The number that now governs access to a career at a frontier AI lab is not a GPA threshold, a publication count, or a Leetcode percentile. It is 78. That is the percentage of new-graduate full-time offers extended by frontier AI labs in the first half of 2026 that went to candidates who had previously completed an internship at the same organisation. In 2022, that figure was 45 percent. The summer internship, once an exploratory program designed to give undergraduates a taste of research-grade work before they returned to campus, has become the primary pipeline through which frontier labs hire the generation that will build their next models.
For the students who secured an internship at Anthropic, OpenAI, Google DeepMind, Meta AI Research, or Microsoft Research in the summer of 2025, this is straightforwardly good news. Their conversion rate was high, their return offers were generous, and their career path into the AI economy is largely settled. For the much larger population of 2026 computer science and ML graduates who did not — because they attended a non-target school, because they were international students navigating visa constraints, because they applied in October and the cohort was already full — the internship bottleneck represents a structural narrowing of a pipeline that was already competitive. The AI lab new-grad market in 2026 is not an open meritocracy. It is, increasingly, a closed-loop system that rewards students who solved the internship problem two or three years before graduation.
This report, drawing on ENTRA Talent Index tracking of 1,200 placed candidates, Levels.fyi compensation data, and direct recruiter conversations at 47 internship programs, maps the full architecture of the AI lab internship economy: what labs are paying, how conversion rates have moved, why the campus war for intern cohorts is intensifying, and what the 2026 graduating class that missed the intern pipeline needs to understand about their options.
The Conversion Numbers, Lab by Lab
The 78 percent aggregate conversion rate obscures meaningful variation across labs. Anthropic sits at the high end of the market with an estimated 84 percent intern-to-full-time conversion rate across its summer 2025 cohort of approximately 120 interns. The company's conversion discipline is intentional: Anthropic runs a deliberately small intern program relative to its headcount — OpenAI hosts roughly 180 interns against a larger employee base — and invests heavily in the intern experience as a recruiting mechanism. Interns are embedded in active model teams, not auxiliary research projects, and the bar for the internship offer is calibrated to approximate the bar for a full-time research scientist or engineer offer. If Anthropic is extending a summer internship, the organisation has effectively already decided that it wants that person.
OpenAI's summer 2025 intern cohort of approximately 180 converted at 79 percent, a figure consistent with the prior year's 81 percent and reflective of the organisation's practice of using the internship as a final-stage evaluation tool. The remaining 21 percent who did not receive return offers were concentrated in the applied engineering tracks rather than the research tracks, where the conversion rate is nearer to 90 percent. OpenAI's intern compensation — $90 to $110 per hour, plus a $12,000 to $15,000 housing stipend for relocating students — has made the program the most-applied-to internship at every target school ENTRA surveyed, including MIT, Stanford, CMU, and Caltech.
Google DeepMind presents the most geographically complex picture. The organisation splits its intern cohort across Mountain View and London, with approximately 250 summer 2025 interns in total. The Mountain View-based cohort converted at 74 percent; the London cohort at 66 percent, reflecting in part the narrower hiring band in DeepMind UK's research science tracks and the UK's higher bar for sponsoring international intern-to-full-time transitions under the Skilled Worker visa route. Aggregate conversion across both sites was 71 percent, the lowest among the five labs tracked but still materially above the 45 percent industry average of 2022.
Meta AI Research hosted over 300 interns across its Fundamental AI Research (FAIR) and Applied Research teams in summer 2025, and converted 68 percent to full-time offers. Meta's lower conversion rate is partly a function of program scale: a larger intern cohort drawn from a wider range of target schools will include a higher proportion of candidates for whom the internship serves as a genuine evaluation rather than a near-certain prequel to an offer. Meta's compensation stack is competitive — $95 to $115 per hour with a $10,000 housing stipend — but the organisation's equity conversion packages for returning interns are where the real differentiation lives. Meta has moved to accelerated vesting cliffs for full-time hires who completed a Meta internship, a mechanism that compresses the liquidity gap with publicly traded stock compared to what an incoming hire from the open market would receive.
Microsoft Research converted 62 percent of its approximately 200 summer 2025 interns, the lowest conversion rate among the five labs but accompanied by the most diverse set of placement pathways. Microsoft Research interns who do not convert to the research organisation directly have a higher-than-average rate of converting to Microsoft's broader AI engineering org, Azure AI, or the GitHub Copilot product team. The effective "Microsoft group" conversion rate — capturing all eventual Microsoft full-time placements from the intern cohort — sits closer to 74 percent, making the Microsoft Research internship a credible pipeline even for candidates whose research ambitions do not perfectly align with the organisation's current project portfolio.
The Compensation Stack
Intern compensation at frontier labs in 2026 is, on a per-hour basis, higher than the full-time salary of a mid-career engineer at many traditional technology companies. The basic arithmetic is not complicated: at $95 to $115 per hour for a 12-week summer program, an Anthropic or OpenAI intern earns between $57,000 and $69,000 in gross wages before accounting for the housing stipend, travel reimbursement, and compute credits that the leading programs now include as standard. Over 12 weeks, a fully loaded Anthropic internship — wages plus $12,000 housing plus $3,500 travel plus $2,000 in AWS research credits — represents total compensation of approximately $74,000 to $87,000. The hourly wage alone exceeds the median US software engineer salary on an annualised basis.
The compensation arms race among intern programs tracks the broader frontier-lab salary escalation documented in ENTRA's earlier Q2 2026 hiring report. In 2022, the top-of-market intern rate at a frontier lab was approximately $65 to $75 per hour. By summer 2024 it had reached $80 to $90 per hour. Summer 2025 landed at $90 to $115 per hour as Anthropic and OpenAI each moved their floors upward in the weeks following the announcement of their respective funding rounds. The pattern is consistent: each major capital raise by a frontier lab is followed within one to two academic terms by an upward adjustment in intern compensation, because the internal logic of the programs links intern pay to full-time pay through a consistent ratio — typically 70 to 80 percent of the annualised first-year total comp — and full-time pay has risen steadily.
Return-offer packages compound the intern economics further. An Anthropic research engineering intern who converts to full-time receives a first-year total compensation package in the range of $220,000 to $280,000, structured as $180,000 to $210,000 base plus an equity grant of $80,000 to $120,000 per year on a standard four-year vest with a one-year cliff, plus a sign-on of $25,000 to $40,000 that partially compensates for the vesting cliff. OpenAI's new-grad band sits at $185,000 to $225,000 base with equity and sign-on bringing total first-year comp to $230,000 to $295,000 at the median. DeepMind UK's return-offer structure is more compressed on base — £85,000 to £110,000 — but includes Google stock grants and benefits that bring total compensation above £160,000 for research science hires in London, a figure that exceeds the published NHS consultant pay band by a factor of nearly two.
The "ghost offer" problem deserves specific attention. Labs now routinely extend return offers to interns in the final week of the internship — sometimes earlier — with decision windows of 48 to 72 hours. The strategic logic is transparent: an intern who accepts before returning to campus cannot be recruited by a competing lab during the fall recruiting season. An intern who declines and returns to campus becomes available for competing offers. The compressed decision window is not primarily a reflection of operational urgency; it is a mechanism for locking in converted talent before the secondary market can operate. University career offices at MIT, Stanford, CMU, and Oxford have collectively documented 34 instances in academic year 2025-26 where a student reported receiving a return offer with a decision window of 72 hours or less.
The Campus War
The competition for intern cohorts has moved upstream from the traditional summer recruiting timeline. Frontier labs now have dedicated university partnership structures — not career fair booths but embedded relationships with specific faculty, departments, and research groups — that function as feeders into the internship pipeline two to three years before the candidate is ready to convert to full-time.
Anthropic's relationship with MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) has formalised since 2024 into a research collaboration that provides Anthropic engineers access to MIT PhD students as co-researchers, and MIT PhD students access to Anthropic's Constitutional AI research agenda and compute infrastructure. The collaboration is not structured as a recruiting program. It functions as one. Students who co-author work with Anthropic researchers during their PhD have a significantly higher rate of internship offers than the general applicant pool, per ENTRA's placement data.
OpenAI's Stanford partnership follows a similar structure, with OpenAI researchers regularly guest-lecturing in Stanford's AI safety and alignment courses — a soft recruiting mechanism that generates familiarity and, for the students who engage, a relationship with OpenAI researchers before they reach the stage of submitting an internship application. Google DeepMind's Oxford connection predates the frontier-lab era and reflects DeepMind's UK origins; the organisation runs a formal research visitor program at Oxford that places doctoral students in short-term research residencies, which convert to internship offers at a high rate. Meta's CMU relationship is the most explicitly transactional: Meta has contributed $25 million to CMU's School of Computer Science since 2022, a figure disclosed in CMU's donor records and widely understood by both parties to include an informal expectation of preferred access to graduating PhDs and summer intern candidates.
The downstream consequence of these partnerships is a funnel structure that advantages students at a small number of target institutions. MIT, Stanford, CMU, UC Berkeley, Caltech, Oxford, Cambridge, ETH Zurich, and the University of Toronto account for a disproportionate share of frontier-lab intern placements. ENTRA's placement data for summer 2025 shows that these nine institutions produced 61 percent of placed interns at the five labs covered in this report, despite accounting for a fraction of total CS and ML graduate enrollment globally. The concentration is not accidental. It is the output of partnership structures, informal relationship networks, and geographic clustering that compound over time.
The offer deadline conflict is the most acute front in the campus war. The National Association of Colleges and Employers (NACE) guidelines recommend that employers give candidates at least three weeks to consider offers. Frontier labs routinely ignore this guideline for intern return offers, operating on the logic that the guideline applies to full-time offers in the fall recruiting season and that summer return offers occupy a different category. MIT's Career Services office sent letters to Anthropic, OpenAI, and Google in October 2025 requesting compliance with a minimum 14-day window for return offers; the responses, per MIT career office communications reviewed by ENTRA, ranged from non-committal to no reply. Stanford's career office has taken a more accommodating posture toward the labs, reflecting the depth of the institutional relationships involved.
The Excluded Middle
The structural consequence of the internship-first pipeline is a compression of access for the majority of computer science and ML graduates who did not secure a frontier-lab internship. Open new-grad pools — recruiting cycles where labs hire from the general graduate market without a prior internship relationship — now account for approximately 22 percent of new-grad offers at the five labs tracked, down from 55 percent in 2022. The arithmetic for a 2026 CS graduate from a non-target school who did not intern at a frontier lab is unfavourable: 78 percent of spots are reserved for returning interns; of the remaining 22 percent, competition is intense and the application-to-offer conversion rate for open-pool candidates is below 1 percent at the most selective organisations.
The equity dimensions are significant and underexamined. International students face compounding constraints: US visa rules create timing mismatches between internship start dates and international student work authorisation, and the OPT clock that begins running after graduation creates urgency around full-time offers that can push international new-grads toward accepting suboptimal offers rather than waiting for a second recruiting cycle. First-generation university students are less likely to have the informal networks — alumni connections, faculty relationships, early awareness of application timelines — that feeder-school students mobilise to secure junior-year internships. Students at teaching-focused universities without research programs have lower exposure to the faculty collaborations that generate the informal intern referrals that are, in practice, the highest-conversion sourcing channel for the labs.
The labs are aware of these dynamics and have created a set of programs designed to address them — partially. Anthropic's "Interpretability Research" open track accepts applications from candidates without a prior Anthropic internship for a small number of research scientist roles, prioritising demonstrated research output over institutional affiliation. OpenAI's Residency program, which accepts applications on a rolling basis and does not require a prior OpenAI internship, has historically had an acceptance rate of under 2 percent from the open applicant pool; it is a credible pathway but not a scalable one. Google DeepMind's Fellowship program provides research funding and lab access to PhD students at non-partner universities, generating a pipeline that sits outside the standard intern-to-offer track. Meta operates an AI Mentorship program for underrepresented students in STEM.
These programs are not substitutes for a scaled open new-grad pipeline. They are, at best, corrective mechanisms that address a small fraction of the population that the intern-first structure excludes. None of the five labs has announced plans to expand open new-grad hiring to the level it represented in 2022. The direction of travel — more intern-first, tighter pipelines, longer embedded relationships with target schools — runs in the opposite direction.
University Structural Response
The internship concentration dynamic has not gone unnoticed by university administrators and CS faculty, who face growing pressure from students, parents, and ranking agencies to demonstrate graduate employment outcomes at frontier organisations. The response has been to restructure curricula and extracurricular programs around the metrics that predict intern placement.
Stanford's CS department introduced a junior-year "AI Systems Lab" course in 2025 that is co-taught by a rotating set of Google DeepMind and Meta FAIR researchers. The practical consequence is that students who perform well in the course have documented relationships with senior lab researchers before they apply for internships — exactly the informal referral mechanism that the labs' sourcing data shows generates the highest conversion rates. MIT's CSAIL has expanded its Undergraduate Research Opportunities Program (UROP) to include Anthropic and OpenAI as partner organisations for the first time in the program's history, placing undergraduates in lab-supervised research roles during the academic year.
CMU's School of Computer Science restructured its junior capstone requirement in 2025 to allow lab-embedded research projects to substitute for the traditional capstone, a change that was explicitly motivated by CMU's recognition that lab-embedded research experience was becoming a near-prerequisite for summer internship offers at frontier organisations. University of Toronto's Vector Institute, Canada's national AI research organisation, has formalised a matching program that places Canadian undergraduates in research roles at Vector-affiliated labs — a structure designed to replicate the embedded-relationship advantage that Stanford and MIT students access through their direct lab partnerships.
The geographic gap is significant. UK and European universities are, as a cohort, approximately 18 months behind their US counterparts in establishing the embedded research partnerships that generate intern placement advantages. Oxford and Cambridge are the exceptions: both have multi-year relationships with DeepMind that predate the frontier-lab era, and both have embedded research collaborations with newer labs. But Imperial College, UCL, Edinburgh, ETH Zurich, TU Munich, and the other strong European ML programmes are competing for intern placements primarily through formal application cycles rather than the informal referral networks that dominate at US target schools. The result is that European students, on average, face a harder path to frontier-lab internships than their American counterparts with equivalent technical preparation — a structural disadvantage that plays out in placement statistics and, ultimately, in the distribution of frontier-lab new-grad offers by geography.
The 2026 Forecast
The summer 2026 intern cohorts are the largest in frontier-lab history. Anthropic is hosting approximately 150 interns, a 25 percent increase over summer 2025. OpenAI has expanded to approximately 220. The expansion is a function of the labs' own growth: each of the five organisations tracked has added significant headcount in the first half of 2026, and the intern program scales with the research organisation's appetite for evaluated talent. Larger intern cohorts mean more return offers — and more of the 2026 new-grad market locked up before the fall recruiting cycle begins.
Return offer decisions for summer 2026 interns will land primarily between August and October 2026. The students who accept those offers in September and October will be largely off the market before the traditional fall new-grad recruiting season reaches peak volume in November and December. For the open-pool new-grad market — the 22 percent of frontier-lab offers that don't go to returning interns — fall 2026 recruiting will be competitive, fast-moving, and heavily weighted toward candidates who can demonstrate lab-adjacent research experience even if not a direct lab internship.
For students currently in their second year of undergraduate study who are targeting frontier-lab employment at graduation in 2028, the critical path looks different than it did three years ago. The junior-year summer internship is now the pivotal event — not the senior-year offer. Securing a sophomore-year "research exposure" slot, a UROP placement, or a lab-embedded course is the mechanism by which a 2028 targeting student should be building the relationship that generates a junior-year internship application with an informal referral rather than a cold submission. The labs are not advertising this as the dominant pathway because doing so would generate political pressure to open the pipeline. It is, nonetheless, what the data describes.
The longer-term structural risk is homogenisation. If 78 percent of new-grad hires are former interns, and intern cohorts are drawn disproportionately from nine institutions, and those institutions are clustered geographically and socioeconomically, then the graduate pipeline feeding the frontier labs in 2026 is substantially less diverse — in background, in perspective, in problem framing — than the pipeline of 2022. The lab that discovers that the most important alignment insight of the next decade required a perspective that its intern-first pipeline systematically excluded will not be able to recover that ground quickly. The internship economy is an extraordinarily efficient talent acquisition mechanism. Efficiency and diversity are, in this context, in tension. The labs have chosen efficiency. The costs of that choice are not yet visible in the output metrics. They may become visible later.
Methodology
ENTRA Talent Index tracked 1,200 placed candidates across 47 frontier AI lab internship programs between January 2025 and May 2026. Conversion rate is defined as the intern receiving and accepting a full-time offer within 90 days of internship completion. Lab-level intern cohort sizes are ENTRA estimates based on aggregated placement data, LinkedIn announcement tracking, and recruiter conversations; labs did not independently confirm program sizes. Compensation data sourced from offer letters submitted to ENTRA by placed candidates, Levels.fyi cross-referenced submissions (updated through May 21, 2026), and the ENTRA direct-survey panel of 340 2026 interns. Data covers US, UK, and European programs. University partnership descriptions are based on public disclosures, course catalog records, and ENTRA recruiter conversations; none of the universities or labs reviewed this report prior to publication. Historical conversion rate figures (2022: 45%) are derived from ENTRA's retrospective analysis of the same candidate tracking methodology applied to a 2022 cohort subset.