Fifty-eight percent of OpenAI's active job postings in Q3 2026 require physical presence in San Francisco — the highest SF-concentration ratio among frontier AI labs (ENTRA Job Signal Index, Q3 2026). The company has leased over one million square feet across Mission Bay, plans to nearly double its workforce from 4,500 to 8,000 by year-end, and runs a policy that gives new hires 45 days to relocate. It is the most explicit in-person bet in frontier AI, and it is costing OpenAI candidates at the top of the funnel every quarter it holds.
The In-Person Bet
OpenAI's office mandate has hardened since 2023. The current policy — three mandatory days in office Monday through Wednesday, with optional remote on Thursday and Friday — sits well above the industry median for technical roles. Sam Altman publicly called remote work "a mistake" when sealing a major Mission Bay lease in late 2023 (Fortune, October 2023). By March 2026, the company had surpassed one million square feet of leased space across three Mission Bay addresses: approximately 486,600 sq ft at 1455 and 1515 Third Street, roughly 330,000 sq ft at 550 Terry Francois Boulevard, and approximately 280,000 sq ft at the former Dropbox headquarters at 1800 Owens Street (The Real Deal San Francisco, March 2026).
The practical consequence is a funnel that starts in San Francisco and, for the most sought-after roles — model research, alignment, post-training, and product engineering — rarely exits it. Of 681 OpenAI job postings analyzed for Q3 2026, only 28 (4%) are listed as remote-eligible (jobsbyculture.com analysis, Q3 2026). The ENTRA Job Signal Index narrows that further: 58% of all active postings carry an explicit San Francisco city requirement rather than a generic "Bay Area" or "hybrid" designation. The gap between 4% remote and 58% hard-SF is not semantic. It reflects a deliberate funnel architecture — you can work from home two days a week, but you cannot work from Austin.
The Comp Math
OpenAI's in-person premium is real, and it is competitive at the base-salary level. ENTRA compensation tracking for Q3 2026 places the L5 Machine Learning Engineer band at $220K–$310K base in San Francisco. L7 Research Scientist roles run $340K–$450K+ base. Those figures track with Levels.fyi self-reported data through Q2 2026, which shows OpenAI L5 software engineers at a median total compensation of $941K — a base around $280K–$336K plus equity accrual (Levels.fyi, data as of September 1, 2026).
The equity story is where the math gets complicated. OpenAI's PPU (Profit Participation Unit) structure — unique among frontier labs — ties equity value to company profit distributions rather than a conventional RSU grant against a fixed share price. With OpenAI's implied secondary-market valuation tracking well above the $157B post-money close from October 2024 (Wall Street Journal tender-offer reporting, 2025), PPU upside is real for senior holders. But it is not as legible to candidates as Anthropic's RSU structure priced against a published round valuation, or as Google's exchange-traded stock. Candidates from outside Silicon Valley, particularly those being asked to relocate, face a harder time stress-testing the offer. That information friction shows in the acceptance data.
For a distributed alternative to calibrate against: a senior ML engineer working remotely for a Series B AI startup out of New York or Seattle in the same role-weight typically clears $200K–$270K base with conventional RSU grants (ENTRA Talent Hub, Q2 2026 recruiter survey, n=178 placements). The OpenAI base premium for SF presence runs roughly 15–20% above that distributed benchmark. Whether a candidate prices the 45-day relocation window into that arithmetic is the variable OpenAI cannot fully control.
The Filter Effect
OpenAI's in-person requirement filters candidate volume at the top of the funnel — and the filter is not evenly distributed across seniority levels.
For entry-level and mid-level roles (L3–L5), the SF mandate narrows the addressable candidate pool by an estimated 30–40%, based on ENTRA's geographic distribution analysis of approximately 54,000 US-based ML engineer LinkedIn profiles filtered to profiles meeting a minimum ML degree or equivalent experience threshold and three or more years of relevant experience, matched against OpenAI's published technical requirements as of Q3 2026 (ENTRA LinkedIn talent mapping, Q3 2026; n=~54,000 US ML engineer profiles). The majority of US-based ML engineers at those levels live outside San Francisco metro: Austin, Seattle, New York, Boston, and distributed workers across smaller metros collectively account for roughly 63% of the relevant labor supply.
Senior and principal-level candidates (L6–L7 Research Scientists, Principal Engineers) show a different pattern. These candidates are more likely to have SF connections, more likely to have already relocated once for a prior lab role, and more likely to price relocation as a solvable problem. The attrition at the top of the senior funnel is narrower in percentage terms — but each declined offer at L7 costs OpenAI a candidate who likely received simultaneous interest from Anthropic or Google DeepMind.
Two talent operators with direct knowledge of frontier lab recruiting, speaking on background, characterized the Q2 2026 acceptance rate for SF-only senior research offers as "materially below" the rate for hybrid-flexible offers at comparable compensation. Neither would attach a specific number on record. The structural signal is visible in a parallel data point: OpenAI's two-year retention rate, at 67%, trails Anthropic's 80% (SignalFire 2025 State of Talent report) — suggesting that the in-person filter affects not just offer acceptance but tenure once candidates arrive.
What This Means
The in-person lab thesis has three experiments running simultaneously in 2026.
OpenAI is the clearest execution of the thesis: dense Mission Bay campus, strong in-person mandate, premium compensation for San Francisco presence. The bet is that proximity produces collaboration velocity that distributed teams cannot match — and that the research pace advantage compounds into model quality over multi-year cycles.
xAI's Memphis campus is the non-SF variant. With an in-office-only posting posture for engineering roles at its Colossus supercomputer facility (ENTRA Job Signal Index Q3 2026), xAI has demonstrated that in-person density can be assembled outside the Bay Area at lower real-estate cost and with a regional labor market that carries less price competition from other frontier labs. Memphis is not competing for the same candidate pool as Mission Bay. That is the point.
Anthropic runs the calibrated middle. With 8% of its 411 open roles designated remote-eligible as of July 2026 (ENTRA tracking, Anthropic Greenhouse board, July 9, 2026), and its remaining roles anchored in San Francisco's SoMa corridor, Anthropic imposes roughly equivalent in-person expectations without requiring the hard 45-day relocation window or the explicit Mon–Wed attendance language. The policy difference is marginal. The perceived difference — candidate-side, in offer conversations — is not. Anthropic's 80% two-year retention rate suggests that candidates who accept the SF bet there price it differently than those who accept it at OpenAI.
Google DeepMind's King's Cross campus in London offers the distributed-flagship counterargument: world-class research output from a non-SF, non-collocated node. DeepMind has never required US consolidation and maintains a strong two-year retention rate. It complicates any clean claim that in-person SF density is a necessary condition for frontier research.
ENTRA's read: OpenAI is making a deliberate bet on collaboration speed over talent access breadth. The bet is internally coherent — Mission Bay gives OpenAI's post-training and safety teams the kind of ambient information density that Slack channels do not replicate. The cost is visible in the funnel: a 30–40% addressable pool compression at mid-levels, a meaningful percentage of senior candidates who price relocation as a dealbreaker, and a retention rate that trails its nearest competitor by 13 points.
Forecast
As the RTO Wars intensify through Q4 2026, OpenAI's talent position will face a specific stress test: the Class of 2027 recruiting cycle opens in September and October, and the labs competing for the same PhD and MS candidates — Anthropic, Google DeepMind, and a structuring xAI campus program — all offer some form of hybrid flexibility that OpenAI's Mon–Wed mandate does not. If OpenAI's planned expansion to 8,000 employees proceeds on the current timeline, the company will need to run its SF-anchored funnel at a scale it has not previously attempted. The mathematical pressure on acceptance rates grows proportionally with the headcount target. Watch whether OpenAI modifies the 45-day relocation window or introduces targeted remote tracks in non-research functions — legal, policy, finance — as a release valve before Q1 2027 recruiting peaks. The in-person bet is not in question. The pressure test is whether the funnel can sustain the volume the growth plan requires.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.