Meta ended H1 2026 with 77,986 total employees — a figure that obscures a sharper shift happening inside its AI divisions. The company simultaneously cut 8,000 roles from non-AI functions and redeployed 7,000 workers into three new AI-focused units, while its research arm, FAIR, added net research engineering headcount at a pace that outran every closed lab in new open-role velocity through June. The net result: a research bench that grew through deliberate internal reallocation and targeted external raids, not simple top-line expansion — and a compensation structure that has forced Anthropic and OpenAI to reckon with an opponent willing to write nine-figure packages for individual researchers.
What Happened
7,000. That is the number of Meta employees moved into AI-focused teams in May 2026, per SiliconAngle reporting on May 19. The three receiving units — Applied AI Engineering, Agent Transformation Accelerator, and Central Analytics — absorbed engineers and product managers from across the company as Mark Zuckerberg restructured headcount around Meta's $125 to $145 billion 2026 capital expenditure commitment, which Meta disclosed at Q1 earnings on April 29 and which represented an 84% jump from 2025's $72 billion spend.
Within Meta Superintelligence Labs, the org that houses FAIR, the talent picture unfolded in four distinct tracks. FAIR itself, led by Director of AI Research Rob Fergus, took cuts in October 2025 — approximately 600 roles across FAIR and MSL infrastructure were eliminated, per RD World Online reporting, a move that coincided with chief AI scientist Yann LeCun's departure from the company. But the TBD Lab, the unit inside MSL charged with training Meta's frontier foundation models and led by Chief AI Officer Alexandr Wang, remained protected and continued hiring throughout the period.
On the external acquisition front, Meta ran one of the most aggressive talent raids in H1 2026. After Mark Zuckerberg made a roughly $1 billion offer to acquire Thinking Machines Lab outright — and was rejected by founder Mira Murati — Meta pivoted to recruiting the founding team individually. Five of Thinking Machines Lab's founders ultimately joined Meta. The most expensive: Andrew Tulloch, whose package was reported at approximately $1.5 billion over six years, making it the largest individual talent acquisition deal on record in the technology industry, per The Next Web and Calcalist reporting.
Three former Thinking Machines Lab founders returned to OpenAI — Barret Zoph, Luke Metz, and Sam Schoenholz among them — and one joined xAI, per TechCrunch's April 24 breakdown of the dispersal. That three-way split of a single founding team captures the zero-sum character of the senior research market in H1 2026: Meta, OpenAI, and xAI effectively divided one startup's talent cohort among themselves.
The FAIR Menlo Park operation continued to post open research roles through June across multimodal AI, video generation, audio-visual understanding, chemistry, and robotics — all areas where FAIR's publications record provides a recruiting draw that TBD Lab and pure-product teams cannot replicate. FAIR Paris, active since 2015, continued as the primary European hub for fundamental research hiring, operating alongside outposts in New York, Seattle, Pittsburgh, Tel Aviv, and Montreal.
Why It Matters
$145 billion. Meta's revised 2026 capex ceiling, announced April 29, is not primarily a data center story — it is a talent story. Every dollar of compute committed requires researchers who can design training runs, engineers who can manage infrastructure at scale, and applied teams who can extract product value from model outputs. The 7,000-person internal redeployment is the supply-side response to that demand: Meta is creating AI-aligned headcount from existing payroll rather than bidding exclusively in a labor market where senior ML engineers are commanding $1.5 million to $2 million in total annual compensation.
The compensation arms race, however, has not abated at the top. Meta's willingness to offer Andrew Tulloch a $1.5 billion package — even if back-weighted with retention vesting and performance gates — reset the reference price for frontier research leadership in a way that changed how Anthropic and OpenAI frame competing offers. For researchers at the L6 and above equivalent, the question is no longer whether Meta pays frontier rates but whether a researcher values the specific technical direction, publication culture, or equity liquidity profile that each lab offers. Meta's RSUs vest against a public stock that trades in real time; OpenAI's Profit Participation Units offer periodic tender liquidity; Anthropic's equity remains fully private with no current tender path.
The open-source dimension adds a specific wrinkle that matters for mid-senior researchers. FAIR's publication record — including Llama model releases, PyTorch stewardship, and the ongoing open-source releases in molecular property prediction, language processing, and neuroscience flagged on Meta's AI blog — attracts a category of researcher for whom public contribution is professionally non-negotiable. Closed labs like Anthropic explicitly limit external publication of frontier work. OpenAI's publication cadence has declined since 2022. FAIR's open-science commitments, even under restructuring, remain a differentiated recruiting pitch for researchers who want academic visibility alongside industry pay.
That pitch, however, is now complicated by the Avocado episode. Meta's next-generation frontier model — codenamed Avocado and targeting a closed-source release, a significant departure from Llama's open-weight model — was delayed from Q1 to a May or June window, per Reuters reporting cited by MLQ.ai. Internal tests showed the model underperforming against Google Gemini 2.5 and Gemini 3 on reasoning, coding, and writing, prompting discussions about potentially licensing Gemini to power parts of Meta's AI product stack. For researchers who joined FAIR or TBD Lab in 2025 on the expectation of building the world's most capable open model, the Avocado delay and the closed-source pivot represent an ambiguity that competing labs have not hesitated to exploit in recruiting conversations.
What to Watch in H2
Three signals will determine whether Meta's research bench build sustains through December.
First, watch the Avocado release and its open-source status. If Meta ships Avocado as a closed model, it marks a structural break from the Llama strategy that has driven FAIR's talent brand for four years. A closed Avocado would narrow FAIR's open-science recruiting advantage at exactly the moment when xAI and OpenAI are both running aggressive outreach to the same senior researcher population. A delayed or cancelled open release should be read as a talent risk, not merely a product risk.
Second, watch the Thinking Machines Lab diaspora settle. Five founders went to Meta; three to OpenAI; one to xAI. The individuals in each camp are now inside the research organizations shaping H2 model training priorities. If the Meta contingent produces visible results — new architectures, model capability jumps, or publication output — it validates the $1.5 billion package logic. If the OpenAI returners contribute to a GPT-5 or successor release that outperforms Avocado, the narrative flips. This is the highest-stakes personnel experiment running in the frontier lab market right now.
Third, watch the 8,000 layoffs' downstream effect on FAIR morale. The October 2025 FAIR cuts included researchers who had been at Meta for four to eight years — longer tenures than most of their Anthropic or OpenAI counterparts. The combination of LeCun's exit, the Avocado closed-source signal, and the May 2026 restructuring has created internal uncertainty that external recruiting teams are actively targeting. If FAIR's publication output slows in Q3 — measurable via Arxiv submission rates and conference acceptance data — it would be an early indicator that the bench is thinning faster than the open-role pipeline can replenish it.
Meta enters H2 2026 as the most capital-committed AI organization on the planet, carrying a research team that was simultaneously downsized, redeployed, and reinforced with the most expensive individual hires the industry has ever seen. Whether those forces cohere into a coherent research strategy, or continue pulling in four directions at once, is the defining talent question of the second half.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.