Anthropic CPO Mike Krieger said it plainly in May 2025: "We haven't had a summer internship program so we've tended less to hire the like kind of fresh college grads." That sentence, from an executive whose company just crossed a $61.5 billion valuation, carries more weight than any recruiter call when a 2026 CS graduate is evaluating an AI lab offer versus a Big Tech rotational slot.
The fork is real. The two paths have materially different comp structures, promotion mechanics, equity risk profiles, and — most consequentially — what they produce in the engineer you become by 2029. Neither answer is universally correct. The data, however, is specific enough to make an informed call rather than a vibes-based one.
The Compensation Table
The first-year total compensation gap between frontier labs and Big Tech entry-level is narrower than the narrative suggests — $60K to $100K at the median. The equity gap in years two through four is where the real divergence lives.
| Company | Role | Year-1 Base | Year-1 Total Comp | Equity Structure | Equity Liquidity | |---|---|---|---|---|---| | Anthropic | MTS (L3 equiv.) | $200K–$240K | $290K–$441K | RSU, 4-yr vest, 1-yr cliff | Illiquid; secondary market only | | OpenAI | SWE L2 | $200K–$240K | $249K–$350K | PPU, 4-yr vest, 2-yr lock | Illiquid; 10x growth cap | | Google | SWE L3 | $138K–$165K | $186K–$230K | RSU, 4-yr vest, 1-yr cliff | Liquid (GOOGL) | | Meta | SWE E3 | $130K–$145K | $183K–$220K | RSU, 4-yr vest, 1-yr cliff | Liquid (META) | | Microsoft | SDE 1 (L59) | $120K–$150K | $180K–$300K | RSU, 4-yr vest, 1-yr cliff | Liquid (MSFT) | | Amazon | SDE I | $127K–$185K | $179K–$211K | RSU, 4-yr vest, back-loaded | Liquid (AMZN) |
Sources: Levels.fyi verified submissions (May 2026), 6figr aggregated data, Glassdoor salary reports, ENTRA cross-reference.
Three things that table does not show.
OpenAI's PPU structure carries a 2-year lock and a 10x growth cap. A new grad joining in June 2026 cannot sell any PPUs before June 2028, regardless of tender offer activity. The 10x cap means that if OpenAI's implied valuation at exit is 15x the grant price, the overage is forfeited. That is not a reason to decline; it is a reason to model the scenario before signing.
Anthropic's RSUs are priced against a $61.5 billion private valuation. If the IPO values the company below that — or the timing slips past 2028 — year-three equity events become unpredictable. Secondary market access exists but is restricted and episodic. Google RSUs settled in dollars on Nasdaq the day they vest.
Amazon's RSU schedule is back-loaded by design: 5% vest in year one, 15% in year two, 40% in year three, 40% in year four. A new grad who leaves after 18 months leaves with almost nothing in equity. This is not a flaw Amazon hides; it is the published vesting schedule. It is also the primary reason Amazon's advertised total comp figures look comparable to peers on year-one but diverge dramatically on a four-year view.
The Career Trajectory Fork
Compensation is year one. Career architecture is the three-to-five year question.
Big Tech path: structured, benchmarked, promoted by committee. Google L3 to L4 currently takes 24 to 36 months under the GRAD performance system. Meta's E3-to-E4 transition has a harder clock — Meta's performance review system historically required E3s to promote within 24 months or face performance management — but placement in a high-visibility team (FAIR, GenAI) versus a maintenance team is not guaranteed at offer. Microsoft SDE 1 to SDE 2 (L61) runs roughly 18 to 24 months with a structured review process and a defined set of competencies that promotion candidates must evidence. Amazon SDE I (L4) to SDE II (L5) typically takes 24 to 30 months, with the timeline subject to manager discretion more than at the other three.
The common thread: at all four companies, a new grad knows the rungs, knows approximately how long each rung takes, and has a calibrated external benchmark to compare against. The career ladder is legible. That has real value for people who need it.
AI lab path: ad hoc, fast for standouts, opaque for everyone else. OpenAI runs promotions "whenever you're already operating at the next level" — not on annual cycles. A standout performer can move from L2 to L3 in 14 months. A solid-but-not-exceptional performer may wait 36 months with no clear signal on timing. Anthropic's internal leveling structure is less publicly documented than any of the Big Tech comparators. The upside of this system is velocity for the top decile. The cost is opacity for the other 90%.
This is the part that Krieger's quote illuminates. Anthropic does not run new-grad cohort programs. It does not have a structured 18-month rotational track. When it does hire a new grad — which it does, on an exceptional basis — that person is expected to function at near-senior velocity from week four. If that sounds exhilarating, you are the candidate they are describing. If it sounds like being dropped in the deep end without a lane marker, Big Tech's rotational architecture was designed for you, and that is not a consolation prize.
The Structural Programs: What Big Tech Actually Offers
Google's Associate Product Manager program received over 8,000 applications for its 2026 cohort and extended roughly 40 to 45 offers — a sub-0.6% acceptance rate. The program runs two years, two rotations, with AI-adjacent teams (DeepMind, Gemini, Google Cloud AI) as the most competitive internal placements. First-year total compensation for a Google APM sits at $188,000 to $295,000 depending on equity negotiation, per Levels.fyi APM salary submissions through early 2026.
Meta's Rotational AI Science and Engineering program (RAISE) is the company's explicit new-grad on-ramp for AI roles. Candidates rotate across applied AI, research, and infrastructure teams over 18 months. RAISE places participants within Meta AI Research and Fundamental AI Research units — the teams building Llama, ImageBind, and Meta AI — making it one of the only Big Tech pathways that delivers research adjacency comparable to a frontier lab in year one.
Microsoft's AI Development Acceleration Program (MAIDAP), based at the New England R&D Center in Cambridge, targets BS/MS/PhD new grads for rotations across AI product lines. The explicit value proposition is deployment scale — models shipped inside Microsoft 365, Azure, and Copilot reach hundreds of millions of users in weeks, a number no frontier lab external product currently matches.
These programs are not consolation tracks. They are structured bets that prioritize breadth, mentorship infrastructure, and legible promotion over ceiling upside. For a new grad who wants to understand how AI gets built, tested, and shipped at scale — not just how it gets trained — these programs deliver something the labs do not.
The Decision Framework
Four questions. Answer them honestly before signing anything.
1. Do you need liquidity in years two and three?
If student debt, family obligations, or a cost-of-living situation in San Francisco requires actual cash — not paper equity — before 2028, lab equity is the wrong asset to anchor on. A Meta E3 RSU vests quarterly into publicly traded shares 12 months after grant. An Anthropic RSU requires an IPO or secondary market event to realize. The base salaries at labs are higher, but if total comp calculations are driving the decision, strip out the illiquid equity and re-run the numbers. On base salary alone, the gap between Anthropic and Google L3 is roughly $35K–$75K in base — real, but not decisive.
2. Do you have a specific technical depth you want to build?
Labs optimize for people who know exactly what problem they want to work on. Krieger's description of the Anthropic ideal hire: someone "defined more about the problems that they want to solve and how they can creatively solve them than a very specific 'I know JavaScript.'" If you have a thesis — mechanistic interpretability, inference efficiency, RLHF alignment — and you want to be within one org layer of the people running those experiments, a lab role is the more direct path. If you want to learn what problems exist at scale before committing to one, a rotational program at Google or Meta is structurally designed for that.
3. What is your risk tolerance for organizational chaos?
AI labs in 2026 are operating at a pace that produces organizational flux on a four-week cycle. Teams restructure around model releases. Priorities reset when evals do. Product timelines compress or extend based on benchmarks that did not exist the prior quarter. For some engineers, this is the most alive they have felt in a professional context. For others, it is a recipe for 18 months of ambiguity followed by burnout. Big Tech rotational programs run on calendars. Labs run on models.
4. Is the specific company your signal, or is the sector?
If the answer is "I want to work at Anthropic" or "I want to work at OpenAI" — that is a specific thesis and worth pursuing at the cost of structured programming. If the answer is "I want to work in AI," every company in the table above is an AI company in 2026. Google DeepMind is running frontier model research. Microsoft's Azure AI team is shipping at a scale that makes "Big Tech" vs. "AI lab" a category distinction that does not survive contact with the actual work.
What to Watch Through Labor Day 2026
Three inflection points that will clarify the choice for the Fall 2026 cycle.
First, whether Anthropic opens a structured new-grad track. The Fellows Program — a paid research fellowship with mentorship — runs through 2026 and has produced full-time conversions, but it is not a formal employment program. If Anthropic formalizes a cohort hire for the Class of 2027, it signals that Krieger's constraint is loosening as headcount scales past 3,000.
Second, whether OpenAI's structural shift to a public-benefit corporation alters PPU mechanics. The conversion process, ongoing through 2026, has implications for how PPUs are taxed, valued, and potentially liquidated. New grads signing PPU grants before the conversion finalizes are accepting a contract that may look different 18 months into the vest.
Third, whether Meta's June headcount picture — following the May 20, 2026 layoff effective date for the 8,000 positions announced in April — confirms that university pipeline offers are protected. If even a handful of new-grad offers are reneged in the June window, the 2027 recruiting cycle will feel it. Labs would gain ground on org-stability messaging they currently cannot credibly make.
The career fork for the Class of 2026 is not between "prestigious" and "less prestigious." It is between structured and unstructured; liquid and illiquid; defined and emergent. The 2026 graduates who want to build AI will find a path on either side. The ones who get it wrong are the ones who chose a logo when they should have chosen an architecture.
Pick the structure that matches who you are, not the brand that sounds best at dinner.