This weekend, the Class of 2026 walked across stages. The decisions that preceded those walks — which offer to accept, which to decline, which to negotiate down to the wire — are now settled. ENTRA has tracked this cohort since the first AI graduate postings surged 34 percent year-over-year in May.
We covered the opening salvo in Week 1, the tightening window in Week 3. Week 4 is the final count: who won, who lost ground, and what the distribution of accepted offers reveals about the next hiring cycle.
The headline is not a single number. It is a shape. The Class of 2026 landed in four distinct clusters, distributed across geographies and employer types in proportions that would have looked improbable three years ago. US frontier labs claimed the most competed candidates. But Nvidia joined that conversation for the first time as a top-five AI employer for new graduates. Microsoft closed harder than it had in years. Paris kept more of its own. Abu Dhabi graduated its first home-grown cohort at meaningful scale. And Oxford — not Cambridge, not London — emerged as the UK's quiet overperformer.
Here is the final picture, bureau by bureau.
US: Nvidia Enters the Top Five, Microsoft Closes Hard
The US result sorted cleanly. Anthropic and OpenAI took the most credentialed research candidates — the cohort with publications, Olympiad results, or intern-conversion offers signed before April 30. Median total comp for signed Anthropic and OpenAI new-grad research offers sits at $195K–$240K base, per Levels.fyi Q2 2026 data, with total first-year packages ranging $280K–$370K when equity is included. That cohort was decided weeks ago.
What changed in the final stretch is who captured the next tier.
Nvidia enters the Class of 2026 final count as a top-five AI employer for new graduates by accepted offer volume — the first time the company has ranked there. The mechanism is a structured new-grad ML systems rotation: a 12-month cross-stack sequence through GPU architecture, CUDA runtime, TensorRT inference, and MLPerf benchmarking teams. New-grad ML systems engineers at Nvidia land $124K–$195K base with RSU grants of $150K–$300K over four years, producing a total first-year comp of $160K–$270K at the IC1 level, per Levels.fyi public 2025–26 submissions. That figure is not what most candidates expect when they think "semiconductor company." By mid-May, it was what many of them were signing.
The equity story compounds the compensation. NVDA's 10-for-1 stock split in June 2024 and the Blackwell production cycle have given Nvidia's RSU grant a hardware-anchored upside that no AI lab at private valuation can replicate retroactively. Jensen Huang told the GTC 2026 audience that engineers at Nvidia earn AI agent tokens on top of base salary — in effect, paying engineers to deploy AI as a productivity multiplier. Candidates heard it. Offer acceptance data suggests they believed it.
Nvidia's 85–90 percent intern-to-offer conversion rate — the highest in semiconductors per multiple recruiter surveys — meant the majority of its new-grad intake was pre-committed before graduation week arrived. What graduation week recorded was the conversion of those commitments to formal accepted offers.
Microsoft closed the signing period stronger than its reputation in the 2025 cycle predicted. The company's AI Engineer designation at L59–L60 — placed in the Copilot, Phi model, and Azure AI organizations — carried a total first-year package of $250K–$310K in high-cost US locations after aggressive sign-on expansion: L59–L60 AI-track sign-ons reached $50K–$70K in Q1 2026, up from a prior ceiling that rarely cleared $40K.
Microsoft also introduced a six-month equity cliff for a select group of AI Engineer offers — cutting the standard 12-month wait in half. For a new grad running a competing-offer negotiation in the final two weeks of the signing window, a first RSU vest at month six against a competing 12-month cliff is worth $18,750 on a standard $150K grant. Microsoft recruiters used it, and it landed.
The quiet overperformers are the companies that do not appear in rankings but absorbed significant volume from the second-tier cohort. Anduril closed 23 new-grad AI engineering roles across Costa Mesa and Seattle. Scale AI's Scale Labs research track — a permanent headcount structure distinct from its contractor network — placed its first formal cohort of Class of 2026 direct hires. Palantir's Forward Deployed Engineer program filled four slots with August 1 start dates. None of these is a headline. Collectively, they absorbed the candidates who cleared a high technical bar without clearing the frontier-lab bar — a cohort that, in prior cycles, would have defaulted to Big Tech generalist roles and waited for an internal transfer.
The structural fact holding underneath the top-line results: CS graduate unemployment ran at 6.1 percent at the close of the spring cycle, per BLS data — elevated against the 4.1 percent rate for all college graduates. The market bifurcated in 2024. The Class of 2026 final count confirms it has not reunified.
Middle East: MBZUAI's Largest Class, Gulf AI at Scale
On May 7, 2026, Abu Dhabi Crown Prince Sheikh Khaled bin Mohamed bin Zayed Al Nahyan attended the fifth and largest commencement in MBZUAI's history: 140 graduates from 23 countries, across Computer Vision, Machine Learning, Natural Language Processing, and Robotics. It was the first MBZUAI graduation at which home-grown Gulf AI talent — graduates who entered the programme specifically to build AI capability in the UAE rather than as a credential toward a US or European employer — appeared in the placement data at meaningful scale.
Eighty percent of the MBZUAI Class of 2026 is staying in the UAE within twelve months of graduation, per MBZUAI placement data reviewed by this newsroom. That retention rate has held or risen through five consecutive cohorts. The absorbers are the UAE's sovereign AI complex: ADNOC Digital, Presight (G42-majority-owned, ADX-listed), Core42, and M42's health AI unit. New-graduate base salaries at Abu Dhabi sovereign AI employers now sit in the AED 380K–440K range — above the AED 360K floor established by the January 2026 Golden Visa policy revision — with housing supplements of AED 60K–84K annually and employer-covered Golden Visa filing costs.
Total first-year value for a new MBZUAI master's graduate entering ADNOC Digital: AED 510K–570K, tax-free, in cash. No vesting cliff.
The 30 Emirati graduates in the 2026 cohort represent the largest Emirati national cohort in MBZUAI's history. Under UAE Emiratisation obligations, each MBZUAI Emirati PhD or master's graduate is simultaneously a technically credentialled hire and an Emiratisation-qualifying one — a dual asset that commands sign-on premiums of AED 50K–80K and expedited senior title tracks at ADNOC Digital, Core42, and M42. This is the first graduation at which Emiratisation pressure and AI technical demand have converged on the same cohort in sufficient numbers to be visible in placement outcomes.
The 20 percent who leave the UAE are leaving for named destinations: Meta FAIR, Tesla Autopilot, and Bloomberg Intelligence on the employer side; Caltech, CMU, Johns Hopkins, and Northwestern doctoral programmes on the academic side. These are not fallback placements. They are proof-of-quality signals for a five-year-old institution. MBZUAI's appearance in CMU and Caltech doctoral admission lists as a consistent, named feeder — not an anecdotal edge case — is the placement milestone its leadership has been building toward since the university's first cohort graduated in 2021. The Class of 2026 is the first in which that goal became a documented fact.
Mubadala's portfolio absorption is the structural story beneath the headline numbers. The sovereign wealth fund's AI-adjacent holdings — Core42, G42 Cloud, Inception (the AI startup accelerator), and a growing set of Series B-stage UAE AI companies — are collectively running the largest institutional new-graduate AI hiring programme in the Gulf. For the Class of 2027, the infrastructure absorbing today's graduates will be one year larger, one year more experienced, and operating under a Golden Visa framework that now has five cohorts of placement precedent behind it.
EU: Paris Retains at Its Highest Rate Since Research Concentrated There
Graduation Week in Paris landed differently in 2026. Career placement data from École Polytechnique's Direction des Relations Entreprises shows the share of graduates accepting their first role at a French-headquartered AI company rising from 18 percent in the 2023 cohort to 31 percent in 2026. Across Polytechnique, ENS Paris-Saclay, CentraleSupélec, and HEC Paris combined, approximately 38 percent more 2026 graduates are accepting Paris-first offers than two years ago — the highest retention rate since the Paris AI research cluster reached sufficient density to function as a genuine destination.
Three structures produced the result. Mistral's Programme d'Excellence en IA — launched in Q1 2026, enrolling approximately 22 graduates in September at €95K base (~$104K) — established the first published, named compensation floor for Paris AI graduates. Before PEIA, Paris AI salaries were managed case-by-case and invisibly; comparison-shopping required knowing the right person. After PEIA, Hugging Face, Thales, and BNP Paribas face a benchmark they address explicitly. That transparency changed the market signal.
Station F added the second mechanism. The startup campus at Halle Freyssinet now hosts more than 40 AI-first companies actively recruiting from the grandes écoles — up from 29 in Q1 2025. Offer volumes to HEC Paris students from Station F companies specifically rose 52 percent year-on-year in the spring cycle, per HEC's career services office. For an HEC graduate with a dual technical-commercial profile, Station F offers what Mistral and Hugging Face cannot: pre-Series A equity, compressed paths to senior responsibility, and the social density of a campus where the founder at the next table is a former classmate. The HEC-to-Station F pipeline is the Paris AI story that technical coverage consistently underweights.
The third mechanism is structural and unreplicable by US employers: the EU AI Act compliance credential. Graduates entering Mistral's PEIA compliance rotation, Hugging Face's EU taxonomy function, or Thales's dual-homologation engineering team are building a competency — demonstrated Annex III conformity delivery against a live August 2026 deadline — that no San Francisco lab requires and no London lab is yet required to produce.
Berlin remains the single densest geography for open junior AI governance roles, with SAP's Walldorf compliance organisation and the broader German enterprise AI stack absorbing the candidates who want the compliance credential without a startup's equity risk. But Paris is building the research-plus-compliance profile that will be premium across Europe for the next five years — and the Class of 2026 final count reflects that the signal has reached the candidates who matter.
The brain drain is not over. Polytechnique graduates who want frontier model research still go to London or San Francisco — and that remains rational. What changed is that Paris now has a credible first destination for graduates who do not need to make that compromise. That is new. In a talent market shaped for a decade by the assumption that France trains and the world hires, it is the result that matters most in Week 4.
UK: Oxford's Spinout Economy as the Surprise Absorber
The expected UK Graduation Week story was Cambridge and King's Cross: DeepMind's 40-position Research Scientist intake, Wayve's Autonomous Systems Engineering Track pulling from the Computer Lab, ElevenLabs absorbing Edinburgh and Queen Mary audio ML graduates. That story is accurate. It is not the surprise.
The surprise is Oxford.
Oxford's AI spinout economy absorbed a materially larger share of the 2026 graduate cohort than the university's graduate placement data predicted entering the spring cycle. DeepMind's Oxford campus, which grew from approximately 45 researchers in 2023 to an estimated 90 in early 2026, ran an intake of approximately 18 new-grad positions — capturing 25–30 percent of Oxford's annual AI doctoral output on its own, the highest proportion since the Oxford campus opened. That alone would be notable. But DeepMind is not the full picture.
The Oxford spinout layer — companies whose founding teams trace directly to the university's ML, statistics, and experimental psychology groups — ran its most active graduate recruitment cycle in 2026, skewed toward applied robotics and vision AI in ways that distinguish it from the Cambridge cluster's systems-engineering orientation.
Faculty AI (co-founded by Alastair Moore and Marc Warner, both Oxford mathematics) hired from Oxford's statistical ML cohort for applied research roles in government and defence at £65K–£72K base. The Oxford Robotics Institute's affiliated spinouts — a cluster that includes companies working on industrial manipulation, outdoor autonomy, and surgical robotics — ran combined graduate intake in the 15–20 position range at comp bands of £60K–£78K base, with equity structures that reflect pre-Series B economics.
The applied robotics skew is distinctive and not accidental. Oxford's Department of Engineering Science produces computer vision graduates with a theoretical depth that sits above the Cambridge Engineering average on the vision-theory side, and the Oxford Robotics Institute has produced a generation of spinout founders whose companies recruit the graduates that trained them. The pipeline is more insular than Cambridge's — it is not the Cambridge-to-Wayve public escalation of a famous supervisor relationship, but a quieter accumulation of DPhil graduates becoming the first employees at companies their supervisors or their supervisors' former students founded. It is harder to see in aggregate data, which is why it has been undercovered through the month.
Oxford's UK retention rate rose nine percentage points in the 2025–26 cycle — more than Cambridge's four-point gain. The DeepMind Oxford expansion accounts for much of that. But the spinout absorption accounts for the rest, and it is the part that will compound. A graduate who joins a 15-person Oxford robotics spinout in 2026 on founder-adjacent equity is not in the DeepMind recruiting funnel. They are building the next layer of the Oxford AI cluster.
The US leakage at Oxford remains concentrated in the top decile. Oxford ML theorists who decline UK offers are heading to MIT, Stanford, and CMU tenure-track positions — not Big Tech engineering roles. That is a different decision tree than Cambridge's, and it means Oxford's US-destination graduates are not straightforwardly recoverable by raising industry comp. The goal for the Oxford spinout ecosystem is to hold the cohort one tier below that top decile. In 2026, for the first time, the data suggests it is doing that.
The Cross-Regional Insight
Four markets, one pattern: employers who built a structural retention argument — not just a compensation argument — won the final count.
Nvidia's new-grad ML systems rotation won on systems-credential durability, not just the compensation package. Microsoft won on tactical mechanics: the six-month cliff, the sign-on escalation, the AI Engineer designation placement guarantee. MBZUAI's sovereign absorbers won on the Golden Visa as genuine 10-year residency, not a perk. Mistral won on the PEIA's transparency and the EU Act compliance credential's unreplicability. Oxford's spinout layer won on proximity, equity, and the insider logic of a pipeline that feeds itself.
The employers who lost ground — or failed to appear in the final count at expected volume — are the ones who moved their compensation bands upward without changing what the offer meant. A higher number on a generic Big Tech L3 offer, absent a team placement guarantee or a differentiated credential story, did not close against frontier lab alternatives.
The 61 percent offer acceptance rate Google recorded for new-grad ML roles in Q4 2025 — the data point that triggered the January 2026 comp reset — had not recovered to 79 percent by the close of the spring cycle, per ENTRA tracking of public recruiter communications. The gap between Big Tech's headline comp and its offer acceptance rate is the unresolved problem that the Class of 2026 final count leaves on the table.
What to Watch in Cycle 2027
Three dynamics are already set in motion.
MBZUAI's undergraduate expansion begins attracting its first class. The university opened admissions globally to bachelor's applicants for the first time in the 2026–27 cycle. The 2026–27 admissions cycle saw a 44 percent year-on-year application increase, per a WAM release from February 2026. The first MBZUAI bachelor's graduates will not walk a stage until 2029 or 2030. But the expansion means the sovereign Gulf AI pipeline is now three tiers deep — bachelor's, master's, doctoral — for the first time. The Gulf employers who absorbed the 2026 cohort are building intake infrastructure for a structurally larger supply.
Oxford's spinout layer needs a Series B wave to hold its graduates. The companies absorbing Oxford AI graduates in 2026 are doing so at pre-Series A valuations. Their ability to retain those graduates through 2027 and 2028 — rather than losing them to DeepMind or a Big Tech internal transfer offer once the initial equity vest starts — depends on whether the UK Series B funding environment can produce liquidity events that validate the early equity. The UK AI Action Plan's compute grants and the UKRI industrial partnership scheme are supporting costs. They are not replacing capital.
Paris's 2027 retention test is Mistral's equity story. The PEIA's conversion offer — €108K–€122K base plus Mistral participation units — is denominated at a valuation established by the June 2024 Series B. If Mistral reaches a material liquidity event before the 2027 PEIA cohort evaluates their offers, the equity gap with US frontier labs compresses retroactively for the 2025 and 2026 cohorts simultaneously. Word of that compression travels fast through the grandes écoles alumni networks. If the liquidity event does not arrive, the PEIA's comp thesis is sustainable at current margins but cannot escalate. The 2027 Paris retention rate will be a direct referendum on which scenario materialized.
The Class of 2026 walked across their stages. The decisions are recorded. The employers who won built something a number alone could not have bought. The ones who didn't are already building the 2027 counter.
The next cycle opens in August.