inference team is not available at comparable scale elsewhere. It is the difference between writing a model and running one at a hundred million requests per day. For the Class of 2026 who expects to spend most of their career in production ML rather than in research, that credential is worth more than the publication credit an academic research role produces in the same time period.
Amazon's position in the graduate AI hiring market is not the consequence of aggressive campus marketing. It is the consequence of building more ML-dependent products than almost anyone else — and needing the engineers to run them.
Compensation figures for L4 ML Engineer and SDE roles at Amazon sourced from Levels.fyi public submission data for 2025–2026 and ValidGrad's published Amazon compensation guide. Applied Scientist I (L5) compensation from Levels.fyi and 6figr public submissions, Q1 2026. Applied Science internship base pay range ($136,000–$222,200) from Amazon public job posting, 2025 Applied Science Internship — United States, Job ID 2806430. Amazon base salary cap history from Axios reporting, February 7, 2022, citing Amazon's announcement of the cap increase from $160,000 to $350,000. Amazon ML University program details from Amazon Science public announcements and Amazon aboutamazon.com Upskilling 2025 disclosures. Amazon AGI SF Lab details — including David Luan's title as VP of Autonomy and the Adept team origin — from Amazon Science blog, 2024. Amazon Nova model suite details from Amazon product announcements, November 2024. RSU vesting schedule structure from public employer disclosures and Levels.fyi community documentation. Open role count of 2,400+ is ENTRA Intelligence's estimate derived from LinkedIn Jobs and Amazon.jobs active posting counts for ML Engineer, Applied Scientist, and Data Scientist titles in the United States, as of May 2026; Amazon does not publish aggregate open role counts. AMZN stock appreciation figure (approximately 27 percent, calendar year 2024) from public market data.