On January 13, 2025, Meta enforced a full five-day return-to-office requirement for its AI and machine learning divisions across the Menlo Park headquarters and the 30 Hudson Yards office in New York — the same week the company announced it would cut approximately 3,600 employees in a workforce reduction Mark Zuckerberg publicly framed as a performance-based reset (ENTRA estimate from public WARN notice filings and job board tracking; Meta internal memo, VP HR Janelle Gale, February 7, 2025). For Meta AI's mid-to-senior researchers and engineers, the timing was not subtle: come back five days a week, or leave alongside the colleagues being shown the door. Hundreds chose a third option. By Q2 2026, ENTRA's recruiter network was tracking a secondary talent market operating at scale — former Meta AI engineers landing at Anthropic, xAI, Mistral's San Francisco office, and boutique inference labs at a 12–18% compensation premium above their Meta ceiling bands (ENTRA Q2 2026 Salary Survey, n=619 combined US and international respondents). The Menlo Park corridor is no longer the dominant gravitational center it held through 2023.
The mandate landscape
Meta's five-day enforcement for its AI divisions built on a companywide RTO posture the company had been tightening since 2023. The broader Meta corporate minimum — three days per week — was supplemented by a five-day requirement pushed into specific AI engineering and product teams. Instagram formalized that requirement company-wide in December 2025 (CNBC, December 1, 2025), but for FAIR and the Meta Superintelligence Labs org, the pressure arrived earlier and with less flexibility. Engineers in those divisions describe a January 2025 policy directive that eliminated the remote-first arrangements that had survived the 2022 workforce normalization period.
The layoffs amplified the mandate's effect. Meta's FY2025 reduction of approximately 3,600 roles hit engineering functions disproportionately, with AI and ML teams accounting for a higher share of eliminated positions relative to their headcount weight inside the company (ENTRA estimate from public job board tracking). Within FAIR specifically, approximately 600 roles were cut in October 2025, coinciding with the departure of chief AI scientist Yann LeCun (RD World Online, October 2025). The result: thousands of mid-to-senior technical staff simultaneously navigating a five-day office requirement they had not agreed to and a company signaling that high-performing engineers were not insulated from the next reduction.
One former Meta ML engineer, speaking to a recruiter in ENTRA's network on background, described the calculus: "The offer from the other lab was almost the same base, the equity looked better, and I could work remotely two days a week. There was no math where staying made sense."
Forbes characterized the broader pattern in an August 2025 analysis titled "Zuckerberg Squandered His AI Talent, Now He's Spending Billions to Replace It" — a framing that has held as the secondary market data accumulates through the first three quarters of 2026.
Where the talent is going
Anthropic is the primary absorber. The company's open roles — 411 as of ENTRA's July 2026 tracking pull — skew senior, and its recruiter network has been particularly active in Menlo Park and the South Bay geography where Meta's AI engineering bench concentrates. ENTRA's Q2 2026 Salary Survey (n=619 combined US and international respondents) documents the 12–18% compensation premium Anthropic pays to close Meta AI candidates above their Meta ceiling bands.
The math on that premium is not abstract. Meta AI's Staff ML Engineer bands run $180K–$280K base (Levels.fyi, Q2 2026). An 18% uplift on a $280K ceiling positions Anthropic's initial offers for Meta Staff engineers in the $305K–$330K base range — before equity, which at Anthropic is priced against a May 2026 Series H round that positioned the company as the most highly valued private AI lab globally by post-money valuation. For a researcher who holds unvested Meta RSUs against a company with a deteriorating internal talent brand, the switch is not financially difficult to model.
xAI's absorption is narrower but targeted. The lab has recruited specifically from Meta's RLHF and fine-tuning teams — the engineers who built training pipelines for Llama 3 and its successors — rather than broad lateral hiring from the Meta bench. ENTRA's recruiter network logged at least 14 confirmed Meta-to-xAI moves in H1 2026, concentrated in inference optimization and post-training roles. xAI's Memphis campus offers in-person density at lower real-estate cost, and its SF-based roles carry remote flexibility that Meta's divisions have curtailed.
Mistral's San Francisco office and a cluster of boutique inference labs — including Nous Research, Together AI, and several sub-50-person models-as-a-service operations — account for the documented remainder. These organizations cannot match Anthropic or xAI on total compensation, but they offer equity structures that are legible in a way that late-stage RSUs are not, and technical mandates narrow enough that a former Meta researcher can see their specific contribution in production within weeks rather than quarters.
The Bay Area geography shift
The SF Bay Area holds approximately 18,000 ML and AI engineers, per ENTRA's Job Signal Index (Q3 2026). That pool is not shrinking — net inflows from academic pipelines and international relocation continue — but its internal geography is shifting. Menlo Park's share of that pool has contracted relative to San Francisco proper, where Anthropic's SoMa corridor expansion and OpenAI's Mission Bay footprint have added density faster than Menlo Park's own lab ecosystem has recovered post-Meta cuts.
The Hudson Yards angle is specific. Meta's New York AI engineering presence at 30 Hudson Yards had been expanding through 2024 as the company positioned the office as an East Coast hub for applied AI and business AI product development. The five-day enforcement hit the New York cohort with particular force, because the alternative options are proximate: Anthropic's NYC hiring footprint, a dense cluster of fintech-adjacent AI roles in Midtown and the financial district, and xAI's continued willingness to hire remotely for non-Memphis engineering roles. ENTRA estimates the NYC-based Meta AI engineering cohort saw attrition running approximately 1.4x the Menlo Park rate in H1 2026, based on LinkedIn departure signals from ENTRA's talent mapping (ENTRA LinkedIn talent mapping, H1 2026; n=approximately 3,200 tracked Meta AI engineers, US-based).
The broader Bay Area signal: 65% of active US AI engineers remain concentrated in the SF Bay Area and New York metro (KORE1 AI/ML Talent Map, 2026). But the Menlo Park sub-cluster — which once anchored that concentration through Meta FAIR's research gravity and the Sand Hill Road corridor — is losing density to SoMa and Mission Bay faster than any comparable 18-month period since the 2016-2017 Googleplex expansion. The geography is moving north and west.
The 2026-2027 forecast
Three data points determine whether this talent redistribution stabilizes or accelerates through H1 2027.
Meta's counter-bid capacity. Meta has demonstrated it can write large individual packages when retention is a priority — when it hired five founding members of Mira Murati's Thinking Machines Lab, Andrew Tulloch's reported $1.5 billion over six years (The Next Web, 2026; Calcalist reporting) set the sector ceiling for external acquisition. But those packages targeted external names, not internal retention. If Meta's Q4 2026 compensation cycle includes above-band equity refreshers for AI division Staff and Senior Staff engineers, the departure rate should slow. If it does not, ENTRA expects the Anthropic premium for Meta AI alumni to widen from 12–18% to 20–25% by Q2 2027 as the supply of recent, high-quality Meta AI alumni grows thin and placement premiums rise.
The five-day policy flex point. Meta's five-day AI division enforcement has no public modification language. The Instagram-wide extension in December 2025 signals the company's direction is toward rigidity, not relaxation. But enforcement inside FAIR — which draws research talent from an academic community with strong remote norms — requires a different retention architecture than enforcement in a product engineering org. Watch for any FAIR-specific carve-outs or off-campus research pilot programs announced through Q1 2027. An absence of such carve-outs will read as an organizational signal that Meta is content to replace departed research talent rather than retain it.
Anthropic's absorption ceiling. At 411 open roles and a hiring velocity running at +370% year-over-year (ENTRA Talent Index, data as of July 9, 2026), Anthropic can absorb meaningful Meta outflow without visible capacity strain. But the company's one-week-per-month in-office model — while flexible by Big Tech standards — is not fully remote, and its SoMa corridor real estate commitments will put gradual pressure on the policy through 2027. If Anthropic moves its in-office minimum toward two weeks per month in early 2027, it partially closes the flexibility gap that currently makes it the natural landing zone for Menlo Park engineers who want out of the five-day grind. That policy shift would redirect some of the Meta AI outflow toward xAI and the inference lab cluster instead.
The 18% premium is the current clearing price for Meta AI talent at Anthropic. Whether it compresses, holds, or widens through 2027 is the single best real-time indicator of whether Meta's five-day office bet is working — or whether Menlo Park has permanently ceded its position as the Bay Area's center of AI talent gravity to a San Francisco it never fully controlled.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.