Meta spent the first half of 2026 running two simultaneous operations: cutting 8,000 employees from non-AI units on May 20, and offering individual AI researchers packages that in at least one disclosed case reportedly topped $1.5 billion over six years. The contrast is not contradictory — it is the strategy. Fundamental AI Research (FAIR), now housed inside Meta Superintelligence Labs (MSL) under Chief AI Officer Alexandr Wang, is the unit Meta is building into, not cutting from. And in the US AI talent market at the H1 2026 mark, Meta's checkbook and its open-source publishing record are doing something Anthropic and OpenAI cannot easily replicate: offering frontier-scale comp alongside a track record of public research that researchers can put on their CVs.
What Happened
The restructuring that defined Meta's AI talent posture in H1 2026 began in October 2025, when Meta dissolved the legacy AGI Foundations team behind Llama 4 and eliminated roughly 600 positions inside the research organization. Yann LeCun, Meta's Chief AI Scientist for 12 years and the founding intellectual architect of FAIR, departed in November 2025 after being asked to report into Wang. He subsequently co-founded Advanced Machine Intelligence Labs (AMI) in Paris.
What replaced that structure is materially different. Wang split MSL into four units: the TBD Lab for frontier model development; FAIR for fundamental research; an applied products arm led by former GitHub CEO Nat Friedman; and an infrastructure division under Aparna Ramani. FAIR retains its Menlo Park base and its New York office — the New York lab, on Park Avenue South, has been the more active recruiting node in H1 2026, with multiple open positions for senior research scientists posted simultaneously across New York and Menlo Park campuses this spring.
The comp bands MSL is deploying to fill those seats are no longer in the same conversation as traditional Big Tech research roles. Meta's H-1B visa filings — a public record — show AI Research Scientist base salaries running from $163,800 to $328,000, with a VP of AI position filed at $650,000 base. Those figures exclude equity and bonuses, which for senior hires represent the majority of total value delivered. Levels.fyi data through May 2026 shows Meta's median AI Researcher total comp at $416,500 and Research Scientist packages ranging from $305,000 at IC4 to $581,000 at IC6 — figures that have moved up materially from the 2024 baselines tracked on the same platform.
The signal hires make the band movement legible. Meta recruited Jack Rae, pre-training tech lead for Gemini 2.5 at Google DeepMind, and Pei Sun, also from DeepMind. From OpenAI, MSL brought in Trapit Bansal and Hongyu Ren. Joel Pobar, who worked on AI inference at Anthropic, crossed to Meta. Ruoming Pang, formerly of Apple's machine learning team, accepted a reported $200 million package to join MSL's superintelligence effort in mid-2025. Andrew Tulloch, who had worked at OpenAI before joining Mira Murati's Thinking Machines Lab, arrived at MSL in late 2025 with a package that The Next Web and multiple downstream sources reported at $1.5 billion over six years — a figure a Meta spokesman called "inaccurate and ridiculous," while declining to provide an alternative number. MSL's total headcount reached 180 researchers across Menlo Park, London, and Tel Aviv by early 2026.
Sam Altman, in public remarks referenced by The Entrepreneur and other outlets, confirmed Meta had offered at least one researcher a signing bonus of up to $100 million, framing it as evidence of an escalating market dynamic, not an isolated case.
FAIR's first major output from the restructured lab — Muse Spark, Meta's first model produced entirely under MSL — launched April 8, 2026. It is competitive with GPT-4o-era benchmarks but does not lead GPT-5.4 or Claude Opus 4.6 on most published leaderboard tasks. The research pipeline behind it, however, is the more important signal for recruits evaluating whether the lab is building something worth joining.
Why It Matters
The US AI talent market at H1 2026 is operating at near-zero elasticity at the senior level. Frontier research scientists — the individuals with published work in pretraining, reasoning, or alignment who the labs are actually fighting for — number in the low thousands globally. Every hire by one lab is a direct loss for another, and the compensation required to move them has escalated from the million-dollar range in 2023 to packages denominated in the hundreds of millions.
Meta's position in that contest is structurally distinct from Anthropic and OpenAI in one critical dimension: the publishing record. FAIR has, since its 2013 founding, maintained a commitment to open research that produced PyTorch, the foundational Llama model series, and hundreds of peer-reviewed papers. For a senior researcher choosing between labs, the question of whether their work will be published — or classified indefinitely as proprietary — is not peripheral. It is often the deciding variable.
Anthropic and OpenAI have both tightened publication norms as their models have become more commercially valuable. The Information reported in 2025 that Meta's own shift toward closed models under the Muse Spark development cycle caused friction inside FAIR — several researchers objected to publishing restrictions that moved FAIR closer to a model they had specifically left OpenAI or DeepMind to avoid. Meta has since confirmed it will release open-source versions of upcoming models, per a Silicon Angle report from April 2026, partially addressing that concern. But the tension is real and has not fully resolved.
The financial dynamic is equally pointed. Meta's $145 billion AI infrastructure commitment for 2026 — capital that funds compute, data center buildout, and the talent compensation required to staff those systems — is funded by advertising revenue from a business generating $200 billion annually. That is a different capital base than Anthropic, which is venture-backed and Series G funded, or OpenAI, which is converting its corporate structure while managing relationships with Microsoft and its broader investor syndicate. For a senior researcher evaluating counterparty risk — will this lab still be here and still be building in 2029 — Meta's balance sheet is an argument.
The attrition data complicates that argument. Tom's Hardware and multiple industry publications reported in H1 2026 that Meta continues to lose AI staff to Anthropic and OpenAI despite offering total compensation packages that in aggregate dwarf what those labs pay. The reasons are not primarily financial. Researchers describe dissatisfaction with organizational scale — Meta employs nearly 79,000 people, and the signal-to-noise ratio for fundamental research inside a company with that headcount is a recurring complaint. The May 2026 restructuring, which redirected 7,000 non-AI employees into AI-adjacent pods, has increased that perception of diffusion rather than reduced it.
What's Next
Three things to watch in H2 2026:
1. Whether FAIR's open-source commitment holds under MSL. The April 2026 reporting that Meta plans open-source releases for upcoming models was not accompanied by a specific timeline or model tier commitment. If MSL's next major model ships as fully closed — the Muse Spark architecture was not open-sourced at launch — FAIR's differentiated talent proposition narrows to comp and compute. That is a fight Meta can win financially but not necessarily in the researcher consideration set that values publication most.
2. Whether the Yann LeCun departure creates a second-order recruiting drag. LeCun's AMI Labs, targeting a $3 billion valuation and headquartered in Paris with a stated focus on world-model architectures, is the first credible European frontier lab with a founder whose name is known to every senior AI researcher globally. AMI will begin recruiting in H2 2026. Several FAIR alumni are likely in LeCun's first contact list. Meta has not publicly addressed whether it considers AMI a talent risk.
3. The Alexandr Wang tenure test. Wang, 28, was installed as Chief AI Officer in June 2025 after Meta paid $14.3 billion for a 49% stake in Scale AI. His structural authority over FAIR — a division that predates him by 12 years and that produced LeCun's legacy — is not universally accepted inside the organization. Multiple restructuring reports from Puck News and WION describe his role as "evolving." The stability of FAIR's leadership structure over the next two quarters will determine whether the comp investment Meta has made in senior hires translates into retention or into another round of exits toward Anthropic and OpenAI.
Meta entered H1 2026 with the largest AI infrastructure budget in the industry, a comp floor that forces every competitor to respond, and a fundamental research division in mid-reinvention — the outcome of that reinvention is the only number in the US AI talent market that matters more right now than the salary figures.